Skip to content

Automate browser based workflows with AI

Preserve interaction fidelity when refining recordings (#8559)

Cindy Li9728e2488bf6
diff --git a/skyvern-frontend/src/api/types.ts b/skyvern-frontend/src/api/types.ts
index ae91df9fc..f875e906c 100644
--- a/skyvern-frontend/src/api/types.ts
+++ b/skyvern-frontend/src/api/types.ts
@@ -958,6 +958,7 @@ export type WorkflowRunStatusApiResponseWithWorkflow = {
workflow_title: string | null;
browser_session_id: string | null;
browser_profile_id: string | null;
+ browser_type?: string | null;
max_screenshot_scrolls: number | null;
run_with: string | null;
workflow: WorkflowApiResponse;
diff --git a/skyvern-frontend/src/components/onboarding/CopilotCTAStep.test.tsx b/skyvern-frontend/src/components/onboarding/CopilotCTAStep.test.tsx
index c15c066ef..4de2186ba 100644
--- a/skyvern-frontend/src/components/onboarding/CopilotCTAStep.test.tsx
+++ b/skyvern-frontend/src/components/onboarding/CopilotCTAStep.test.tsx
@@ -122,25 +122,6 @@ describe("CopilotCTAStep", () => {
);
});
- it("hands the prompt off to /build when the studio preview is off", async () => {
- studioState.enabled = false;
- mockPost.mockResolvedValue({
- data: { workflow_permanent_id: "wpid_x" },
- });
- setup();
-
- fireEvent.change(screen.getByRole("textbox"), {
- target: { value: "fill out my form" },
- });
- fireEvent.click(screen.getByRole("button", { name: "Create with AI" }));
-
- await waitFor(() =>
- expect(mockNavigate).toHaveBeenCalledWith("/agents/wpid_x/build", {
- state: { copilotMessage: "fill out my form" },
- }),
- );
- });
-
it("reports busy to the parent while the handoff is in flight", async () => {
mockPost.mockReturnValue(new Promise(() => {}));
const onBusyChange = vi.fn();
diff --git a/skyvern-frontend/src/components/onboarding/CredentialSetupPrompt.test.tsx b/skyvern-frontend/src/components/onboarding/CredentialSetupPrompt.test.tsx
index 1be1140ad..0a600c7cf 100644
--- a/skyvern-frontend/src/components/onboarding/CredentialSetupPrompt.test.tsx
+++ b/skyvern-frontend/src/components/onboarding/CredentialSetupPrompt.test.tsx
@@ -53,16 +53,6 @@ describe("CredentialSetupPrompt", () => {
expect(cta.getAttribute("href")).toBe("/agents/wpid_123/studio");
});
- it("routes the CTA to /build when the studio preview is off", () => {
- studioState.enabled = false;
- renderPrompt([{ label: "Login" }]);
-
- const cta = screen.getByRole("link", {
- name: /set up credentials in the editor/i,
- });
- expect(cta.getAttribute("href")).toBe("/agents/wpid_123/build");
- });
-
it("fires credentialSetupShown once on mount with the block count", () => {
renderPrompt([{ label: "Login" }]);
diff --git a/skyvern-frontend/src/hooks/useWorkflowStudioEnabled.ts b/skyvern-frontend/src/hooks/useWorkflowStudioEnabled.ts
index a73d3c14e..a82f24e86 100644
--- a/skyvern-frontend/src/hooks/useWorkflowStudioEnabled.ts
+++ b/skyvern-frontend/src/hooks/useWorkflowStudioEnabled.ts
@@ -1,15 +1,7 @@
-import { useFeatureFlagEnabled } from "posthog-js/react";
-
-import { WORKFLOW_STUDIO_FLAG } from "@/util/featureFlags";
-
-// Client-side eval so the flag reflects per-user opt-in (PostHog person property),
-// not the per-org server-evaluated path. Gates the whole studio redesign.
-// undefined means flags have not resolved yet — callers that act irreversibly on
-// "off" (e.g. URL rewrites) must distinguish it from a resolved false.
-export function useWorkflowStudioFlagState(): boolean | undefined {
- return useFeatureFlagEnabled(WORKFLOW_STUDIO_FLAG);
+export function useWorkflowStudioFlagState(): boolean {
+ return true;
}
export function useWorkflowStudioEnabled(): boolean {
- return useWorkflowStudioFlagState() ?? false;
+ return true;
}
diff --git a/skyvern-frontend/src/router.tsx b/skyvern-frontend/src/router.tsx
index 5dce61453..1df94bda0 100644
--- a/skyvern-frontend/src/router.tsx
+++ b/skyvern-frontend/src/router.tsx
@@ -4,6 +4,7 @@ import {
DebugRoute,
EditRoute,
StudioRoute,
+ WorkflowRunRoute,
} from "@/routes/workflows/StudioRouteGates";
import { LegacyWorkflowsRedirect } from "@/routes/workflows/LegacyWorkflowsRedirect";
import { BrowserSession } from "@/routes/browserSessions/BrowserSession";
@@ -25,15 +26,9 @@ import { TasksPage } from "./routes/tasks/list/TasksPage";
import { WorkflowPage } from "./routes/workflows/WorkflowPage";
import { WorkflowScriptDetailPage } from "./routes/workflows/WorkflowScriptDetailPage";
import { WorkflowScriptsPage } from "./routes/workflows/WorkflowScriptsPage";
-import { WorkflowRun } from "./routes/workflows/WorkflowRun";
import { WorkflowRunParameters } from "./routes/workflows/WorkflowRunParameters";
import { Workflows } from "./routes/workflows/Workflows";
import { WorkflowsPageLayout } from "./routes/workflows/WorkflowsPageLayout";
-import { WorkflowPostRunParameters } from "./routes/workflows/workflowRun/WorkflowPostRunParameters";
-import { WorkflowRunOutput } from "./routes/workflows/workflowRun/WorkflowRunOutput";
-import { WorkflowRunOverview } from "./routes/workflows/workflowRun/WorkflowRunOverview";
-import { WorkflowRunRecording } from "./routes/workflows/workflowRun/WorkflowRunRecording";
-import { WorkflowRunCode } from "@/routes/workflows/workflowRun/WorkflowRunCode";
import { DebugStoreProvider } from "@/store/DebugStoreContext";
import { BrowserProfileDetailPage } from "@/routes/browserProfiles/BrowserProfileDetailPage.tsx";
import { BrowserProfilesPage } from "@/routes/browserProfiles/BrowserProfilesPage.tsx";
@@ -285,41 +280,8 @@ const router = createBrowserRouter([
],
},
{
- path: ":workflowRunId",
- element: <WorkflowRun />,
- children: [
- {
- index: true,
- element: <Navigate to="overview" />,
- },
- {
- path: "blocks",
- element: <Navigate to="overview" />,
- },
- {
- path: "overview",
- element: <WorkflowRunOverview />,
- },
- {
- path: "output",
- element: <WorkflowRunOutput />,
- },
- {
- path: "parameters",
- element: <WorkflowPostRunParameters />,
- },
-
- {
- path: "recording",
- element: <WorkflowRunRecording />,
- },
- {
- path: "code",
- element: (
- <WorkflowRunCode showCacheKeyValueSelector={true} />
- ),
- },
- ],
+ path: ":workflowRunId/*",
+ element: <WorkflowRunRoute />,
},
],
},
diff --git a/skyvern-frontend/src/routes/browserSessions/BrowserSessionWorkflowRuns.tsx b/skyvern-frontend/src/routes/browserSessions/BrowserSessionWorkflowRuns.tsx
index 934374aea..68c99069c 100644
--- a/skyvern-frontend/src/routes/browserSessions/BrowserSessionWorkflowRuns.tsx
+++ b/skyvern-frontend/src/routes/browserSessions/BrowserSessionWorkflowRuns.tsx
@@ -18,7 +18,6 @@ import {
TableHeader,
TableRow,
} from "@/components/ui/table";
-import * as env from "@/util/env";
import { basicLocalTimeFormat, basicTimeFormat } from "@/util/timeFormat";
import { cn } from "@/util/utils";
@@ -99,9 +98,7 @@ function BrowserSessionWorkflowRuns() {
</TableHeader>
<TableBody>
{runs.map((workflowRun) => {
- const url = env.useNewRunsUrl
- ? `/runs/${workflowRun.workflow_run_id}`
- : `/agents/${workflowRun.workflow_permanent_id}/${workflowRun.workflow_run_id}/overview`;
+ const url = `/runs/${workflowRun.workflow_run_id}`;
return (
<TableRow
key={workflowRun.workflow_run_id}
diff --git a/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.test.tsx b/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.test.tsx
index d1123ee69..707ea6414 100644
--- a/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.test.tsx
+++ b/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.test.tsx
@@ -22,7 +22,6 @@ const mocks = vi.hoisted(() => ({
captureRecordBrowser: vi.fn(),
markRecordBrowserProcessed: vi.fn(),
post: vi.fn(),
- useFeatureFlagEnabled: vi.fn(() => false),
}));
vi.mock("@/api/AxiosClient", () => ({
@@ -35,10 +34,6 @@ vi.mock("@/hooks/useCredentialGetter", () => ({
useCredentialGetter: () => vi.fn(async () => "test-token"),
}));
-vi.mock("posthog-js/react", () => ({
- useFeatureFlagEnabled: mocks.useFeatureFlagEnabled,
-}));
-
vi.mock("@/util/recordBrowserTelemetry", () => ({
captureRecordBrowser: mocks.captureRecordBrowser,
markRecordBrowserProcessed: mocks.markRecordBrowserProcessed,
@@ -82,7 +77,6 @@ describe("useProcessRecordingMutation telemetry", () => {
vi.useFakeTimers({ shouldAdvanceTime: true });
vi.setSystemTime(1_000);
vi.clearAllMocks();
- mocks.useFeatureFlagEnabled.mockReturnValue(false);
useRecordingRefinementEvidenceStore.setState({ armed: null });
useRecordingStore.getState().reset();
useRecordingStore.getState().setRecordingTransport("cdp");
@@ -127,7 +121,6 @@ describe("useProcessRecordingMutation telemetry", () => {
});
it("arms the refine_recording copilot turn with the response evidence packet", async () => {
- mocks.useFeatureFlagEnabled.mockReturnValue(true);
const evidence = {
schema_version: 1,
recording: { browser_session_id: "pbs-1" },
@@ -184,12 +177,39 @@ describe("useProcessRecordingMutation telemetry", () => {
expect(mocks.post).toHaveBeenCalledWith(
"/browser_sessions/pbs-1/process_recording",
expect.objectContaining({
+ code_first: true,
+ supports_credential_tokens: true,
recording_attempt_id: "attempt-1",
interpretation_session_id: "interpretation-1",
}),
);
});
+ it("keeps the finalized recording identity available for a retry", async () => {
+ mocks.post.mockRejectedValue(new Error("temporary processing failure"));
+ useRecordingStore.setState({
+ finishRequested: true,
+ recordingAttemptId: "attempt-1",
+ interpretationSessionId: "interpretation-1",
+ });
+ const { result } = renderHook(
+ () =>
+ useProcessRecordingMutation({
+ browserSessionId: "pbs-1",
+ }),
+ { wrapper },
+ );
+
+ act(() => result.current.mutate({ draftSteps: [draftStep] }));
+
+ await waitFor(() => expect(result.current.isError).toBe(true));
+ expect(useRecordingStore.getState()).toMatchObject({
+ finishRequested: true,
+ recordingAttemptId: "attempt-1",
+ interpretationSessionId: "interpretation-1",
+ });
+ });
+
it("keeps the durable recording id for the workflow save that follows", async () => {
mocks.post.mockResolvedValue({
data: {
diff --git a/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.ts b/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.ts
index 40505c5e2..b3dac8863 100644
--- a/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.ts
+++ b/skyvern-frontend/src/routes/browserSessions/hooks/useProcessRecordingMutation.ts
@@ -1,12 +1,10 @@
import { useMutation } from "@tanstack/react-query";
import { useRef } from "react";
-import { useFeatureFlagEnabled } from "posthog-js/react";
import { getClient } from "@/api/AxiosClient";
import { toast } from "@/components/ui/use-toast";
import { useCredentialGetter } from "@/hooks/useCredentialGetter";
import { useWorkflowPermanentId } from "@/routes/workflows/WorkflowPermanentIdContext";
-import { RECORD_BROWSER_CODE_FIRST_FLAG } from "@/util/featureFlags";
import {
useRecordingStore,
type RecordingDraftStep,
@@ -46,10 +44,6 @@ const useProcessRecordingMutation = ({
eventCount: number;
optimisticStepCount: number;
} | null>(null);
- // Per-user opt-in preview; not enrolled reads as false (agent blocks).
- const codeFirst =
- useFeatureFlagEnabled(RECORD_BROWSER_CODE_FIRST_FLAG) ?? false;
-
const processRecordingMutation = useMutation({
mutationFn: async (
variables: {
@@ -132,7 +126,8 @@ const useProcessRecordingMutation = ({
>(`/browser_sessions/${browserSessionId}/process_recording`, {
compressed_chunks: compressedChunks,
workflow_permanent_id: workflowPermanentId,
- code_first: codeFirst,
+ // Keep opting in explicitly while older backends still honor this field.
+ code_first: true,
// This build substitutes credential tokens in a code block's code; a build that
// does not must not be handed blocks whose code reads a token it cannot rename.
supports_credential_tokens: true,
@@ -192,13 +187,15 @@ const useProcessRecordingMutation = ({
}
toast({
variant: "success",
- title: "Recording Processed",
- description: "The recording has been successfully processed.",
+ title: evidence ? "Recorded steps added" : "Recording processed",
+ description: evidence
+ ? "Copilot is refining the workflow now. Follow its progress in the Copilot pane."
+ : "The recording has been successfully processed.",
});
onSuccess?.({ recordingId, blocks, parameters: parameters });
- if (codeFirst && evidence) {
+ if (evidence) {
// One replace-navigation stores the packet and arms the copilot turn that
// reads it, the same handoff RunTab makes for diagnose_run.
const nonce = crypto.randomUUID();
@@ -248,8 +245,6 @@ const useProcessRecordingMutation = ({
latency_ms: latencyMs,
});
- useRecordingStore.setState({ finishRequested: false });
-
toast({
variant: "destructive",
title: "Error Processing Recording",
diff --git a/skyvern-frontend/src/routes/history/RunHistory.tsx b/skyvern-frontend/src/routes/history/RunHistory.tsx
index d478c0c86..afd6d4c5c 100644
--- a/skyvern-frontend/src/routes/history/RunHistory.tsx
+++ b/skyvern-frontend/src/routes/history/RunHistory.tsx
@@ -154,8 +154,6 @@ function inferTriggerType(run: TaskRunListItem): TriggerType | null {
function getRunNavigationPath(run: TaskRunListItem): string {
switch (run.task_run_type) {
case TaskRunType.WorkflowRun:
- // /runs/{wr} renders the studio run view when the preview is on and the
- // legacy run page when it is off, so one short path serves both.
return `/runs/${run.run_id}`;
case TaskRunType.TaskV2:
return `/runs/${run.run_id}`;
diff --git a/skyvern-frontend/src/routes/runs/RunRouter.test.tsx b/skyvern-frontend/src/routes/runs/RunRouter.test.tsx
index d01341588..f2e436573 100644
--- a/skyvern-frontend/src/routes/runs/RunRouter.test.tsx
+++ b/skyvern-frontend/src/routes/runs/RunRouter.test.tsx
@@ -1,7 +1,7 @@
// @vitest-environment jsdom
import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
-import { render, screen } from "@testing-library/react";
+import { render, screen, waitFor } from "@testing-library/react";
import { MemoryRouter, Route, Routes, useLocation } from "react-router-dom";
import { afterEach, beforeEach, describe, expect, test, vi } from "vitest";
@@ -30,7 +30,6 @@ const { getClientMock, realRunQuery } = vi.hoisted(() => ({
}));
const mocks = vi.hoisted(() => ({
- studioFlagState: vi.fn<() => boolean | undefined>(() => true),
taskV2: vi.fn(() => ({ data: undefined, isLoading: false })),
runQuery: vi.fn<
(options?: { workflowRunId?: string; enabled?: boolean }) => RunQueryResult
@@ -43,10 +42,6 @@ const mocks = vi.hoisted(() => ({
})),
}));
-vi.mock("@/hooks/useWorkflowStudioEnabled", () => ({
- useWorkflowStudioFlagState: () => mocks.studioFlagState(),
- useWorkflowStudioEnabled: () => mocks.studioFlagState() ?? false,
-}));
vi.mock("@/routes/runs/useTaskV2Query", () => ({
useTaskV2Query: () => mocks.taskV2(),
}));
@@ -79,9 +74,6 @@ vi.mock("@/routes/workflows/editor/WorkflowEditor", () => ({
<div data-testid="studio">studio:{useWorkflowPermanentId()}</div>
),
}));
-vi.mock("@/routes/workflows/WorkflowRun", () => ({
- WorkflowRun: () => <div data-testid="legacy">legacy</div>,
-}));
function LocationProbe() {
const location = useLocation();
@@ -91,6 +83,11 @@ function LocationProbe() {
}
function renderAt(entry: string, client?: QueryClient) {
+ const queryClient =
+ client ??
+ new QueryClient({
+ defaultOptions: { queries: { retry: false } },
+ });
const tree = (
<MemoryRouter initialEntries={[entry]}>
<Routes>
@@ -101,11 +98,7 @@ function renderAt(entry: string, client?: QueryClient) {
</MemoryRouter>
);
return render(
- client ? (
- <QueryClientProvider client={client}>{tree}</QueryClientProvider>
- ) : (
- tree
- ),
+ <QueryClientProvider client={queryClient}>{tree}</QueryClientProvider>,
);
}
@@ -125,7 +118,6 @@ function expectCenteredLoadingIndicator() {
describe("RunRouter", () => {
beforeEach(() => {
- mocks.studioFlagState.mockReturnValue(true);
mocks.taskV2.mockReturnValue({ data: undefined, isLoading: false });
mocks.runQuery.mockReturnValue({ data: resolvedRun, isLoading: false });
});
@@ -139,10 +131,38 @@ describe("RunRouter", () => {
renderAt("/runs/wr_1");
expect(screen.getByTestId("studio").textContent).toBe("studio:wpid_123");
expect(screen.queryByTestId("redirected")).toBeNull();
- expect(screen.queryByTestId("legacy")).toBeNull();
});
- test("studio on: shows the fetching treatment while the run resolves", () => {
+ test.each([
+ ["overview", "timeline", "overview,browser"],
+ ["blocks", "timeline", "overview,browser"],
+ ["output", "timeline", "overview,browser"],
+ ["parameters", "timeline", "overview,browser"],
+ ["recording", "recording", "browser,overview"],
+ ["code", "code", "overview,browser"],
+ ])("preserves the direct %s subview", (legacySubview, studioView, panes) => {
+ renderAt(`/runs/wr_1/${legacySubview}?active=act_1`);
+
+ expect(screen.getByTestId("location").textContent).toBe(
+ `/runs/wr_1?active=act_1&view=${studioView}&panes=${panes}`,
+ );
+ });
+
+ test.each([
+ ["output", "outputs"],
+ ["parameters", "inputs"],
+ ])(
+ "preserves the run-level %s subview without an active block",
+ (legacySubview, studioView) => {
+ renderAt(`/runs/wr_1/${legacySubview}`);
+
+ expect(screen.getByTestId("location").textContent).toBe(
+ `/runs/wr_1?view=${studioView}&panes=overview,browser`,
+ );
+ },
+ );
+
+ test("shows the fetching treatment while the workflow run resolves", () => {
mocks.runQuery.mockReturnValue({ data: undefined, isLoading: true });
renderAt("/runs/wr_1");
expectCenteredLoadingIndicator();
@@ -173,122 +193,63 @@ describe("RunRouter", () => {
expect(screen.queryByTestId("studio")).toBeNull();
});
- test("studio off: keeps the legacy run view", () => {
- mocks.studioFlagState.mockReturnValue(false);
- renderAt("/runs/wr_1");
- expect(screen.getByTestId("legacy")).toBeTruthy();
- expect(screen.queryByTestId("studio")).toBeNull();
- });
-
- test("studio on: enables the run-resolver query for the wr_ id", () => {
+ test("enables the run-resolver query for the wr_ id", () => {
renderAt("/runs/wr_1");
expect(mocks.runQuery).toHaveBeenCalledWith(
expect.objectContaining({ workflowRunId: "wr_1", enabled: true }),
);
});
- test("studio off: disables the run-resolver query so non-studio routes don't fetch a workflow run", () => {
- mocks.studioFlagState.mockReturnValue(false);
- renderAt("/runs/wr_1");
+ test("does not fetch a workflow run for task routes", () => {
+ renderAt("/runs/tsk_1");
expect(mocks.runQuery).toHaveBeenCalledWith(
expect.objectContaining({ enabled: false }),
);
});
- test("embed=true keeps the chrome-free legacy view, not the studio shell", () => {
+ test("embed=true renders a chrome-free Overview-only studio run", async () => {
renderAt("/runs/wr_1?embed=true");
- expect(screen.getByTestId("legacy")).toBeTruthy();
- expect(screen.queryByTestId("studio")).toBeNull();
- });
-
- test("embed=true honors ?wr= even while the studio flag is unresolved", () => {
- // Embed never renders the studio shell, so there is no URL state to protect
- // and the redirect need not wait for the flag.
- mocks.studioFlagState.mockReturnValue(undefined);
- renderAt("/runs/wr_1?embed=true&wr=wr_2&active=act_9");
expect(screen.getByTestId("location").textContent).toBe(
- "/runs/wr_2/overview?embed=true&active=act_9",
+ "/runs/wr_1?embed=true&panes=overview",
);
- expect(screen.getByTestId("legacy")).toBeTruthy();
- });
-
- test("a permanently failed run fetch lands on 404, not an endless spinner", () => {
- mocks.runQuery.mockReturnValue({
- data: undefined,
- isLoading: false,
- isError: true,
+ await waitFor(() => {
+ expect(screen.getByTestId("studio").textContent).toBe("studio:wpid_123");
});
- renderAt("/runs/wr_1");
- expect(screen.queryByTestId("studio")).toBeNull();
- expect(screen.queryByAltText("Minimized Logo")).toBeNull();
- expect(screen.queryByTestId("legacy")).toBeNull();
});
- test("studio off: a studio-shared link resolves to the ?wr= run, selection intact", () => {
- // The studio switches runs by rewriting ?wr= and leaving the path alone, so
- // its URLs can carry a stale run id in the path. Flag-off must land on the
- // ?wr= run — straight onto the sub-path so the index redirect can't drop
- // ?active= (the shared selection).
- mocks.studioFlagState.mockReturnValue(false);
- renderAt(
- "/runs/wr_1?wr=wr_2&panes=editor,overview&active=act_9&selected-block=Payment",
- );
+ test("embedded recording links focus only the Browser pane", () => {
+ renderAt("/runs/wr_1/recording?embed=true");
expect(screen.getByTestId("location").textContent).toBe(
- "/runs/wr_2/overview?panes=editor,overview&active=act_9&selected-block=Payment",
+ "/runs/wr_1?embed=true&view=recording&panes=browser",
);
- expect(screen.getByTestId("legacy")).toBeTruthy();
- expect(screen.queryByTestId("studio")).toBeNull();
});
- test("studio off: ?wr= naming the path run is not a redirect", () => {
- mocks.studioFlagState.mockReturnValue(false);
- renderAt("/runs/wr_1?wr=wr_1&active=act_9");
- expect(screen.getByTestId("location").textContent).toBe(
- "/runs/wr_1?wr=wr_1&active=act_9",
- );
- expect(screen.getByTestId("legacy")).toBeTruthy();
- });
+ test("embedded in-app pane changes are not normalized away", () => {
+ renderAt("/runs/wr_1?embed=true&panes=browser&active=wrb_1");
- test("flag unresolved: no rewrite — a studio user's URL state must survive cold load", () => {
- mocks.studioFlagState.mockReturnValue(undefined);
- renderAt("/runs/wr_1?wr=wr_2&panes=editor,overview&active=act_9");
+ expect(screen.getByTestId("studio").textContent).toBe("studio:wpid_123");
expect(screen.getByTestId("location").textContent).toBe(
- "/runs/wr_1?wr=wr_2&panes=editor,overview&active=act_9",
+ "/runs/wr_1?embed=true&panes=browser&active=wrb_1",
);
- expect(screen.getByTestId("legacy")).toBeTruthy();
});
- test("studio off: /blocks maps to /overview and the studio-internal ?wrs=/?bl= are scrubbed", () => {
- // /blocks immediately re-navigates to /overview without the search, and
- // ?wrs=/?bl= are companions of the ?wr= being promoted into the path.
- mocks.studioFlagState.mockReturnValue(false);
- renderAt("/runs/wr_1/blocks?wr=wr_2&wrs=copilot&bl=Login&active=act_9");
- expect(screen.getByTestId("location").textContent).toBe(
- "/runs/wr_2/overview?active=act_9",
- );
- });
+ test("embedded URLs cannot open authoring panes", () => {
+ renderAt("/runs/wr_1?embed=true&panes=editor,browser");
- test("studio off: the splat is whitelisted — known sub-paths forward, anything else lands on overview", () => {
- mocks.studioFlagState.mockReturnValue(false);
- const first = renderAt("/runs/wr_1/recording?wr=wr_2");
expect(screen.getByTestId("location").textContent).toBe(
- "/runs/wr_2/recording",
- );
- first.unmount();
- renderAt("/runs/wr_1/%2E%2E%2Fagents?wr=wr_2&active=act_9");
- expect(screen.getByTestId("location").textContent).toBe(
- "/runs/wr_2/overview?active=act_9",
+ "/runs/wr_1?embed=true&panes=overview",
);
});
- test("studio off: a malformed ?wr= is ignored, not spliced into the path", () => {
- mocks.studioFlagState.mockReturnValue(false);
- const wr = encodeURIComponent("wr_2/../../agents/wpid_x/studio");
- renderAt(`/runs/wr_1?wr=${wr}`);
- expect(screen.getByTestId("location").textContent).toBe(
- `/runs/wr_1?wr=${wr}`,
- );
- expect(screen.getByTestId("legacy")).toBeTruthy();
+ test("a permanently failed run fetch lands on 404, not an endless spinner", () => {
+ mocks.runQuery.mockReturnValue({
+ data: undefined,
+ isLoading: false,
+ isError: true,
+ });
+ renderAt("/runs/wr_1");
+ expect(screen.queryByTestId("studio")).toBeNull();
+ expect(screen.queryByAltText("Minimized Logo")).toBeNull();
});
test("studio on: the studio owns ?wr= — the path is never rewritten under it", () => {
diff --git a/skyvern-frontend/src/routes/runs/RunRouter.tsx b/skyvern-frontend/src/routes/runs/RunRouter.tsx
index 84b10a3c8..aa9695c9c 100644
--- a/skyvern-frontend/src/routes/runs/RunRouter.tsx
+++ b/skyvern-frontend/src/routes/runs/RunRouter.tsx
@@ -7,8 +7,8 @@ import {
Navigate,
Route,
Routes,
+ useLocation,
useParams,
- useSearchParams,
} from "react-router-dom";
import { useMemo } from "react";
@@ -20,32 +20,16 @@ import { TaskActions } from "@/routes/tasks/detail/TaskActions";
import { TaskDetails } from "@/routes/tasks/detail/TaskDetails";
import { TaskParameters } from "@/routes/tasks/detail/TaskParameters";
import { TaskRecording } from "@/routes/tasks/detail/TaskRecording";
-import { WorkflowRun } from "@/routes/workflows/WorkflowRun";
-import { WorkflowPostRunParameters } from "@/routes/workflows/workflowRun/WorkflowPostRunParameters";
-import { WorkflowRunOutput } from "@/routes/workflows/workflowRun/WorkflowRunOutput";
-import { WorkflowRunOverview } from "@/routes/workflows/workflowRun/WorkflowRunOverview";
-import { WorkflowRunRecording } from "@/routes/workflows/workflowRun/WorkflowRunRecording";
-import { WorkflowRunCode } from "@/routes/workflows/workflowRun/WorkflowRunCode";
-import { WorkflowsPageLayout } from "@/routes/workflows/WorkflowsPageLayout";
import { WorkflowEditor } from "@/routes/workflows/editor/WorkflowEditor";
import { WorkflowPermanentIdContext } from "@/routes/workflows/WorkflowPermanentIdContext";
import { useWorkflowRunWithWorkflowQuery } from "@/routes/workflows/hooks/useWorkflowRunWithWorkflowQuery";
-import { useWorkflowStudioFlagState } from "@/hooks/useWorkflowStudioEnabled";
-import { useTaskV2Query } from "@/routes/runs/useTaskV2Query";
import {
- SYSTEM_RUN_FOCUS_PARAM,
+ parsePanesParam,
+ RUN_APPEND_PANES,
+ STUDIO_PANES_PARAM,
toReadableSearch,
} from "@/routes/workflows/studio/panes";
-
-// Sub-paths the ?wr= redirect below may forward; anything else (including the
-// self-redirecting blocks route and crafted splats) lands on overview.
-const REDIRECTABLE_RUN_SUBPATHS = new Set([
- "overview",
- "output",
- "parameters",
- "recording",
- "code",
-]);
+import { useTaskV2Query } from "@/routes/runs/useTaskV2Query";
const loadingIndicator = (
<div
@@ -60,66 +44,36 @@ const loadingIndicator = (
);
function RunRouter() {
- const { runId, "*": subPath } = useParams();
- const studioFlagState = useWorkflowStudioFlagState();
- const studioEnabled = studioFlagState ?? false;
+ const params = useParams();
+ const runId = params.runId;
+ const location = useLocation();
+ const isTaskV2Run = Boolean(runId?.startsWith("tsk_v2"));
+ const isWorkflowRun = Boolean(runId?.startsWith("wr_"));
const { data: task_v2, isLoading } = useTaskV2Query({
- id: runId?.startsWith("tsk_v2") ? runId : undefined,
+ id: isTaskV2Run ? runId : undefined,
});
- // With the studio on, a workflow-run short URL renders the studio run view in
- // place (short URL stays in the address bar) rather than redirecting to the
- // long /agents/{wpid}/studio?wr= form. The studio components read the workflow
- // id, which the path lacks, so resolve it from the run first — this query is
- // shared/cached with the shell that renders next. An embedded run (?embed=true)
- // keeps the legacy chrome-free view instead of the full studio shell.
- const [searchParams] = useSearchParams();
- const isEmbedded = searchParams.get("embed") === "true";
- const renderStudioRun =
- studioEnabled && Boolean(runId?.startsWith("wr_")) && !isEmbedded;
+ // Workflow-run short URLs render the studio run view in place (short URL
+ // stays in the address bar) rather than redirecting to the long
+ // /agents/{wpid}/studio?wr= form. The studio components read the workflow id,
+ // which the path lacks, so resolve it from the run first — this query is
+ // shared/cached with the shell that renders next.
const { data: studioRun, isError: studioRunFailed } =
useWorkflowRunWithWorkflowQuery({
- workflowRunId: renderStudioRun ? runId : undefined,
- enabled: renderStudioRun,
+ workflowRunId: isWorkflowRun ? runId : undefined,
+ enabled: isWorkflowRun,
});
- const runType = runId?.startsWith("tsk_v2")
+ const runType = isTaskV2Run
? "redirect"
- : runId?.startsWith("wr_")
+ : isWorkflowRun
? "workflow"
: runId?.startsWith("tsk_")
? "task"
: null;
const routes = useMemo(() => {
- if (runType === "workflow") {
- return (
- <Routes>
- <Route element={<WorkflowsPageLayout />}>
- <Route element={<WorkflowRun />}>
- <Route index element={<Navigate to="overview" replace />} />
- <Route
- path="blocks"
- element={<Navigate to={`/runs/${runId}/overview`} replace />}
- />
- <Route path="overview" element={<WorkflowRunOverview />} />
- <Route path="output" element={<WorkflowRunOutput />} />
- <Route
- path="parameters"
- element={<WorkflowPostRunParameters />}
- />
- <Route path="recording" element={<WorkflowRunRecording />} />
- <Route
- path="code"
- element={<WorkflowRunCode showCacheKeyValueSelector={true} />}
- />
- </Route>
- </Route>
- </Routes>
- );
- }
-
if (runType === "task") {
return (
<Routes>
@@ -137,7 +91,63 @@ function RunRouter() {
}
return <Status404 />;
- }, [runType, runId]);
+ }, [runType]);
+
+ const legacySubview = params["*"]?.split("/")[0] || undefined;
+ const studioView = legacySubview
+ ? {
+ overview: "timeline",
+ blocks: "timeline",
+ output: "outputs",
+ parameters: "inputs",
+ recording: "recording",
+ code: "code",
+ }[legacySubview]
+ : undefined;
+ const searchParams = new URLSearchParams(location.search);
+ const routedStudioView =
+ searchParams.has("active") &&
+ (studioView === "outputs" || studioView === "inputs")
+ ? "timeline"
+ : studioView;
+ const embedded = searchParams.get("embed") === "true";
+ const requestedView = routedStudioView ?? searchParams.get("view");
+ const requiredPane = requestedView === "recording" ? "browser" : "overview";
+ const embeddedPanes = parsePanesParam(searchParams.get(STUDIO_PANES_PARAM));
+ const embeddedLayoutInvalid =
+ embeddedPanes === null ||
+ embeddedPanes.length === 0 ||
+ embeddedPanes.some((pane) => pane !== "overview" && pane !== "browser");
+ const normalizeEmbeddedLayout =
+ embedded && (legacySubview !== undefined || embeddedLayoutInvalid);
+ if (isWorkflowRun && (routedStudioView || normalizeEmbeddedLayout)) {
+ if (routedStudioView) {
+ searchParams.set("view", routedStudioView);
+ }
+ const panes = embedded
+ ? [requiredPane]
+ : (parsePanesParam(searchParams.get(STUDIO_PANES_PARAM)) ?? [
+ ...RUN_APPEND_PANES,
+ ]);
+ searchParams.set(
+ STUDIO_PANES_PARAM,
+ embedded
+ ? requiredPane
+ : [requiredPane, ...panes.filter((pane) => pane !== requiredPane)].join(
+ ",",
+ ),
+ );
+ return (
+ <Navigate
+ to={{
+ pathname: `/runs/${runId}`,
+ search: toReadableSearch(searchParams),
+ hash: location.hash,
+ }}
+ replace
+ />
+ );
+ }
if (runId?.startsWith("tsk_v2")) {
if (isLoading) {
@@ -159,11 +169,11 @@ function RunRouter() {
return <Navigate to={`/runs/${workflowRunId}`} replace />;
}
- if (renderStudioRun) {
+ if (isWorkflowRun) {
if (!studioRun) {
- // A permanently failed initial fetch (deleted, foreign-org, or garbage run
- // id) lands on 404 like the legacy run view; a failed background poll of a
- // live run retains its data, so it never flashes 404 over a working view.
+ // A permanently failed initial fetch (foreign-org or garbage run id)
+ // lands on 404; a failed background poll of a live run retains its data,
+ // so it never flashes 404 over a working view.
if (studioRunFailed) {
return <Status404 />;
}
@@ -181,34 +191,6 @@ function RunRouter() {
);
}
- // A studio URL carries the viewed run in ?wr= and a possibly stale one in the
- // path, so honor ?wr= here — but only once the flag is known off (or under
- // ?embed=true, which never renders the studio), since rewriting while flags
- // resolve would strip a studio user's URL state. Land on the sub-path
- // directly because the index and /blocks redirects drop the search (?active=
- // must survive), and scrub ?wrs=/?bl=, studio-internal companions of ?wr=.
- const sharedStudioRunId = searchParams.get("wr");
- if (
- runType === "workflow" &&
- (studioFlagState === false || isEmbedded) &&
- sharedStudioRunId !== null &&
- /^wr_\w+$/.test(sharedStudioRunId) &&
- sharedStudioRunId !== runId
- ) {
- const params = new URLSearchParams(searchParams);
- params.delete("wr");
- params.delete(SYSTEM_RUN_FOCUS_PARAM);
- params.delete("bl");
- const targetSubPath =
- subPath && REDIRECTABLE_RUN_SUBPATHS.has(subPath) ? subPath : "overview";
- return (
- <Navigate
- to={`/runs/${sharedStudioRunId}/${targetSubPath}${toReadableSearch(params)}`}
- replace
- />
- );
- }
-
return routes;
}
diff --git a/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.test.ts b/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.test.ts
index 39fe9cd48..2f214091e 100644
--- a/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.test.ts
+++ b/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.test.ts
@@ -130,7 +130,79 @@ describe("useRecordingMessageChannel", () => {
socket.send.mock.calls.map((call) => JSON.parse(String(call[0]))),
).toEqual([
expect.objectContaining({ kind: "begin-exfiltration" }),
- { kind: "end-exfiltration" },
+ { kind: "end-exfiltration", discard: true },
+ ]);
+ });
+
+ it("preserves finalized actions when Done ends exfiltration", async () => {
+ useRecordingStore.setState({
+ isRecording: true,
+ finishRequested: false,
+ });
+ const { rerender } = renderHook(
+ ({ exfiltrate }) =>
+ useRecordingMessageChannel({
+ browserSessionId: "pbs-1",
+ enabled: true,
+ exfiltrate,
+ workflowPermanentId: "workflow-1",
+ clipboard: "none",
+ }),
+ { initialProps: { exfiltrate: true } },
+ );
+ const socket = await openSocket();
+ socket.send.mockClear();
+
+ act(() => useRecordingStore.setState({ finishRequested: true }));
+ rerender({ exfiltrate: false });
+
+ expect(
+ socket.send.mock.calls.map((call) => JSON.parse(String(call[0]))),
+ ).toContainEqual(
+ expect.objectContaining({ kind: "end-exfiltration", discard: false }),
+ );
+ });
+
+ it("purges finalized actions when Discard follows Done", async () => {
+ useRecordingStore.setState({
+ isRecording: true,
+ finishRequested: false,
+ interpretationSessionId: "interpretation-1",
+ });
+ const { rerender, unmount } = renderHook(
+ ({ exfiltrate }) =>
+ useRecordingMessageChannel({
+ browserSessionId: "pbs-1",
+ enabled: true,
+ exfiltrate,
+ workflowPermanentId: "workflow-1",
+ clipboard: "none",
+ }),
+ { initialProps: { exfiltrate: true } },
+ );
+ const socket = await openSocket();
+ socket.send.mockClear();
+
+ act(() => useRecordingStore.setState({ finishRequested: true }));
+ rerender({ exfiltrate: false });
+ act(() => useRecordingStore.getState().reset());
+ unmount();
+
+ expect(
+ socket.send.mock.calls.map((call) => JSON.parse(String(call[0]))),
+ ).toEqual([
+ {
+ kind: "end-exfiltration",
+ discard: false,
+ interpretation_session_id: "interpretation-1",
+ workflow_permanent_id: "workflow-1",
+ },
+ {
+ kind: "end-exfiltration",
+ discard: true,
+ interpretation_session_id: "interpretation-1",
+ workflow_permanent_id: "workflow-1",
+ },
]);
});
diff --git a/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.ts b/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.ts
index 5524a4563..86de0edb8 100644
--- a/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.ts
+++ b/skyvern-frontend/src/routes/streaming/useRecordingMessageChannel.ts
@@ -45,6 +45,9 @@ interface CommandCedeControl {
interface CommandEndExfiltration {
kind: "end-exfiltration";
+ discard: boolean;
+ interpretation_session_id?: string;
+ workflow_permanent_id?: string;
}
interface CommandTakeControl {
@@ -410,6 +413,10 @@ export function useRecordingMessageChannel(
const optionsRef = useRef(options);
const exfiltrateRef = useRef(exfiltrate);
const beganRef = useRef(false);
+ const retainedFinalizationRef = useRef<{
+ interpretationSessionId: string;
+ workflowPermanentId: string;
+ } | null>(null);
const previousExfiltrateRef = useRef(false);
const beginRetryAttemptsRef = useRef(0);
const beginRetryTimerRef = useRef<ReturnType<typeof setTimeout> | null>(null);
@@ -423,6 +430,38 @@ export function useRecordingMessageChannel(
}
}, []);
+ const sendEndExfiltration = useCallback(
+ (socket: WebSocket, discard: boolean) => {
+ const store = useRecordingStore.getState();
+ const finalizationIdentity = retainedFinalizationRef.current ?? {
+ interpretationSessionId: store.interpretationSessionId,
+ workflowPermanentId: optionsRef.current.workflowPermanentId,
+ };
+ socket.send(
+ JSON.stringify({
+ kind: "end-exfiltration",
+ discard,
+ interpretation_session_id:
+ finalizationIdentity.interpretationSessionId ?? undefined,
+ workflow_permanent_id:
+ finalizationIdentity.workflowPermanentId ?? undefined,
+ }),
+ );
+ if (discard) {
+ retainedFinalizationRef.current = null;
+ } else if (
+ finalizationIdentity.interpretationSessionId &&
+ finalizationIdentity.workflowPermanentId
+ ) {
+ retainedFinalizationRef.current = {
+ interpretationSessionId: finalizationIdentity.interpretationSessionId,
+ workflowPermanentId: finalizationIdentity.workflowPermanentId,
+ };
+ }
+ },
+ [],
+ );
+
const sendBeginExfiltration = useCallback(() => {
const socket = messageSocketRef.current;
if (
@@ -549,13 +588,14 @@ export function useRecordingMessageChannel(
// session resumable (SKY-12429), so it does not send END. Once the store
// says recording ended, cleanup flushes END for a same-commit panel
// unmount that prevented the exfiltrate falling-edge effect from running.
- if (
- ws?.readyState === WebSocket.OPEN &&
- beganRef.current &&
- !useRecordingStore.getState().isRecording
- ) {
- ws.send(JSON.stringify({ kind: "end-exfiltration" }));
- beganRef.current = false;
+ const store = useRecordingStore.getState();
+ if (ws?.readyState === WebSocket.OPEN && !store.isRecording) {
+ if (beganRef.current) {
+ sendEndExfiltration(ws, !store.finishRequested);
+ beganRef.current = false;
+ } else if (retainedFinalizationRef.current && !store.finishRequested) {
+ sendEndExfiltration(ws, true);
+ }
}
messageSocketRef.current = null;
try {
@@ -573,6 +613,7 @@ export function useRecordingMessageChannel(
enabled,
reconnectTrigger,
scheduleBeginRetry,
+ sendEndExfiltration,
socketUrl,
]);
@@ -600,7 +641,10 @@ export function useRecordingMessageChannel(
beginRetryAttemptsRef.current = 0;
const socket = messageSocketRef.current;
if (socket?.readyState === WebSocket.OPEN && beganRef.current) {
- socket.send(JSON.stringify({ kind: "end-exfiltration" }));
+ sendEndExfiltration(
+ socket,
+ !useRecordingStore.getState().finishRequested,
+ );
beganRef.current = false;
}
}
@@ -610,6 +654,7 @@ export function useRecordingMessageChannel(
messageSocket,
recordingAttemptId,
sendBeginExfiltration,
+ sendEndExfiltration,
workflowPermanentId,
]);
diff --git a/skyvern-frontend/src/routes/tasks/create/PromptBox.test.tsx b/skyvern-frontend/src/routes/tasks/create/PromptBox.test.tsx
index c8c177e53..17c70beee 100644
--- a/skyvern-frontend/src/routes/tasks/create/PromptBox.test.tsx
+++ b/skyvern-frontend/src/routes/tasks/create/PromptBox.test.tsx
@@ -160,7 +160,7 @@ function renderPromptBox(
afterEach(() => {
cleanup();
sessionStorage.clear();
- studioState.enabled = false;
+ studioState.enabled = true;
mockNavigate.mockReset();
mockPost.mockReset();
mockSetAutoplay.mockReset();
@@ -302,46 +302,10 @@ describe("PromptBox", () => {
await submitPrompt("Visit the docs");
await waitFor(() => expect(mockNavigate).toHaveBeenCalledTimes(1));
- expect(mockNavigate).toHaveBeenCalledWith("/agents/wpid_empty/build");
+ expect(mockNavigate).toHaveBeenCalledWith("/agents/wpid_empty/studio");
expect(mockSetAutoplay).not.toHaveBeenCalled();
});
- test("hands Discover prompts to the legacy build path via route state only", async () => {
- mockPost.mockResolvedValue({
- data: {
- workflow_permanent_id: "wpid_copilot",
- workflow_definition: { blocks: [] },
- },
- });
-
- renderPromptBox(true);
-
- fireEvent.change(screen.getByPlaceholderText("Enter your prompt..."), {
- target: { value: "Build this workflow" },
- });
- fireEvent.click(screen.getByLabelText("submit-prompt"));
-
- await waitFor(() => expect(mockPost).toHaveBeenCalledTimes(1));
- const call = mockPost.mock.calls[0];
- expect(call).toBeDefined();
- const [path, yaml] = call!;
-
- expect(path).toBe("/workflows");
- expect(yaml).toContain("run_with: agent");
- expect(yaml).not.toContain("run_with: code");
- expect(mockNavigate).toHaveBeenCalledWith(
- "/agents/wpid_copilot/build?via=discover",
- {
- state: { copilotMessage: "Build this workflow" },
- },
- );
- // The legacy /build (Debugger) surface never mounts the recovery hook, so
- // no session-scoped copy is written — writing one would strand a dead entry.
- expect(
- sessionStorage.getItem("skyvern.discoverCopilotHandoff:wpid_copilot"),
- ).toBeNull();
- });
-
test("hands Discover prompts to workflow studio with recoverable prompt state", async () => {
studioState.enabled = true;
mockPost.mockResolvedValue({
diff --git a/skyvern-frontend/src/routes/tasks/detail/TaskDetails.tsx b/skyvern-frontend/src/routes/tasks/detail/TaskDetails.tsx
index 8633f1c96..31ebde092 100644
--- a/skyvern-frontend/src/routes/tasks/detail/TaskDetails.tsx
+++ b/skyvern-frontend/src/routes/tasks/detail/TaskDetails.tsx
@@ -41,7 +41,7 @@ import { statusIsCancellable, statusIsFinalized } from "../types";
import { MAX_STEPS_DEFAULT } from "../constants";
import { useTaskQuery } from "./hooks/useTaskQuery";
import { useFirstParam } from "@/hooks/useFirstParam";
-import * as env from "@/util/env";
+import { runsApiBaseUrl } from "@/util/env";
import { TaskRunVerificationCodeForm } from "./TaskRunVerificationCodeForm";
import { RunTagsEditor } from "../components/tagging/RunTagsEditor";
@@ -257,7 +257,7 @@ function TaskDetails() {
return {
method: "POST",
- url: `${env.runsApiBaseUrl}/run/tasks`,
+ url: `${runsApiBaseUrl}/run/tasks`,
body: buildTaskRunPayload(
createTaskRequestObject(task),
RunEngine.SkyvernV1,
@@ -323,13 +323,7 @@ function TaskDetails() {
) : (
workflow &&
workflowRun && (
- <Link
- to={
- env.useNewRunsUrl
- ? `/runs/${workflowRun.workflow_run_id}`
- : `/agents/${workflow.workflow_permanent_id}/${workflowRun.workflow_run_id}/overview`
- }
- >
+ <Link to={`/runs/${workflowRun.workflow_run_id}`}>
{workflow.title}
</Link>
)
diff --git a/skyvern-frontend/src/routes/workflows/LegacyWorkflowsRedirect.test.tsx b/skyvern-frontend/src/routes/workflows/LegacyWorkflowsRedirect.test.tsx
index d90659700..dc8f0c09a 100644
--- a/skyvern-frontend/src/routes/workflows/LegacyWorkflowsRedirect.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/LegacyWorkflowsRedirect.test.tsx
@@ -116,25 +116,4 @@ describe("LegacyWorkflowsRedirect", () => {
),
);
});
-
- it("composes the 3-hop chain: legacy /debug → /agents/debug → /agents/build when studio is off", async () => {
- studioState.enabled = false;
- renderAt("/workflows/wpid_1/debug?wr=wr_1#frame");
- await waitFor(() =>
- expect(screen.getByTestId("location").textContent).toBe(
- "/agents/wpid_1/build?wr=wr_1#frame",
- ),
- );
- });
-
- it("stays on /agents build (no further hop) when studio is off", async () => {
- studioState.enabled = false;
- renderAt("/workflows/wpid_1/build?wr=wr_1");
- await waitFor(() =>
- expect(screen.getByTestId("location").textContent).toBe(
- "/agents/wpid_1/build?wr=wr_1",
- ),
- );
- expect(screen.getByText("debugger")).toBeTruthy();
- });
});
diff --git a/skyvern-frontend/src/routes/workflows/RunWorkflowForm.tsx b/skyvern-frontend/src/routes/workflows/RunWorkflowForm.tsx
index ca884b83d..9bc57831e 100644
--- a/skyvern-frontend/src/routes/workflows/RunWorkflowForm.tsx
+++ b/skyvern-frontend/src/routes/workflows/RunWorkflowForm.tsx
@@ -57,6 +57,7 @@ import { EXPERIMENT } from "@/util/onboarding/experimentConfig";
import { isActivationRun } from "@/util/onboarding/rolloutGating";
import { useOnboardingStateOptional } from "@/store/onboarding/useOnboardingState";
import { type ApiCommandOptions } from "@/util/apiCommands";
+import { runsApiBaseUrl } from "@/util/env";
import { parseHeaderJson } from "@/util/secretHeaders";
import {
getRecoveryGuidanceRetryContext,
@@ -87,9 +88,13 @@ import {
} from "@/components/ui/tooltip";
import { RotatingCredentialField } from "./components/RotatingCredentialField";
import { TestWebhookDialog } from "@/components/TestWebhookDialog";
-import * as env from "@/util/env";
import {
+ browserTypeSelectionDisabled,
+ extractBrowserTypeSetting,
parseJsonWorkflowParameterValue,
+ hasBrowserTypeOptions,
+ RESERVED_BROWSER_TYPE_FIELD,
+ useBrowserTypeOptionsQuery,
validateJsonWorkflowParameterValue,
} from "./utils";
import {
@@ -179,6 +184,7 @@ type Props = {
browserProfileId: string | null;
cdpConnectHeaders: Record<string, string> | null;
runWith: string | null;
+ browserType?: string | null;
};
};
@@ -257,6 +263,7 @@ type RunWorkflowRequestBody = {
cdp_connect_headers?: Record<string, string> | null;
browser_address?: string | null;
run_with?: "agent" | "code";
+ browser_type?: string | null;
ai_fallback?: boolean;
};
@@ -295,9 +302,14 @@ export function getRunWorkflowRequestBody(
cdpConnectHeaders,
runWith,
aiFallback,
- ...parameters
+ ...rest
} = values;
+ // The internal browser-type setting lives under a reserved field name so it never collides with
+ // a workflow input parameter literally named `browserType`; that user parameter stays in `rest`
+ // and reaches the request `data` unchanged.
+ const { browserType, rest: parameters } = extractBrowserTypeSetting(rest);
+
const parsedParameters = parseValuesForWorkflowRun(
parameters,
workflowParameters,
@@ -306,6 +318,9 @@ export function getRunWorkflowRequestBody(
const bsi = browserSessionId?.trim() === "" ? null : browserSessionId;
const bpi = browserProfileId?.trim() === "" ? null : browserProfileId;
+ // A cleared address is an empty string, which is not an attachment; normalize it to
+ // null so the backend does not read "" as an attachment and 422 against browser_type.
+ const cda = cdpAddress?.trim() === "" ? null : cdpAddress;
// A live session is the browser for the run and the backend rejects fresh +
// session together, so an attached session wins and suppresses the fresh flag.
const startFresh = Boolean(startFreshBrowser) && !bsi;
@@ -323,7 +338,7 @@ export function getRunWorkflowRequestBody(
// Backend ranks an explicit profile override above start_fresh_browser, so a
// fresh run must drop the (possibly settings-derived) override to take effect.
browser_profile_id: startFresh || perInputAgent ? null : bpi,
- browser_address: cdpAddress,
+ browser_address: cda,
run_with: runWith,
ai_fallback: aiFallback ?? true,
};
@@ -342,6 +357,19 @@ export function getRunWorkflowRequestBody(
body.webhook_callback_url = webhookCallbackUrl;
}
+ // Only send an explicit engine; omitting it inherits the workflow / system default and keeps
+ // the legacy wire shape byte-identical for runs that don't set one. An attached browser (session
+ // or remote address) owns its engine, so browser_type is suppressed even if the field held one.
+ if (
+ browserType &&
+ !browserTypeSelectionDisabled({
+ browserSessionId: bsi,
+ browserAddress: cda,
+ })
+ ) {
+ body.browser_type = browserType as string;
+ }
+
if (extraHttpHeaders) {
try {
body.extra_http_headers = parseHeaderJson(extraHttpHeaders);
@@ -545,6 +573,7 @@ function RunWorkflowForm({
location.state,
);
const { data: workflow } = useWorkflowQuery({ workflowPermanentId });
+ const { data: browserTypeOptions } = useBrowserTypeOptionsQuery();
const loginCredentialInputs = useMemo(
() => getLoginCredentialInputs({ workflow, workflowParameters }),
[workflow, workflowParameters],
@@ -620,6 +649,7 @@ function RunWorkflowForm({
? JSON.stringify(initialSettings.cdpConnectHeaders)
: null,
runWith: deriveRunWith(workflow, initialSettings.runWith),
+ [RESERVED_BROWSER_TYPE_FIELD]: initialSettings.browserType ?? null,
aiFallback: workflow?.ai_fallback ?? true,
},
});
@@ -635,6 +665,10 @@ function RunWorkflowForm({
const explicitBrowserSessionPicked = Boolean(
form.watch("browserSessionId")?.trim(),
);
+ const browserTypeDisabled = browserTypeSelectionDisabled({
+ browserSessionId: form.watch("browserSessionId"),
+ browserAddress: form.watch("cdpAddress"),
+ });
const hasBlockingParameterError = workflowParameters.some(
(param) =>
blockingParameterTypes.has(param.workflow_parameter_type) &&
@@ -674,11 +708,7 @@ function RunWorkflowForm({
});
},
onNavigate: (workflowRunId, recoveryGuidanceRetry) => {
- const runPath = studioEnabled
- ? `/runs/${workflowRunId}`
- : env.useNewRunsUrl
- ? `/runs/${workflowRunId}`
- : `/agents/${workflowPermanentId}/${workflowRunId}/overview`;
+ const runPath = `/runs/${workflowRunId}`;
if (recoveryGuidanceRetry) {
navigate(runPath, {
state: { recoveryGuidanceRetry },
@@ -767,6 +797,7 @@ function RunWorkflowForm({
? JSON.stringify(initialSettings.cdpConnectHeaders)
: null,
runWith: deriveRunWith(workflow, initialSettings.runWith),
+ [RESERVED_BROWSER_TYPE_FIELD]: initialSettings.browserType ?? null,
aiFallback: workflow?.ai_fallback ?? true,
});
setIsFormReset(true);
@@ -804,9 +835,13 @@ function RunWorkflowForm({
cdpAddress,
runWith,
aiFallback,
- ...parameters
+ ...rest
} = values;
+ // Keep the reserved browser-type setting separate from `parameters` so the request builder can
+ // extract it; a workflow param literally named `browserType` stays in `parameters`.
+ const { browserType, rest: parameters } = extractBrowserTypeSetting(rest);
+
const parsedParameters = parseValuesForWorkflowRun(
parameters,
workflowParameters,
@@ -824,6 +859,7 @@ function RunWorkflowForm({
cdpConnectHeaders,
cdpAddress,
runWith,
+ [RESERVED_BROWSER_TYPE_FIELD]: browserType,
aiFallback,
});
}
@@ -842,6 +878,7 @@ function RunWorkflowForm({
"cdpAddress",
"runWith",
"aiFallback",
+ RESERVED_BROWSER_TYPE_FIELD,
]);
// A settings-only edit still fires this subscription, and a fresh object
@@ -910,7 +947,7 @@ function RunWorkflowForm({
return {
method: "POST",
- url: `${env.runsApiBaseUrl}/run/workflows`,
+ url: `${runsApiBaseUrl}/run/workflows`,
body: transformedBody,
headers,
} satisfies ApiCommandOptions;
@@ -1485,6 +1522,76 @@ function RunWorkflowForm({
</AccordionTrigger>
<AccordionContent className="pl-6 pr-1 pt-1">
<div className="space-y-8 pt-5">
+ {hasBrowserTypeOptions(browserTypeOptions) ? (
+ <FormField
+ key={RESERVED_BROWSER_TYPE_FIELD}
+ control={form.control}
+ name={RESERVED_BROWSER_TYPE_FIELD}
+ render={({ field }) => {
+ const current =
+ field.value === null || field.value === undefined
+ ? "default"
+ : (field.value as string);
+ return (
+ <FormItem>
+ <div className="flex gap-16">
+ <FormLabel>
+ <div className="w-72">
+ <div className="flex items-center gap-2 text-lg">
+ Browser Type
+ </div>
+ <h2 className="text-sm text-muted-foreground">
+ Browser engine for this run, overriding the
+ agent's setting. Default keeps the agent's
+ setting. Google Chrome does not support the
+ captcha-solver extension.
+ </h2>
+ </div>
+ </FormLabel>
+ <div className="w-full space-y-2">
+ <FormControl>
+ <Select
+ value={current}
+ onValueChange={(v) =>
+ field.onChange(v === "default" ? null : v)
+ }
+ disabled={browserTypeDisabled}
+ >
+ <SelectTrigger className="w-48">
+ <SelectValue placeholder="Browser Type" />
+ </SelectTrigger>
+ <SelectContent>
+ <SelectItem value="default">
+ Default
+ </SelectItem>
+ {(browserTypeOptions ?? []).map(
+ (option) => (
+ <SelectItem
+ key={option.value}
+ value={option.value}
+ >
+ {option.label}
+ </SelectItem>
+ ),
+ )}
+ </SelectContent>
+ </Select>
+ </FormControl>
+ {browserTypeDisabled ? (
+ <p className="text-sm text-muted-foreground">
+ An attached browser session or remote
+ address provides its own engine, so the
+ browser type can&apos;t be set for this run.
+ </p>
+ ) : null}
+ <FormMessage />
+ </div>
+ </div>
+ </FormItem>
+ );
+ }}
+ />
+ ) : null}
<FormField
key="browserSessionId"
control={form.control}
diff --git a/skyvern-frontend/src/routes/workflows/StudioRouteGates.test.tsx b/skyvern-frontend/src/routes/workflows/StudioRouteGates.test.tsx
new file mode 100644
index 000000000..174b5a42a
--- /dev/null
+++ b/skyvern-frontend/src/routes/workflows/StudioRouteGates.test.tsx
@@ -0,0 +1,148 @@
+// @vitest-environment jsdom
+
+import { render, screen } from "@testing-library/react";
+import { MemoryRouter, Route, Routes, useLocation } from "react-router-dom";
+import { describe, expect, test, vi } from "vitest";
+
+import { WorkflowRunRoute } from "./StudioRouteGates";
+
+vi.mock("./editor/WorkflowEditor", () => ({
+ WorkflowEditor: () => <div data-testid="workflow-editor" />,
+}));
+
+vi.mock("./LegacyBuildRedirect", () => ({
+ LegacyBuildRedirect: () => <div data-testid="legacy-build-redirect" />,
+}));
+
+function LocationProbe() {
+ const location = useLocation();
+ return (
+ <div data-testid="location">
+ {location.pathname + location.search + location.hash}
+ </div>
+ );
+}
+
+describe("WorkflowRunRoute", () => {
+ test("redirects legacy per-agent run URLs to the short run route", () => {
+ render(
+ <MemoryRouter
+ initialEntries={[
+ "/agents/wpid_1/wr_1/overview?active=act_1&iteration=2#frame",
+ ]}
+ >
+ <Routes>
+ <Route
+ path="/agents/:workflowPermanentId/:workflowRunId/*"
+ element={<WorkflowRunRoute />}
+ />
+ <Route path="/runs/:runId" element={<LocationProbe />} />
+ </Routes>
+ </MemoryRouter>,
+ );
+
+ expect(screen.getByTestId("location").textContent).toBe(
+ "/runs/wr_1?active=act_1&iteration=2&view=timeline&panes=overview,browser#frame",
+ );
+ });
+
+ test.each([
+ ["blocks", "timeline", "overview,browser"],
+ ["output", "timeline", "overview,browser"],
+ ["parameters", "timeline", "overview,browser"],
+ ["recording", "recording", "browser,overview"],
+ ["code", "code", "overview,browser"],
+ ])("preserves the legacy %s subview", (legacySubview, studioView, panes) => {
+ render(
+ <MemoryRouter
+ initialEntries={[
+ `/agents/wpid_1/wr_1/${legacySubview}?active=act_1#frame`,
+ ]}
+ >
+ <Routes>
+ <Route
+ path="/agents/:workflowPermanentId/:workflowRunId/*"
+ element={<WorkflowRunRoute />}
+ />
+ <Route path="/runs/:runId" element={<LocationProbe />} />
+ </Routes>
+ </MemoryRouter>,
+ );
+
+ expect(screen.getByTestId("location").textContent).toBe(
+ `/runs/wr_1?active=act_1&view=${studioView}&panes=${panes}#frame`,
+ );
+ });
+
+ test.each([
+ ["code", "browser", "overview,browser"],
+ ["recording", "overview", "browser,overview"],
+ ])(
+ "adds the pane needed for %s to an explicit layout",
+ (legacySubview, initialPanes, expectedPanes) => {
+ render(
+ <MemoryRouter
+ initialEntries={[
+ `/agents/wpid_1/wr_1/${legacySubview}?panes=${initialPanes}`,
+ ]}
+ >
+ <Routes>
+ <Route
+ path="/agents/:workflowPermanentId/:workflowRunId/*"
+ element={<WorkflowRunRoute />}
+ />
+ <Route path="/runs/:runId" element={<LocationProbe />} />
+ </Routes>
+ </MemoryRouter>,
+ );
+
+ expect(screen.getByTestId("location").textContent).toContain(
+ `panes=${expectedPanes}`,
+ );
+ },
+ );
+
+ test.each([
+ ["output", "outputs"],
+ ["parameters", "inputs"],
+ ])(
+ "preserves the run-level legacy %s subview",
+ (legacySubview, studioView) => {
+ render(
+ <MemoryRouter initialEntries={[`/agents/wpid_1/wr_1/${legacySubview}`]}>
+ <Routes>
+ <Route
+ path="/agents/:workflowPermanentId/:workflowRunId/*"
+ element={<WorkflowRunRoute />}
+ />
+ <Route path="/runs/:runId" element={<LocationProbe />} />
+ </Routes>
+ </MemoryRouter>,
+ );
+
+ expect(screen.getByTestId("location").textContent).toBe(
+ `/runs/wr_1?view=${studioView}&panes=overview,browser`,
+ );
+ },
+ );
+
+ test("keeps embedded legacy links focused on their requested pane", () => {
+ render(
+ <MemoryRouter
+ initialEntries={["/agents/wpid_1/wr_1/code?embed=true&panes=browser"]}
+ >
+ <Routes>
+ <Route
+ path="/agents/:workflowPermanentId/:workflowRunId/*"
+ element={<WorkflowRunRoute />}
+ />
+ <Route path="/runs/:runId" element={<LocationProbe />} />
+ </Routes>
+ </MemoryRouter>,
+ );
+
+ expect(screen.getByTestId("location").textContent).toBe(
+ "/runs/wr_1?embed=true&panes=overview&view=code",
+ );
+ });
+});
diff --git a/skyvern-frontend/src/routes/workflows/StudioRouteGates.tsx b/skyvern-frontend/src/routes/workflows/StudioRouteGates.tsx
index 7e6c7f3a8..5395a61f2 100644
--- a/skyvern-frontend/src/routes/workflows/StudioRouteGates.tsx
+++ b/skyvern-frontend/src/routes/workflows/StudioRouteGates.tsx
@@ -1,44 +1,32 @@
import { Navigate, useLocation, useParams } from "react-router-dom";
-import { useWorkflowStudioEnabled } from "@/hooks/useWorkflowStudioEnabled";
-
-import { DebugToBuildRedirect } from "./DebugToBuildRedirect";
import { LegacyBuildRedirect } from "./LegacyBuildRedirect";
-import { Debugger } from "./debugger/Debugger";
import { WorkflowEditor } from "./editor/WorkflowEditor";
+import {
+ parsePanesParam,
+ RUN_APPEND_PANES,
+ STUDIO_PANES_PARAM,
+ toReadableSearch,
+} from "./studio/panes";
-/**
- * /build (and block-build) routes. With the studio preview on, these redirect
- * into the editor; off, they render the legacy Debugger.
- */
export function BuildRoute() {
- const studioEnabled = useWorkflowStudioEnabled();
- return studioEnabled ? <LegacyBuildRedirect /> : <Debugger />;
+ return <LegacyBuildRedirect />;
}
-/**
- * /debug (and block-debug) routes. With the studio preview on, these redirect
- * into the editor; off, they fall back to the legacy debug→build redirect.
- */
export function DebugRoute() {
- const studioEnabled = useWorkflowStudioEnabled();
- return studioEnabled ? <LegacyBuildRedirect /> : <DebugToBuildRedirect />;
+ return <LegacyBuildRedirect />;
}
-/**
- * /studio — the studio shell when the preview is on; off, bounce to the legacy
- * /edit editor so a shared studio link never strands a non-flagged user.
- */
export function StudioRoute() {
- const studioEnabled = useWorkflowStudioEnabled();
+ return <WorkflowEditor />;
+}
+
+export function EditRoute() {
const { workflowPermanentId } = useParams();
const location = useLocation();
- if (studioEnabled) {
- return <WorkflowEditor />;
- }
return (
<Navigate
- to={`/agents/${workflowPermanentId}/edit${location.search}`}
+ to={`/agents/${workflowPermanentId}/studio${location.search}`}
state={location.state}
replace
/>
@@ -46,21 +34,64 @@ export function StudioRoute() {
}
/**
- * /edit — the legacy editor when the preview is off; on, the studio lives at
- * /studio, so redirect there (preserving query + state) for old links.
+ * Legacy per-agent run URLs now land on the short run URL, where workflow runs
+ * render inside the studio.
*/
-export function EditRoute() {
- const studioEnabled = useWorkflowStudioEnabled();
- const { workflowPermanentId } = useParams();
+export function WorkflowRunRoute() {
+ const params = useParams();
+ const workflowRunId = params.workflowRunId;
const location = useLocation();
- if (studioEnabled) {
- return (
- <Navigate
- to={`/agents/${workflowPermanentId}/studio${location.search}`}
- state={location.state}
- replace
- />
+
+ if (!workflowRunId) {
+ return <Navigate to="/runs" replace />;
+ }
+
+ const legacySubview = params["*"]?.split("/")[0] || undefined;
+ const studioView = legacySubview
+ ? {
+ overview: "timeline",
+ blocks: "timeline",
+ output: "outputs",
+ parameters: "inputs",
+ recording: "recording",
+ code: "code",
+ }[legacySubview]
+ : undefined;
+ const searchParams = new URLSearchParams(location.search);
+ const routedStudioView =
+ searchParams.has("active") &&
+ (studioView === "outputs" || studioView === "inputs")
+ ? "timeline"
+ : studioView;
+ if (routedStudioView) {
+ searchParams.set("view", routedStudioView);
+ const embedded = searchParams.get("embed") === "true";
+ const requiredPane =
+ routedStudioView === "recording" ? "browser" : "overview";
+ const panes = embedded
+ ? [requiredPane]
+ : (parsePanesParam(searchParams.get(STUDIO_PANES_PARAM)) ?? [
+ ...RUN_APPEND_PANES,
+ ]);
+ searchParams.set(
+ STUDIO_PANES_PARAM,
+ embedded
+ ? requiredPane
+ : [requiredPane, ...panes.filter((pane) => pane !== requiredPane)].join(
+ ",",
+ ),
);
}
- return <WorkflowEditor />;
+
+ return (
+ <Navigate
+ to={{
+ pathname: `/runs/${encodeURIComponent(workflowRunId)}`,
+ search: toReadableSearch(searchParams),
+ hash: location.hash,
+ }}
+ state={location.state}
+ replace
+ />
+ );
}
diff --git a/skyvern-frontend/src/routes/workflows/WorkflowPage.tsx b/skyvern-frontend/src/routes/workflows/WorkflowPage.tsx
index 591197f63..59dbcadbf 100644
--- a/skyvern-frontend/src/routes/workflows/WorkflowPage.tsx
+++ b/skyvern-frontend/src/routes/workflows/WorkflowPage.tsx
@@ -75,7 +75,7 @@ import { useQuery } from "@tanstack/react-query";
import { useCredentialGetter } from "@/hooks/useCredentialGetter";
import { useGlobalWorkflowsQuery } from "./hooks/useGlobalWorkflowsQuery";
import { useWorkflowStudioEnabled } from "@/hooks/useWorkflowStudioEnabled";
-import { legacyRunDetailPath, workflowEditorPath } from "./studioNavigation";
+import { workflowEditorPath, workflowRunDetailPath } from "./studioNavigation";
import { TableSearchInput } from "@/components/TableSearchInput";
import { useKeywordSearch } from "./hooks/useKeywordSearch";
import { useParameterExpansion } from "./hooks/useParameterExpansion";
@@ -468,12 +468,9 @@ function WorkflowPage() {
const isRowSelected = isSelected(
workflowRun.workflow_run_id,
);
- const runPath = studioEnabled
- ? `/runs/${workflowRun.workflow_run_id}`
- : legacyRunDetailPath(
- workflowPermanentId,
- workflowRun.workflow_run_id,
- );
+ const runPath = workflowRunDetailPath(
+ workflowRun.workflow_run_id,
+ );
const mainRow = (kebab: React.ReactNode) => (
<TableRow
diff --git a/skyvern-frontend/src/routes/workflows/WorkflowRun.tsx b/skyvern-frontend/src/routes/workflows/WorkflowRun.tsx
deleted file mode 100644
index 293dfa8d6..000000000
--- a/skyvern-frontend/src/routes/workflows/WorkflowRun.tsx
+++ /dev/null
@@ -1,998 +0,0 @@
-import { useRunCompletionToast } from "./workflowRun/useRunCompletionToast";
-import { claimRunCompletionNotice } from "@/routes/workflows/workflowRun/runCompletionNotices";
-import {
- getRunAttempt,
- runIsRetryWaiting,
- runIsCancellable,
- runIsLogicallyFinal,
-} from "@/routes/workflows/workflowRun/runRetryState";
-import { AxiosError } from "axios";
-import { useCallback, useEffect, useMemo, useRef, useState } from "react";
-import { getClient } from "@/api/AxiosClient";
-import { ProxyLocation, Status } from "@/api/types";
-import { FailureCategoryBadge } from "@/components/FailureCategoryBadge";
-import { StatusBadge } from "@/components/StatusBadge";
-import { WorkflowRunAttemptChip } from "@/components/WorkflowRunAttemptChip";
-import { CredentialFallbackRetryBadge } from "@/components/CredentialFallbackRetryBadge";
-import {
- SwitchBarNavigation,
- type SwitchBarNavigationOption,
-} from "@/components/SwitchBarNavigation";
-import { Button } from "@/components/ui/button";
-import { Status404 } from "@/components/Status404";
-import {
- Dialog,
- DialogClose,
- DialogContent,
- DialogDescription,
- DialogFooter,
- DialogHeader,
- DialogTitle,
- DialogTrigger,
-} from "@/components/ui/dialog";
-import { Skeleton } from "@/components/ui/skeleton";
-import { toast } from "@/components/ui/use-toast";
-import { useApiCredential } from "@/hooks/useApiCredential";
-import { useCredentialGetter } from "@/hooks/useCredentialGetter";
-import { runsApiBaseUrl } from "@/util/env";
-import { basicLocalTimeFormat, basicTimeFormat } from "@/util/timeFormat";
-import {
- CodeIcon,
- FileIcon,
- Pencil2Icon,
- PlayIcon,
- ReloadIcon,
-} from "@radix-ui/react-icons";
-import { useMutation, useQueryClient } from "@tanstack/react-query";
-import {
- Link,
- Navigate,
- Outlet,
- useLocation,
- useNavigate,
- useParams,
- useSearchParams,
-} from "react-router-dom";
-import {
- runOverviewScreenshotLocation,
- runViewTabBasePath,
-} from "@/routes/runs/runViewTabBasePath";
-import { useWorkflowRunWithWorkflowQuery } from "./hooks/useWorkflowRunWithWorkflowQuery";
-import { useRefreshOnboardingOnRunCompletion } from "./hooks/useRefreshOnboardingOnRunCompletion";
-import { ResizableTimelineSplit } from "./workflowRun/ResizableTimelineSplit";
-import { WorkflowRunBlockDetail } from "./workflowRun/WorkflowRunBlockDetail";
-import { WorkflowRunTimeline } from "./workflowRun/WorkflowRunTimeline";
-import { useWorkflowRunTimelineQuery } from "./hooks/useWorkflowRunTimelineQuery";
-import {
- filterTimelineToAttempt,
- findActiveItem,
- parseActiveIterationParam,
-} from "./workflowRun/workflowTimelineUtils";
-import { ArtifactDownloadLink } from "@/components/ArtifactDownloadLink";
-import { pickDownloadedFileFilename } from "./workflowRun/blockDownloadedFiles";
-import { findRunCodeBlockFailure } from "./workflowRun/codeBlockFailure";
-import { matchFailureTips } from "./studio/runview/failureTips";
-import { isBlockItem } from "./types/workflowRunTypes";
-import { Label } from "@/components/ui/label";
-import { CodeEditor } from "./components/CodeEditor";
-import { cn } from "@/util/utils";
-import { ScrollArea, ScrollAreaViewport } from "@/components/ui/scroll-area";
-import { ApiWebhookActionsMenu } from "@/components/ApiWebhookActionsMenu";
-import { WebhookReplayDialog } from "@/components/WebhookReplayDialog";
-import { useFirstParam } from "@/hooks/useFirstParam";
-import { type ApiCommandOptions } from "@/util/apiCommands";
-import { useBlockScriptsQuery } from "@/routes/workflows/hooks/useBlockScriptsQuery";
-import { constructCacheKeyValue } from "@/routes/workflows/editor/utils";
-import { useCacheKeyValuesQuery } from "@/routes/workflows/hooks/useCacheKeyValuesQuery";
-import { WorkflowRunStatusAlert } from "@/routes/workflows/workflowRun/WorkflowRunStatusAlert";
-import { WorkflowRunVerificationCodeForm } from "@/routes/workflows/workflowRun/WorkflowRunVerificationCodeForm";
-import { ScriptUpdateCard } from "@/routes/workflows/workflowRun/ScriptUpdateCard";
-import { useFallbackEpisodesQuery } from "@/routes/workflows/hooks/useFallbackEpisodesQuery";
-import { usePageSlots } from "@/store/PageSlots";
-import { useOnboardingStateOptional } from "@/store/onboarding/useOnboardingState";
-import { useWorkflowStudioEnabled } from "@/hooks/useWorkflowStudioEnabled";
-import { workflowEditorPath } from "@/routes/workflows/studioNavigation";
-import {
- FirstRunRecoveryGuidance,
- shouldShowRecoveryGuidance,
-} from "@/components/onboarding/FirstRunRecoveryGuidance";
-import { useFeatureFlag } from "@/hooks/useFeatureFlag";
-import {
- getRecoveryGuidanceRetryNavigation,
- RecoveryGuidanceTelemetry,
- retryRunHasStarted,
- type RecoveryGuidanceTelemetryContext,
-} from "@/util/onboarding/recoveryGuidanceTelemetry";
-import { RunTagsEditor } from "@/routes/tasks/components/tagging/RunTagsEditor";
-import { getRerunNavigationState, shouldPollForGeneratedCode } from "./utils";
-
-const RECOVERY_GUIDANCE_TREATMENT_SURFACE_FLAG =
- "RECOVERY_GUIDANCE_TREATMENT_SURFACE";
-
-function WorkflowRunRightColumn({
- workflowRunId,
- activeItem,
- activeIteration,
- timeline,
- timelineReady,
- onSetActiveItem,
- onSetActiveIteration,
- onViewScreenshot,
-}: {
- workflowRunId: string | undefined;
- activeItem: ReturnType<typeof findActiveItem>;
- activeIteration: number | null;
- timeline: NonNullable<ReturnType<typeof useWorkflowRunTimelineQuery>["data"]>;
- timelineReady: boolean;
- onSetActiveItem: (id: string) => void;
- onSetActiveIteration: (loopBlockId: string, iterationIndex: number) => void;
- onViewScreenshot: (workflowRunBlockId: string) => void;
-}) {
- return (
- <ResizableTimelineSplit
- className="w-[clamp(28rem,34vw,36rem)] shrink-0"
- top={
- <div className="min-h-0 w-full overflow-hidden">
- <WorkflowRunTimeline
- workflowRunId={workflowRunId}
- activeItem={activeItem}
- activeIteration={activeIteration}
- onActionItemSelected={(item) => {
- onSetActiveItem(item.action.action_id);
- }}
- onBlockItemSelected={(item) => {
- onSetActiveItem(item.workflow_run_block_id);
- }}
- onThoughtItemSelected={(item) => {
- onSetActiveItem(item.thought_id);
- }}
- onLiveStreamSelected={() => {
- onSetActiveItem("stream");
- }}
- onIterationSelected={(loopBlock, iterationIndex) => {
- onSetActiveIteration(
- loopBlock.workflow_run_block_id,
- iterationIndex,
- );
- }}
- />
- </div>
- }
- bottom={
- <div className="flex min-h-0 w-full flex-col overflow-hidden rounded-md border border-border bg-slate-elevation1">
- <WorkflowRunBlockDetail
- workflowRunId={workflowRunId}
- activeItem={activeItem}
- activeIteration={activeIteration}
- timeline={timeline}
- timelineReady={timelineReady}
- onViewScreenshot={onViewScreenshot}
- onThoughtSelect={(thought) => {
- onSetActiveItem(thought.thought_id);
- }}
- />
- </div>
- }
- />
- );
-}
-
-function WorkflowRun() {
- const [searchParams, setSearchParams] = useSearchParams();
- const searchParamsRef = useRef(searchParams);
- searchParamsRef.current = searchParams;
- const embed = searchParams.get("embed");
- const isEmbedded = embed === "true";
- const active = searchParams.get("active");
- const iterationParam = searchParams.get("iteration");
- const activeIteration = parseActiveIterationParam(iterationParam);
- const workflowRunId = useFirstParam("workflowRunId", "runId");
- const runBasePath =
- runViewTabBasePath(useParams()) ?? `/runs/${workflowRunId}`;
- const workflowPermanentIdParam = useFirstParam("workflowPermanentId");
- const credentialGetter = useCredentialGetter();
- const apiCredential = useApiCredential();
- const queryClient = useQueryClient();
- const navigate = useNavigate();
- const location = useLocation();
- const studioEnabled = useWorkflowStudioEnabled();
- const onboarding = useOnboardingStateOptional();
- const { workflowRunMilestoneCard: WorkflowRunMilestoneCard } = usePageSlots();
-
- const {
- data: workflowRun,
- isLoading: workflowRunIsLoading,
- isPlaceholderData: workflowRunIsPlaceholder,
- isFetched,
- error,
- } = useWorkflowRunWithWorkflowQuery();
-
- useRefreshOnboardingOnRunCompletion(workflowRun);
- useRunCompletionToast(workflowRunIsPlaceholder ? undefined : workflowRun);
- const recoveryGuidanceRetry = getRecoveryGuidanceRetryNavigation(
- location.state,
- );
- const reportedRecoveryRetryStartRef = useRef<string | null>(null);
-
- useEffect(() => {
- if (
- !recoveryGuidanceRetry ||
- !retryRunHasStarted({
- retryRunId: recoveryGuidanceRetry.retryRunId,
- observedRunId: workflowRun?.workflow_run_id,
- status: workflowRun?.status,
- startedAt: workflowRun?.started_at,
- }) ||
- reportedRecoveryRetryStartRef.current === recoveryGuidanceRetry.retryRunId
- ) {
- return;
- }
- reportedRecoveryRetryStartRef.current = recoveryGuidanceRetry.retryRunId;
- RecoveryGuidanceTelemetry.retryStarted(
- recoveryGuidanceRetry,
- recoveryGuidanceRetry.retryRunId,
- );
- }, [
- recoveryGuidanceRetry,
- workflowRun?.started_at,
- workflowRun?.status,
- workflowRun?.workflow_run_id,
- ]);
-
- const status = (error as AxiosError | undefined)?.response?.status;
- const workflow = workflowRun?.workflow;
- const workflowPermanentId =
- workflowPermanentIdParam ?? workflow?.workflow_permanent_id;
- const cacheKey = workflow?.cache_key ?? "";
- const isFinalized = workflowRun ? runIsLogicallyFinal(workflowRun) : null;
- const isWorkflowDeleted = Boolean(workflow?.deleted_at);
-
- const [hasPublishedCode, setHasPublishedCode] = useState(false);
- const isGeneratingCode = shouldPollForGeneratedCode(
- workflow,
- isFinalized,
- hasPublishedCode,
- );
-
- const [cacheKeyValue, setCacheKeyValue] = useState(
- cacheKey === ""
- ? ""
- : constructCacheKeyValue({ codeKey: cacheKey, workflow, workflowRun }),
- );
-
- const { data: cacheKeyValues } = useCacheKeyValuesQuery({
- cacheKey,
- debounceMs: 100,
- page: 1,
- workflowPermanentId: isWorkflowDeleted ? undefined : workflowPermanentId,
- });
-
- useEffect(() => {
- setCacheKeyValue(
- constructCacheKeyValue({ codeKey: cacheKey, workflow, workflowRun }) ??
- cacheKeyValues?.values[0],
- );
- }, [cacheKey, cacheKeyValues, setCacheKeyValue, workflow, workflowRun]);
-
- const { data: blockScriptsPublished } = useBlockScriptsQuery({
- cacheKey,
- cacheKeyValue,
- workflowPermanentId: isWorkflowDeleted ? undefined : workflowPermanentId,
- pollIntervalMs: isGeneratingCode ? 3000 : undefined,
- status: "published",
- workflowRunId: workflowRun?.workflow_run_id,
- });
-
- useEffect(() => {
- const keys = Object.keys(blockScriptsPublished?.blocks ?? {});
- setHasPublishedCode(
- keys.length > 0 || Boolean(blockScriptsPublished?.main_script),
- );
- }, [blockScriptsPublished, setHasPublishedCode]);
-
- const { data: retainedTimeline, isPlaceholderData: timelineIsPlaceholder } =
- useWorkflowRunTimelineQuery();
- const workflowRunTimeline = timelineIsPlaceholder
- ? undefined
- : retainedTimeline;
- const [replayOpen, setReplayOpen] = useState(false);
-
- const cancelWorkflowMutation = useMutation({
- mutationFn: async () => {
- const client = await getClient(credentialGetter);
- return client
- .post(`/workflows/runs/${workflowRunId}/cancel`)
- .then((response) => response.data);
- },
- onSuccess: () => {
- queryClient.invalidateQueries({
- queryKey: ["workflowRun", workflowRunId],
- });
- queryClient.invalidateQueries({
- queryKey: ["workflowRun", workflowPermanentId, workflowRunId],
- });
- if (workflowRunId) {
- claimRunCompletionNotice(workflowRunId, () => {
- toast({
- variant: "success",
- title: "Agent Canceled",
- description: "The agent has been successfully canceled.",
- });
- });
- }
- },
- onError: (error) => {
- toast({
- variant: "destructive",
- title: "Error",
- description: error.message,
- });
- },
- });
-
- const statusUnavailable = Boolean(error);
- const workflowRunIsCancellable =
- !statusUnavailable && workflowRun && runIsCancellable(workflowRun);
-
- const workflowRunIsFinalized =
- workflowRun && runIsLogicallyFinal(workflowRun);
-
- const { data: fallbackEpisodes } = useFallbackEpisodesQuery({
- workflowPermanentId,
- workflowRunId: workflowRun?.workflow_run_id,
- enabled: workflowRunIsFinalized === true && !isWorkflowDeleted,
- });
- const finallyBlockLabel =
- workflow?.workflow_definition?.finally_block_label ?? null;
- const currentAttemptTimeline = filterTimelineToAttempt(
- workflowRunTimeline ?? [],
- workflowRun?.attempts ?? [],
- getRunAttempt(workflowRun ?? {}),
- );
- const selection = findActiveItem(
- active === null ? currentAttemptTimeline : (workflowRunTimeline ?? []),
- active,
- !!workflowRunIsFinalized,
- finallyBlockLabel,
- );
- const parameters = workflowRun?.parameters ?? {};
- const proxyLocation =
- workflowRun?.proxy_location ?? ProxyLocation.Residential;
- const maxScreenshotScrolls = workflowRun?.max_screenshot_scrolls ?? null;
-
- const title = workflowRunIsLoading ? (
- <Skeleton className="h-9 w-48" />
- ) : isWorkflowDeleted ? (
- <h1 className="text-3xl">{workflow!.title}</h1>
- ) : (
- <h1 className="text-3xl">
- <Link
- className="hover:underline hover:underline-offset-2"
- to={`/agents/${workflowPermanentId}/runs`}
- >
- {workflow?.title}
- </Link>
- </h1>
- );
-
- const failureReason = workflowRun?.failure_reason;
-
- const matchedTips = matchFailureTips(failureReason ?? null).map((tip) => (
- <div key={tip} className="text-sm italic text-red-700">
- {tip}
- </div>
- ));
-
- const codeFailure = findRunCodeBlockFailure(
- failureReason,
- currentAttemptTimeline,
- finallyBlockLabel,
- );
-
- const failureReasonTitle =
- workflowRun?.status === Status.Terminated
- ? "Termination Reason"
- : "Failure Reason";
- const siblingRunLinkClassName =
- "font-mono text-sm text-neutral-600 hover:text-neutral-950 hover:underline hover:underline-offset-2 dark:text-slate-400 dark:hover:text-slate-200";
-
- const finallyBlockInTimeline = finallyBlockLabel
- ? currentAttemptTimeline.find(
- (item) => isBlockItem(item) && item.block.label === finallyBlockLabel,
- )
- : null;
-
- const finallyBlockStatus =
- finallyBlockInTimeline && isBlockItem(finallyBlockInTimeline)
- ? finallyBlockInTimeline.block.status
- : null;
-
- const shouldShowFinallyNote =
- (workflowRun?.status === Status.Terminated ||
- workflowRun?.status === Status.Failed) &&
- finallyBlockLabel &&
- finallyBlockInTimeline;
-
- const recoveryGuidanceAssignment =
- onboarding?.recoveryGuidanceAssignment ?? null;
- const recoveryGuidanceTreatmentSurfaceEnabled =
- useFeatureFlag(RECOVERY_GUIDANCE_TREATMENT_SURFACE_FLAG) === true;
- const recoveryGuidanceTelemetryContext =
- useMemo<RecoveryGuidanceTelemetryContext | null>(() => {
- if (!recoveryGuidanceAssignment || !workflowRun) {
- return null;
- }
- return {
- organizationId: recoveryGuidanceAssignment.organization_id,
- experimentVersion: recoveryGuidanceAssignment.experiment_version,
- arm: recoveryGuidanceAssignment.arm,
- eligibleRunId: recoveryGuidanceAssignment.eligible_run_id,
- failureCategory: workflowRun.failure_category?.[0]?.category ?? null,
- };
- }, [recoveryGuidanceAssignment, workflowRun]);
- // This is a new, intentionally unconfigured flag. Its unknown/default value
- // is false, so neither arm gains a visible surface during instrumentation.
- const showFirstFailedRunRecovery = shouldShowRecoveryGuidance({
- assignment: recoveryGuidanceAssignment,
- workflowRunId: workflowRun?.workflow_run_id,
- treatmentSurfaceEnabled: recoveryGuidanceTreatmentSurfaceEnabled,
- });
-
- const handleFirstFailedRunRetry = useCallback(() => {
- navigate(`/agents/${workflowPermanentId}/run`, {
- state: {
- data: workflowRun?.parameters ?? {},
- proxyLocation,
- webhookCallbackUrl: workflowRun?.webhook_callback_url ?? "",
- maxScreenshotScrolls,
- runWith: workflowRun?.run_with ?? "agent",
- browserProfileId: workflowRun?.browser_profile_id ?? null,
- ...(recoveryGuidanceTelemetryContext
- ? { recoveryGuidanceRetry: recoveryGuidanceTelemetryContext }
- : {}),
- },
- });
- }, [
- navigate,
- workflowPermanentId,
- proxyLocation,
- maxScreenshotScrolls,
- recoveryGuidanceTelemetryContext,
- workflowRun?.parameters,
- workflowRun?.webhook_callback_url,
- workflowRun?.run_with,
- workflowRun?.browser_profile_id,
- ]);
-
- const workflowFailureReason = workflowRun?.failure_reason ? (
- <div className="space-y-2 rounded-md border border-red-600 bg-error-light p-4">
- <div className="flex items-center gap-2">
- <div className="font-bold">{failureReasonTitle}</div>
- <FailureCategoryBadge failureCategory={workflowRun.failure_category} />
- </div>
- {codeFailure ? (
- <div className="space-y-1">
- <div className="flex flex-wrap items-center gap-1.5 text-sm font-medium">
- <span>{codeFailure.title}</span>
- {codeFailure.line !== null ? (
- <span className="shrink-0 rounded border border-red-600/40 px-1.5 py-0.5 text-[10px] tabular-nums">
- line {codeFailure.line}
- </span>
- ) : null}
- {codeFailure.code ? (
- <span className="shrink-0 rounded border border-red-600/40 px-1.5 py-0.5 font-mono text-[10px]">
- {codeFailure.code}
- </span>
- ) : null}
- </div>
- <div className="text-sm">{codeFailure.guidance}</div>
- </div>
- ) : null}
- <div className="text-sm">{workflowRun.failure_reason}</div>
- {matchedTips}
- {showFirstFailedRunRecovery && recoveryGuidanceTelemetryContext && (
- <FirstRunRecoveryGuidance
- telemetryContext={recoveryGuidanceTelemetryContext}
- workflowPermanentId={workflowPermanentId}
- onRetry={handleFirstFailedRunRetry}
- />
- )}
- {shouldShowFinallyNote && (
- <div className="mt-2 flex items-center gap-2 rounded bg-amber-500/20 px-3 py-2 text-sm text-amber-700 dark:text-amber-200">
- <span className="font-medium">Note:</span>
- <span>
- "Execute on any outcome" block ({finallyBlockLabel}){" "}
- {finallyBlockStatus === Status.Completed
- ? "completed successfully"
- : finallyBlockStatus === Status.Failed
- ? "failed"
- : "ran"}
- .
- </span>
- </div>
- )}
- </div>
- ) : null;
-
- const updateSearchParams = useCallback(
- (mutate: (params: URLSearchParams) => void) => {
- setSearchParams(
- () => {
- const next = new URLSearchParams(searchParamsRef.current);
- mutate(next);
- searchParamsRef.current = next;
- return next;
- },
- { replace: true },
- );
- },
- [setSearchParams],
- );
-
- function handleSetActiveItem(id: string) {
- updateSearchParams((next) => {
- next.set("active", id);
- next.delete("iteration");
- });
- }
-
- const handleViewScreenshot = useCallback(
- (workflowRunBlockId: string) => {
- const destination = runOverviewScreenshotLocation(
- runBasePath,
- searchParamsRef.current.toString(),
- workflowRunBlockId,
- );
- searchParamsRef.current = new URLSearchParams(destination.search);
- navigate(destination, { replace: true });
- },
- [navigate, runBasePath],
- );
-
- function handleSetActiveIteration(
- loopBlockId: string,
- iterationIndex: number,
- ) {
- updateSearchParams((next) => {
- next.set("active", loopBlockId);
- next.set("iteration", String(iterationIndex));
- });
- }
-
- const isTaskv2Run = workflowRun && workflowRun.task_v2 !== null;
-
- const webhookFailureReasonData =
- workflowRun?.task_v2?.webhook_failure_reason ??
- workflowRun?.webhook_failure_reason;
-
- const webhookFailureReason = webhookFailureReasonData ? (
- <div className="space-y-4">
- <Label>Webhook Failure Reason</Label>
- <div className="rounded-md border border-yellow-600 p-4 text-sm">
- {webhookFailureReasonData}
- </div>
- </div>
- ) : null;
-
- const outputs = workflowRun?.outputs;
- const extractedInformation =
- typeof outputs === "object" &&
- outputs !== null &&
- "extracted_information" in outputs
- ? (outputs.extracted_information as Record<string, unknown>)
- : null;
-
- const hasSomeExtractedInformation = extractedInformation
- ? Object.values(extractedInformation).some((value) => value !== null)
- : false;
-
- const hasTaskv2Output = Boolean(isTaskv2Run && workflowRun.task_v2?.output);
-
- const hasFileUrls =
- isFetched &&
- workflowRun &&
- workflowRun.downloaded_file_urls &&
- workflowRun.downloaded_file_urls.length > 0;
- const fileUrls = hasFileUrls
- ? (workflowRun.downloaded_file_urls as string[])
- : [];
- // Prefer the rich downloaded_files array (carries filename, checksum, size)
- // when the backend sends it; falls back to URL parsing otherwise.
- const filenameByUrl = new Map<string, string>();
- if (workflowRun?.downloaded_files) {
- for (const file of workflowRun.downloaded_files) {
- if (file.filename) {
- filenameByUrl.set(file.url, file.filename);
- }
- }
- }
-
- const showBoth =
- (hasSomeExtractedInformation || hasTaskv2Output) && hasFileUrls;
-
- const showOutputSection =
- workflowRunIsFinalized &&
- (hasSomeExtractedInformation ||
- hasFileUrls ||
- hasTaskv2Output ||
- webhookFailureReasonData) &&
- workflowRun.status === Status.Completed;
-
- const switchBarOptions: SwitchBarNavigationOption[] = [
- {
- label: "Overview",
- to: `${runBasePath}/overview`,
- },
- {
- label: "Output",
- to: `${runBasePath}/output`,
- },
- {
- label: "Inputs",
- to: `${runBasePath}/parameters`,
- },
- {
- label: "Recording",
- to: `${runBasePath}/recording`,
- },
- {
- label: "Code",
- to: `${runBasePath}/code`,
- icon: !isGeneratingCode ? (
- <CodeIcon className="inline-block size-5" />
- ) : (
- <ReloadIcon className="inline-block size-5 animate-spin" />
- ),
- },
- ];
-
- if (status === 404) {
- return <Status404 />;
- }
-
- // With the preview on, route legacy run links into the studio run view under
- // the short /runs/{wr} URL (preserving the selected item); flag-off keeps this
- // legacy run view. The run id lives in the path, so it stays out of the query.
- if (studioEnabled && !isEmbedded && workflowRunId && workflowPermanentId) {
- const studioParams = new URLSearchParams();
- if (active) {
- studioParams.set("active", active);
- }
- if (iterationParam) {
- studioParams.set("iteration", iterationParam);
- }
- const search = studioParams.toString();
- return (
- <Navigate
- to={`/runs/${workflowRunId}${search ? `?${search}` : ""}`}
- replace
- />
- );
- }
-
- return (
- <div className="space-y-8">
- {!isEmbedded && (
- <header className="flex justify-between">
- <div className="space-y-3">
- <div className="mr-2 flex items-start gap-5">
- {title}
- {workflowRunIsLoading ? (
- <Skeleton className="h-8 w-28" />
- ) : workflowRun && !statusUnavailable ? (
- <div className="mt-[0.27rem] flex items-center gap-2">
- <StatusBadge status={workflowRun?.status} />
- <WorkflowRunAttemptChip
- attempt={workflowRun.attempt}
- retryPending={workflowRun.retry_pending}
- nextAttemptAt={workflowRun.next_attempt_at}
- />
- <CredentialFallbackRetryBadge
- retriedFromWorkflowRunId={
- workflowRun.retried_from_workflow_run_id
- }
- />
- </div>
- ) : null}
- </div>
- <h2 className="text-2xl text-neutral-600 dark:text-slate-400">
- {workflowRunId}
- </h2>
- {workflowRunId ? (
- <RunTagsEditor workflowRunId={workflowRunId} />
- ) : null}
- {workflowRun && workflowPermanentId && (
- <div className="flex flex-wrap gap-x-4 gap-y-1">
- {workflowRun.retried_from_workflow_run_id && (
- <Link
- className={siblingRunLinkClassName}
- to={`/agents/${workflowPermanentId}/${workflowRun.retried_from_workflow_run_id}/overview`}
- >
- Retried from {workflowRun.retried_from_workflow_run_id}
- </Link>
- )}
- {workflowRun.retried_by_workflow_run_id && (
- <Link
- className={siblingRunLinkClassName}
- to={`/agents/${workflowPermanentId}/${workflowRun.retried_by_workflow_run_id}/overview`}
- >
- Retried by {workflowRun.retried_by_workflow_run_id}
- </Link>
- )}
- </div>
- )}
- {workflowRun &&
- (workflowRun.started_at ||
- workflowRun.finished_at ||
- isWorkflowDeleted) && (
- <div className="flex flex-wrap gap-x-4 gap-y-1 text-sm text-neutral-600 dark:text-slate-400">
- {workflowRun.started_at && (
- <span title={basicTimeFormat(workflowRun.started_at)}>
- Started: {basicLocalTimeFormat(workflowRun.started_at)}
- </span>
- )}
- {workflowRun.finished_at && (
- <span title={basicTimeFormat(workflowRun.finished_at)}>
- Finished: {basicLocalTimeFormat(workflowRun.finished_at)}
- </span>
- )}
- {isWorkflowDeleted && (
- <span title={basicTimeFormat(workflow!.deleted_at!)}>
- Agent deleted on{" "}
- {basicLocalTimeFormat(workflow!.deleted_at!)}
- </span>
- )}
- </div>
- )}
- {workflowRun?.browser_session_id && (
- <Link
- className="font-mono text-sm text-neutral-600 hover:text-neutral-950 hover:underline hover:underline-offset-2 dark:text-slate-400 dark:hover:text-slate-200"
- to={`/browser-session/${workflowRun.browser_session_id}/stream`}
- >
- Browser Session: {workflowRun.browser_session_id}
- </Link>
- )}
- {workflowRun?.browser_profile_id && (
- <Link
- className="font-mono text-sm text-neutral-600 hover:text-neutral-950 hover:underline hover:underline-offset-2 dark:text-slate-400 dark:hover:text-slate-200"
- to={`/browser-profiles/${workflowRun.browser_profile_id}`}
- >
- Browser Profile: {workflowRun.browser_profile_id}
- </Link>
- )}
- </div>
-
- <div className="flex gap-2">
- {!isWorkflowDeleted && (
- <>
- <ApiWebhookActionsMenu
- getOptions={() => {
- // Build headers - x-max-steps-override is optional and can be added manually if needed
- const headers: Record<string, string> = {
- "Content-Type": "application/json",
- "x-api-key": apiCredential ?? "<your-api-key>",
- };
-
- const body: Record<string, unknown> = {
- workflow_id: workflowPermanentId,
- parameters: workflowRun?.parameters,
- proxy_location: proxyLocation,
- };
-
- if (maxScreenshotScrolls !== null) {
- body.max_screenshot_scrolls = maxScreenshotScrolls;
- }
-
- if (workflowRun?.webhook_callback_url) {
- body.webhook_url = workflowRun.webhook_callback_url;
- }
-
- return {
- method: "POST",
- url: `${runsApiBaseUrl}/run/workflows`,
- body,
- headers,
- } satisfies ApiCommandOptions;
- }}
- disabledReason={
- workflowRun && runIsRetryWaiting(workflowRun)
- ? "Unavailable while a retry is pending"
- : undefined
- }
- webhookDisabled={
- workflowRunIsLoading || !workflowRunIsFinalized
- }
- onTestWebhook={() => setReplayOpen(true)}
- />
- <WebhookReplayDialog
- runId={workflowRunId ?? ""}
- disabled={workflowRunIsLoading || !workflowRunIsFinalized}
- open={replayOpen}
- onOpenChange={setReplayOpen}
- hideTrigger
- />
- <Button asChild variant="secondary">
- <Link
- to={workflowEditorPath(
- workflowPermanentId ?? "",
- studioEnabled,
- )}
- data-testid="workflow-open-editor-link"
- >
- <Pencil2Icon className="mr-2 h-4 w-4" />
- Edit
- </Link>
- </Button>
- </>
- )}
- {workflowRunIsCancellable && (
- <Dialog>
- <DialogTrigger asChild>
- <Button variant="destructive">Cancel</Button>
- </DialogTrigger>
- <DialogContent>
- <DialogHeader>
- <DialogTitle>Are you sure?</DialogTitle>
- <DialogDescription>
- Are you sure you want to cancel this agent run?
- </DialogDescription>
- </DialogHeader>
- <DialogFooter>
- <DialogClose asChild>
- <Button variant="secondary">Back</Button>
- </DialogClose>
- <Button
- variant="destructive"
- onClick={() => {
- cancelWorkflowMutation.mutate();
- }}
- disabled={cancelWorkflowMutation.isPending}
- >
- {cancelWorkflowMutation.isPending && (
- <ReloadIcon className="mr-2 h-4 w-4 animate-spin" />
- )}
- Cancel Agent Run
- </Button>
- </DialogFooter>
- </DialogContent>
- </Dialog>
- )}
- {workflowRunIsFinalized && !isTaskv2Run && !isWorkflowDeleted && (
- <Button asChild>
- <Link
- to={`/agents/${workflowPermanentId}/run`}
- state={{
- data: parameters,
- proxyLocation,
- webhookCallbackUrl: workflowRun?.webhook_callback_url ?? "",
- maxScreenshotScrolls,
- runWith: workflowRun?.run_with ?? "agent",
- browserProfileId: workflowRun?.browser_profile_id ?? null,
- }}
- >
- <PlayIcon className="mr-2 h-4 w-4" />
- Rerun
- </Link>
- </Button>
- )}
- </div>
- </header>
- )}
- {WorkflowRunMilestoneCard &&
- workflowRun &&
- !workflowRunIsPlaceholder &&
- runIsLogicallyFinal(workflowRun) &&
- workflowRun.status === Status.Completed ? (
- <WorkflowRunMilestoneCard
- workflowRunId={workflowRun.workflow_run_id}
- rerun={
- !isEmbedded &&
- !isTaskv2Run &&
- !isWorkflowDeleted &&
- workflowPermanentId
- ? {
- to: `/agents/${workflowPermanentId}/run`,
- state: getRerunNavigationState(workflowRun),
- }
- : undefined
- }
- />
- ) : null}
- {/* 2FA Verification Code Form - shown when workflow is waiting for a code */}
- <WorkflowRunVerificationCodeForm />
- {showOutputSection && (
- <div
- className={cn("grid gap-4 rounded-lg bg-slate-elevation1 p-4", {
- "grid-cols-2": showBoth,
- })}
- >
- {(hasSomeExtractedInformation || hasTaskv2Output) && (
- <div className="space-y-4">
- <Label>
- {hasTaskv2Output ? "Output" : "Extracted Information"}
- </Label>
- <CodeEditor
- language="json"
- value={
- hasTaskv2Output
- ? JSON.stringify(workflowRun.task_v2?.output, null, 2)
- : JSON.stringify(extractedInformation, null, 2)
- }
- readOnly
- maxHeight="250px"
- />
- </div>
- )}
- {hasFileUrls && (
- <div className="space-y-4">
- <Label>Downloaded Files</Label>
- <ScrollArea>
- <ScrollAreaViewport className="max-h-[250px] space-y-2">
- {fileUrls.length > 0 ? (
- fileUrls.map((url) => {
- const filename = pickDownloadedFileFilename(
- url,
- filenameByUrl,
- );
- return (
- <div key={url} title={url} className="flex gap-2">
- <FileIcon className="size-6" />
- <ArtifactDownloadLink
- href={url}
- className="underline underline-offset-4"
- >
- <span>{filename}</span>
- </ArtifactDownloadLink>
- </div>
- );
- })
- ) : (
- <div className="text-sm">No files downloaded</div>
- )}
- </ScrollAreaViewport>
- </ScrollArea>
- </div>
- )}
- {webhookFailureReason}
- </div>
- )}
- {workflowFailureReason}
- {fallbackEpisodes && fallbackEpisodes.episodes.length > 0 && (
- <ScriptUpdateCard
- episodes={fallbackEpisodes.episodes}
- scriptId={blockScriptsPublished?.script_id}
- />
- )}
- {!isEmbedded && (
- <div className="flex items-center justify-between">
- <SwitchBarNavigation options={switchBarOptions} />
- {workflowRun && (
- <WorkflowRunStatusAlert
- run={workflowRun}
- title={workflow?.title}
- visible={!statusUnavailable && !isFinalized}
- />
- )}
- </div>
- )}
- {/* 18rem accounts for nav, run metadata, tabs, and page gutters above this work area. */}
- <div className="flex h-[calc(100vh-18rem)] max-h-[52rem] min-h-[34rem] gap-6">
- <div className="min-w-0 flex-[2]">
- <Outlet />
- </div>
- <WorkflowRunRightColumn
- workflowRunId={workflowRunId}
- activeItem={selection}
- activeIteration={activeIteration}
- timeline={workflowRunTimeline ?? []}
- timelineReady={workflowRunTimeline !== undefined}
- onSetActiveItem={handleSetActiveItem}
- onSetActiveIteration={handleSetActiveIteration}
- onViewScreenshot={handleViewScreenshot}
- />
- </div>
- </div>
- );
-}
-
-export { WorkflowRun };
diff --git a/skyvern-frontend/src/routes/workflows/WorkflowRunParameters.tsx b/skyvern-frontend/src/routes/workflows/WorkflowRunParameters.tsx
index e0a160c82..384317288 100644
--- a/skyvern-frontend/src/routes/workflows/WorkflowRunParameters.tsx
+++ b/skyvern-frontend/src/routes/workflows/WorkflowRunParameters.tsx
@@ -6,7 +6,7 @@ import { RunWorkflowForm } from "./RunWorkflowForm";
import { WorkflowApiResponse } from "./types/workflowTypes";
import { Skeleton } from "@/components/ui/skeleton";
import { ProxyLocation } from "@/api/types";
-import { getInitialValues } from "./utils";
+import { getInitialValues, resolveInitialBrowserType } from "./utils";
import { isMaskedHeaders } from "@/util/secretHeaders";
import { useFeatureFlag } from "@/hooks/useFeatureFlag";
@@ -103,6 +103,13 @@ function WorkflowRunParameters() {
cdpConnectHeaders: cdpConnectHeaders ?? storedCdpConnectHeaders,
cdpAddress: null,
runWith,
+ // A rerun/retry carries the executed run's effective browser_type in location state. A
+ // non-null value preselects what actually ran; a null OR absent value inherits the workflow's
+ // current browser_type (a run-level null means "inherit the workflow setting").
+ browserType: resolveInitialBrowserType(
+ location.state,
+ workflow.browser_type,
+ ),
}}
/>
);
diff --git a/skyvern-frontend/src/routes/workflows/WorkflowScriptDetailPage.tsx b/skyvern-frontend/src/routes/workflows/WorkflowScriptDetailPage.tsx
index 681dd17a2..5fb4d93b4 100644
--- a/skyvern-frontend/src/routes/workflows/WorkflowScriptDetailPage.tsx
+++ b/skyvern-frontend/src/routes/workflows/WorkflowScriptDetailPage.tsx
@@ -404,7 +404,7 @@ function WorkflowScriptDetailPage() {
<TableRow key={run.workflow_run_id}>
<TableCell>
<Link
- to={`/agents/${workflowPermanentId}/${run.workflow_run_id}/overview`}
+ to={`/runs/${run.workflow_run_id}`}
className="font-mono text-sm text-blue-700 hover:underline dark:text-blue-400"
>
{run.workflow_run_id}
diff --git a/skyvern-frontend/src/routes/workflows/copilot/RecordingRefinementProgressCard.tsx b/skyvern-frontend/src/routes/workflows/copilot/RecordingRefinementProgressCard.tsx
new file mode 100644
index 000000000..86c61a840
--- /dev/null
+++ b/skyvern-frontend/src/routes/workflows/copilot/RecordingRefinementProgressCard.tsx
@@ -0,0 +1,162 @@
+import {
+ CheckCircledIcon,
+ CrossCircledIcon,
+ ReloadIcon,
+} from "@radix-ui/react-icons";
+import { useEffect, useState } from "react";
+
+export type RecordingRefinementStatus =
+ | "working"
+ | "complete"
+ | "failed"
+ | "cancelled";
+
+type Props = {
+ actionCount: number;
+ awaitingReview: boolean;
+ startedAtMs: number;
+ status: RecordingRefinementStatus;
+};
+
+function CompletedStep({ children }: { children: string }) {
+ return (
+ <div className="flex items-center gap-2 text-xs text-foreground">
+ <CheckCircledIcon
+ aria-hidden="true"
+ className="size-4 shrink-0 text-emerald-500"
+ />
+ <span>{children}</span>
+ </div>
+ );
+}
+
+function RecordingRefinementProgressCard({
+ actionCount,
+ awaitingReview,
+ startedAtMs,
+ status,
+}: Props) {
+ const [elapsedSeconds, setElapsedSeconds] = useState(() =>
+ Math.max(0, Math.floor((Date.now() - startedAtMs) / 1000)),
+ );
+ const working = status === "working";
+ const failed = status === "failed" || status === "cancelled";
+
+ useEffect(() => {
+ if (!working) {
+ return;
+ }
+ const updateElapsed = () =>
+ setElapsedSeconds(
+ Math.max(0, Math.floor((Date.now() - startedAtMs) / 1000)),
+ );
+ const timer = window.setInterval(updateElapsed, 1000);
+ return () => window.clearInterval(timer);
+ }, [startedAtMs, working]);
+
+ const actionLabel = `${actionCount} recorded action${actionCount === 1 ? "" : "s"}`;
+ const title = working
+ ? `Turning ${actionLabel} into a workflow`
+ : status === "complete"
+ ? awaitingReview
+ ? "Workflow ready to review"
+ : "Workflow refinement complete"
+ : "Workflow refinement stopped";
+
+ return (
+ <div
+ className={`rounded-lg border p-3.5 ${
+ status === "complete"
+ ? "border-emerald-500/30 bg-emerald-500/[0.06]"
+ : failed
+ ? "border-red-500/30 bg-red-500/[0.06]"
+ : "border-sky-400/35 bg-sky-500/[0.07]"
+ }`}
+ role="status"
+ aria-live="polite"
+ data-testid="recording-refinement-progress"
+ >
+ <div className="flex items-start gap-2.5">
+ <div
+ className={`mt-0.5 flex size-7 shrink-0 items-center justify-center rounded-full ${
+ status === "complete"
+ ? "bg-emerald-500/15 text-emerald-500"
+ : failed
+ ? "bg-red-500/15 text-red-500"
+ : "bg-sky-500/15 text-sky-500"
+ }`}
+ >
+ {working ? (
+ <ReloadIcon
+ aria-hidden="true"
+ className="size-4 motion-safe:animate-spin motion-reduce:animate-none"
+ />
+ ) : status === "complete" ? (
+ <CheckCircledIcon aria-hidden="true" className="size-4" />
+ ) : (
+ <CrossCircledIcon aria-hidden="true" className="size-4" />
+ )}
+ </div>
+ <div className="min-w-0 flex-1">
+ <div className="text-[13px] font-semibold text-foreground">
+ {title}
+ </div>
+ <div className="mt-0.5 text-[11.5px] leading-relaxed text-muted-foreground">
+ {working
+ ? "Copilot is reviewing the recording and preparing changes for you to review."
+ : status === "complete"
+ ? awaitingReview
+ ? "Review the changes below, then save when you’re ready."
+ : "Copilot finished refining this workflow."
+ : status === "cancelled"
+ ? "The refinement was cancelled. Start it again when you’re ready."
+ : "Copilot could not finish the refinement. Review the error below and try again."}
+ </div>
+ </div>
+ {working ? (
+ <span
+ className="shrink-0 font-mono text-[10.5px] text-muted-foreground"
+ aria-hidden="true"
+ >
+ {elapsedSeconds}s
+ </span>
+ ) : null}
+ </div>
+
+ <div className="ml-9 mt-3 flex flex-col gap-2 border-l border-border/70 pl-3">
+ <CompletedStep>Recording saved</CompletedStep>
+ <CompletedStep>Recorded actions prepared</CompletedStep>
+ <div
+ className={`flex items-center gap-2 text-xs ${
+ failed ? "text-red-600 dark:text-red-400" : "text-foreground"
+ }`}
+ >
+ {working ? (
+ <ReloadIcon
+ aria-hidden="true"
+ className="size-4 shrink-0 text-sky-500 motion-safe:animate-spin motion-reduce:animate-none"
+ />
+ ) : status === "complete" ? (
+ <CheckCircledIcon
+ aria-hidden="true"
+ className="size-4 shrink-0 text-emerald-500"
+ />
+ ) : (
+ <CrossCircledIcon aria-hidden="true" className="size-4 shrink-0" />
+ )}
+ <span>
+ {working
+ ? "Refining the workflow…"
+ : status === "complete"
+ ? "Refined into a reusable workflow"
+ : status === "cancelled"
+ ? "Refinement cancelled"
+ : "Refinement needs attention"}
+ </span>
+ </div>
+ </div>
+ </div>
+ );
+}
+
+export { RecordingRefinementProgressCard };
diff --git a/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.runGrounding.test.tsx b/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.runGrounding.test.tsx
index 132655f86..a3d742105 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.runGrounding.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.runGrounding.test.tsx
@@ -12,6 +12,7 @@ import { FeatureFlagContext } from "@/hooks/useFeatureFlag";
import { useRecordingRefinementEvidenceStore } from "@/store/RecordingRefinementEvidenceStore";
type StreamBody = {
+ cancel_token?: string | null;
code_block?: boolean | null;
message: string;
mode?: string | null;
@@ -29,6 +30,7 @@ type StreamCall = {
const {
streamCalls,
postStreaming,
+ getSseClient,
cancelPost,
historyGet,
historyResponse,
@@ -47,7 +49,18 @@ const {
calls.push({ body, onMessage, resolve, reject });
}),
);
- const history = {
+ const getStreamingClient = vi
+ .fn()
+ .mockResolvedValue({ postStreaming: streaming });
+ const history: {
+ data: {
+ workflow_copilot_chat_id: string | null;
+ request_turn_id?: string | null;
+ chat_history: unknown[];
+ proposed_workflow: Record<string, unknown> | null;
+ auto_accept: boolean;
+ };
+ } = {
data: {
workflow_copilot_chat_id: null as string | null,
chat_history: [] as unknown[],
@@ -66,6 +79,7 @@ const {
return {
streamCalls: calls,
postStreaming: streaming,
+ getSseClient: getStreamingClient,
cancelPost: post,
historyGet: get,
historyResponse: history,
@@ -75,7 +89,7 @@ const {
});
vi.mock("@/api/sse", () => ({
- getSseClient: vi.fn().mockResolvedValue({ postStreaming }),
+ getSseClient,
}));
vi.mock("@/api/AxiosClient", () => ({
@@ -145,7 +159,10 @@ const saveData = {
};
vi.mock("@/store/WorkflowHasChangesStore", () => ({
- useWorkflowHasChangesStore: () => ({ getSaveData: () => saveData }),
+ useWorkflowHasChangesStore: Object.assign(
+ () => ({ getSaveData: () => saveData }),
+ { getState: () => ({ hasChanges: false }) },
+ ),
}));
// Unrelated to this file's tests; the real hook needs a QueryClientProvider
@@ -197,7 +214,10 @@ beforeEach(() => {
HTMLElement.prototype.scrollTo = vi.fn();
streamCalls.length = 0;
postStreaming.mockClear();
+ getSseClient.mockReset();
+ getSseClient.mockResolvedValue({ postStreaming });
cancelPost.mockClear();
+ cancelPost.mockResolvedValue({});
historyGet.mockReset();
historyGet.mockImplementation(() => Promise.resolve(historyResponse));
toast.mockClear();
@@ -298,6 +318,910 @@ describe("WorkflowCopilotChat — run grounding bridge", () => {
expect(consumed).toHaveBeenCalledTimes(1);
});
+ it("keeps recording refinement progress visible through completion", async () => {
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }, { action_id: "a002" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+
+ expect(
+ screen.getByText("Turning 2 recorded actions into a workflow"),
+ ).toBeTruthy();
+ expect(screen.getByText("Recording saved")).toBeTruthy();
+ expect(screen.getByText("Recorded actions prepared")).toBeTruthy();
+ expect(screen.getByText("Refining the workflow…")).toBeTruthy();
+
+ await act(async () => {
+ streamCalls[0]?.onMessage({
+ type: "turn_start",
+ turn_id: "turn-1",
+ turn_index: 0,
+ mode: "build",
+ timestamp: "2026-06-10T00:00:00Z",
+ });
+ });
+
+ expect(
+ screen.getByText("Turning 2 recorded actions into a workflow"),
+ ).toBeTruthy();
+
+ await act(async () => {
+ streamCalls[0]?.onMessage({
+ type: "response",
+ turn_id: "turn-1",
+ message: "The recorded workflow is ready.",
+ workflow_copilot_chat_id: "chat-1",
+ response_time: "2026-06-10T00:00:02Z",
+ updated_workflow: { workflow_id: "wf-draft" },
+ proposal_disposition: "review_required",
+ });
+ streamCalls[0]?.resolve();
+ });
+
+ expect(screen.getByText("Workflow ready to review")).toBeTruthy();
+ expect(screen.getByText("Refined into a reusable workflow")).toBeTruthy();
+ });
+
+ it("keeps recording refinement active while a disconnected stream recovers", async () => {
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+ vi.useFakeTimers();
+
+ await act(async () => {
+ streamCalls[0]?.onMessage({
+ type: "turn_start",
+ turn_id: "turn-1",
+ turn_index: 0,
+ mode: "build",
+ timestamp: "2026-06-10T00:00:00Z",
+ });
+ streamCalls[0]?.reject(new Error("stream disconnected"));
+ });
+
+ expect(
+ screen.getByText("Turning 1 recorded action into a workflow"),
+ ).toBeTruthy();
+ expect(screen.queryByText("Workflow refinement stopped")).toBeNull();
+
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (1 actions) into a reusable workflow.",
+ turn_id: "turn-1",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "ai",
+ content: "This turn was interrupted before it could finish.",
+ created_at: "2026-06-10T00:00:02Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-1",
+ terminal_reason: "interrupted",
+ },
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000);
+ });
+
+ expect(
+ screen.getByText("Turning 1 recorded action into a workflow"),
+ ).toBeTruthy();
+ expect(screen.queryByText("Workflow ready to review")).toBeNull();
+
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (1 actions) into a reusable workflow.",
+ turn_id: "turn-1",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "ai",
+ content: "The recorded workflow is ready.",
+ created_at: "2026-06-10T00:00:03Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-1",
+ terminal_reason: "completed",
+ },
+ narrative_payload: {
+ turnId: "turn-1",
+ proposalDisposition: "review_untested",
+ draft: {
+ blockCount: 1,
+ blockLabels: ["Open page"],
+ summary: "Built from the recording",
+ },
+ terminal: "response",
+ },
+ },
+ ],
+ proposed_workflow: { workflow_id: "wf-draft" },
+ auto_accept: false,
+ };
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(3_000);
+ });
+
+ expect(screen.getByText("Workflow ready to review")).toBeTruthy();
+ expect(screen.getByText("Refined into a reusable workflow")).toBeTruthy();
+ });
+
+ it("recovers recording refinement when the turn_start frame is lost", async () => {
+ const consumed = vi.fn();
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ onInitialMessageConsumed: consumed,
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+ vi.useFakeTimers();
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ request_turn_id: null,
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (1 actions) into a reusable workflow.",
+ turn_id: "turn-earlier-same-receipt",
+ created_at: new Date().toISOString(),
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+
+ await act(async () => {
+ streamCalls[0]?.reject(new Error("stream disconnected"));
+ await vi.advanceTimersByTimeAsync(0);
+ });
+
+ expect(historyGet).toHaveBeenCalledWith(
+ "/workflow/copilot/chat-history",
+ expect.objectContaining({
+ params: {
+ request_cancel_token: streamCalls[0]?.body.cancel_token,
+ workflow_permanent_id: "wpid_1",
+ },
+ }),
+ );
+
+ expect(
+ screen.getByText("Turning 1 recorded action into a workflow"),
+ ).toBeTruthy();
+ expect(
+ screen.getByText(
+ "The connection dropped, so Copilot is checking whether this turn finished.",
+ ),
+ ).toBeTruthy();
+
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ request_turn_id: "turn-lost-start",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (1 actions) into a reusable workflow.",
+ turn_id: "turn-lost-start",
+ created_at: new Date().toISOString(),
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000);
+ });
+ expect(
+ historyGet.mock.calls.filter(
+ ([, config]) => config?.params?.request_cancel_token,
+ ),
+ ).toHaveLength(2);
+ expect(useRecordingRefinementEvidenceStore.getState().armed).toBeNull();
+ expect(consumed).toHaveBeenCalledTimes(1);
+
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ request_turn_id: "turn-lost-start",
+ chat_history: [
+ historyResponse.data.chat_history[0]!,
+ {
+ sender: "ai",
+ content: "The recorded workflow is ready.",
+ created_at: new Date().toISOString(),
+ turn_outcome: {
+ copilot_turn_id: "turn-lost-start",
+ terminal_reason: "completed",
+ },
+ narrative_payload: {
+ turnId: "turn-lost-start",
+ proposalDisposition: "review_untested",
+ draft: {
+ blockCount: 1,
+ blockLabels: ["Open page"],
+ },
+ terminal: "response",
+ },
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000);
+ });
+
+ expect(screen.getByText("Workflow refinement complete")).toBeTruthy();
+ });
+
+ it("scopes lost-turn recovery to the chat that sent the request", async () => {
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-existing",
+ chat_history: [],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+ historyGet.mockClear();
+ vi.useFakeTimers();
+ historyResponse.data.chat_history = [
+ {
+ sender: "product",
+ content: "Refine the recording (1 actions) into a reusable workflow.",
+ turn_id: "turn-lost-start",
+ created_at: new Date().toISOString(),
+ },
+ ];
+
+ await act(async () => {
+ streamCalls[0]?.reject(new Error("stream disconnected"));
+ await vi.advanceTimersByTimeAsync(0);
+ });
+
+ expect(historyGet).toHaveBeenCalledWith(
+ "/workflow/copilot/chat-history",
+ expect.objectContaining({
+ params: {
+ request_cancel_token: streamCalls[0]?.body.cancel_token,
+ workflow_copilot_chat_id: "chat-existing",
+ },
+ }),
+ );
+ });
+
+ it("bounds a stalled lookup after losing turn_start", async () => {
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+ historyGet.mockImplementation(
+ (_path: string, config?: { signal?: AbortSignal }) =>
+ new Promise((_resolve, reject) => {
+ config?.signal?.addEventListener("abort", () =>
+ reject(new Error("history lookup aborted")),
+ );
+ }),
+ );
+ vi.useFakeTimers();
+
+ await act(async () => {
+ streamCalls[0]?.reject(new Error("stream disconnected"));
+ await vi.advanceTimersByTimeAsync(0);
+ });
+ expect(
+ screen.getByText(
+ "The connection dropped, so Copilot is checking whether this turn finished.",
+ ),
+ ).toBeTruthy();
+
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(1_500_000);
+ });
+
+ expect(screen.getByText("Workflow refinement stopped")).toBeTruthy();
+ expect(
+ screen.getByText("Sorry, I encountered an error. Please try again."),
+ ).toBeTruthy();
+ });
+
+ it("only fails the recovery notice owned by the poll that gives up", async () => {
+ await renderChat();
+ vi.useFakeTimers();
+
+ await submit("First request");
+ await act(async () => {
+ streamCalls[0]?.onMessage({
+ type: "turn_start",
+ turn_id: "turn-1",
+ turn_index: 0,
+ mode: "build",
+ timestamp: "2026-06-10T00:00:00Z",
+ });
+ streamCalls[0]?.reject(new Error("stream disconnected"));
+ });
+ await submit("Second request");
+ await act(async () => {
+ streamCalls[1]?.onMessage({
+ type: "turn_start",
+ turn_id: "turn-2",
+ turn_index: 1,
+ mode: "build",
+ timestamp: "2026-06-10T00:00:01Z",
+ });
+ streamCalls[1]?.reject(new Error("stream disconnected"));
+ });
+
+ const emptyHistory = {
+ data: {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [],
+ proposed_workflow: null,
+ auto_accept: false,
+ },
+ };
+ historyGet.mockReset();
+ for (let cycle = 0; cycle < 3; cycle += 1) {
+ historyGet.mockRejectedValueOnce(new Error("turn-1 offline"));
+ historyGet.mockResolvedValueOnce(emptyHistory);
+ }
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000 + 3_000 + 5_000);
+ });
+
+ expect(
+ screen.getAllByText(
+ "The connection dropped, so Copilot is checking whether this turn finished.",
+ ),
+ ).toHaveLength(1);
+ expect(
+ screen.getAllByText("Sorry, I encountered an error. Please try again."),
+ ).toHaveLength(1);
+ });
+
+ it("hydrates recording refinement progress from turn-correlated history", async () => {
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (3 actions) into a reusable workflow.",
+ turn_id: "turn-refine",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "user",
+ content: "Add a final confirmation step.",
+ turn_id: "turn-later",
+ created_at: "2026-06-10T00:00:01Z",
+ },
+ {
+ sender: "ai",
+ content: "Added the confirmation step.",
+ turn_id: "turn-later",
+ created_at: "2026-06-10T00:00:02Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-later",
+ terminal_reason: "completed",
+ },
+ },
+ {
+ sender: "ai",
+ content: "The recorded workflow is ready.",
+ created_at: "2026-06-10T00:00:03Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-refine",
+ terminal_reason: "completed",
+ },
+ narrative_payload: {
+ turnId: "turn-refine",
+ proposalDisposition: "review_untested",
+ draft: {
+ blockCount: 3,
+ blockLabels: ["Open page", "Enter search", "Select result"],
+ summary: "Built from the recording",
+ },
+ terminal: "response",
+ },
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+
+ await renderChat();
+
+ expect(screen.getByText("Workflow refinement complete")).toBeTruthy();
+ expect(screen.getByText("Refined into a reusable workflow")).toBeTruthy();
+ expect(
+ screen.queryByText(
+ "Review the changes below, then save when you’re ready.",
+ ),
+ ).toBeNull();
+ expect(screen.getByTestId("recording-refinement-progress")).toBeTruthy();
+ });
+
+ it("hydrates an agent-loop cancellation as cancelled", async () => {
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (2 actions) into a reusable workflow.",
+ turn_id: "turn-refine",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "ai",
+ content: "The refinement was cancelled.",
+ created_at: "2026-06-10T00:00:01Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-refine",
+ terminal_reason: "cancel",
+ },
+ narrative_payload: {
+ turnId: "turn-refine",
+ proposalDisposition: "review_untested",
+ draft: {
+ blockCount: 2,
+ blockLabels: ["Open page", "Enter search"],
+ },
+ terminal: "response",
+ cancelled: true,
+ },
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+
+ await renderChat();
+
+ expect(screen.getByText("Workflow refinement stopped")).toBeTruthy();
+ expect(screen.getByText("Refinement cancelled")).toBeTruthy();
+ });
+
+ it("polls a rehydrated interrupted recording refinement to completion", async () => {
+ vi.useFakeTimers();
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (2 actions) into a reusable workflow.",
+ turn_id: "turn-refine",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "ai",
+ content: "This turn was interrupted before it could finish.",
+ created_at: "2026-06-10T00:00:01Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-refine",
+ terminal_reason: "interrupted",
+ },
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+
+ render(chatUi({}));
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(0);
+ });
+ expect(screen.getByText("Workflow refinement stopped")).toBeTruthy();
+
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (2 actions) into a reusable workflow.",
+ turn_id: "turn-refine",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "ai",
+ content: "The recorded workflow is ready.",
+ turn_id: "turn-refine",
+ created_at: "2026-06-10T00:00:03Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-refine",
+ terminal_reason: "completed",
+ },
+ narrative_payload: {
+ turnId: "turn-refine",
+ proposalDisposition: "review_untested",
+ draft: {
+ blockCount: 2,
+ blockLabels: ["Open page", "Enter search"],
+ summary: "Built from the recording",
+ },
+ terminal: "response",
+ },
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000);
+ });
+
+ expect(screen.getByText("Workflow refinement complete")).toBeTruthy();
+ expect(screen.getByText("Refined into a reusable workflow")).toBeTruthy();
+ });
+
+ it("fails a rehydrated refinement when a restarted recovery poll gives up", async () => {
+ vi.useFakeTimers();
+ const question = {
+ interaction_id: "question-1",
+ turn_id: "turn-refine",
+ tool_call_id: "call-1",
+ status: "pending",
+ response: null,
+ created_at: "2026-06-10T00:00:01Z",
+ resolved_at: null,
+ parts: [
+ {
+ part_id: "format",
+ prompt: "Which format?",
+ choices: [{ choice_id: "csv", text: "CSV" }],
+ },
+ ],
+ };
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (2 actions) into a reusable workflow.",
+ turn_id: "turn-refine",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "ai",
+ content: "This turn was interrupted before it could finish.",
+ created_at: "2026-06-10T00:00:01Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-refine",
+ terminal_reason: "interrupted",
+ },
+ },
+ ],
+ proposed_workflow: null,
+ auto_accept: false,
+ };
+ render(chatUi({}));
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(0);
+ await vi.advanceTimersByTimeAsync(2_000);
+ });
+ expect(
+ screen.getByText("Turning 2 recorded actions into a workflow"),
+ ).toBeTruthy();
+
+ Object.assign(historyResponse.data, {
+ question_interactions: [question],
+ pending_question_cancel_token: "stop",
+ });
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(3_000);
+ });
+
+ cancelPost.mockResolvedValueOnce({
+ data: { ...question, status: "resolved", response: { skipped: true } },
+ });
+ await act(async () => {
+ fireEvent.click(screen.getByRole("button", { name: "Skip" }));
+ });
+ historyGet.mockRejectedValue(new Error("offline"));
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000 + 3_000 + 5_000);
+ });
+
+ expect(screen.getByText("Workflow refinement stopped")).toBeTruthy();
+ expect(screen.getByText("Refinement needs attention")).toBeTruthy();
+ });
+
+ it("shows current-turn feedback while an older refinement recovers", async () => {
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+ vi.useFakeTimers();
+ await act(async () => {
+ streamCalls[0]?.onMessage({
+ type: "turn_start",
+ turn_id: "turn-refine",
+ turn_index: 0,
+ mode: "build",
+ timestamp: "2026-06-10T00:00:00Z",
+ });
+ streamCalls[0]?.reject(new Error("stream disconnected"));
+ });
+ expect(
+ screen.getByText("Turning 1 recorded action into a workflow"),
+ ).toBeTruthy();
+
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(1);
+ });
+ await submit("Add a final confirmation step");
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(0);
+ });
+ expect(postStreaming).toHaveBeenCalledTimes(2);
+
+ expect(
+ screen.getByText("Copilot is working on your request…"),
+ ).toBeTruthy();
+ expect(
+ screen.getByRole("button", { name: "History" }).hasAttribute("disabled"),
+ ).toBe(true);
+
+ Object.assign(historyResponse.data, {
+ question_interactions: [
+ {
+ interaction_id: "question-old",
+ turn_id: "turn-refine",
+ tool_call_id: "call-old",
+ status: "pending",
+ response: null,
+ created_at: "2026-06-10T00:00:02Z",
+ resolved_at: null,
+ parts: [
+ {
+ part_id: "format",
+ prompt: "Which format?",
+ choices: [{ choice_id: "csv", text: "CSV" }],
+ },
+ ],
+ },
+ ],
+ pending_question_cancel_token: "stop-old",
+ });
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000);
+ });
+
+ expect(
+ screen.getByRole("button", { name: "History" }).hasAttribute("disabled"),
+ ).toBe(true);
+ });
+
+ it("does not use another turn's proposal as recovered refinement success", async () => {
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+ vi.useFakeTimers();
+ await act(async () => {
+ streamCalls[0]?.onMessage({
+ type: "turn_start",
+ turn_id: "turn-refine",
+ turn_index: 0,
+ mode: "build",
+ timestamp: "2026-06-10T00:00:00Z",
+ });
+ streamCalls[0]?.reject(new Error("stream disconnected"));
+ });
+
+ historyResponse.data = {
+ workflow_copilot_chat_id: "chat-1",
+ chat_history: [
+ {
+ sender: "product",
+ content: "Refine the recording (1 actions) into a reusable workflow.",
+ turn_id: "turn-refine",
+ created_at: "2026-06-10T00:00:00Z",
+ },
+ {
+ sender: "ai",
+ content: "I could not produce a workflow.",
+ created_at: "2026-06-10T00:00:02Z",
+ turn_outcome: {
+ copilot_turn_id: "turn-refine",
+ terminal_reason: "completed",
+ },
+ narrative_payload: {
+ turnId: "turn-refine",
+ proposalDisposition: "no_proposal",
+ terminal: "response",
+ },
+ },
+ ],
+ proposed_workflow: { workflow_id: "older-draft" },
+ auto_accept: false,
+ };
+ Object.assign(historyResponse.data, {
+ proposed_workflow_metadata: { owner_turn_id: "turn-older" },
+ });
+ await act(async () => {
+ await vi.advanceTimersByTimeAsync(2_000);
+ });
+
+ expect(screen.getByText("Workflow refinement stopped")).toBeTruthy();
+ expect(screen.getByText("Refinement needs attention")).toBeTruthy();
+ expect(screen.queryByText("Workflow ready to review")).toBeNull();
+ });
+
+ it("does not report recording refinement complete without a workflow result", async () => {
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(postStreaming).toHaveBeenCalledTimes(1));
+
+ await act(async () => {
+ streamCalls[0]?.onMessage({
+ type: "turn_start",
+ turn_id: "turn-1",
+ turn_index: 0,
+ mode: "build",
+ timestamp: "2026-06-10T00:00:00Z",
+ });
+ streamCalls[0]?.onMessage({
+ type: "response",
+ turn_id: "turn-1",
+ message: "I could not produce a workflow.",
+ workflow_copilot_chat_id: "chat-1",
+ response_time: "2026-06-10T00:00:02Z",
+ updated_workflow: null,
+ proposal_disposition: "no_proposal",
+ });
+ streamCalls[0]?.resolve();
+ });
+
+ expect(screen.getByText("Workflow refinement stopped")).toBeTruthy();
+ expect(screen.getByText("Refinement needs attention")).toBeTruthy();
+ expect(screen.queryByText("Workflow ready to review")).toBeNull();
+ });
+
+ it("cancels recording refinement when the send aborts before the request", async () => {
+ let resolveClient:
+ | ((client: { postStreaming: typeof postStreaming }) => void)
+ | null = null;
+ getSseClient.mockImplementationOnce(
+ () =>
+ new Promise((resolve) => {
+ resolveClient = resolve;
+ }),
+ );
+ cancelPost.mockRejectedValueOnce(new Error("cancel unavailable"));
+ useRecordingRefinementEvidenceStore.getState().set({
+ nonce: "n-refine",
+ evidence: {
+ schema_version: 1,
+ recording: { browser_session_id: "pbs-1" },
+ actions: [{ action_id: "a001" }],
+ deleted_action_ids: [],
+ truncated_action_count: 0,
+ provenance: { source: "browser_recording" },
+ },
+ });
+
+ await renderChat({
+ initialAction: { kind: "refine_recording", nonce: "n-refine" },
+ });
+ await waitFor(() => expect(getSseClient).toHaveBeenCalledTimes(1));
+
+ fireEvent.keyDown(window, { key: "Escape" });
+ await waitFor(() => expect(cancelPost).toHaveBeenCalledTimes(1));
+ await act(async () => {
+ resolveClient?.({ postStreaming });
+ });
+
+ await waitFor(() =>
+ expect(screen.getByText("Workflow refinement stopped")).toBeTruthy(),
+ );
+ expect(screen.getByText("Refinement cancelled")).toBeTruthy();
+ expect(postStreaming).not.toHaveBeenCalled();
+ });
+
it("keeps code authoring enabled for a typed diagnose action on a code workflow", async () => {
BOOLEAN_FLAGS.WORKFLOW_COPILOT_CODE_BLOCK_MODE = true;
BOOLEAN_FLAGS.CODE_BLOCK_ACCESS = true;
diff --git a/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.tsx b/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.tsx
index e3e0d406c..9fe6c6510 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.tsx
+++ b/skyvern-frontend/src/routes/workflows/copilot/WorkflowCopilotChat.tsx
@@ -92,6 +92,10 @@ import { shouldArmDraftingGapTimer } from "./copilotPhases";
import { InstantAckPlaceholder, NarrativeView } from "./NarrativeView";
import { CopilotMarkdown } from "./CopilotMarkdown";
import { CopilotWorkingStatus } from "./CopilotWorkingStatus";
+import {
+ RecordingRefinementProgressCard,
+ type RecordingRefinementStatus,
+} from "./RecordingRefinementProgressCard";
import { useRunLifecycleAnnouncements } from "./useRunLifecycleAnnouncements";
import { ConfirmCard, shouldShowConfirmCard } from "./cards/ConfirmCard";
import { ConnectedAccountChoiceCard } from "./cards/ConnectedAccountChoiceCard";
@@ -188,6 +192,26 @@ const diagnoseRunReceipt = (runId: string) =>
// rewrites the message, so a different wording here would change the row on history reload.
const refineRecordingReceipt = (actionCount: number) =>
`Refine the recording (${actionCount} actions) into a reusable workflow.`;
+const REFINE_RECORDING_RECEIPT_PATTERN =
+ /^Refine the recording \((\d+) actions\) into a reusable workflow\.$/;
+
+function refineRecordingActionCount(content: string): number | null {
+ const match = REFINE_RECORDING_RECEIPT_PATTERN.exec(content);
+ if (!match) return null;
+ const count = Number(match[1]);
+ return Number.isSafeInteger(count) && count >= 0 ? count : null;
+}
+
+function isCancelledRefinementTurn(
+ terminalReason: string | null | undefined,
+ narrative: TurnNarrativeState | undefined,
+): boolean {
+ return (
+ terminalReason === "cancel" ||
+ terminalReason === "user_cancelled" ||
+ narrative?.cancelled === true
+ );
+}
// diagnose_run and refine_recording both open the turn with a server-authored receipt.
const isProductAuthoredAction = (action: ArmedProductAction | null): boolean =>
@@ -424,7 +448,14 @@ export interface ChatMessage {
// turns. Live in-flight narrative is rendered separately at the bottom.
narrative?: TurnNarrativeState;
// FE-synthetic rows (never persisted, never sent to the LLM).
- kind?: "run_lifecycle" | "status_notice";
+ kind?: "run_lifecycle" | "status_notice" | "recording_refinement";
+ recoveryTurnId?: string;
+ recordingRefinement?: {
+ actionCount: number;
+ startedAtMs: number;
+ status: RecordingRefinementStatus;
+ turnId?: string;
+ };
attachedFiles?: CopilotAttachedFile[];
}
@@ -1184,6 +1215,7 @@ export function WorkflowCopilotChat({
// the cancel POST and the watcher firing, so the frontend must remember it.
const cancelInFlightController = useRef<AbortController | null>(null);
const recoveryPolls = useRef(new Map<string, () => void>());
+ const recoverySnapshotStamps = useRef(new Map<string, number>());
const recoveryGeneration = useRef(0);
// A turn resumed from saved human input has no live SSE controller, but it
// still owns the composer until history proves that continuation finished.
@@ -1859,37 +1891,104 @@ export function WorkflowCopilotChat({
(
data: WorkflowCopilotChatHistoryResponse,
carryForwardLifecycle = true,
+ recoveredRecordingRefinement?: {
+ responseIndex: number;
+ turnId: string;
+ messageId: string;
+ status: RecordingRefinementStatus;
+ },
) => {
setRecoveredPauseFrames(data.pending_credential_requests ?? []);
setQuestionInteractions(data.question_interactions ?? []);
setQuestionCancelToken(data.pending_question_cancel_token ?? null);
- const historyMessages = data.chat_history.map((message, index) => ({
- id: `${index}-${Date.now()}`,
- sender: message.sender,
- content: message.content,
- timestamp: message.created_at,
- attachedFiles:
- message.attached_files && message.attached_files.length > 0
- ? message.attached_files
- : undefined,
- narrative: (() => {
- const hydrated = hydrateHistoryNarrative(
- message.narrative_payload,
- message.turn_outcome,
- );
- if (!hydrated) return undefined;
- // Fall back to the legacy message body when the persisted payload
- // predates terminal-text capture.
- if (!hydrated.terminalMessage && message.content) {
- return {
- ...hydrated,
- terminalMessage: message.content,
- narrativeSummary: hydrated.narrativeSummary ?? message.content,
- };
- }
- return hydrated;
- })(),
- }));
+ const historyMessages: ChatMessage[] = data.chat_history.map(
+ (message, index) => ({
+ id: `${index}-${Date.now()}`,
+ sender: message.sender,
+ content: message.content,
+ timestamp: message.created_at,
+ attachedFiles:
+ message.attached_files && message.attached_files.length > 0
+ ? message.attached_files
+ : undefined,
+ narrative: (() => {
+ const hydrated = hydrateHistoryNarrative(
+ message.narrative_payload,
+ message.turn_outcome,
+ );
+ if (!hydrated) return undefined;
+ // Fall back to the legacy message body when the persisted payload
+ // predates terminal-text capture.
+ if (!hydrated.terminalMessage && message.content) {
+ return {
+ ...hydrated,
+ terminalMessage: message.content,
+ narrativeSummary: hydrated.narrativeSummary ?? message.content,
+ };
+ }
+ return hydrated;
+ })(),
+ }),
+ );
+ for (let index = 0; index < data.chat_history.length; index += 1) {
+ const productRow = data.chat_history[index];
+ const productMessage = historyMessages[index];
+ if (!productRow || !productMessage) continue;
+ const actionCount = refineRecordingActionCount(productRow.content);
+ const turnId = productRow.turn_id ?? undefined;
+ if (
+ productRow.sender !== "product" ||
+ actionCount === null ||
+ turnId === undefined
+ ) {
+ continue;
+ }
+
+ const responseIndex = data.chat_history.findIndex(
+ (message) =>
+ message.sender === "ai" &&
+ message.turn_outcome?.copilot_turn_id === turnId,
+ );
+ const responseRow =
+ responseIndex >= 0 ? data.chat_history[responseIndex] : undefined;
+ const responseMessage =
+ responseIndex >= 0 ? historyMessages[responseIndex] : undefined;
+ const terminalReason = responseRow?.turn_outcome?.terminal_reason;
+ const ownsChatProposal = Boolean(
+ turnId &&
+ data.proposed_workflow &&
+ data.proposed_workflow_metadata?.owner_turn_id === turnId,
+ );
+ const producedWorkflow = Boolean(
+ responseMessage?.narrative?.proposalDisposition !== "no_proposal" &&
+ (responseMessage?.narrative?.draft || ownsChatProposal),
+ );
+ const status: RecordingRefinementStatus = !responseRow
+ ? "working"
+ : isCancelledRefinementTurn(
+ terminalReason,
+ responseMessage?.narrative,
+ )
+ ? "cancelled"
+ : terminalReason === INTERRUPTED_TERMINAL_REASON ||
+ responseMessage?.narrative?.terminal === "error" ||
+ (responseMessage?.narrative &&
+ notConfirmedOutcome(responseMessage.narrative) !== null) ||
+ !producedWorkflow
+ ? "failed"
+ : "complete";
+ historyMessages[index] = {
+ ...productMessage,
+ id: `recording-refinement-${turnId}`,
+ kind: "recording_refinement",
+ recordingRefinement: {
+ actionCount,
+ startedAtMs: parseServerStamp(productRow.created_at) || Date.now(),
+ status,
+ turnId,
+ },
+ };
+ }
// A rehydrated turn still owns the run its records name, so the
// lifecycle hook keeps announcing nothing about it after a reload.
for (const message of historyMessages) {
@@ -1903,12 +2002,53 @@ export function WorkflowCopilotChat({
latestTurnId.current = restoredPendingProposalTurnId;
// History never carries run_lifecycle lines (local-only); carry them
// forward only for the mount-race caller, not an explicit chat switch.
- setMessages((prev) => [
- ...historyMessages,
- ...(carryForwardLifecycle
- ? prev.filter((message) => message.kind === "run_lifecycle")
- : []),
- ]);
+ setMessages((prev) => {
+ const nextMessages: ChatMessage[] = [
+ ...historyMessages,
+ ...(carryForwardLifecycle
+ ? prev.filter((message) => message.kind === "run_lifecycle")
+ : []),
+ ];
+ const recovery = recoveredRecordingRefinement;
+ const hydratedProgressIndex = recovery
+ ? nextMessages.findIndex(
+ (message) =>
+ message.kind === "recording_refinement" &&
+ message.recordingRefinement?.turnId === recovery.turnId,
+ )
+ : -1;
+ const progressMessage = recovery
+ ? (prev.find(
+ (message) =>
+ message.id === recovery.messageId &&
+ message.kind === "recording_refinement" &&
+ message.recordingRefinement?.turnId === recovery.turnId,
+ ) ??
+ (hydratedProgressIndex >= 0
+ ? nextMessages[hydratedProgressIndex]
+ : undefined))
+ : undefined;
+ if (!recovery || !progressMessage?.recordingRefinement) {
+ return nextMessages;
+ }
+ const completedProgress: ChatMessage = {
+ ...progressMessage,
+ recordingRefinement: {
+ ...progressMessage.recordingRefinement,
+ status: recovery.status,
+ },
+ };
+ if (hydratedProgressIndex >= 0) {
+ nextMessages[hydratedProgressIndex] = completedProgress;
+ } else {
+ nextMessages.splice(
+ Math.max(0, recovery.responseIndex),
+ 0,
+ completedProgress,
+ );
+ }
+ return nextMessages;
+ });
setWorkflowCopilotChatId(data.workflow_copilot_chat_id);
setProposedWorkflow(data.proposed_workflow ?? null);
setPendingProposalMetadata(data.proposed_workflow_metadata ?? null);
@@ -1929,11 +2069,17 @@ export function WorkflowCopilotChat({
recoveryGeneration.current += 1;
recoveryPolls.current.forEach((stop) => stop());
recoveryPolls.current.clear();
+ recoverySnapshotStamps.current.clear();
recoveredTurnOwnerRef.current = null;
}, []);
const startRecoveryPoll = useCallback(
- (chatId: string | null, turnId: string, ownsTurn = false) => {
+ (
+ chatId: string | null,
+ turnId: string,
+ ownsTurn = false,
+ recordingRefinementMessageId?: string,
+ ) => {
if (!workflowPermanentId) {
return;
}
@@ -1985,9 +2131,19 @@ export function WorkflowCopilotChat({
finish();
setMessages((prev) =>
prev.map((message) =>
- message.content === RECOVERY_IN_PROGRESS_MESSAGE
- ? { ...message, content: SEND_FAILED_MESSAGE }
- : message,
+ message.recordingRefinement?.turnId === turnId &&
+ message.recordingRefinement.status === "working"
+ ? {
+ ...message,
+ recordingRefinement: {
+ ...message.recordingRefinement,
+ status: "failed",
+ },
+ }
+ : message.recoveryTurnId === turnId &&
+ message.content === RECOVERY_IN_PROGRESS_MESSAGE
+ ? { ...message, content: SEND_FAILED_MESSAGE }
+ : message,
),
);
}
@@ -2079,7 +2235,6 @@ export function WorkflowCopilotChat({
setQuestionCancelToken(
response.data.pending_question_cancel_token ?? null,
);
- setIsLoading(false);
finish();
return;
}
@@ -2114,9 +2269,71 @@ export function WorkflowCopilotChat({
schedule();
return;
}
+ const recoveryChatId =
+ response.data.workflow_copilot_chat_id ?? chatId;
+ const snapshotStamp = response.data.chat_history.reduce(
+ (newest, message) =>
+ Math.max(
+ newest,
+ parseServerStamp(message.modified_at ?? message.created_at),
+ ),
+ Number.NEGATIVE_INFINITY,
+ );
+ if (recoveryChatId) {
+ // Concurrent polls replace the full history. A slower response
+ // captured before a newer terminal row must not erase that row.
+ const latestAppliedStamp =
+ recoverySnapshotStamps.current.get(recoveryChatId) ??
+ Number.NEGATIVE_INFINITY;
+ if (snapshotStamp < latestAppliedStamp) {
+ schedule();
+ return;
+ }
+ recoverySnapshotStamps.current.set(recoveryChatId, snapshotStamp);
+ }
if (row.content !== appliedContent) {
appliedContent = row.content;
- applyHistoryResponse(response.data);
+ const recoveredNarrative = hydrateHistoryNarrative(
+ row.narrative_payload,
+ row.turn_outcome,
+ );
+ const interrupted =
+ row.turn_outcome?.terminal_reason ===
+ INTERRUPTED_TERMINAL_REASON;
+ const ownsChatProposal = Boolean(
+ response.data.proposed_workflow &&
+ response.data.proposed_workflow_metadata?.owner_turn_id ===
+ turnId,
+ );
+ const recoveredProducedWorkflow = Boolean(
+ recoveredNarrative?.proposalDisposition !== "no_proposal" &&
+ (recoveredNarrative?.draft || ownsChatProposal),
+ );
+ const recoveredStatus: RecordingRefinementStatus = interrupted
+ ? "working"
+ : isCancelledRefinementTurn(
+ row.turn_outcome?.terminal_reason,
+ recoveredNarrative,
+ )
+ ? "cancelled"
+ : recoveredNarrative?.terminal === "error" ||
+ (recoveredNarrative &&
+ notConfirmedOutcome(recoveredNarrative) !== null) ||
+ !recoveredProducedWorkflow
+ ? "failed"
+ : "complete";
+ applyHistoryResponse(
+ response.data,
+ true,
+ recordingRefinementMessageId
+ ? {
+ responseIndex: response.data.chat_history.indexOf(row),
+ turnId,
+ messageId: recordingRefinementMessageId,
+ status: recoveredStatus,
+ }
+ : undefined,
+ );
setIsLoading(false);
}
// An interrupted row is replaced by the real reply if the turn
@@ -2134,7 +2351,7 @@ export function WorkflowCopilotChat({
// has a truthful row; only a few reads are spent on a supersede.
supersedeReadsLeft -= 1;
if (supersedeReadsLeft <= 0) {
- finish();
+ giveUp();
return;
}
}
@@ -2165,12 +2382,50 @@ export function WorkflowCopilotChat({
[applyHistoryResponse, credentialGetter, workflowPermanentId],
);
- const adoptRecoveredCredentialPause = useCallback(
+ const adoptRecoveredTurns = useCallback(
(data: WorkflowCopilotChatHistoryResponse) => {
const pendingRequests = data.pending_credential_requests ?? [];
const pending = pendingRequests[pendingRequests.length - 1];
- if (!pending) return;
- startRecoveryPoll(data.workflow_copilot_chat_id, pending.turn_id, true);
+ const completedTurnIds = new Set(
+ data.chat_history.flatMap((message) =>
+ message.sender === "ai" &&
+ message.turn_outcome?.copilot_turn_id &&
+ message.turn_outcome.terminal_reason !== INTERRUPTED_TERMINAL_REASON
+ ? [message.turn_outcome.copilot_turn_id]
+ : [],
+ ),
+ );
+ const pendingRefinements = data.chat_history.filter(
+ (message) =>
+ message.sender === "product" &&
+ message.turn_id &&
+ refineRecordingActionCount(message.content) !== null &&
+ !completedTurnIds.has(message.turn_id),
+ );
+ const pendingRefinementIds = new Map(
+ pendingRefinements.map((message) => [
+ message.turn_id as string,
+ `recording-refinement-${message.turn_id}`,
+ ]),
+ );
+ if (pending) {
+ startRecoveryPoll(
+ data.workflow_copilot_chat_id,
+ pending.turn_id,
+ true,
+ pendingRefinementIds.get(pending.turn_id),
+ );
+ }
+ for (const [turnId, messageId] of pendingRefinementIds) {
+ if (turnId !== pending?.turn_id) {
+ startRecoveryPoll(
+ data.workflow_copilot_chat_id,
+ turnId,
+ false,
+ messageId,
+ );
+ }
+ }
},
[startRecoveryPoll],
);
@@ -2208,7 +2463,7 @@ export function WorkflowCopilotChat({
},
);
applyHistoryResponse(response.data, false);
- adoptRecoveredCredentialPause(response.data);
+ adoptRecoveredTurns(response.data);
// Mark history loaded for this workflow so the mount effect won't reload
// the latest chat over the one the user just selected.
historyLoadedForRef.current = workflowPermanentId;
@@ -2223,7 +2478,7 @@ export function WorkflowCopilotChat({
credentialGetter,
workflowPermanentId,
applyHistoryResponse,
- adoptRecoveredCredentialPause,
+ adoptRecoveredTurns,
stopRecoveryPolls,
repin,
discardQueuedPrompt,
@@ -2929,7 +3184,7 @@ export function WorkflowCopilotChat({
if (!isMounted) return;
applyHistoryResponse(response.data);
- adoptRecoveredCredentialPause(response.data);
+ adoptRecoveredTurns(response.data);
historyLoadedForRef.current = workflowPermanentId;
} catch (error) {
console.error("Failed to load chat history:", error);
@@ -2948,7 +3203,7 @@ export function WorkflowCopilotChat({
}, [
credentialGetter,
initialAction?.nonce,
- adoptRecoveredCredentialPause,
+ adoptRecoveredTurns,
repin,
stopRecoveryPolls,
updateQueuedPrompt,
@@ -3536,10 +3791,26 @@ export function WorkflowCopilotChat({
const userMessageId = options.queuedMessageId ?? Date.now().toString();
const sendOwnsTray = composerSend;
+ const recordingRefinementAction =
+ productActionRef.current?.action === "refine_recording"
+ ? productActionRef.current
+ : null;
+ const recordingRefinement = recordingRefinementAction
+ ? {
+ actionCount:
+ useRecordingRefinementEvidenceStore
+ .getState()
+ .peek(recordingRefinementAction.nonce)?.actions.length ?? 0,
+ startedAtMs: Date.now(),
+ status: "working" as const,
+ }
+ : undefined;
const userMessage: ChatMessage = {
id: userMessageId,
sender: echoSenderForArmedAction(),
content: candidate,
+ kind: recordingRefinement ? "recording_refinement" : undefined,
+ recordingRefinement,
attachedFiles: sentAttachments.length > 0 ? sentAttachments : undefined,
};
if (sendOwnsTray) {
@@ -3564,6 +3835,11 @@ export function WorkflowCopilotChat({
message.id === userMessageId
? {
...message,
+ kind: recordingRefinement
+ ? "recording_refinement"
+ : message.kind,
+ recordingRefinement:
+ recordingRefinement ?? message.recordingRefinement,
attachedFiles:
sentAttachments.length > 0 ? sentAttachments : undefined,
}
@@ -3572,6 +3848,7 @@ export function WorkflowCopilotChat({
);
}
const messageContent = candidate;
+ let chatIdForRequest = workflowCopilotChatId;
if (messageOverride === undefined && !options.queuedMessageId) {
setInputValue("");
}
@@ -3622,12 +3899,207 @@ export function WorkflowCopilotChat({
let streamChatId: string | null = null;
let sawTerminalFrame = false;
let sawCredentialPause = false;
+ let recoveryNoticeId: string | null = null;
+ const finishRecordingRefinement = (
+ status: Exclude<RecordingRefinementStatus, "working">,
+ ) => {
+ setMessages((current) =>
+ current.map((message) =>
+ message.id === userMessageId && message.recordingRefinement
+ ? {
+ ...message,
+ recordingRefinement: {
+ ...message.recordingRefinement,
+ status,
+ },
+ }
+ : message,
+ ),
+ );
+ };
const shouldArmRecovery = () =>
streamTurnId !== null &&
!sawTerminalFrame &&
!abortController.signal.aborted &&
cancelInFlightController.current !== abortController &&
recoveryGeneration.current === sendGeneration;
+ const recoverPersistedRecordingTurn = async (
+ signal = abortController.signal,
+ ): Promise<boolean> => {
+ if (!recordingRefinement || !requestStarted || streamTurnId !== null) {
+ return false;
+ }
+ try {
+ const client = await getClient(credentialGetter, "sans-api-v1");
+ const response = await client.get<WorkflowCopilotChatHistoryResponse>(
+ "/workflow/copilot/chat-history",
+ {
+ params: chatIdForRequest
+ ? {
+ workflow_copilot_chat_id: chatIdForRequest,
+ request_cancel_token: cancelToken,
+ }
+ : {
+ workflow_permanent_id: workflowPermanentId,
+ request_cancel_token: cancelToken,
+ },
+ signal,
+ },
+ );
+ const earliestMatchingCreatedAt =
+ recordingRefinement.startedAtMs - 5 * 60 * 1000;
+ const opener = [...response.data.chat_history]
+ .reverse()
+ .find(
+ (message) =>
+ message.sender === "product" &&
+ message.turn_id &&
+ message.content === messageContent &&
+ parseServerStamp(message.created_at) >=
+ earliestMatchingCreatedAt,
+ );
+ const recoveredTurnId =
+ response.data.request_turn_id === undefined
+ ? opener?.turn_id
+ : (response.data.request_turn_id ?? undefined);
+ if (!recoveredTurnId) {
+ return false;
+ }
+ streamTurnId = recoveredTurnId;
+ streamChatId = response.data.workflow_copilot_chat_id;
+ if (recordingRefinementAction) {
+ useRecordingRefinementEvidenceStore
+ .getState()
+ .take(recordingRefinementAction.nonce);
+ onInitialMessageConsumedRef.current?.();
+ }
+ setWorkflowCopilotChatId(streamChatId);
+ workflowCopilotChatIdRef.current = streamChatId;
+ setMessages((current) =>
+ current.map((message) =>
+ message.id === userMessageId && message.recordingRefinement
+ ? {
+ ...message,
+ recordingRefinement: {
+ ...message.recordingRefinement,
+ turnId: recoveredTurnId,
+ },
+ }
+ : message,
+ ),
+ );
+ return true;
+ } catch (historyError) {
+ console.warn(
+ "Failed to recover recording refinement after losing turn_start:",
+ historyError,
+ );
+ return false;
+ }
+ };
+ const startPersistedRecordingTurnLookup = () => {
+ const recoveryKey = `recording-request:${userMessageId}`;
+ const generation = sendGeneration;
+ const deadline = Date.now() + RECOVERY_POLL_BUDGET_MS;
+ let step = 0;
+ let timer: ReturnType<typeof setTimeout> | null = null;
+ let deadlineTimer: ReturnType<typeof setTimeout> | null = null;
+ let inFlight: AbortController | null = null;
+ let stopped = false;
+
+ const finish = () => {
+ stopped = true;
+ if (timer !== null) {
+ clearTimeout(timer);
+ timer = null;
+ }
+ inFlight?.abort();
+ inFlight = null;
+ if (deadlineTimer !== null) {
+ clearTimeout(deadlineTimer);
+ deadlineTimer = null;
+ }
+ if (recoveryPolls.current.get(recoveryKey) === finish) {
+ recoveryPolls.current.delete(recoveryKey);
+ }
+ };
+ const giveUp = () => {
+ finish();
+ finishRecordingRefinement("failed");
+ if (recoveryNoticeId) {
+ setMessages((current) =>
+ current.map((message) =>
+ message.id === recoveryNoticeId
+ ? { ...message, content: SEND_FAILED_MESSAGE }
+ : message,
+ ),
+ );
+ }
+ };
+ const schedule = () => {
+ if (stopped) return;
+ if (Date.now() >= deadline) {
+ giveUp();
+ return;
+ }
+ const delay =
+ RECOVERY_POLL_DELAYS_MS[step] ?? RECOVERY_POLL_STEADY_MS;
+ step += 1;
+ timer = setTimeout(() => {
+ timer = null;
+ void tick();
+ }, delay);
+ };
+ const tick = async () => {
+ if (
+ stopped ||
+ recoveryGeneration.current !== generation ||
+ abortController.signal.aborted
+ ) {
+ finish();
+ return;
+ }
+ const controller = new AbortController();
+ inFlight = controller;
+ const recovered = await recoverPersistedRecordingTurn(
+ controller.signal,
+ );
+ if (inFlight === controller) {
+ inFlight = null;
+ }
+ if (stopped) return;
+ if (recovered) {
+ finish();
+ if (streamTurnId && shouldArmRecovery()) {
+ if (recoveryNoticeId) {
+ setMessages((current) =>
+ current.map((message) =>
+ message.id === recoveryNoticeId
+ ? {
+ ...message,
+ recoveryTurnId: streamTurnId ?? undefined,
+ }
+ : message,
+ ),
+ );
+ }
+ startRecoveryPoll(
+ streamChatId ?? workflowCopilotChatIdRef.current,
+ streamTurnId,
+ sawCredentialPause,
+ userMessageId,
+ );
+ }
+ return;
+ }
+ schedule();
+ };
+
+ recoveryPolls.current.get(recoveryKey)?.();
+ recoveryPolls.current.set(recoveryKey, finish);
+ deadlineTimer = setTimeout(giveUp, RECOVERY_POLL_BUDGET_MS);
+ void tick();
+ };
setStopArmed(false);
if (stopArmTimer.current !== null) {
@@ -3645,7 +4117,6 @@ export function WorkflowCopilotChat({
const saveData = getSaveData();
const workflowId = saveData?.workflow.workflow_id;
let workflowYaml = "";
- let chatIdForRequest = workflowCopilotChatId;
let audioArtifactId: string | null = null;
if (!workflowId) {
@@ -3707,6 +4178,7 @@ export function WorkflowCopilotChat({
totp_verification_url: saveData.workflow.totp_verification_url,
extra_http_headers: extraHttpHeaders,
run_with: saveData.settings.runWith,
+ browser_type: saveData.settings.browserType ?? null,
cache_key: normalizedKey,
ai_fallback: saveData.settings.aiFallback ?? true,
enable_self_healing: saveData.settings.enableSelfHealing ?? false,
@@ -3744,6 +4216,7 @@ export function WorkflowCopilotChat({
mask_secrets: saveData.settings.maskSecrets,
browser_profile_id: saveData.settings.browserProfileId,
browser_profile_key: saveData.settings.browserProfileKey,
+ browser_type: saveData.settings.browserType ?? null,
model: saveData.settings.model,
workflow_definition: {
...saveData.workflow.workflow_definition,
@@ -3855,6 +4328,19 @@ export function WorkflowCopilotChat({
const runFailed =
frozenNarrative !== undefined &&
notConfirmedOutcome(frozenNarrative) !== null;
+ const refinementProducedWorkflow = Boolean(
+ response.updated_workflow &&
+ response.proposal_disposition !== "no_proposal",
+ );
+ finishRecordingRefinement(
+ response.cancelled || userCancelledThisTurn
+ ? "cancelled"
+ : frozenNarrative?.terminal === "error" ||
+ runFailed ||
+ !refinementProducedWorkflow
+ ? "failed"
+ : "complete",
+ );
if (
lastTurnRef.current &&
!response.cancelled &&
@@ -3928,6 +4414,7 @@ export function WorkflowCopilotChat({
payload: WorkflowCopilotStreamErrorUpdate,
errorNarrative?: TurnNarrativeState,
) => {
+ finishRecordingRefinement("failed");
pendingCancelToken.current = null;
// A terminal error carries no credentialPause payload, so the dead
// frame must be cleared explicitly or its card would stay actionable.
@@ -3983,6 +4470,9 @@ export function WorkflowCopilotChat({
if (!composerMountedRef.current || abortController.signal.aborted) {
if (composerMountedRef.current) {
returnFilesToTray(sentAttachments);
+ if (abortController.signal.aborted) {
+ finishRecordingRefinement("cancelled");
+ }
}
return;
}
@@ -4148,6 +4638,20 @@ export function WorkflowCopilotChat({
.getState()
.take(productAction.nonce);
onInitialMessageConsumedRef.current?.();
+ setMessages((prev) =>
+ prev.map((message) =>
+ message.id === userMessageId &&
+ message.recordingRefinement
+ ? {
+ ...message,
+ recordingRefinement: {
+ ...message.recordingRefinement,
+ turnId: payload.turn_id,
+ },
+ }
+ : message,
+ ),
+ );
}
// A new turn can't carry the prior turn's dead resume_token.
setLivePauseFrame(null);
@@ -4233,15 +4737,18 @@ export function WorkflowCopilotChat({
// The streaming client resolves rather than throws on abort, so New chat, a chat switch, or
// Stop before turn_start ends here, not in the catch.
if (abortController.signal.aborted) {
+ finishRecordingRefinement("cancelled");
returnPossiblySavedFiles();
}
} catch (error) {
// A stream severed before turn_start may still have been saved by the server. An error frame
// before turn_start is definitive and leaves the files deletable.
- returnPossiblySavedFiles();
if (abortController.signal.aborted) {
+ returnPossiblySavedFiles();
+ finishRecordingRefinement("cancelled");
return;
}
+ returnPossiblySavedFiles();
console.error("Failed to send message:", error);
if (options.idempotencyKey !== undefined && workflowCopilotChatId) {
toast({
@@ -4262,14 +4769,27 @@ export function WorkflowCopilotChat({
sendGeneration = recoveryGeneration.current;
}
} else {
+ const recovering =
+ shouldArmRecovery() ||
+ Boolean(recordingRefinement && requestStarted);
+ if (!recovering) {
+ finishRecordingRefinement("failed");
+ }
const errorMessage: ChatMessage = {
id: Date.now().toString(),
sender: "ai",
- content: shouldArmRecovery()
+ content: recovering
? RECOVERY_IN_PROGRESS_MESSAGE
: SEND_FAILED_MESSAGE,
+ recoveryTurnId: recovering
+ ? (streamTurnId ?? undefined)
+ : undefined,
};
+ recoveryNoticeId = errorMessage.id;
setMessages((prev) => [...prev, errorMessage]);
+ if (recovering && streamTurnId === null && recordingRefinement) {
+ startPersistedRecordingTurnLookup();
+ }
}
// A thrown stream never emits a terminal narrative event, so clear the
// bubble or its Working/elapsed indicator would tick forever.
@@ -4320,6 +4840,7 @@ export function WorkflowCopilotChat({
streamChatId ?? workflowCopilotChatIdRef.current,
streamTurnId,
sawCredentialPause,
+ recordingRefinement ? userMessageId : undefined,
);
}
}
@@ -4889,9 +5410,18 @@ export function WorkflowCopilotChat({
inputValue.trim().length > 0 ||
attachments.length > 0 ||
pendingAttachments.some((item) => item.status === "uploading");
- // The cycling verb row plus the stop button's orbiting ring carry the
- // working state, so the prose status line and the queued chip stand down.
- const showWorkingRow = isLoading;
+ const recordingRefinementInFlight = messages.some((message) => {
+ const refinement = message.recordingRefinement;
+ if (refinement?.status !== "working") return false;
+ return (
+ message.id === pendingMessageId.current ||
+ (recoveredTurnOwnerRef.current !== null &&
+ refinement.turnId === recoveredTurnOwnerRef.current)
+ );
+ });
+ // Recording refinement owns a more specific progress card, so the generic
+ // cycling verb stands down while that product-authored turn is active.
+ const showWorkingRow = isLoading && !recordingRefinementInFlight;
// A live_browser-reason queued prompt parks with no active turn to stop, so
// an empty composer's morph button would render as a guaranteed no-op "Send".
// With text typed it does act — it rewrites the parked prompt.
@@ -5104,6 +5634,22 @@ export function WorkflowCopilotChat({
{messages.flatMap((message, index) => {
const rendered = (() => {
const isLastMessage = index === lastTurnIndex;
+ if (
+ message.kind === "recording_refinement" &&
+ message.recordingRefinement
+ ) {
+ return (
+ <RecordingRefinementProgressCard
+ key={message.id}
+ awaitingReview={Boolean(
+ proposedWorkflow &&
+ pendingProposalTurnId ===
+ message.recordingRefinement.turnId,
+ )}
+ {...message.recordingRefinement}
+ />
+ );
+ }
if (
message.kind === "run_lifecycle" ||
(message.sender === "product" &&
@@ -5502,7 +6048,9 @@ export function WorkflowCopilotChat({
every send, including queued-then-drained follow-ups.
*/}
{isLoading && !isLoadingHistory && narrative.turnId === null ? (
- <InstantAckPlaceholder />
+ recordingRefinementInFlight ? null : (
+ <InstantAckPlaceholder />
+ )
) : null}
{/*
Bottom in-flight narrative bubble. Suppressed once the terminal
diff --git a/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts b/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts
index c4834f81a..ca55f37dc 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts
+++ b/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts
@@ -170,9 +170,11 @@ export interface WorkflowCopilotCancelRequest {
export interface WorkflowCopilotChatHistoryMessage {
sender: WorkflowCopilotChatSender;
content: string;
+ turn_id?: string | null;
audio_artifact_id?: string | null;
attached_files?: CopilotAttachedFile[];
created_at: string;
+ modified_at?: string;
// Typed turn outcome persisted on assistant rows; optional so the FE
// tolerates an older backend that does not serve it.
turn_outcome?:
@@ -193,6 +195,7 @@ export interface WorkflowCopilotChatHistoryResponse {
question_interactions?: QuestionInteraction[];
pending_question_cancel_token?: string | null;
workflow_copilot_chat_id: string | null;
+ request_turn_id?: string | null;
chat_history: WorkflowCopilotChatHistoryMessage[];
proposed_workflow?: WorkflowApiResponse | null;
proposed_workflow_metadata?: CopilotProposalMetadata | null;
diff --git a/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx b/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx
index debfd9019..e9d34b772 100644
--- a/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx
+++ b/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx
@@ -86,6 +86,7 @@ function Debugger() {
? JSON.stringify(workflow.cdp_connect_headers)
: null,
runWith: workflow.run_with ?? "agent",
+ browserType: workflow.browser_type ?? null,
codeVersion: workflow.code_version ?? null,
scriptCacheKey: workflow.cache_key,
aiFallback: workflow.ai_fallback ?? true,
diff --git a/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx b/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx
index 23a8cea45..7cebb814e 100644
--- a/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx
@@ -114,6 +114,7 @@ function WorkflowEditor() {
? JSON.stringify(workflow.cdp_connect_headers)
: null,
runWith: workflow.run_with ?? "agent",
+ browserType: workflow.browser_type ?? null,
codeVersion: workflow.code_version ?? null,
scriptCacheKey: workflow.cache_key,
aiFallback: workflow.ai_fallback ?? true,
diff --git a/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx b/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx
index 7a4d4111c..a819cc912 100644
--- a/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx
@@ -1719,6 +1719,7 @@ function Workspace({
? JSON.stringify(workflowData.cdp_connect_headers)
: null,
runWith: workflowData.run_with ?? "agent",
+ browserType: workflowData.browser_type ?? null,
codeVersion: workflowData.code_version ?? null,
scriptCacheKey: workflowData.cache_key ?? null,
aiFallback: workflowData.ai_fallback ?? true,
@@ -2062,6 +2063,7 @@ function Workspace({
? JSON.stringify(selectedVersion.cdp_connect_headers)
: null,
runWith: selectedVersion.run_with ?? "agent",
+ browserType: selectedVersion.browser_type ?? null,
codeVersion: selectedVersion.code_version ?? null,
scriptCacheKey: selectedVersion.cache_key,
aiFallback: selectedVersion.ai_fallback ?? true,
@@ -3076,6 +3078,7 @@ function Workspace({
modified_at: new Date().toISOString(),
deleted_at: null,
run_with: saveData.settings.runWith,
+ browser_type: saveData.settings.browserType ?? null,
cache_key: saveData.settings.scriptCacheKey,
ai_fallback: saveData.settings.aiFallback,
enable_self_healing: saveData.settings.enableSelfHealing ?? false,
diff --git a/skyvern-frontend/src/routes/workflows/editor/edges/EdgeWithAddButton.tsx b/skyvern-frontend/src/routes/workflows/editor/edges/EdgeWithAddButton.tsx
index d0112c19e..9b8b748dc 100644
--- a/skyvern-frontend/src/routes/workflows/editor/edges/EdgeWithAddButton.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/edges/EdgeWithAddButton.tsx
@@ -98,7 +98,9 @@ function EdgeWithAddButton({
const isUploadingSOP = sopToBlocksMutation.isPending;
const isProcessing =
- processRecordingMutation.isPending || recordingStore.isCommitting;
+ processRecordingMutation.isPending ||
+ recordingStore.finishRequested ||
+ recordingStore.isCommitting;
const isBusy =
(isProcessing || recordingStore.isRecording) &&
@@ -152,6 +154,10 @@ function EdgeWithAddButton({
};
const onEndRecord = () => {
+ if (isProcessing) {
+ return;
+ }
+
// With live interpretation, the recording panel owns the commit so user
// edits/deletes to draft steps are honored.
if (recordingStore.isRecording && recordingStore.sessionRevision > 0) {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.test.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.test.tsx
index 9b844e1c5..cef34de37 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.test.tsx
@@ -59,10 +59,13 @@ const initialSettings = useSettingsStore.getState();
const initialRecording = useRecordingStore.getState();
const initialWorkflowSettings = useWorkflowSettingsStore.getState();
-function renderNodeAdder(props: Partial<NodeProps<NodeAdderNodeType>>) {
+function renderNodeAdder(
+ props: Partial<NodeProps<NodeAdderNodeType>>,
+ { debug = false }: { debug?: boolean } = {},
+) {
return render(
<DebugStoreContext.Provider
- value={{ isDebugMode: false, blockRunsEnabled: false }}
+ value={{ isDebugMode: debug, blockRunsEnabled: false }}
>
<NodeAdderNode
{...({
@@ -158,4 +161,48 @@ describe("NodeAdderNode", () => {
});
expect(state.data?.branchContext).toBeUndefined();
});
+
+ it("shows processing instead of a second stop action after Done", () => {
+ useEdgesMock.mockReturnValue([
+ { id: "edge", source: "start", target: "adder" },
+ ]);
+ useNodesMock.mockReturnValue([
+ { id: "start", parentId: "loop", type: "start", data: {} },
+ ]);
+ useSettingsStore.getState().setIsUsingABrowser(true);
+ useRecordingStore.setState({
+ isRecording: true,
+ finishRequested: true,
+ });
+ useWorkflowPanelStore.setState({
+ workflowPanelState: {
+ active: false,
+ content: "nodeLibrary",
+ data: {
+ previous: "start",
+ next: "adder",
+ parent: "loop",
+ connectingEdgeType: "default",
+ },
+ },
+ });
+
+ renderNodeAdder({}, { debug: true });
+
+ const busyControl =
+ screen.getByTestId("node-adder-button").parentElement?.parentElement;
+ expect(busyControl).toBeTruthy();
+ fireEvent.mouseEnter(busyControl!);
+
+ expect(screen.getAllByText("Processing...").length).toBeGreaterThan(0);
+ expect(
+ screen.queryAllByText("Stop recording & generate blocks"),
+ ).toHaveLength(0);
+
+ fireEvent.click(busyControl!.firstElementChild!);
+ expect(useRecordingStore.getState()).toMatchObject({
+ isRecording: true,
+ finishRequested: true,
+ });
+ });
});
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.tsx
index c21d3c7f3..e15a57e4c 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/NodeAdderNode/NodeAdderNode.tsx
@@ -114,7 +114,9 @@ function NodeAdderNode({ id, parentId }: NodeProps<NodeAdderNode>) {
const isUploadingSOP = sopToBlocksMutation.isPending;
const isProcessing =
- processRecordingMutation.isPending || recordingStore.isCommitting;
+ processRecordingMutation.isPending ||
+ recordingStore.finishRequested ||
+ recordingStore.isCommitting;
const isBusy =
(isProcessing || recordingStore.isRecording) &&
@@ -181,6 +183,10 @@ function NodeAdderNode({ id, parentId }: NodeProps<NodeAdderNode>) {
};
const onEndRecord = () => {
+ if (isProcessing) {
+ return;
+ }
+
// With live interpretation, the recording panel owns the commit so user
// edits/deletes to draft steps are honored.
if (recordingStore.isRecording && recordingStore.sessionRevision > 0) {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/WorkflowSettingsEditor.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/WorkflowSettingsEditor.tsx
index 1dbaf79cb..e1504cc75 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/WorkflowSettingsEditor.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/WorkflowSettingsEditor.tsx
@@ -36,6 +36,10 @@ import {
} from "@/components/ui/select";
import { Switch } from "@/components/ui/switch";
+import {
+ hasBrowserTypeOptions,
+ useBrowserTypeOptionsQuery,
+} from "@/routes/workflows/utils";
import { useResetProfileMutation } from "@/routes/workflows/hooks/useResetProfileMutation";
import { useWorkflowQuery } from "@/routes/workflows/hooks/useWorkflowQuery";
import { useWorkflowStudioEnabled } from "@/hooks/useWorkflowStudioEnabled";
@@ -107,6 +111,7 @@ function WorkflowSettingsEditorBody({
const studioEnabled = useWorkflowStudioEnabled();
const browserMemoryEnabled = useFeatureFlag("browser_memory_v1");
const { data: workflow } = useWorkflowQuery({ workflowPermanentId });
+ const { data: browserTypeOptions } = useBrowserTypeOptionsQuery();
// Self-healing is restricted to copilot-authored workflows; hide the toggle
// elsewhere so it never reads as a switch that silently does nothing.
// copilot_authored is lineage-derived server-side — the current version's
@@ -232,6 +237,32 @@ function WorkflowSettingsEditorBody({
onChange={(value) => update({ proxyLocation: value })}
/>
</div>
+ {hasBrowserTypeOptions(browserTypeOptions) ? (
+ <div className="space-y-2">
+ <div className="flex gap-2">
+ <Label>Browser Type</Label>
+ <HelpTooltip content="Browser engine used for this agent's runs. Leave as Default to let Skyvern choose. A workflow-run setting overrides this. Note: Google Chrome does not support the captcha-solver extension." />
+ </div>
+ <Select
+ value={data.browserType ?? "default"}
+ onValueChange={(value) =>
+ update({ browserType: value === "default" ? null : value })
+ }
+ >
+ <SelectTrigger className="w-full">
+ <SelectValue placeholder="Browser Type" />
+ </SelectTrigger>
+ <SelectContent>
+ <SelectItem value="default">Default</SelectItem>
+ {browserTypeOptions.map((option) => (
+ <SelectItem key={option.value} value={option.value}>
+ {option.label}
+ </SelectItem>
+ ))}
+ </SelectContent>
+ </Select>
+ </div>
+ ) : null}
<div className="flex flex-col gap-4 rounded-md bg-slate-elevation5 p-4 pl-4">
<div className="flex flex-col gap-4">
<div className="flex justify-between">
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/types.ts
index df7bd104f..ebf0f429f 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/types.ts
@@ -23,6 +23,7 @@ export type WorkflowStartNodeData = {
cdpConnectHeaders: string | Record<string, unknown> | null;
editable: boolean;
runWith: string;
+ browserType?: string | null;
codeVersion: number | null;
scriptCacheKey: string | null;
aiFallback: boolean;
diff --git a/skyvern-frontend/src/routes/workflows/editor/panels/BlockConfigForm/WorkflowSettingsBlockForm.tsx b/skyvern-frontend/src/routes/workflows/editor/panels/BlockConfigForm/WorkflowSettingsBlockForm.tsx
index a4209a192..956152eb7 100644
--- a/skyvern-frontend/src/routes/workflows/editor/panels/BlockConfigForm/WorkflowSettingsBlockForm.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/panels/BlockConfigForm/WorkflowSettingsBlockForm.tsx
@@ -44,6 +44,7 @@ function WorkflowSettingsBlockFormBody({
pinSavedSessionIp: data.pinSavedSessionIp,
browserProfileId: data.browserProfileId,
browserProfileKey: data.browserProfileKey,
+ browserType: data.browserType,
extraHttpHeaders: data.extraHttpHeaders,
maxScreenshotScrolls: data.maxScreenshotScrolls,
maxElapsedTimeMinutes: data.maxElapsedTimeMinutes,
diff --git a/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx b/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx
index 196357ecb..bbe98dd7f 100644
--- a/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx
@@ -158,6 +158,7 @@ function getWorkflowElements(version: WorkflowVersion) {
? JSON.stringify(version.cdp_connect_headers)
: null,
runWith: version.run_with ?? "agent",
+ browserType: version.browser_type ?? null,
codeVersion: version.code_version ?? null,
scriptCacheKey: version.cache_key,
aiFallback: version.ai_fallback ?? true,
diff --git a/skyvern-frontend/src/routes/workflows/editor/workflowChangesSummary.ts b/skyvern-frontend/src/routes/workflows/editor/workflowChangesSummary.ts
index a5b6c12f3..a170b5c70 100644
--- a/skyvern-frontend/src/routes/workflows/editor/workflowChangesSummary.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/workflowChangesSummary.ts
@@ -212,6 +212,7 @@ const SETTINGS_LABELS: Array<{
fields: ["browserProfileId", "browserProfileKey"],
label: "Changed browser profile",
},
+ { fields: ["browserType"], label: "Changed browser type" },
{ fields: ["model"], label: "Changed model" },
{ fields: ["maxScreenshotScrolls"], label: "Changed max screenshot scrolls" },
{ fields: ["maxElapsedTimeMinutes"], label: "Changed max run time" },
diff --git a/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts b/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts
index fef878f00..5fe234c95 100644
--- a/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts
@@ -2191,6 +2191,7 @@ function getElements(
cdpConnectHeaders: settings.cdpConnectHeaders,
editable,
runWith: settings.runWith,
+ browserType: settings.browserType ?? null,
codeVersion: settings.codeVersion,
scriptCacheKey: settings.scriptCacheKey,
aiFallback: settings.aiFallback ?? true,
@@ -3827,6 +3828,7 @@ function getWorkflowSettings(nodes: Array<AppNode>): WorkflowSettings {
? JSON.stringify(data.cdpConnectHeaders)
: data.cdpConnectHeaders,
runWith: data.runWith,
+ browserType: data.browserType ?? null,
codeVersion: data.codeVersion,
scriptCacheKey: data.scriptCacheKey,
aiFallback: data.aiFallback,
@@ -5102,6 +5104,7 @@ function convert(workflow: WorkflowApiResponse): WorkflowCreateYAMLRequest {
is_saved_task: workflow.is_saved_task,
status: workflow.status,
run_with: workflow.run_with ?? "agent",
+ browser_type: workflow.browser_type ?? null,
adaptive_caching: workflow.adaptive_caching ?? undefined,
code_version: workflow.code_version ?? undefined,
cache_key: workflow.cache_key,
diff --git a/skyvern-frontend/src/routes/workflows/editor/workflowVersionFromSaveData.ts b/skyvern-frontend/src/routes/workflows/editor/workflowVersionFromSaveData.ts
index d54c25023..d26f27163 100644
--- a/skyvern-frontend/src/routes/workflows/editor/workflowVersionFromSaveData.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/workflowVersionFromSaveData.ts
@@ -64,6 +64,7 @@ export function workflowVersionFromSaveData(
modified_at: workflow.modified_at,
deleted_at: workflow.deleted_at ?? null,
run_with: settings.runWith,
+ browser_type: settings.browserType ?? null,
cache_key: settings.scriptCacheKey,
ai_fallback: settings.aiFallback,
enable_self_healing: settings.enableSelfHealing ?? false,
diff --git a/skyvern-frontend/src/routes/workflows/studio/BrowserTab.test.tsx b/skyvern-frontend/src/routes/workflows/studio/BrowserTab.test.tsx
index 1e4ebb433..6edf56c55 100644
--- a/skyvern-frontend/src/routes/workflows/studio/BrowserTab.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/BrowserTab.test.tsx
@@ -6,8 +6,9 @@ import {
fireEvent,
render,
screen,
+ waitFor,
} from "@testing-library/react";
-import { MemoryRouter, Route, Routes } from "react-router-dom";
+import { MemoryRouter, Route, Routes, useLocation } from "react-router-dom";
import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
@@ -114,7 +115,6 @@ vi.mock("./runview/RunLiveStream", () => ({
/>
),
}));
-
const initialBrowserState = useStudioBrowserStore.getState();
const initialRunViewState = useRunViewStore.getState();
const initialRecordingState = useRecordingStore.getState();
@@ -236,6 +236,7 @@ function renderBrowserPane(initialPath: string) {
<BrowserPaneActions />
</div>
<BrowserTab />
+ <LocationProbe />
</>
}
/>
@@ -254,6 +255,11 @@ function renderBrowserPane(initialPath: string) {
};
}
+function LocationProbe() {
+ const location = useLocation();
+ return <div data-testid="location-search">{location.search}</div>;
+}
+
const STUDIO_PATH = "/workflows/wpid_test/studio?panes=copilot,browser";
beforeEach(() => {
@@ -273,6 +279,64 @@ afterEach(() => {
});
describe("BrowserTab view machine", () => {
+ it("keeps the recording view synchronized with the URL", async () => {
+ seedRun({
+ status: Status.Completed,
+ recordingUrl: "https://example.com/recording.webm",
+ });
+ renderBrowserPane("/workflows/wpid_test/studio?wr=wr_1&view=recording");
+
+ await waitFor(() => {
+ expect(
+ screen
+ .getByRole("button", { name: "Recording" })
+ .getAttribute("aria-pressed"),
+ ).toBe("true");
+ });
+
+ fireEvent.click(screen.getByRole("button", { name: "Screenshots" }));
+ await waitFor(() => {
+ expect(screen.getByTestId("location-search").textContent).not.toContain(
+ "view=recording",
+ );
+ });
+
+ fireEvent.click(screen.getByRole("button", { name: "Recording" }));
+ await waitFor(() => {
+ expect(screen.getByTestId("location-search").textContent).toContain(
+ "view=recording",
+ );
+ expect(screen.getByTestId("location-search").textContent).toContain(
+ "panes=browser,overview",
+ );
+ });
+ });
+
+ it("keeps a recording deep link pinned when frame hydration updates", async () => {
+ seedRun({ status: Status.Failed });
+ renderBrowserPane(
+ "/workflows/wpid_test/studio?wr=wr_1&active=wrb_1&view=recording",
+ );
+
+ await waitFor(() => {
+ expect(
+ screen
+ .getByRole("button", { name: "Recording" })
+ .getAttribute("aria-pressed"),
+ ).toBe("true");
+ });
+
+ act(() => useRunViewStore.getState().pinFrame("wrb_1"));
+
+ await waitFor(() => {
+ expect(
+ screen
+ .getByRole("button", { name: "Recording" })
+ .getAttribute("aria-pressed"),
+ ).toBe("true");
+ });
+ });
+
it("keeps Live selected during a retry wait and pauses artifact polling until execution resumes", () => {
mocks.realScreenshot = true;
seedRun({ status: Status.Running });
@@ -575,6 +639,18 @@ describe("BrowserTab view machine", () => {
expect(screen.getByText("No recording for this run")).toBeTruthy();
});
+ it("opens the recording requested by a legacy deep link", () => {
+ seedRun({
+ status: Status.Failed,
+ recordingUrl: "https://r.test/1.mp4",
+ });
+ mocks.debugSession = { browser_session_id: "pbs_test" };
+ renderBrowserPane(`${STUDIO_PATH}&wr=wr_1&view=recording`);
+
+ expect(screen.getByTestId("hero-recording")).toBeTruthy();
+ expect(useStudioBrowserStore.getState().view).toBe("recording");
+ });
+
it("flags a queued block run on the live debug stream", () => {
seedRun({ status: Status.Queued, browserSessionId: "pbs_test" });
mocks.debugSession = { browser_session_id: "pbs_test" };
diff --git a/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.test.tsx b/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.test.tsx
index 86a95d6cf..5064e223e 100644
--- a/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.test.tsx
@@ -14,7 +14,12 @@ const runtimeConfigMock = vi.hoisted(() => ({
const browserStreamProps = vi.hoisted(
() =>
- ({ last: null }) as { last: { resetRecordingOnUnmount?: boolean } | null },
+ ({ last: null }) as {
+ last: {
+ exfiltrate?: boolean;
+ resetRecordingOnUnmount?: boolean;
+ } | null;
+ },
);
const cdpStreamProps = vi.hoisted(
@@ -40,7 +45,10 @@ vi.mock("@/hooks/useRuntimeConfig", () => ({
}));
vi.mock("@/components/BrowserStream", () => ({
- BrowserStream: (props: { resetRecordingOnUnmount?: boolean }) => {
+ BrowserStream: (props: {
+ exfiltrate?: boolean;
+ resetRecordingOnUnmount?: boolean;
+ }) => {
browserStreamProps.last = props;
return <div data-testid="vnc-stream" />;
},
@@ -101,6 +109,21 @@ describe("StreamPresenter transport-swap recording", () => {
expect(screen.queryByTestId("vnc-stream")).not.toBeNull();
expect(browserStreamProps.last?.resetRecordingOnUnmount).toBe(false);
});
+
+ it("stops VNC recording when Done requests finalization", () => {
+ runtimeConfigMock.browserStreamingMode = "vnc";
+ useRecordingStore.getState().setIsRecording(true);
+ const { rerender } = render(
+ <StreamPresenter browserSessionId="pbs_test" isRecording />,
+ );
+
+ expect(browserStreamProps.last?.exfiltrate).toBe(true);
+
+ useRecordingStore.getState().requestFinish();
+ rerender(<StreamPresenter browserSessionId="pbs_test" isRecording />);
+
+ expect(browserStreamProps.last?.exfiltrate).toBe(false);
+ });
});
describe("StreamPresenter recording on a cdp-transport session", () => {
diff --git a/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.tsx b/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.tsx
index d4add4347..447a7ae62 100644
--- a/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/StreamPresenter.tsx
@@ -92,7 +92,7 @@ export function StreamPresenter({
browserSessionId={browserSessionId}
interactive={interactive}
showControlButtons={showControlButtons}
- exfiltrate={isRecording}
+ exfiltrate={isRecording && !finishRequested}
hideRecordingIndicator={hideRecordingIndicator}
// StrictMode remounts this component; the recording must survive that.
// StudioBrowserStream owns the session-level reset instead.
diff --git a/skyvern-frontend/src/routes/workflows/studio/StudioShell.test.tsx b/skyvern-frontend/src/routes/workflows/studio/StudioShell.test.tsx
index 44beee2d5..544b04c51 100644
--- a/skyvern-frontend/src/routes/workflows/studio/StudioShell.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/StudioShell.test.tsx
@@ -9,7 +9,65 @@ import { afterEach, beforeEach, describe, expect, test, vi } from "vitest";
import { type StudioPaneId } from "./panes";
import { paneAccessibleName } from "./paneMeta";
import { paneExpansionKeyframes } from "./paneLayout";
-import { StudioPane } from "./StudioShell";
+import { panesAfterRecordingTransition } from "./recordingPaneLifecycle";
+import { EmbeddedBrowserOverlays, StudioPane } from "./StudioShell";
+
+vi.mock("./useRunVisuals", () => ({
+ useRunVisuals: () => ({
+ recordingArchived: true,
+ recordingUrls: [],
+ }),
+}));
+vi.mock("../workflowRun/WorkflowRunVerificationCodeForm", () => ({
+ WorkflowRunVerificationCodeForm: ({
+ workflowRunId,
+ }: {
+ workflowRunId: string | undefined;
+ }) => <div data-testid="verification-form">{workflowRunId}</div>,
+}));
+
+describe("EmbeddedBrowserOverlays", () => {
+ test("keeps verification available in screenshot-only browser views", () => {
+ render(
+ <EmbeddedBrowserOverlays runId="wr_1" showArchivedRecording={false} />,
+ );
+
+ expect(screen.getByTestId("verification-form").textContent).toBe("wr_1");
+ expect(screen.queryByText("Recording archived")).toBeNull();
+ });
+});
+
+describe("panesAfterRecordingTransition", () => {
+ test("replaces the Editor with Browser and Copilot when recording starts", () => {
+ expect(
+ panesAfterRecordingTransition(["overview", "editor"], "started"),
+ ).toEqual(["overview", "copilot", "browser"]);
+ });
+
+ test("replaces Browser with Editor when a recording starts processing", () => {
+ expect(
+ panesAfterRecordingTransition(
+ ["overview", "copilot", "browser"],
+ "processing",
+ ),
+ ).toEqual(["overview", "copilot", "editor"]);
+ });
+
+ test("restores Editor but leaves Browser open when recording ends without processing", () => {
+ expect(
+ panesAfterRecordingTransition(["copilot", "browser"], "ended"),
+ ).toEqual(["copilot", "browser", "editor"]);
+ });
+
+ test("does not duplicate panes already in the requested layout", () => {
+ expect(
+ panesAfterRecordingTransition(
+ ["copilot", "browser", "editor"],
+ "started",
+ ),
+ ).toEqual(["copilot", "browser"]);
+ });
+});
describe("paneExpansionKeyframes", () => {
test("grows from the pane's current bounds into its final bounds", () => {
diff --git a/skyvern-frontend/src/routes/workflows/studio/StudioShell.tsx b/skyvern-frontend/src/routes/workflows/studio/StudioShell.tsx
index 546626eea..4ca8c5ac1 100644
--- a/skyvern-frontend/src/routes/workflows/studio/StudioShell.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/StudioShell.tsx
@@ -8,6 +8,7 @@ import {
type ReactNode,
} from "react";
import { createPortal } from "react-dom";
+import { useSearchParams } from "react-router-dom";
import { useWorkflowPermanentId } from "@/routes/workflows/WorkflowPermanentIdContext";
import {
Cross2Icon,
@@ -16,6 +17,7 @@ import {
} from "@radix-ui/react-icons";
import { CopyButton } from "@/components/CopyButton";
+import { StreamStatusPanel } from "@/routes/streaming/StreamDiagnostics";
import {
Tooltip,
TooltipContent,
@@ -31,6 +33,7 @@ import { cn } from "@/util/utils";
import { deriveDropIndicator } from "../editor/sortable/dropIndicator";
import { useDebugSessionQuery } from "../hooks/useDebugSessionQuery";
+import { WorkflowRunVerificationCodeForm } from "../workflowRun/WorkflowRunVerificationCodeForm";
import { BrowserPaneActions, BrowserPaneViewPills } from "./BrowserPaneHeader";
import { CopilotActiveDot, CopilotPaneControls } from "./CopilotPaneHeader";
@@ -68,6 +71,7 @@ import {
} from "./panes";
import { StudioPaneDefaultsProvider } from "./StudioPaneDefaults";
import { useStudioPaneDefaults } from "./StudioPaneDefaultsContext";
+import { panesAfterRecordingTransition } from "./recordingPaneLifecycle";
import {
StudioPaneCompactContext,
StudioShellContext,
@@ -78,6 +82,7 @@ import { StudioTopBar } from "./StudioTopBar";
import { StudioWorkflowPanels } from "./StudioWorkflowPanels";
import { useStudioPanes } from "./useStudioPanes";
import { useStudioRunId } from "./useStudioRunId";
+import { useRunVisuals } from "./useRunVisuals";
// Below this header width, pane header chrome (view pills, badges) collapses
// to icons — same idea as the run hero, measured per pane, not per viewport.
@@ -156,6 +161,7 @@ export function StudioPane({
headerExtras,
headerActions,
iconBadge,
+ chromeless = false,
children,
}: {
id: StudioPaneId;
@@ -177,6 +183,7 @@ export function StudioPane({
headerActions?: ReactNode;
// Presence-badge overlay on the pane icon (e.g. the Copilot active dot).
iconBadge?: ReactNode;
+ chromeless?: boolean;
children: ReactNode;
}) {
const { icon: Icon } = STUDIO_PANE_META[id];
@@ -189,7 +196,8 @@ export function StudioPane({
const headerRef = useRef<HTMLDivElement>(null);
const onTransitionEndRef = useRef(onTransitionEnd);
onTransitionEndRef.current = onTransitionEnd;
- const hasChrome = headerExtras != null || headerActions != null;
+ const hasChrome =
+ !chromeless && (headerExtras != null || headerActions != null);
const [compact, setCompact] = useState(false);
useLayoutEffect(() => {
const pane = paneRef.current;
@@ -280,7 +288,8 @@ export function StudioPane({
aria-label={accessibleLabel}
style={{ order, minWidth: STUDIO_PANE_MIN_WIDTH[id], flex }}
className={cn(
- "relative min-h-0 flex-col overflow-hidden rounded-lg border-2 border-border bg-slate-elevation1",
+ "relative min-h-0 flex-col overflow-hidden bg-slate-elevation1",
+ !chromeless && "rounded-lg border-2 border-border",
expanded && "absolute inset-3 z-30",
transitionFromBounds !== null &&
"pointer-events-none z-30 will-change-transform",
@@ -335,7 +344,10 @@ export function StudioPane({
event.preventDefault();
reorder.onMove(event.key === "ArrowLeft" ? -1 : 1);
}}
- className="flex h-11 shrink-0 cursor-grab select-none items-center gap-2 border-b border-border px-2 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-inset focus-visible:ring-ring active:cursor-grabbing"
+ className={cn(
+ "flex h-11 shrink-0 cursor-grab select-none items-center gap-2 border-b border-border px-2 focus-visible:outline-none focus-visible:ring-1 focus-visible:ring-inset focus-visible:ring-ring active:cursor-grabbing",
+ chromeless && "hidden",
+ )}
>
{/* The drag hint lives on the grip (icon + label) only, so the header's
buttons keep their own tooltips instead of inheriting this one. */}
@@ -694,8 +706,43 @@ function WorkflowDeletedPaneNotice() {
);
}
+export function EmbeddedBrowserOverlays({
+ runId,
+ showArchivedRecording,
+}: {
+ runId: string | undefined;
+ showArchivedRecording: boolean;
+}) {
+ const visuals = useRunVisuals(runId);
+ const archived =
+ showArchivedRecording &&
+ visuals.recordingArchived &&
+ visuals.recordingUrls.length === 0;
+
+ return (
+ <>
+ {archived ? (
+ <div className="absolute inset-0 z-[1] flex items-center justify-center bg-slate-950">
+ <StreamStatusPanel
+ diagnostic={{
+ title: "Recording archived",
+ detail: "To request restoration, contact [email protected].",
+ }}
+ />
+ </div>
+ ) : null}
+ <div className="absolute inset-x-0 top-0 z-10">
+ <WorkflowRunVerificationCodeForm workflowRunId={runId} />
+ </div>
+ </>
+ );
+}
+
function StudioStage(props: StudioWorkspaceProps) {
- const { panes, closePane, openPane, setPanesOrder } = useStudioPanes();
+ const [searchParams] = useSearchParams();
+ const embedded = searchParams.get("embed") === "true";
+ const { panes, closePane, openPane, setOpenPanes, setPanesOrder } =
+ useStudioPanes();
const { registerStageElement } = useStudioPaneDefaults();
const workflowPermanentId = useWorkflowPermanentId();
const runId = useStudioRunId();
@@ -859,15 +906,17 @@ function StudioStage(props: StudioWorkspaceProps) {
};
// Recording lives in the Browser pane (the live stream is there) with the
- // live-drafts panel taking over the Copilot pane. Once a commit is in flight
- // or its blocks are landing, reveal the Editor pane (it shows the loading
- // overlay). Gated on lifecycle transitions so manual pane changes made
+ // live-drafts panel taking over the Copilot pane. The Editor is not useful
+ // during capture, and the Browser is not useful while the recording is being
+ // applied. Gated on lifecycle transitions so manual pane changes made
// mid-recording are preserved.
+ const finishRequested = useRecordingStore((s) => s.finishRequested);
const isCommitting = useRecordingStore((s) => s.isCommitting);
const recordedBlocksPending = useRecordedBlocksStore(
(s) => (s.blocks?.length ?? 0) > 0,
);
- const processingRecording = isCommitting || recordedBlocksPending;
+ const processingRecording =
+ finishRequested || isCommitting || recordedBlocksPending;
const prevIsRecordingRef = useRef(false);
const prevProcessingRef = useRef(false);
useEffect(() => {
@@ -875,26 +924,40 @@ function StudioStage(props: StudioWorkspaceProps) {
const wasProcessing = prevProcessingRef.current;
if (processingRecording && !wasProcessing) {
restoreExpandedPane();
- openPane("editor");
+ if (finishRequested || isCommitting || wasRecording) {
+ setOpenPanes(panesAfterRecordingTransition(panes, "processing"));
+ } else {
+ // Other block imports share this loading state but are not recordings.
+ openPane("editor");
+ }
} else if (isRecording && !wasRecording) {
- // Recording or finalizing → live browser + drafts.
+ // Recording → live browser + Copilot guidance.
restoreExpandedPane();
- openPane("copilot");
- openPane("browser");
+ setOpenPanes(panesAfterRecordingTransition(panes, "started"));
} else if (!isRecording && wasRecording && !processingRecording) {
- // Recording ended (commit or discard) → back to the canvas.
+ // Discarded recording → restore the canvas without closing Browser.
restoreExpandedPane();
- openPane("editor");
+ setOpenPanes(panesAfterRecordingTransition(panes, "ended"));
}
prevIsRecordingRef.current = isRecording;
prevProcessingRef.current = processingRecording;
- }, [isRecording, processingRecording, openPane, restoreExpandedPane]);
+ }, [
+ isRecording,
+ finishRequested,
+ isCommitting,
+ openPane,
+ panes,
+ processingRecording,
+ restoreExpandedPane,
+ setOpenPanes,
+ ]);
const paneProps = (id: StudioPaneId) => {
const index = panes.indexOf(id);
return {
id,
runId,
+ chromeless: embedded,
open: index >= 0 && visiblePanes.includes(id),
// Panes take even slots and the dividers between them take odd slots.
order: index >= 0 ? index * 2 : undefined,
@@ -937,7 +1000,7 @@ function StudioStage(props: StudioWorkspaceProps) {
only affects Radix tooltips inside the shell. */}
<TooltipProvider delayDuration={200} skipDelayDuration={300}>
<div className="flex h-full w-full flex-col">
- <StudioTopBar />
+ {embedded ? null : <StudioTopBar />}
<div className="flex min-h-0 min-w-0 flex-1">
{/* Panes keep a fixed DOM order (stable mounts for the canvas, chat and
stream slots); the CSS order carries the layout order instead, so
@@ -946,7 +1009,10 @@ function StudioStage(props: StudioWorkspaceProps) {
the stream singleton never remount. */}
<div
ref={registerStageElement}
- className="relative flex min-h-0 min-w-0 flex-1 overflow-hidden p-3"
+ className={cn(
+ "relative flex min-h-0 min-w-0 flex-1 overflow-hidden",
+ embedded ? "p-0" : "p-3",
+ )}
>
<StudioPane
{...paneProps("copilot")}
@@ -980,10 +1046,10 @@ function StudioStage(props: StudioWorkspaceProps) {
headerExtras={<EditorPaneModeToggle />}
headerActions={<EditorPaneBlockSearch />}
>
- {/* The embedded Workspace boots debug sessions, copilot chats
- and block-script queries against the workflow — none of
- which exist once the agent is deleted, so it never mounts. */}
- {workflowDeleted ? (
+ {/* Embedded run views never mount the authoring canvas: aside
+ from doing unnecessary work, its data-tour targets can
+ launch the editor onboarding overlay over the iframe. */}
+ {embedded ? null : workflowDeleted ? (
<WorkflowDeletedPaneNotice />
) : (
<EditorTab {...props} />
@@ -997,6 +1063,14 @@ function StudioStage(props: StudioWorkspaceProps) {
headerActions={<BrowserPaneActions />}
>
<BrowserTab />
+ {embedded ? (
+ <EmbeddedBrowserOverlays
+ runId={runId}
+ showArchivedRecording={
+ searchParams.get("view") === "recording"
+ }
+ />
+ ) : null}
</StudioPane>
<StudioPane
{...paneProps("overview")}
@@ -1022,9 +1096,9 @@ function StudioStage(props: StudioWorkspaceProps) {
/>
))
: null}
- {panes.length === 0 ? <StudioStageLauncher /> : null}
- <StudioCoachMark />
- <StudioWorkflowPanels />
+ {!embedded && panes.length === 0 ? <StudioStageLauncher /> : null}
+ {embedded ? null : <StudioCoachMark />}
+ {embedded ? null : <StudioWorkflowPanels />}
{/* Overlay target for Workspace-wired panels (pointer-events off
while empty so the stage stays clickable). */}
<div
diff --git a/skyvern-frontend/src/routes/workflows/studio/recordingPaneLifecycle.ts b/skyvern-frontend/src/routes/workflows/studio/recordingPaneLifecycle.ts
new file mode 100644
index 000000000..a73bdccab
--- /dev/null
+++ b/skyvern-frontend/src/routes/workflows/studio/recordingPaneLifecycle.ts
@@ -0,0 +1,28 @@
+import { type StudioPaneId } from "./panes";
+
+type RecordingPaneTransition = "started" | "processing" | "ended";
+
+export function panesAfterRecordingTransition(
+ panes: readonly StudioPaneId[],
+ transition: RecordingPaneTransition,
+): StudioPaneId[] {
+ const next = panes.filter((pane) => {
+ if (transition === "started") {
+ return pane !== "editor";
+ }
+ if (transition === "processing") {
+ return pane !== "browser";
+ }
+ return true;
+ });
+ const requiredPanes: StudioPaneId[] =
+ transition === "started" ? ["copilot", "browser"] : ["editor"];
+
+ for (const pane of requiredPanes) {
+ if (!next.includes(pane)) {
+ next.push(pane);
+ }
+ }
+
+ return next;
+}
diff --git a/skyvern-frontend/src/routes/workflows/studio/runSwitchNavigation.ts b/skyvern-frontend/src/routes/workflows/studio/runSwitchNavigation.ts
index 98e6f09b6..748785350 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runSwitchNavigation.ts
+++ b/skyvern-frontend/src/routes/workflows/studio/runSwitchNavigation.ts
@@ -38,9 +38,9 @@ export function searchWithRunSwitched(
}
params.delete("active");
params.delete("bl");
- // A loop-iteration scope belongs to the run being left (WorkflowRun.tsx
- // forwards it into studio URLs); inert in studio today, but clearing it keeps
- // this the single, complete home for run-switch navigation.
+ // A loop-iteration scope belongs to the run being left; inert in studio
+ // today, but clearing it keeps this the single, complete home for run-switch
+ // navigation.
params.delete("iteration");
return toReadableSearch(params);
}
diff --git a/skyvern-frontend/src/routes/workflows/studio/runview/HeroScreenshot.tsx b/skyvern-frontend/src/routes/workflows/studio/runview/HeroScreenshot.tsx
index 4800379aa..a5f85acc9 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runview/HeroScreenshot.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/runview/HeroScreenshot.tsx
@@ -30,9 +30,9 @@ export type HeroSelection =
/**
* The selected element's screenshot, fit to the run-hero width and scrollable for
- * long captures. Mirrors the legacy run page: an action shows its own post-action
- * screenshot (by artifact id, falling back to the step's action screenshots), a
- * block shows its representative screenshot via `selectBlockScreenshot`.
+ * long captures. An action shows its own post-action screenshot (by artifact id,
+ * falling back to the step's action screenshots), and a block shows its
+ * representative screenshot via `selectBlockScreenshot`.
*/
export function HeroScreenshot({
selection,
diff --git a/skyvern-frontend/src/routes/workflows/studio/runview/RunPaneHeader.tsx b/skyvern-frontend/src/routes/workflows/studio/runview/RunPaneHeader.tsx
index de82e842e..7c7735541 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runview/RunPaneHeader.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/runview/RunPaneHeader.tsx
@@ -4,6 +4,7 @@ import {
runIsLogicallyFinal,
} from "@/routes/workflows/workflowRun/runRetryState";
import { useCallback, useState } from "react";
+import { useLocation, useNavigate } from "react-router-dom";
import {
ActivityLogIcon,
CodeIcon,
@@ -41,11 +42,17 @@ import { cn } from "@/util/utils";
import { useIsGeneratingCode } from "../../editor/hooks/useIsGeneratingCode";
import { constructCacheKeyValue } from "../../editor/utils";
import { useWorkflowRunWithWorkflowQuery } from "../../hooks/useWorkflowRunWithWorkflowQuery";
+import { WorkflowRunStatusAlert } from "../../workflowRun/WorkflowRunStatusAlert";
import { studioPanelId } from "../constants";
import { useStudioPaneCompact } from "../StudioShellContext";
import { useStudioInspectedRun } from "../useStudioInspectedRun";
import { useStudioPanes } from "../useStudioPanes";
import { ViewToggle } from "../ViewToggle";
+import {
+ parsePanesParam,
+ STUDIO_PANES_PARAM,
+ toReadableSearch,
+} from "../panes";
/**
* Left header cluster of the Overview pane: the Timeline / Inputs / Outputs
@@ -61,9 +68,11 @@ export function RunPaneViewToggles() {
useWorkflowRunWithWorkflowQuery({ workflowRunId: runId });
const view = useRunPaneViewStore((s) => s.view);
const setView = useRunPaneViewStore((s) => s.setView);
+ const location = useLocation();
+ const navigate = useNavigate();
const jumpToLive = useRunViewStore((s) => s.jumpToLive);
const setBrowserPaneView = useStudioBrowserStore((s) => s.setView);
- const { openPane } = useStudioPanes();
+ const { openPane, resolveLivePanes } = useStudioPanes();
const cacheKey = workflowRun?.workflow?.cache_key ?? "";
const codeGenerating = useIsGeneratingCode({
cacheKey,
@@ -72,13 +81,27 @@ export function RunPaneViewToggles() {
workflow: workflowRun?.workflow,
workflowRun,
}),
- // Same guard as the legacy run page: script queries 404 once the source
- // agent is deleted.
+ // Script queries 404 once the source agent is deleted.
workflowPermanentId: workflowRun?.workflow?.deleted_at
? undefined
: workflowPermanentId,
workflowRunId: runId,
});
+ const selectView = useCallback(
+ (nextView: Parameters<typeof setView>[0]) => {
+ setView(nextView);
+ const next = new URLSearchParams(location.search);
+ next.set("view", nextView);
+ const panes =
+ parsePanesParam(next.get(STUDIO_PANES_PARAM)) ?? resolveLivePanes();
+ next.set(
+ STUDIO_PANES_PARAM,
+ ["overview", ...panes.filter((pane) => pane !== "overview")].join(","),
+ );
+ navigate({ search: toReadableSearch(next) }, { replace: true });
+ },
+ [location.search, navigate, resolveLivePanes, setView],
+ );
const focusBrowserPane = useCallback(() => {
// An explicit "watch live": unpin back to the live edge and hand the
@@ -117,21 +140,21 @@ export function RunPaneViewToggles() {
>
<ViewToggle
active={view === "timeline"}
- onClick={() => setView("timeline")}
+ onClick={() => selectView("timeline")}
compact={compact}
label="Timeline"
icon={<ActivityLogIcon className="h-3 w-3" />}
/>
<ViewToggle
active={view === "inputs"}
- onClick={() => setView("inputs")}
+ onClick={() => selectView("inputs")}
compact={compact}
label="Inputs"
icon={<ListBulletIcon className="h-3 w-3" />}
/>
<ViewToggle
active={view === "outputs"}
- onClick={() => setView("outputs")}
+ onClick={() => selectView("outputs")}
compact={compact}
label="Outputs"
icon={<FileTextIcon className="h-3 w-3" />}
@@ -173,7 +196,7 @@ export function RunPaneViewToggles() {
menu reads as part of the toggle cluster, not a full-size menu. */}
<DropdownMenuContent align="start" sideOffset={6} className="min-w-0">
<DropdownMenuItem
- onSelect={() => setView("code")}
+ onSelect={() => selectView("code")}
className="cursor-pointer gap-1.5 rounded px-2 py-1.5 pr-3 text-xs font-medium text-muted-foreground focus:text-foreground"
>
<CodeIcon className="h-3 w-3" />
@@ -222,21 +245,28 @@ export function RunPaneActions() {
const compact = useStudioPaneCompact();
const apiCredential = useApiCredential();
const { runId } = useStudioInspectedRun();
- const { data: workflowRun } = useWorkflowRunWithWorkflowQuery({
- workflowRunId: runId,
- });
+ const { data: workflowRun, isError: statusUnavailable } =
+ useWorkflowRunWithWorkflowQuery({ workflowRunId: runId });
const [replayOpen, setReplayOpen] = useState(false);
if (!workflowRun) {
return null;
}
+ const finalized = runIsLogicallyFinal(workflowRun);
+ const statusAlert = (
+ <WorkflowRunStatusAlert
+ run={workflowRun}
+ title={workflowRun.workflow?.title}
+ visible={!statusUnavailable && !finalized}
+ />
+ );
// Legacy parity: the menu re-runs the workflow via API, which is gone with
- // the source agent.
+ // the source agent. Status notifications still apply to its in-flight run.
if (workflowRun.workflow?.deleted_at) {
- return null;
+ return statusAlert;
}
- const finalized = runIsLogicallyFinal(workflowRun);
return (
<>
+ {statusAlert}
<ApiWebhookActionsMenu
triggerTooltip={compact ? "API & Webhooks" : undefined}
trigger={
diff --git a/skyvern-frontend/src/routes/workflows/studio/runview/RunView.test.tsx b/skyvern-frontend/src/routes/workflows/studio/runview/RunView.test.tsx
index eedd2adc4..5fb12b233 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runview/RunView.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/runview/RunView.test.tsx
@@ -43,17 +43,32 @@ import type {
} from "../../types/workflowRunTypes";
import { WorkflowCopilotChat } from "../../copilot/WorkflowCopilotChat";
import { StudioPaneCompactContext } from "../StudioShellContext";
-import { RunPaneViewToggles } from "./RunPaneHeader";
+import { RunPaneActions, RunPaneViewToggles } from "./RunPaneHeader";
import capturedRun from "../__fixtures__/completed-code-block-run.json";
import { RunView } from "./RunView";
const mocks = vi.hoisted(() => ({
workflowRun: undefined as unknown,
timeline: undefined as unknown,
+ fallbackEpisodes: undefined as unknown,
+ blockScripts: undefined as unknown,
+ onboarding: null as null | {
+ recoveryGuidanceAssignment: {
+ organization_id: string;
+ experiment_version: string;
+ arm: "control" | "treatment";
+ eligible_run_id: string;
+ } | null;
+ },
+ recoverySurfaceEnabled: false,
codeGenerating: false,
isPlaceholderData: false,
statusUnavailable: false,
refetchRunStatus: vi.fn(),
+ refreshOnboarding: vi.fn(),
+ fallbackEpisodesQuery: vi.fn(),
+ blockScriptsQuery: vi.fn(),
+ posthogCapture: vi.fn(),
}));
const { getSpy } = vi.hoisted(() => ({ getSpy: vi.fn() }));
@@ -80,6 +95,43 @@ vi.mock("../../hooks/useWorkflowRunTimelineQuery", () => ({
isPlaceholderData: mocks.isPlaceholderData,
}),
}));
+vi.mock("../../hooks/useRefreshOnboardingOnRunCompletion", () => ({
+ useRefreshOnboardingOnRunCompletion: (run: unknown) =>
+ mocks.refreshOnboarding(run),
+}));
+vi.mock("../../hooks/useFallbackEpisodesQuery", () => ({
+ useFallbackEpisodesQuery: (options: unknown) => {
+ mocks.fallbackEpisodesQuery(options);
+ return { data: mocks.fallbackEpisodes };
+ },
+}));
+vi.mock("../../hooks/useBlockScriptsQuery", () => ({
+ useBlockScriptsQuery: (options: unknown) => {
+ mocks.blockScriptsQuery(options);
+ return { data: mocks.blockScripts };
+ },
+}));
+vi.mock("@/store/onboarding/useOnboardingState", () => ({
+ useOnboardingStateOptional: () => mocks.onboarding,
+}));
+vi.mock("@/hooks/useFeatureFlag", () => ({
+ useFeatureFlag: () => mocks.recoverySurfaceEnabled,
+}));
+vi.mock("../../workflowRun/ScriptUpdateCard", () => ({
+ ScriptUpdateCard: ({
+ episodes,
+ scriptId,
+ }: {
+ episodes: unknown[];
+ scriptId?: string | null;
+ }) => (
+ <div
+ data-testid="script-update-card"
+ data-episode-count={episodes.length}
+ data-script-id={scriptId}
+ />
+ ),
+}));
vi.mock("../../editor/hooks/useIsGeneratingCode", () => ({
useIsGeneratingCode: () => mocks.codeGenerating,
}));
@@ -115,6 +167,9 @@ if (typeof globalThis.ResizeObserver === "undefined") {
vi.mock("posthog-js/react", () => ({
usePostHog: () => ({ capture: vi.fn() }),
}));
+vi.mock("posthog-js", () => ({
+ default: { capture: (...args: unknown[]) => mocks.posthogCapture(...args) },
+}));
// The header toggles resolve the inspected run themselves; pin it to the same
// run the RunView under test renders (avoids the latest-run fallback query).
vi.mock("../useStudioInspectedRun", () => ({
@@ -295,7 +350,13 @@ function seedRunningRun() {
function LocationSpy() {
const location = useLocation();
- return <div data-testid="location-search">{location.search}</div>;
+ return (
+ <>
+ <div data-testid="location-pathname">{location.pathname}</div>
+ <div data-testid="location-search">{location.search}</div>
+ <div data-testid="location-state">{JSON.stringify(location.state)}</div>
+ </>
+ );
}
// Subscribes to the failure card's artifact query so a test can wait for its
@@ -345,9 +406,39 @@ function ReleaseCopilotRunFocus() {
);
}
+function SameRunSelectionNavigation() {
+ const [params, setParams] = useSearchParams();
+ return (
+ <>
+ <button
+ onClick={() => {
+ const next = new URLSearchParams(params);
+ next.set("active", "wrb_second");
+ next.set("iteration", "2");
+ setParams(next);
+ }}
+ >
+ select second frame in URL
+ </button>
+ <button
+ onClick={() => {
+ const next = new URLSearchParams(params);
+ next.delete("active");
+ next.delete("iteration");
+ setParams(next);
+ }}
+ >
+ clear frame from URL
+ </button>
+ </>
+ );
+}
+
function renderRunView(
props: Partial<Parameters<typeof RunView>[0]> = {},
- initialEntry = "/",
+ initialEntry:
+ | string
+ | { pathname: string; search?: string; state?: unknown } = "/",
compact = false,
extra?: ReactNode,
pageSlots: PageSlots = {},
@@ -369,7 +460,10 @@ function renderRunView(
<StudioPaneCompactContext.Provider value={compact}>
<RunPaneViewToggles />
</StudioPaneCompactContext.Provider>
- {initialEntry.startsWith("/runs/") ? (
+ {(typeof initialEntry === "string"
+ ? initialEntry
+ : initialEntry.pathname
+ ).startsWith("/runs/") ? (
<Routes>
<Route
path="/runs/:runId"
@@ -394,10 +488,18 @@ afterEach(() => {
cleanup();
mocks.workflowRun = undefined;
mocks.timeline = undefined;
+ mocks.fallbackEpisodes = undefined;
+ mocks.blockScripts = undefined;
+ mocks.onboarding = null;
+ mocks.recoverySurfaceEnabled = false;
mocks.codeGenerating = false;
mocks.isPlaceholderData = false;
mocks.statusUnavailable = false;
mocks.refetchRunStatus.mockReset();
+ mocks.refreshOnboarding.mockReset();
+ mocks.fallbackEpisodesQuery.mockReset();
+ mocks.blockScriptsQuery.mockReset();
+ mocks.posthogCapture.mockReset();
});
beforeEach(() => {
getSpy.mockReset();
@@ -407,7 +509,234 @@ beforeEach(() => {
useStudioBrowserStore.setState({ view: "auto" });
});
+describe("RunView legacy run-page parity", () => {
+ test("refreshes onboarding and carries recovery context into a retry", async () => {
+ seedCompletedRun({
+ status: Status.Failed,
+ failure_reason: "The network request timed out",
+ failure_category: [{ category: "NETWORK_TIMEOUT" }],
+ workflow: {
+ workflow_permanent_id: "wpid_1",
+ cache_key: "cache-key",
+ deleted_at: null,
+ workflow_definition: { blocks: [], finally_block_label: null },
+ },
+ });
+ mocks.onboarding = {
+ recoveryGuidanceAssignment: {
+ organization_id: "org_1",
+ experiment_version: "recovery-v1",
+ arm: "treatment",
+ eligible_run_id: "wr_1",
+ },
+ };
+ mocks.recoverySurfaceEnabled = true;
+
+ renderRunView({ onRetry: vi.fn() }, "/?panes=overview");
+
+ expect(mocks.refreshOnboarding).toHaveBeenCalledWith(mocks.workflowRun);
+ expect(
+ await screen.findByTestId("first-run-recovery-guidance"),
+ ).not.toBeNull();
+
+ fireEvent.click(screen.getByRole("button", { name: "Retry run" }));
+
+ await waitFor(() =>
+ expect(screen.getByTestId("location-pathname").textContent).toBe(
+ "/agents/wpid_1/run",
+ ),
+ );
+ expect(screen.getByTestId("location-state").textContent).toContain(
+ '"recoveryGuidanceRetry":{"organizationId":"org_1","experimentVersion":"recovery-v1","arm":"treatment","eligibleRunId":"wr_1","failureCategory":"NETWORK_TIMEOUT"}',
+ );
+ });
+
+ test("records retry_started after the recovery retry begins", async () => {
+ seedCompletedRun({
+ workflow_run_id: "wr_retry",
+ status: Status.Running,
+ started_at: "2026-09-15T20:00:00Z",
+ });
+
+ renderRunView(
+ {},
+ {
+ pathname: "/runs/wr_retry",
+ state: {
+ recoveryGuidanceRetry: {
+ organizationId: "org_1",
+ experimentVersion: "recovery-v1",
+ arm: "treatment",
+ eligibleRunId: "wr_1",
+ failureCategory: "NETWORK_TIMEOUT",
+ retryRunId: "wr_retry",
+ },
+ },
+ },
+ );
+
+ await waitFor(() =>
+ expect(mocks.posthogCapture).toHaveBeenCalledWith(
+ "retry_started",
+ expect.objectContaining({
+ eligible_run_id: "wr_1",
+ retry_run_id: "wr_retry",
+ }),
+ ),
+ );
+ });
+
+ test("shows the fallback script update with the published script id", () => {
+ seedCompletedRun({
+ workflow: {
+ workflow_permanent_id: "wpid_1",
+ cache_key: "cache-key",
+ deleted_at: null,
+ workflow_definition: { blocks: [], finally_block_label: null },
+ },
+ });
+ mocks.fallbackEpisodes = {
+ episodes: [
+ {
+ episode_id: "episode_1",
+ new_script_revision_id: "revision_2",
+ },
+ ],
+ };
+ mocks.blockScripts = { script_id: "script_1" };
+
+ renderRunView({}, "/?panes=overview");
+
+ const scriptUpdateCard = screen.getByTestId("script-update-card");
+ expect(scriptUpdateCard.getAttribute("data-episode-count")).toBe("1");
+ expect(scriptUpdateCard.getAttribute("data-script-id")).toBe("script_1");
+ expect(mocks.fallbackEpisodesQuery).toHaveBeenCalledWith({
+ workflowPermanentId: "wpid_1",
+ workflowRunId: "wr_1",
+ enabled: true,
+ });
+ });
+});
+
describe("RunView view toggles", () => {
+ test("keeps status notification controls for an active run", () => {
+ seedRunningRun();
+ renderRunView(
+ {},
+ "/?wr=wr_1",
+ false,
+ <TooltipProvider delayDuration={0}>
+ <RunPaneActions />
+ </TooltipProvider>,
+ );
+
+ expect(screen.getByRole("button", { name: /Notify/ })).toBeTruthy();
+ });
+
+ test("links to the runs before and after a retry", () => {
+ seedCompletedRun({
+ retried_from_workflow_run_id: "wr_previous",
+ retried_by_workflow_run_id: "wr_next",
+ });
+ renderRunView({}, "/?wr=wr_1");
+
+ expect(
+ screen
+ .getByRole("link", { name: "Retried from wr_previous" })
+ .getAttribute("href"),
+ ).toBe("/runs/wr_previous");
+ expect(
+ screen
+ .getByRole("link", { name: "Retried by wr_next" })
+ .getAttribute("href"),
+ ).toBe("/runs/wr_next");
+ });
+
+ test("opens and replaces the run subview requested by a legacy deep link", async () => {
+ seedCompletedRun();
+ const { container } = renderRunView(
+ {},
+ "/?panes=overview&view=code&active=wrb_loop&iteration=2",
+ );
+
+ expect(within(container).queryByTestId("workflow-run-code")).not.toBeNull();
+ expect(useRunPaneViewStore.getState().view).toBe("code");
+ expect(useRunViewStore.getState().pinnedFrameId).toBe("wrb_loop");
+ expect(useRunViewStore.getState().activeIteration).toBe(2);
+
+ fireEvent.click(screen.getByRole("button", { name: "Inputs" }));
+ await waitFor(() => {
+ expect(screen.getByTestId("location-search").textContent).toContain(
+ "view=inputs",
+ );
+ });
+
+ fireEvent.click(screen.getByRole("button", { name: "Timeline" }));
+ await waitFor(() => {
+ expect(screen.getByTestId("location-search").textContent).toContain(
+ "view=timeline",
+ );
+ });
+ });
+
+ test("reapplies the requested view when switching runs", () => {
+ seedCompletedRun();
+ const props: Partial<Parameters<typeof RunView>[0]> = {
+ workflowRunId: "wr_1",
+ };
+ const view = renderRunView(props, "/?view=inputs");
+ expect(useRunPaneViewStore.getState().view).toBe("inputs");
+
+ props.workflowRunId = "wr_2";
+ view.rerenderRunView();
+
+ expect(useRunPaneViewStore.getState().view).toBe("inputs");
+ });
+
+ test("rehydrates frame selection when same-run URL parameters change", async () => {
+ seedCompletedRun();
+ renderRunView(
+ {},
+ "/?wr=wr_1&active=wrb_first&iteration=1",
+ false,
+ <SameRunSelectionNavigation />,
+ );
+
+ await waitFor(() => {
+ expect(useRunViewStore.getState().pinnedFrameId).toBe("wrb_first");
+ expect(useRunViewStore.getState().activeIteration).toBe(1);
+ });
+
+ fireEvent.click(
+ screen.getByRole("button", { name: "select second frame in URL" }),
+ );
+ await waitFor(() => {
+ expect(useRunViewStore.getState().pinnedFrameId).toBe("wrb_second");
+ expect(useRunViewStore.getState().activeIteration).toBe(2);
+ });
+
+ fireEvent.click(
+ screen.getByRole("button", { name: "clear frame from URL" }),
+ );
+ await waitFor(() => {
+ expect(useRunViewStore.getState().pinnedFrameId).toBeNull();
+ expect(useRunViewStore.getState().activeIteration).toBeNull();
+ });
+ });
+
+ test("writes the required Overview pane when selecting a run view", async () => {
+ seedCompletedRun();
+ renderRunView({}, "/?wr=wr_1");
+
+ fireEvent.click(screen.getByRole("button", { name: "Inputs" }));
+
+ await waitFor(() => {
+ expect(screen.getByTestId("location-search").textContent).toContain(
+ "panes=overview,browser",
+ );
+ });
+ });
+
test("mounts the completion milestone only after a pending retry is final", () => {
seedCompletedRun({ retry_pending: true });
const MilestoneCard = vi.fn(() => <div data-testid="milestone-card" />);
@@ -459,10 +788,19 @@ describe("RunView view toggles", () => {
}
});
- test("does not render run tags in Studio Overview", () => {
+ test("keeps workflow run tags reachable in Studio Overview", () => {
seedCompletedRun();
const { queryByTestId } = renderRunView();
+ expect(queryByTestId("run-tags-editor")?.dataset.workflowRunId).toBe(
+ "wr_1",
+ );
+ });
+
+ test("keeps workflow run tags out of embedded surfaces", () => {
+ seedCompletedRun();
+ const { queryByTestId } = renderRunView({}, "/?embed=true");
+
expect(queryByTestId("run-tags-editor")).toBeNull();
});
@@ -1101,6 +1439,21 @@ describe("RunView failure presentation", () => {
expect(within(container).queryByRole("alert")).toBeNull();
});
+ test("keeps mutation controls out of embedded failure views", () => {
+ seedCompletedRun({
+ status: Status.Failed,
+ failure_reason: "Login page rejected the credentials",
+ });
+ const { getByTestId } = renderRunView(
+ { onFix: vi.fn(), onRetry: vi.fn() },
+ "/?embed=true",
+ );
+ const line = within(getByTestId("run-failure-line"));
+
+ expect(line.queryByRole("button", { name: "Fix with Copilot" })).toBeNull();
+ expect(line.queryByRole("button", { name: "Retry" })).toBeNull();
+ });
+
test("Fix passes the failing block's label", () => {
seedCompletedRun({
status: Status.Failed,
@@ -1345,6 +1698,46 @@ describe("RunView failure presentation", () => {
);
});
+ test("the embedded screenshot action replaces Overview with Browser", async () => {
+ getSpy.mockResolvedValue({
+ data: [
+ {
+ artifact_id: "art_screenshot",
+ artifact_type: ArtifactType.ActionScreenshot,
+ created_at: "2026-08-27T00:00:00Z",
+ modified_at: "2026-08-27T00:00:00Z",
+ organization_id: "org_1",
+ task_id: "task_1",
+ step_id: "step_1",
+ uri: "s3://bucket/screenshot.png",
+ },
+ ],
+ });
+ seedFailedCodeRun(
+ "CodeBlock failed because a browser operation failed at line 4.",
+ "browser_operation_failed",
+ );
+ const { container, getByTestId } = renderRunView(
+ {},
+ "/?wr=wr_1&embed=true&panes=overview",
+ );
+
+ fireEvent.click(
+ await within(container).findByRole("button", {
+ name: "View block screenshot",
+ }),
+ );
+
+ await waitFor(() =>
+ expect(getByTestId("location-search").textContent).toContain(
+ "panes=browser",
+ ),
+ );
+ expect(getByTestId("location-search").textContent).not.toContain(
+ "panes=overview",
+ );
+ });
+
test("hides the screenshot link when the failed block has no capture", async () => {
getSpy.mockResolvedValue({ data: [] });
seedCompletedRun({
@@ -1654,7 +2047,7 @@ describe("RunView live affordances", () => {
});
describe("RunView iteration selection", () => {
- test("selecting the loop block after an iteration clears the iteration scope", () => {
+ test("selecting the loop block after an iteration clears the iteration scope", async () => {
seedForLoopRun();
// Seed ?active= so the loop is the selected (and expanded) item on mount,
// making its iteration rows visible.
@@ -1676,6 +2069,11 @@ describe("RunView iteration selection", () => {
expect(scope.queryByText("alpha")).toBeNull();
// The iteration scope is shared with the Browser pane via the store.
expect(useRunViewStore.getState().activeIteration).toBe(1);
+ await waitFor(() => {
+ expect(screen.getByTestId("location-search").textContent).toBe(
+ "?active=wrb_loop&wr=wr_1&iteration=1",
+ );
+ });
// Click the loop block row (descriptor text is timeline-only). The detail
// must fall back to the loop's own iteration instead of staying on 2.
@@ -1683,6 +2081,11 @@ describe("RunView iteration selection", () => {
expect(headerMeta()).toContain("Iteration 1");
expect(scope.getByText("alpha")).not.toBeNull();
expect(useRunViewStore.getState().activeIteration).toBeNull();
+ await waitFor(() => {
+ expect(screen.getByTestId("location-search").textContent).toBe(
+ "?active=wrb_loop&wr=wr_1",
+ );
+ });
}, 20_000);
});
@@ -2083,6 +2486,7 @@ describe("finished-run landing", () => {
`/?wr=wr_1&active=${capturedRun.blocks[1]!.actions[0]!.action_id}`,
],
["?bl=", "/?wr=wr_1&bl=code_block_2"],
+ ["?view=", "/?wr=wr_1&view=timeline"],
])(
"a completed run deep-linked with %s stays on the timeline",
async (_label, route) => {
diff --git a/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx b/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx
index 7b4c5f4e0..60c066083 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx
@@ -3,9 +3,15 @@ import {
runIsLogicallyFinal,
runIsRetryWaiting,
} from "@/routes/workflows/workflowRun/runRetryState";
-import { useEffect, useMemo, useRef, useState } from "react";
+import { useCallback, useEffect, useMemo, useRef, useState } from "react";
import { MagnifyingGlassIcon } from "@radix-ui/react-icons";
-import { useNavigate, useParams, useSearchParams } from "react-router-dom";
+import {
+ Link,
+ useLocation,
+ useNavigate,
+ useParams,
+ useSearchParams,
+} from "react-router-dom";
import { Status } from "@/api/types";
import {
@@ -26,7 +32,10 @@ import {
SELECTED_BLOCK_SEARCH_PARAM,
SYSTEM_BLOCK_FOCUS_PARAM,
} from "@/routes/workflows/editor/hooks/useSelectedBlockUrlSync";
-import { useRunPaneViewStore } from "@/store/useRunPaneViewStore";
+import {
+ type RunPaneView,
+ useRunPaneViewStore,
+} from "@/store/useRunPaneViewStore";
import { useRunViewStore } from "@/store/RunViewStore";
import { useStudioBrowserStore } from "@/store/useStudioBrowserStore";
import { useWorkflowBlockSearchStore } from "@/store/WorkflowBlockSearchStore";
@@ -35,12 +44,30 @@ import {
usePageSlots,
} from "@/store/PageSlots";
import { isRecord } from "@/util/utils";
+import {
+ FirstRunRecoveryGuidance,
+ shouldShowRecoveryGuidance,
+} from "@/components/onboarding/FirstRunRecoveryGuidance";
+import { useFeatureFlag } from "@/hooks/useFeatureFlag";
+import { RunTagsEditor } from "@/routes/tasks/components/tagging/RunTagsEditor";
+import { useOnboardingStateOptional } from "@/store/onboarding/useOnboardingState";
+import {
+ getRecoveryGuidanceRetryNavigation,
+ RecoveryGuidanceTelemetry,
+ retryRunHasStarted,
+ type RecoveryGuidanceTelemetryContext,
+} from "@/util/onboarding/recoveryGuidanceTelemetry";
+import { constructCacheKeyValue } from "../../editor/utils";
+import { useBlockScriptsQuery } from "../../hooks/useBlockScriptsQuery";
+import { useFallbackEpisodesQuery } from "../../hooks/useFallbackEpisodesQuery";
+import { useRefreshOnboardingOnRunCompletion } from "../../hooks/useRefreshOnboardingOnRunCompletion";
import { useWorkflowRunTimelineQuery } from "../../hooks/useWorkflowRunTimelineQuery";
import { useWorkflowRunWithWorkflowQuery } from "../../hooks/useWorkflowRunWithWorkflowQuery";
import { ResizableTimelineSplit } from "../../workflowRun/ResizableTimelineSplit";
import { WorkflowRunBlockDetail } from "../../workflowRun/WorkflowRunBlockDetail";
import { WorkflowRunCode } from "../../workflowRun/WorkflowRunCode";
+import { ScriptUpdateCard } from "../../workflowRun/ScriptUpdateCard";
import { WorkflowRunTimeline } from "../../workflowRun/WorkflowRunTimeline";
import { WorkflowRunVerificationCodeForm } from "../../workflowRun/WorkflowRunVerificationCodeForm";
import { findRunCodeBlockFailure } from "../../workflowRun/codeBlockFailure";
@@ -51,9 +78,10 @@ import {
filterTimelineToAttempt,
findActiveItem,
flattenTimelineChronologically,
+ parseActiveIterationParam,
type TimelineSearchTarget,
} from "../../workflowRun/workflowTimelineUtils";
-import { getOrderedRunParameters } from "../../utils";
+import { getOrderedRunParameters, getRerunNavigationState } from "../../utils";
import {
buildFilmstrip,
ELAPSED_NEVER_STARTED,
@@ -88,6 +116,18 @@ import {
import { resolveEditorSelectionPin } from "./editorSelectionPin";
import { resolveTimelineBlockJumpNodeId } from "./timelineBlockJump";
+const RECOVERY_GUIDANCE_TREATMENT_SURFACE_FLAG =
+ "RECOVERY_GUIDANCE_TREATMENT_SURFACE";
+
+function isRunPaneView(value: string | null): value is RunPaneView {
+ return (
+ value === "timeline" ||
+ value === "inputs" ||
+ value === "outputs" ||
+ value === "code"
+ );
+}
+
type RunViewProps = {
workflowRunId?: string;
// The caller is still resolving which run to show; keep the placeholder in its
@@ -232,6 +272,9 @@ export function RunView({
isPlaceholderData: runIsPlaceholder,
isError: statusUnavailable,
} = useWorkflowRunWithWorkflowQuery(queryOptions);
+ useRefreshOnboardingOnRunCompletion(
+ runIsPlaceholder ? undefined : workflowRun,
+ );
const { data: retainedTimeline, isPlaceholderData: timelineIsPlaceholder } =
useWorkflowRunTimelineQuery(queryOptions);
// The timeline payload carries no run id of its own, so keepPreviousData serves
@@ -252,38 +295,193 @@ export function RunView({
const pinnedFrameId = useRunViewStore((s) => s.pinnedFrameId);
const activeIteration = useRunViewStore((s) => s.activeIteration);
const pinFrame = useRunViewStore((s) => s.pinFrame);
+ const jumpToLive = useRunViewStore((s) => s.jumpToLive);
const resetRunView = useRunViewStore((s) => s.reset);
- const { panes: studioPanes, openPane } = useStudioPanes();
+ const { panes: studioPanes, openPane, setOpenPanes } = useStudioPanes();
const runPaneOpen = studioPanes.includes("overview");
const navigate = useNavigate();
+ const location = useLocation();
const [searchParams] = useSearchParams();
+ const embedded = searchParams.get("embed") === "true";
+ const requestedView = searchParams.get("view");
+ const requestedActive = searchParams.get("active");
+ const requestedIteration = requestedActive
+ ? parseActiveIterationParam(searchParams.get("iteration"))
+ : null;
const searchParamsRef = useRef(searchParams);
searchParamsRef.current = searchParams;
const view = useRunPaneViewStore((s) => s.view);
const setPaneView = useRunPaneViewStore((s) => s.setView);
const resetPaneView = useRunPaneViewStore((s) => s.reset);
const [outputSummary, setOutputSummary] = useState<string | null>(null);
+ const finalized =
+ !statusUnavailable && workflowRun
+ ? runIsLogicallyFinal(workflowRun)
+ : false;
+ const workflow = workflowRun?.workflow;
+ const workflowPermanentId = workflow?.workflow_permanent_id;
+ const isWorkflowDeleted = Boolean(workflow?.deleted_at);
+ const cacheKey = workflow?.cache_key ?? "";
+ const cacheKeyValue =
+ constructCacheKeyValue({ codeKey: cacheKey, workflow, workflowRun }) ?? "";
+ const { data: blockScriptsPublished } = useBlockScriptsQuery({
+ cacheKey,
+ cacheKeyValue,
+ enabled: finalized && !isWorkflowDeleted,
+ workflowPermanentId,
+ status: "published",
+ workflowRunId: workflowRun?.workflow_run_id,
+ });
+ const { data: fallbackEpisodes } = useFallbackEpisodesQuery({
+ workflowPermanentId,
+ workflowRunId: workflowRun?.workflow_run_id,
+ enabled: finalized && !isWorkflowDeleted,
+ });
+ const onboarding = useOnboardingStateOptional();
+ const recoveryGuidanceTreatmentSurfaceEnabled =
+ useFeatureFlag(RECOVERY_GUIDANCE_TREATMENT_SURFACE_FLAG) === true;
+ const recoveryGuidanceAssignment =
+ onboarding?.recoveryGuidanceAssignment ?? null;
+ const recoveryGuidanceTelemetryContext =
+ useMemo<RecoveryGuidanceTelemetryContext | null>(() => {
+ if (!recoveryGuidanceAssignment || !workflowRun) {
+ return null;
+ }
+ return {
+ organizationId: recoveryGuidanceAssignment.organization_id,
+ experimentVersion: recoveryGuidanceAssignment.experiment_version,
+ arm: recoveryGuidanceAssignment.arm,
+ eligibleRunId: recoveryGuidanceAssignment.eligible_run_id,
+ failureCategory: workflowRun.failure_category?.[0]?.category ?? null,
+ };
+ }, [recoveryGuidanceAssignment, workflowRun]);
+ const showFirstFailedRunRecovery = shouldShowRecoveryGuidance({
+ assignment: recoveryGuidanceAssignment,
+ workflowRunId: workflowRun?.workflow_run_id,
+ treatmentSurfaceEnabled: recoveryGuidanceTreatmentSurfaceEnabled,
+ });
+ const recoveryGuidanceRetry = getRecoveryGuidanceRetryNavigation(
+ location.state,
+ );
+ const reportedRecoveryRetryStartRef = useRef<string | null>(null);
+
+ useEffect(() => {
+ if (
+ !recoveryGuidanceRetry ||
+ !retryRunHasStarted({
+ retryRunId: recoveryGuidanceRetry.retryRunId,
+ observedRunId: workflowRun?.workflow_run_id,
+ status: workflowRun?.status,
+ startedAt: workflowRun?.started_at,
+ }) ||
+ reportedRecoveryRetryStartRef.current === recoveryGuidanceRetry.retryRunId
+ ) {
+ return;
+ }
+ reportedRecoveryRetryStartRef.current = recoveryGuidanceRetry.retryRunId;
+ RecoveryGuidanceTelemetry.retryStarted(
+ recoveryGuidanceRetry,
+ recoveryGuidanceRetry.retryRunId,
+ );
+ }, [
+ recoveryGuidanceRetry,
+ workflowRun?.started_at,
+ workflowRun?.status,
+ workflowRun?.workflow_run_id,
+ ]);
+
+ const handleFirstFailedRunRetry = useCallback(() => {
+ if (
+ !workflowRun ||
+ !workflowPermanentId ||
+ !recoveryGuidanceTelemetryContext
+ ) {
+ return;
+ }
+ navigate(`/agents/${workflowPermanentId}/run`, {
+ state: {
+ ...getRerunNavigationState(workflowRun),
+ recoveryGuidanceRetry: recoveryGuidanceTelemetryContext,
+ },
+ });
+ }, [
+ navigate,
+ recoveryGuidanceTelemetryContext,
+ workflowPermanentId,
+ workflowRun,
+ ]);
// Last editor selection the canvas→run sync below acted on.
const syncedBlockLabelRef = useRef<string | null>(null);
- // A pinned frame belongs to one run; drop it when the run changes, then re-seed
- // from ?active= to restore a deep-linked selection.
+ // A pinned frame belongs to one run; drop it when the run changes.
useEffect(() => {
resetRunView();
setOutputSummary(null);
resetPaneView();
- const active = searchParamsRef.current.get("active");
- if (active) {
- pinFrame(active);
- }
// Adopt the standing editor selection without acting on it, so a cold open
// and a run switch land via ?active= / the auto-pin one-shot rather than
// racing them; the sync applies from the next selection change onward.
syncedBlockLabelRef.current = searchParamsRef.current.get(
SELECTED_BLOCK_SEARCH_PARAM,
);
- }, [workflowRunId, resetRunView, resetPaneView, pinFrame]);
+ }, [workflowRunId, resetRunView, resetPaneView]);
+
+ // Rehydrate deep-linked selections on cold load, run switches, and browser
+ // history changes that update ?active= / ?iteration= for the same run.
+ const lastHydratedSelectionRef = useRef<{
+ runId: string | undefined;
+ active: string | null;
+ iteration: number | null;
+ } | null>(null);
+ useEffect(() => {
+ const previous = lastHydratedSelectionRef.current;
+ if (
+ previous !== null &&
+ previous.runId === workflowRunId &&
+ previous.active === requestedActive &&
+ previous.iteration === requestedIteration
+ ) {
+ return;
+ }
+ lastHydratedSelectionRef.current = {
+ runId: workflowRunId,
+ active: requestedActive,
+ iteration: requestedIteration,
+ };
+
+ const current = useRunViewStore.getState();
+ if (requestedActive) {
+ if (
+ previous?.runId !== workflowRunId ||
+ current.pinnedFrameId !== requestedActive ||
+ current.activeIteration !== requestedIteration
+ ) {
+ pinFrame(requestedActive, requestedIteration);
+ }
+ } else if (
+ previous !== null &&
+ previous.runId === workflowRunId &&
+ previous.active !== null &&
+ (current.pinnedFrameId !== null || current.activeIteration !== null)
+ ) {
+ jumpToLive();
+ }
+ }, [
+ jumpToLive,
+ pinFrame,
+ requestedActive,
+ requestedIteration,
+ workflowRunId,
+ ]);
+
+ useEffect(() => {
+ if (isRunPaneView(requestedView)) {
+ setPaneView(requestedView);
+ } else if (requestedView !== "recording") {
+ resetPaneView();
+ }
+ }, [requestedView, resetPaneView, setPaneView, workflowRunId]);
// Mirror the pinned item to ?active= so selection survives reload. Skip the first
// pass after a run change so the seed above doesn't fight the URL.
@@ -300,7 +498,12 @@ export function RunView({
);
const desired =
pinnedFrameId && !/:\d+$/.test(pinnedFrameId) ? pinnedFrameId : null;
- if ((next.get("active") ?? null) === desired) {
+ const desiredIteration =
+ desired && activeIteration !== null ? String(activeIteration) : null;
+ if (
+ (next.get("active") ?? null) === desired &&
+ (next.get("iteration") ?? null) === desiredIteration
+ ) {
return;
}
if (desired) {
@@ -308,8 +511,13 @@ export function RunView({
} else {
next.delete("active");
}
+ if (desiredIteration) {
+ next.set("iteration", desiredIteration);
+ } else {
+ next.delete("iteration");
+ }
navigate({ search: toReadableSearch(next) }, { replace: true });
- }, [pinnedFrameId, workflowRunId, navigate]);
+ }, [activeIteration, pinnedFrameId, workflowRunId, navigate]);
// Stabilize an ?active=-only deep link by ADDING ?wr= when it's absent. Gated on
// the Overview pane being open: RunView stays mounted while its pane is closed.
@@ -370,10 +578,6 @@ export function RunView({
!runIsPlaceholder &&
outcome === "failed" &&
!canceled;
- const finalized =
- !statusUnavailable && workflowRun
- ? runIsLogicallyFinal(workflowRun)
- : false;
useLiveClock(Boolean(workflowRun) && !finalized && !statusUnavailable);
const finallyBlockLabel =
workflowRun?.workflow?.workflow_definition?.finally_block_label ?? null;
@@ -496,12 +700,12 @@ export function RunView({
if (!finalized || outcome !== "success") {
return;
}
+ const landingSearchParams = new URLSearchParams(
+ window.location.search || searchParamsRef.current.toString(),
+ );
if (
- hasExplicitSelection(
- new URLSearchParams(
- window.location.search || searchParamsRef.current.toString(),
- ),
- )
+ landingSearchParams.has("view") ||
+ hasExplicitSelection(landingSearchParams)
) {
return;
}
@@ -746,16 +950,28 @@ export function RunView({
// would rewrite working code.
const showFix = codeFailure === null || codeFailure.recovery !== "retry";
const hasFixAction = Boolean(onFix && showFix);
- const recovery = failed ? (
- <FailureRecoveryActions
- onFix={
- onFix && hasFixAction
- ? () => onFix(failedBlock?.label ?? null)
- : undefined
- }
- onRetry={onRetry}
- />
- ) : null;
+ const recovery =
+ failed && !embedded ? (
+ <>
+ <FailureRecoveryActions
+ onFix={
+ onFix && hasFixAction
+ ? () => onFix(failedBlock?.label ?? null)
+ : undefined
+ }
+ onRetry={onRetry}
+ />
+ {onRetry &&
+ showFirstFailedRunRecovery &&
+ recoveryGuidanceTelemetryContext ? (
+ <FirstRunRecoveryGuidance
+ telemetryContext={recoveryGuidanceTelemetryContext}
+ workflowPermanentId={workflowPermanentId}
+ onRetry={handleFirstFailedRunRetry}
+ />
+ ) : null}
+ </>
+ ) : null;
const jumpToFailedBlock = failedBlock
? () => selectTimelineBlock(failedBlock)
: undefined;
@@ -765,6 +981,27 @@ export function RunView({
<WorkflowRunVerificationCodeForm
workflowRunId={workflowRun.workflow_run_id}
/>
+ {embedded ? null : (
+ <div className="flex shrink-0 flex-wrap items-center gap-x-3 gap-y-1">
+ <RunTagsEditor workflowRunId={workflowRun.workflow_run_id} />
+ {workflowRun.retried_from_workflow_run_id ? (
+ <Link
+ className="text-xs text-muted-foreground underline-offset-2 hover:text-foreground hover:underline"
+ to={`/runs/${workflowRun.retried_from_workflow_run_id}`}
+ >
+ Retried from {workflowRun.retried_from_workflow_run_id}
+ </Link>
+ ) : null}
+ {workflowRun.retried_by_workflow_run_id ? (
+ <Link
+ className="text-xs text-muted-foreground underline-offset-2 hover:text-foreground hover:underline"
+ to={`/runs/${workflowRun.retried_by_workflow_run_id}`}
+ >
+ Retried by {workflowRun.retried_by_workflow_run_id}
+ </Link>
+ ) : null}
+ </div>
+ )}
{WorkflowRunMilestoneCard &&
runPaneOpen &&
!runIsPlaceholder &&
@@ -772,7 +1009,13 @@ export function RunView({
workflowRun.status === Status.Completed ? (
<WorkflowRunMilestoneCard
workflowRunId={workflowRun.workflow_run_id}
- rerun={milestoneRerun}
+ rerun={embedded ? undefined : milestoneRerun}
+ />
+ ) : null}
+ {fallbackEpisodes && fallbackEpisodes.episodes.length > 0 ? (
+ <ScriptUpdateCard
+ episodes={fallbackEpisodes.episodes}
+ scriptId={blockScriptsPublished?.script_id}
/>
) : null}
@@ -847,7 +1090,11 @@ export function RunView({
workflowRunId={workflowRunId}
onThoughtSelect={(thought) => pinFrame(thought.thought_id)}
onViewScreenshot={(workflowRunBlockId) => {
- openPane("browser");
+ if (embedded) {
+ setOpenPanes(["browser"]);
+ } else {
+ openPane("browser");
+ }
// Let the pane-open navigation commit before mirroring the
// pin; otherwise the pin's URL writer can drop `browser`.
requestAnimationFrame(() => pinFrame(workflowRunBlockId));
diff --git a/skyvern-frontend/src/routes/workflows/studio/useBrowserPaneView.ts b/skyvern-frontend/src/routes/workflows/studio/useBrowserPaneView.ts
index 395ab184d..3584e5ff3 100644
--- a/skyvern-frontend/src/routes/workflows/studio/useBrowserPaneView.ts
+++ b/skyvern-frontend/src/routes/workflows/studio/useBrowserPaneView.ts
@@ -1,5 +1,5 @@
-import { useEffect, useRef } from "react";
-import { useSearchParams } from "react-router-dom";
+import { useCallback, useEffect, useRef } from "react";
+import { useNavigate, useSearchParams } from "react-router-dom";
import { useWorkflowPermanentId } from "@/routes/workflows/WorkflowPermanentIdContext";
import { useRecordingStore } from "@/store/useRecordingStore";
@@ -15,9 +15,15 @@ import {
resolveLiveSurface,
type BrowserPaneView,
} from "./browserPaneView";
-import { SYSTEM_RUN_FOCUS_PARAM } from "./panes";
+import {
+ parsePanesParam,
+ STUDIO_PANES_PARAM,
+ SYSTEM_RUN_FOCUS_PARAM,
+ toReadableSearch,
+} from "./panes";
import { useRunVisuals, type RunVisuals } from "./useRunVisuals";
import { useStudioInspectedRun } from "./useStudioInspectedRun";
+import { useStudioPanes } from "./useStudioPanes";
type BrowserPaneViewState = {
view: BrowserPaneView;
@@ -44,6 +50,8 @@ type BrowserPaneViewState = {
export function useBrowserPaneView(): BrowserPaneViewState {
const workflowPermanentId = useWorkflowPermanentId();
const [searchParams] = useSearchParams();
+ const navigate = useNavigate();
+ const { resolveLivePanes } = useStudioPanes();
const { runId, explicit } = useStudioInspectedRun();
const visuals = useRunVisuals(runId);
const { data: debugSession } = useDebugSessionQuery({
@@ -52,9 +60,10 @@ export function useBrowserPaneView(): BrowserPaneViewState {
});
const debugBrowserSessionId = debugSession?.browser_session_id ?? null;
const intent = useStudioBrowserStore((s) => s.view);
- const setView = useStudioBrowserStore((s) => s.setView);
+ const setViewIntent = useStudioBrowserStore((s) => s.setView);
const pinNonce = useRunViewStore((s) => s.pinNonce);
const isRecording = useRecordingStore((s) => s.isRecording);
+ const recordingDeepLink = searchParams.get("view") === "recording";
// Timeline selection outranks a pinned pill: a step click (?active= change or
// a re-pin of the same step) or a run swap hands the pane back to the machine,
@@ -71,11 +80,53 @@ export function useBrowserPaneView(): BrowserPaneViewState {
recordingStarted
) {
syncRef.current = { activeParam, pinNonce, runId, isRecording };
- setView("auto");
+ if (!recordingDeepLink) {
+ setViewIntent("auto");
+ }
} else if (prev.isRecording !== isRecording) {
syncRef.current = { activeParam, pinNonce, runId, isRecording };
}
- }, [activeParam, pinNonce, runId, isRecording, setView]);
+ }, [
+ activeParam,
+ pinNonce,
+ runId,
+ isRecording,
+ recordingDeepLink,
+ setViewIntent,
+ ]);
+
+ const appliedRecordingDeepLinkRef = useRef<string | undefined>(undefined);
+ useEffect(() => {
+ if (!recordingDeepLink) {
+ appliedRecordingDeepLinkRef.current = undefined;
+ return;
+ }
+ if (!runId || appliedRecordingDeepLinkRef.current === runId) {
+ return;
+ }
+ appliedRecordingDeepLinkRef.current = runId;
+ setViewIntent("recording");
+ }, [recordingDeepLink, runId, setViewIntent]);
+
+ const setView = useCallback(
+ (nextView: BrowserPaneViewIntent) => {
+ setViewIntent(nextView);
+ const next = new URLSearchParams(searchParams);
+ if (nextView === "recording") {
+ next.set("view", "recording");
+ const panes =
+ parsePanesParam(next.get(STUDIO_PANES_PARAM)) ?? resolveLivePanes();
+ next.set(
+ STUDIO_PANES_PARAM,
+ ["browser", ...panes.filter((pane) => pane !== "browser")].join(","),
+ );
+ } else if (next.get("view") === "recording") {
+ next.delete("view");
+ }
+ navigate({ search: toReadableSearch(next) }, { replace: true });
+ },
+ [navigate, resolveLivePanes, searchParams, setViewIntent],
+ );
const runInDebugSession =
visuals.workflowRun?.browser_session_id != null &&
diff --git a/skyvern-frontend/src/routes/workflows/studioNavigation.ts b/skyvern-frontend/src/routes/workflows/studioNavigation.ts
index 3907295b2..eb81ede3b 100644
--- a/skyvern-frontend/src/routes/workflows/studioNavigation.ts
+++ b/skyvern-frontend/src/routes/workflows/studioNavigation.ts
@@ -1,27 +1,14 @@
-import * as env from "@/util/env";
-
/**
- * Where "open the editor" links point. With the studio preview on, the editor
- * lives at /studio; off, it is the legacy /build debugger surface.
+ * Where "open the editor" links point.
*/
export function workflowEditorPath(
workflowPermanentId: string,
- studioEnabled: boolean,
+ _studioEnabled: boolean,
search = "",
): string {
- const leaf = studioEnabled ? "studio" : "build";
- return `/agents/${workflowPermanentId}/${leaf}${search}`;
+ return `/agents/${workflowPermanentId}/studio${search}`;
}
-/**
- * Legacy (studio-off) destination for viewing a finished workflow run, honoring
- * the existing useNewRunsUrl split between the global and per-workflow pages.
- */
-export function legacyRunDetailPath(
- workflowPermanentId: string,
- workflowRunId: string,
-): string {
- return env.useNewRunsUrl
- ? `/runs/${workflowRunId}`
- : `/agents/${workflowPermanentId}/${workflowRunId}/overview`;
+export function workflowRunDetailPath(workflowRunId: string): string {
+ return `/runs/${workflowRunId}`;
}
diff --git a/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts b/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts
index 84aa48d85..b3964c493 100644
--- a/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts
+++ b/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts
@@ -786,6 +786,7 @@ export type WorkflowApiResponse = {
modified_at: string;
deleted_at: string | null;
run_with: string; // 'agent' or 'code'
+ browser_type?: string | null; // BrowserType value; null = system default
cache_key: string | null;
ai_fallback: boolean | null;
enable_self_healing: boolean | null;
@@ -816,6 +817,7 @@ export type WorkflowSettings = {
extraHttpHeaders: string | null;
cdpConnectHeaders: string | null;
runWith: string; // 'agent' or 'code'
+ browserType?: string | null; // BrowserType value; null = system default
codeVersion: number | null;
scriptCacheKey: string | null;
aiFallback: boolean | null;
diff --git a/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts b/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts
index d6f9fd3de..46f33f779 100644
--- a/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts
+++ b/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts
@@ -30,6 +30,7 @@ export type WorkflowCreateYAMLRequest = {
cdp_connect_headers?: Record<string, string> | null;
status?: string | null;
run_with?: string | null;
+ browser_type?: string | null;
cache_key?: string | null;
ai_fallback?: boolean;
enable_self_healing?: boolean;
diff --git a/skyvern-frontend/src/routes/workflows/useStudioRunRouteMatch.ts b/skyvern-frontend/src/routes/workflows/useStudioRunRouteMatch.ts
index fc787d6f0..f3802b092 100644
--- a/skyvern-frontend/src/routes/workflows/useStudioRunRouteMatch.ts
+++ b/skyvern-frontend/src/routes/workflows/useStudioRunRouteMatch.ts
@@ -1,11 +1,9 @@
import { type PathMatch, useMatch } from "react-router-dom";
/**
- * The short run URL (/runs/{wr}) renders the studio run view when the preview is
- * on, so route classifiers that recognize /agents/.../studio must also treat it
- * as a studio surface. Only workflow runs (wr_) route into the studio; task runs
- * keep their own detail view. Callers still gate on the studio flag themselves,
- * since flag-off /runs/{wr} renders the legacy run page.
+ * The short run URL (/runs/{wr}) renders the studio run view, so route
+ * classifiers that recognize /agents/.../studio must also treat it as a studio
+ * surface. Task runs keep their own detail view.
*/
export function useStudioRunRouteMatch(): PathMatch<string> | null {
const match = useMatch("/runs/:runId/*");
diff --git a/skyvern-frontend/src/routes/workflows/utils.test.ts b/skyvern-frontend/src/routes/workflows/utils.test.ts
index 4d390fb5d..366931ec0 100644
--- a/skyvern-frontend/src/routes/workflows/utils.test.ts
+++ b/skyvern-frontend/src/routes/workflows/utils.test.ts
@@ -1,3 +1,6 @@
+import { renderHook } from "@testing-library/react";
+import { act } from "react";
+import { useForm } from "react-hook-form";
import { describe, expect, it } from "vitest";
import {
@@ -8,16 +11,29 @@ import type {
Parameter,
WorkflowApiResponse,
WorkflowParameter,
+ WorkflowSettings,
} from "./types/workflowTypes";
import {
+ browserTypeSelectionDisabled,
+ extractBrowserTypeSetting,
getInitialValues,
getOrderedRunParameters,
getRerunNavigationState,
+ hasBrowserTypeOptions,
+ RESERVED_BROWSER_TYPE_FIELD,
+ rerunBrowserTypeState,
+ resolveInitialBrowserType,
normalizeJsonParameterFormValue,
parseJsonWorkflowParameterValue,
shouldPollForGeneratedCode,
validateJsonWorkflowParameterValue,
} from "./utils";
+import { getRunWorkflowRequestBody } from "./RunWorkflowForm";
+import {
+ snapshotOf,
+ summarizeWorkflowChanges,
+} from "./editor/workflowChangesSummary";
+import type { WorkflowSaveData } from "@/store/WorkflowHasChangesStore";
function buildWorkflowRun(
overrides: Partial<WorkflowRunStatusApiResponseWithWorkflow> = {},
@@ -29,6 +45,7 @@ function buildWorkflowRun(
max_screenshot_scrolls: 8,
run_with: "code",
browser_profile_id: "profile_synthetic",
+ browser_type: "msedge",
extra_http_headers: { "X-Test": "synthetic" },
...overrides,
} as WorkflowRunStatusApiResponseWithWorkflow;
@@ -70,8 +87,297 @@ describe("shouldPollForGeneratedCode", () => {
});
});
+describe("rerunBrowserTypeState (shared by all Rerun/Retry callers)", () => {
+ it("carries a real browser_type value", () => {
+ expect(rerunBrowserTypeState({ browser_type: "msedge" })).toEqual({
+ browserType: "msedge",
+ });
+ });
+ it("carries explicit null as browserType: null", () => {
+ expect(rerunBrowserTypeState({ browser_type: null })).toEqual({
+ browserType: null,
+ });
+ });
+ it("omits browserType entirely when the property is absent (older backend/legacy)", () => {
+ const state = rerunBrowserTypeState({});
+ expect(state).not.toHaveProperty("browserType");
+ expect(state).toEqual({});
+ });
+});
+
+describe("resolveInitialBrowserType (null inherits the workflow setting)", () => {
+ it("preselects the executed run's non-null browser_type carried in state", () => {
+ expect(resolveInitialBrowserType({ browserType: "msedge" }, "chrome")).toBe(
+ "msedge",
+ );
+ });
+
+ it("inherits the current workflow default when the run's browser_type is null", () => {
+ // Contract: a run-level null means inherit the workflow setting (no "force default" sentinel).
+ // The source run ran with null; the workflow default is now chrome, so the rerun uses chrome —
+ // matching how the backend resolves an omitted/null run value.
+ expect(resolveInitialBrowserType({ browserType: null }, "chrome")).toBe(
+ "chrome",
+ );
+ });
+
+ it("inherits the workflow default when state never carried the property (older backend)", () => {
+ expect(resolveInitialBrowserType({}, "chrome")).toBe("chrome");
+ expect(resolveInitialBrowserType(null, "chrome")).toBe("chrome");
+ expect(resolveInitialBrowserType(undefined, "chrome")).toBe("chrome");
+ });
+
+ it("resolves to null only when both the run value and the workflow are null (dynamic routing)", () => {
+ expect(resolveInitialBrowserType({ browserType: null }, null)).toBeNull();
+ expect(resolveInitialBrowserType({}, null)).toBeNull();
+ expect(resolveInitialBrowserType({}, undefined)).toBeNull();
+ });
+});
+
+describe("reserved browser-type form field", () => {
+ it("react-hook-form treats the hyphenated reserved name as a flat top-level key", () => {
+ const { result } = renderHook(() =>
+ useForm<Record<string, unknown>>({
+ defaultValues: { [RESERVED_BROWSER_TYPE_FIELD]: null },
+ }),
+ );
+
+ act(() => {
+ result.current.register(RESERVED_BROWSER_TYPE_FIELD);
+ result.current.setValue(RESERVED_BROWSER_TYPE_FIELD, "msedge");
+ });
+
+ expect(result.current.getValues(RESERVED_BROWSER_TYPE_FIELD)).toBe(
+ "msedge",
+ );
+ expect(result.current.watch(RESERVED_BROWSER_TYPE_FIELD)).toBe("msedge");
+ // Flat key, not a nested path: no accidental `__skyvern-browser-type` object graph.
+ const all = result.current.getValues();
+ expect(all[RESERVED_BROWSER_TYPE_FIELD]).toBe("msedge");
+ });
+
+ it("extracts the reserved setting while preserving a user parameter literally named browserType", () => {
+ const { browserType, rest } = extractBrowserTypeSetting({
+ [RESERVED_BROWSER_TYPE_FIELD]: "msedge",
+ browserType: "business-value",
+ other: 1,
+ });
+
+ // The internal setting becomes the API browser_type source...
+ expect(browserType).toBe("msedge");
+ // ...and the reserved key is stripped from the request data...
+ expect(rest).not.toHaveProperty(RESERVED_BROWSER_TYPE_FIELD);
+ // ...while the legitimate workflow parameter survives unchanged.
+ expect(rest.browserType).toBe("business-value");
+ expect(rest.other).toBe(1);
+ });
+
+ it("treats an absent or null reserved setting as Default (null)", () => {
+ expect(extractBrowserTypeSetting({}).browserType).toBeNull();
+ expect(
+ extractBrowserTypeSetting({ [RESERVED_BROWSER_TYPE_FIELD]: null })
+ .browserType,
+ ).toBeNull();
+ });
+});
+
+describe("browserTypeSelectionDisabled (an attachment owns the engine)", () => {
+ it("is true when only a session is set", () => {
+ expect(browserTypeSelectionDisabled({ browserSessionId: "pbs_abc" })).toBe(
+ true,
+ );
+ });
+
+ it("is true when only a remote address is set", () => {
+ expect(
+ browserTypeSelectionDisabled({ browserAddress: "http://host:9222" }),
+ ).toBe(true);
+ });
+
+ it("is true when both are set", () => {
+ expect(
+ browserTypeSelectionDisabled({
+ browserSessionId: "pbs_abc",
+ browserAddress: "http://host:9222",
+ }),
+ ).toBe(true);
+ });
+
+ it("is false when neither is set (including blank/whitespace); a browser profile is not an attachment", () => {
+ expect(browserTypeSelectionDisabled({})).toBe(false);
+ expect(
+ browserTypeSelectionDisabled({
+ browserSessionId: null,
+ browserAddress: null,
+ }),
+ ).toBe(false);
+ expect(
+ browserTypeSelectionDisabled({
+ browserSessionId: " ",
+ browserAddress: " ",
+ }),
+ ).toBe(false);
+ });
+});
+
+describe("hasBrowserTypeOptions (render the selector only with real options)", () => {
+ it("is false while the query is loading or errored (options undefined)", () => {
+ expect(hasBrowserTypeOptions(undefined)).toBe(false);
+ });
+
+ it("is false when the backend returns an empty list", () => {
+ expect(hasBrowserTypeOptions([])).toBe(false);
+ });
+
+ it("is true once at least one option exists", () => {
+ expect(
+ hasBrowserTypeOptions([{ value: "msedge", label: "Microsoft Edge" }]),
+ ).toBe(true);
+ });
+});
+
+describe("getRunWorkflowRequestBody browser_type suppression rule", () => {
+ function buildFormValues(
+ overrides: Record<string, unknown> = {},
+ ): Parameters<typeof getRunWorkflowRequestBody>[0] {
+ return {
+ webhookCallbackUrl: "",
+ proxyLocation: ProxyLocation.Residential,
+ browserSessionId: null,
+ reuseBrowserSession: null,
+ browserProfileId: null,
+ startFreshBrowser: false,
+ cdpAddress: null,
+ maxScreenshotScrolls: null,
+ extraHttpHeaders: null,
+ cdpConnectHeaders: null,
+ runWith: "agent",
+ aiFallback: true,
+ [RESERVED_BROWSER_TYPE_FIELD]: "msedge",
+ ...overrides,
+ } as Parameters<typeof getRunWorkflowRequestBody>[0];
+ }
+
+ it("sends browser_type when neither a session nor an address is attached", () => {
+ const body = getRunWorkflowRequestBody(buildFormValues(), []);
+ expect(body.browser_type).toBe("msedge");
+ });
+
+ it("omits browser_type when a session is attached even if the field held a value", () => {
+ const body = getRunWorkflowRequestBody(
+ buildFormValues({ browserSessionId: "pbs_abc" }),
+ [],
+ );
+ expect(body).not.toHaveProperty("browser_type");
+ });
+
+ it("omits browser_type when a remote address is attached", () => {
+ const body = getRunWorkflowRequestBody(
+ buildFormValues({ cdpAddress: "http://host:9222" }),
+ [],
+ );
+ expect(body).not.toHaveProperty("browser_type");
+ });
+
+ it("normalizes a cleared browser address to null so it does not conflict with a selected browser_type", () => {
+ const body = getRunWorkflowRequestBody(
+ buildFormValues({ cdpAddress: "" }),
+ [],
+ );
+ // A cleared address is not an attachment, so the selected engine still applies...
+ expect(body.browser_type).toBe("msedge");
+ // ...and the empty string must become null; otherwise the backend treats ""
+ // as an attachment (browser_address is not None) and rejects the run with 422.
+ expect(body.browser_address).toBeNull();
+ });
+
+ it("keeps a legitimate `browserType` workflow parameter in request data unchanged", () => {
+ const workflowParameters = [
+ {
+ parameter_type: "workflow",
+ key: "browserType",
+ workflow_parameter_type: "string",
+ },
+ ] as WorkflowParameter[];
+ const body = getRunWorkflowRequestBody(
+ buildFormValues({
+ browserSessionId: "pbs_abc",
+ browserType: "business-value",
+ }),
+ workflowParameters,
+ );
+ // The reserved internal setting is suppressed by the attachment...
+ expect(body).not.toHaveProperty("browser_type");
+ // ...while the user's workflow parameter literally named browserType survives.
+ expect(body.data.browserType).toBe("business-value");
+ });
+});
+
+describe("browser-type change summary", () => {
+ function buildSettings(
+ overrides: Partial<WorkflowSettings> = {},
+ ): WorkflowSettings {
+ return {
+ retryPolicy: null,
+ proxyLocation: null,
+ webhookCallbackUrl: null,
+ persistBrowserSession: false,
+ reuseBrowserSession: false,
+ pinSavedSessionIp: false,
+ browserProfileId: null,
+ browserProfileKey: null,
+ model: null,
+ maxScreenshotScrolls: null,
+ maxElapsedTimeMinutes: null,
+ extraHttpHeaders: null,
+ cdpConnectHeaders: null,
+ runWith: "agent",
+ browserType: null,
+ codeVersion: null,
+ scriptCacheKey: null,
+ aiFallback: null,
+ enableSelfHealing: null,
+ maskSecrets: false,
+ runSequentially: false,
+ sequentialKey: null,
+ finallyBlockLabel: null,
+ workflowSystemPrompt: null,
+ errorCodeMapping: null,
+ ...overrides,
+ };
+ }
+
+ function buildSaveData(settings: WorkflowSettings): WorkflowSaveData {
+ return {
+ blocks: [],
+ parameters: [],
+ settings,
+ title: "wf",
+ } as unknown as WorkflowSaveData;
+ }
+
+ it("emits the browser-type-specific label when only browserType changes", () => {
+ const baseline = buildSaveData(buildSettings({ browserType: null }));
+ const draft = buildSaveData(buildSettings({ browserType: "msedge" }));
+
+ const changes = summarizeWorkflowChanges(draft, snapshotOf(baseline));
+
+ expect(changes).toContain("Changed browser type");
+ expect(changes).not.toContain("Other workflow changes");
+ });
+
+ it("does not report a browser-type change when it is unchanged", () => {
+ const baseline = buildSaveData(buildSettings({ browserType: "chrome" }));
+ const draft = buildSaveData(buildSettings({ browserType: "chrome" }));
+
+ const changes = summarizeWorkflowChanges(draft, snapshotOf(baseline));
+
+ expect(changes).not.toContain("Changed browser type");
+ });
+});
+
describe("getRerunNavigationState", () => {
- it("maps exactly the six legacy rerun fields", () => {
+ it("maps the legacy rerun fields plus the executed run's browser_type", () => {
const state = getRerunNavigationState(buildWorkflowRun());
expect(state).toEqual({
@@ -81,6 +387,7 @@ describe("getRerunNavigationState", () => {
maxScreenshotScrolls: 8,
runWith: "code",
browserProfileId: "profile_synthetic",
+ browserType: "msedge",
});
expect(state).not.toHaveProperty("extraHttpHeaders");
expect(state).not.toHaveProperty("cdpConnectHeaders");
@@ -95,6 +402,7 @@ describe("getRerunNavigationState", () => {
max_screenshot_scrolls: null,
run_with: null,
browser_profile_id: null,
+ browser_type: null,
} as unknown as WorkflowRunStatusApiResponseWithWorkflow;
expect(getRerunNavigationState(workflowRun)).toEqual({
@@ -104,6 +412,30 @@ describe("getRerunNavigationState", () => {
maxScreenshotScrolls: null,
runWith: "agent",
browserProfileId: null,
+ browserType: null,
+ });
+ });
+
+ it("omits browserType entirely when the source has no browser_type property (legacy/older backend)", () => {
+ const workflowRun = {
+ parameters: null,
+ proxy_location: null,
+ webhook_callback_url: null,
+ max_screenshot_scrolls: null,
+ run_with: null,
+ browser_profile_id: null,
+ // no browser_type property at all — an older backend or legacy fixture
+ } as unknown as WorkflowRunStatusApiResponseWithWorkflow;
+
+ const state = getRerunNavigationState(workflowRun);
+ expect(state).not.toHaveProperty("browserType");
+ expect(state).toEqual({
+ data: {},
+ proxyLocation: ProxyLocation.Residential,
+ webhookCallbackUrl: "",
+ maxScreenshotScrolls: null,
+ runWith: "agent",
+ browserProfileId: null,
});
});
diff --git a/skyvern-frontend/src/routes/workflows/utils.ts b/skyvern-frontend/src/routes/workflows/utils.ts
index 7bad1a43c..953327f51 100644
--- a/skyvern-frontend/src/routes/workflows/utils.ts
+++ b/skyvern-frontend/src/routes/workflows/utils.ts
@@ -1,16 +1,145 @@
import { useLocation } from "react-router-dom";
-import {
- ProxyLocation,
- type WorkflowRunStatusApiResponseWithWorkflow,
-} from "@/api/types";
+import { useQuery } from "@tanstack/react-query";
+import { ProxyLocation } from "@/api/types";
+import { getClient } from "@/api/AxiosClient";
+import { useCredentialGetter } from "@/hooks/useCredentialGetter";
import type { Parameter, WorkflowParameter } from "./types/workflowTypes";
import { WorkflowApiResponse } from "@/routes/workflows/types/workflowTypes";
type Location = ReturnType<typeof useLocation>;
-export function getRerunNavigationState(
- workflowRun: WorkflowRunStatusApiResponseWithWorkflow,
-) {
+/**
+ * Internal react-hook-form field name for the browser-type SETTING. Hyphens make it an invalid
+ * workflow-parameter identifier, so it can never collide with a user parameter literally named
+ * `browserType`. RHF stores names that aren't valid dotted paths as flat keys, so this reads/writes
+ * as a single top-level field.
+ */
+export const RESERVED_BROWSER_TYPE_FIELD = "__skyvern-browser-type";
+
+/**
+ * Split the reserved internal browser-type setting out of the run-form values: returns the selected
+ * setting (null = Default) and the remaining values (still carrying any user parameter literally
+ * named `browserType`, whose value must reach the request `data` unchanged).
+ */
+export function extractBrowserTypeSetting(values: Record<string, unknown>): {
+ browserType: string | null;
+ rest: Record<string, unknown>;
+} {
+ const { [RESERVED_BROWSER_TYPE_FIELD]: browserTypeSetting, ...rest } = values;
+ return {
+ browserType:
+ browserTypeSetting === undefined || browserTypeSetting === null
+ ? null
+ : (browserTypeSetting as string),
+ rest,
+ };
+}
+
+/**
+ * An attached browser owns its engine: a live session (`browserSessionId`) or a remote CDP address
+ * (`browserAddress`) is the browser for the run, so the backend rejects a run-level browser_type
+ * alongside it. This is the single source of truth for suppressing the Browser Type selector and
+ * omitting browser_type from the request. A browser PROFILE is not an attachment.
+ */
+export function browserTypeSelectionDisabled(attachment: {
+ browserSessionId?: string | null;
+ browserAddress?: string | null;
+}): boolean {
+ return Boolean(
+ attachment.browserSessionId?.trim() || attachment.browserAddress?.trim(),
+ );
+}
+
+export type BrowserTypeOption = {
+ value: string;
+ label: string;
+};
+
+/**
+ * Whether the Browser Type selector should render at all. The options come from `/browser_types`,
+ * which is `undefined` while loading or on error and `[]` if the backend returns nothing; in every
+ * one of those states the control would degrade to a lone misleading `Default` item, so both the
+ * workflow-settings and run-form selectors gate on a real non-empty option set instead.
+ */
+export function hasBrowserTypeOptions(
+ options: Array<BrowserTypeOption> | undefined,
+): options is Array<BrowserTypeOption> {
+ return Array.isArray(options) && options.length > 0;
+}
+
+/**
+ * Selectable browser engines, served by the backend `GET /browser_types` endpoint (the workflow/run
+ * `BrowserType` domain source). Fetching them keeps the picker in sync with the backend without a
+ * hand-maintained frontend option list — a newly supported backend type appears here automatically.
+ */
+export function useBrowserTypeOptionsQuery() {
+ const credentialGetter = useCredentialGetter();
+
+ return useQuery<Array<BrowserTypeOption>>({
+ queryKey: ["browser-types"],
+ queryFn: async () => {
+ const client = await getClient(credentialGetter);
+ return client
+ .get<Array<BrowserTypeOption>>("/browser_types")
+ .then((response) => response.data);
+ },
+ staleTime: 60 * 60 * 1000,
+ });
+}
+
+/**
+ * The rerun/retry navigation-state fragment carrying the executed run's browser_type — shared by
+ * every Rerun/Retry caller (the studio callers via getRerunNavigationState, and the run-overview
+ * recovery-guidance Retry) so the destination run form preselects what actually ran. Included only
+ * when the source exposes the property, so an older backend / legacy fixture without it keeps the
+ * prior navigation-state shape; the current backend always emits persisted browser_type, so a real
+ * rerun carries it and an Edge override wins over the workflow default.
+ */
+export function rerunBrowserTypeState(source: {
+ browser_type?: string | null;
+}): { browserType?: string | null } {
+ return "browser_type" in source
+ ? { browserType: source.browser_type ?? null }
+ : {};
+}
+
+type RerunNavigationSource = {
+ parameters?: Record<string, unknown> | null;
+ proxy_location?: ProxyLocation | null;
+ webhook_callback_url?: string | null;
+ max_screenshot_scrolls?: number | null;
+ run_with?: string | null;
+ browser_profile_id?: string | null;
+ browser_type?: string | null;
+};
+
+/**
+ * Resolve the browser-type a rerun/retry form should preselect. Contract: a run-level
+ * browser_type of `null` means INHERIT the workflow setting — only workflow+run both null yields
+ * dynamic routing, and there is no "force system default" sentinel. So a non-null executed-run value
+ * wins, but a `null` OR absent rerun value falls back to the workflow's CURRENT browser_type. A source
+ * run that ran with null, after the workflow later changed to Chrome, must rerun as Chrome — matching
+ * how the backend resolves an omitted/null run value against the workflow default.
+ */
+export function resolveInitialBrowserType(
+ locationState: unknown,
+ workflowBrowserType?: string | null,
+): string | null {
+ if (
+ locationState !== null &&
+ typeof locationState === "object" &&
+ "browserType" in locationState
+ ) {
+ const stateBrowserType = (locationState as { browserType?: string | null })
+ .browserType;
+ if (stateBrowserType != null) {
+ return stateBrowserType;
+ }
+ }
+ return workflowBrowserType ?? null;
+}
+
+export function getRerunNavigationState(workflowRun: RerunNavigationSource) {
return {
data: workflowRun.parameters ?? {},
proxyLocation: workflowRun.proxy_location ?? ProxyLocation.Residential,
@@ -18,6 +147,7 @@ export function getRerunNavigationState(
maxScreenshotScrolls: workflowRun.max_screenshot_scrolls ?? null,
runWith: workflowRun.run_with ?? "agent",
browserProfileId: workflowRun.browser_profile_id ?? null,
+ ...rerunBrowserTypeState(workflowRun),
};
}
diff --git a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowPostRunParameters.test.tsx b/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowPostRunParameters.test.tsx
deleted file mode 100644
index 30bf8a5de..000000000
--- a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowPostRunParameters.test.tsx
+++ /dev/null
@@ -1,93 +0,0 @@
-// @vitest-environment jsdom
-
-vi.mock("../hooks/useWorkflowRunWithWorkflowQuery", () => ({
- useWorkflowRunWithWorkflowQuery: () => ({
- data: mocks.workflowRun,
- isLoading: false,
- }),
-}));
-
-vi.mock("../hooks/useWorkflowRunTimelineQuery", () => ({
- useWorkflowRunTimelineQuery: () => ({
- data: [],
- isLoading: false,
- }),
-}));
-
-vi.mock("./useActiveWorkflowRunItem", () => ({
- useActiveWorkflowRunItem: () => [null, vi.fn()],
-}));
-
-vi.mock("@/components/KeyValueInput", () => ({
- KeyValueInput: ({ value }: { value: string | null }) => (
- <div data-testid="key-value-input">{value}</div>
- ),
-}));
-
-vi.mock("@/components/ProxySelector", () => ({
- ProxySelector: () => <div data-testid="proxy-selector" />,
-}));
-
-import { cleanup, render, screen } from "@testing-library/react";
-import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
-
-import { ProxyLocation } from "@/api/types";
-
-import { WorkflowPostRunParameters } from "./WorkflowPostRunParameters";
-
-const mocks = vi.hoisted(() => ({
- workflowRun: null as unknown,
-}));
-
-function buildWorkflowRun(extraHttpHeaders: Record<string, string> | null) {
- return {
- workflow_run_id: "wr_1",
- parameters: {},
- workflow: {
- workflow_definition: {
- blocks: [],
- parameters: [],
- },
- run_with: null,
- },
- task_v2: null,
- webhook_callback_url: null,
- proxy_location: ProxyLocation.Residential,
- extra_http_headers: extraHttpHeaders,
- browser_session_id: null,
- run_with: null,
- max_screenshot_scrolls: null,
- };
-}
-
-beforeEach(() => {
- mocks.workflowRun = buildWorkflowRun(null);
-});
-
-afterEach(() => {
- cleanup();
- vi.clearAllMocks();
-});
-
-describe("WorkflowPostRunParameters extra HTTP headers", () => {
- it("hides default empty extra HTTP headers", () => {
- mocks.workflowRun = buildWorkflowRun({});
-
- render(<WorkflowPostRunParameters />);
-
- expect(screen.getByText("Other Agent Inputs")).toBeDefined();
- expect(screen.queryByText("Extra HTTP Headers")).toBeNull();
- expect(screen.queryByTestId("key-value-input")).toBeNull();
- });
-
- it("shows non-empty extra HTTP headers", () => {
- mocks.workflowRun = buildWorkflowRun({ "X-Trace-Id": "abc123" });
-
- render(<WorkflowPostRunParameters />);
-
- expect(screen.getByText("Extra HTTP Headers")).toBeDefined();
- expect(screen.getByTestId("key-value-input").textContent).toBe(
- '{"X-Trace-Id":"abc123"}',
- );
- });
-});
diff --git a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowPostRunParameters.tsx b/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowPostRunParameters.tsx
deleted file mode 100644
index d94242c0f..000000000
--- a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowPostRunParameters.tsx
+++ /dev/null
@@ -1,482 +0,0 @@
-import { useMemo } from "react";
-import { useWorkflowRunWithWorkflowQuery } from "../hooks/useWorkflowRunWithWorkflowQuery";
-import { CodeEditor } from "../components/CodeEditor";
-import { AutoResizingTextarea } from "@/components/AutoResizingTextarea/AutoResizingTextarea";
-import { useActiveWorkflowRunItem } from "./useActiveWorkflowRunItem";
-import { useWorkflowRunTimelineQuery } from "../hooks/useWorkflowRunTimelineQuery";
-import { isAction, isWorkflowRunBlock } from "../types/workflowRunTypes";
-import { findBlockSurroundingAction } from "./workflowTimelineUtils";
-import { TaskBlockParameters } from "./TaskBlockParameters";
-import { isTaskVariantBlock, WorkflowBlockTypes } from "../types/workflowTypes";
-import { findWorkflowBlockByLabel, isBlockOfType } from "../workflowBlockUtils";
-import { Input } from "@/components/ui/input";
-import { ProxySelector } from "@/components/ProxySelector";
-import { SendEmailBlockParameters } from "./blockInfo/SendEmailBlockInfo";
-import { ProxyLocation } from "@/api/types";
-import { getOrderedRunParameters } from "../utils";
-import { KeyValueInput } from "@/components/KeyValueInput";
-import { CodeBlockParameters } from "./blockInfo/CodeBlockParameters";
-import { TextPromptBlockParameters } from "./blockInfo/TextPromptBlockParameters";
-import { GotoUrlBlockParameters } from "./blockInfo/GotoUrlBlockParameters";
-import { FileDownloadBlockParameters } from "./blockInfo/FileDownloadBlockParameters";
-import { WaitBlockParameters } from "./blockInfo/WaitBlockParameters";
-import { HttpRequestBlockParameters } from "./blockInfo/HttpRequestBlockParameters";
-import { PrintPageBlockParameters } from "./blockInfo/PrintPageBlockParameters";
-import { HumanInteractionBlockParameters } from "./blockInfo/HumanInteractionBlockParameters";
-import { ConditionalBlockParameters } from "./blockInfo/ConditionalBlockParameters";
-import { Taskv2BlockParameters } from "./blockInfo/Taskv2BlockParameters";
-import { WhileLoopPostRunFields } from "../components/WhileLoopPostRunFields";
-import { hasExtraHttpHeaders } from "@/util/extraHttpHeaders";
-
-function WorkflowPostRunParameters() {
- const { data: workflowRunTimeline, isLoading: workflowRunTimelineIsLoading } =
- useWorkflowRunTimelineQuery();
- const [activeItem] = useActiveWorkflowRunItem();
- const {
- data: workflowRun,
- isLoading: workflowRunIsLoading,
- isPlaceholderData: runIsWithheld,
- } = useWorkflowRunWithWorkflowQuery();
- const parameters = workflowRun?.parameters ?? {};
- const workflow = workflowRun?.workflow;
-
- function getActiveBlock() {
- if (!workflowRunTimeline) {
- return;
- }
- if (isWorkflowRunBlock(activeItem)) {
- return activeItem;
- }
- if (isAction(activeItem)) {
- return findBlockSurroundingAction(
- workflowRunTimeline,
- activeItem.action_id,
- );
- }
- }
-
- const activeBlock = getActiveBlock();
- const activeBlockLabel = activeBlock?.label ?? null;
- const definitionBlock = useMemo(() => {
- if (!workflow || !activeBlockLabel) {
- return null;
- }
- return findWorkflowBlockByLabel(
- workflow.workflow_definition.blocks,
- activeBlockLabel,
- );
- }, [workflow, activeBlockLabel]);
- const isTaskV2 = Boolean(workflowRun?.task_v2);
-
- const webhookCallbackUrl = isTaskV2
- ? (workflowRun?.task_v2?.webhook_callback_url ?? null)
- : (workflowRun?.webhook_callback_url ?? null);
-
- const proxyLocation = isTaskV2
- ? (workflowRun?.task_v2?.proxy_location ?? null)
- : (workflowRun?.proxy_location ?? null);
-
- const extraHttpHeaders = isTaskV2
- ? (workflowRun?.task_v2?.extra_http_headers ?? null)
- : (workflowRun?.extra_http_headers ?? null);
- const showExtraHttpHeaders = hasExtraHttpHeaders(extraHttpHeaders);
-
- if (workflowRunIsLoading || workflowRunTimelineIsLoading || runIsWithheld) {
- return <div>Loading workflow inputs...</div>;
- }
-
- if (!workflowRun || !workflowRunTimeline) {
- return null;
- }
-
- return (
- <div className="space-y-5">
- {activeBlock && isTaskVariantBlock(activeBlock) ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Block Inputs</h1>
- <TaskBlockParameters
- block={activeBlock}
- definitionBlock={definitionBlock}
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.FileDownload &&
- isBlockOfType(definitionBlock, WorkflowBlockTypes.FileDownload) ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">File Download Settings</h1>
- <FileDownloadBlockParameters
- prompt={
- activeBlock.navigation_goal ??
- definitionBlock.navigation_goal ??
- null
- }
- downloadSuffix={definitionBlock.download_suffix ?? null}
- downloadTimeout={definitionBlock.download_timeout ?? null}
- errorCodeMapping={definitionBlock.error_code_mapping ?? null}
- maxRetries={definitionBlock.max_retries ?? null}
- maxStepsPerRun={definitionBlock.max_steps_per_run ?? null}
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.SendEmail ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Block Inputs</h1>
- <SendEmailBlockParameters
- body={activeBlock.body ?? ""}
- recipients={activeBlock.recipients ?? []}
- subject={activeBlock.subject ?? ""}
- />
- </div>
- </div>
- ) : null}
- {activeBlock && activeBlock.block_type === WorkflowBlockTypes.ForLoop ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-sm font-bold">Block Inputs</h1>
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-sm">Loop Values</h1>
- <h2 className="text-sm text-muted-foreground">
- The values that are being looped over
- </h2>
- </div>
- <CodeEditor
- className="w-full"
- language="json"
- value={JSON.stringify(activeBlock.loop_values ?? [], null, 2)}
- readOnly
- minHeight="96px"
- maxHeight="200px"
- />
- </div>
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.WhileLoop ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-sm font-bold">Block Inputs</h1>
- <WhileLoopPostRunFields
- layout="sidebar"
- definitionBlock={definitionBlock}
- loopOutput={activeBlock.output}
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.Code &&
- isBlockOfType(definitionBlock, WorkflowBlockTypes.Code) ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Code Block</h1>
- <CodeBlockParameters
- code={definitionBlock.code}
- parameters={definitionBlock.parameters}
- prompt={definitionBlock.prompt}
- steps={definitionBlock.steps}
- blockStatus={activeBlock.status}
- failureReason={activeBlock.failure_reason}
- actions={activeBlock.actions}
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.TextPrompt &&
- isBlockOfType(definitionBlock, WorkflowBlockTypes.TextPrompt) ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Text Prompt Block</h1>
- <TextPromptBlockParameters
- prompt={activeBlock.prompt ?? definitionBlock.prompt ?? ""}
- llmKey={definitionBlock.llm_key}
- jsonSchema={definitionBlock.json_schema}
- parameters={definitionBlock.parameters}
- />
- </div>
- </div>
- ) : null}
- {activeBlock && activeBlock.block_type === WorkflowBlockTypes.URL ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Go To URL Block</h1>
- <GotoUrlBlockParameters
- url={
- activeBlock.url ??
- (isBlockOfType(definitionBlock, WorkflowBlockTypes.URL)
- ? definitionBlock.url
- : "")
- }
- continueOnFailure={activeBlock.continue_on_failure}
- />
- </div>
- </div>
- ) : null}
- {activeBlock && activeBlock.block_type === WorkflowBlockTypes.Wait ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Wait Block</h1>
- <WaitBlockParameters
- waitSec={
- activeBlock.wait_sec ??
- (isBlockOfType(definitionBlock, WorkflowBlockTypes.Wait)
- ? (definitionBlock.wait_sec ?? null)
- : null)
- }
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.HttpRequest &&
- isBlockOfType(definitionBlock, WorkflowBlockTypes.HttpRequest) ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">HTTP Request Block</h1>
- <HttpRequestBlockParameters
- method={definitionBlock.method}
- url={definitionBlock.url}
- headers={definitionBlock.headers}
- body={definitionBlock.body}
- files={definitionBlock.files}
- timeout={definitionBlock.timeout}
- followRedirects={definitionBlock.follow_redirects}
- downloadFilename={definitionBlock.download_filename}
- saveResponseAsFile={definitionBlock.save_response_as_file}
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.PrintPage &&
- isBlockOfType(definitionBlock, WorkflowBlockTypes.PrintPage) ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Print Page Block</h1>
- <PrintPageBlockParameters
- format={definitionBlock.format}
- landscape={definitionBlock.landscape}
- printBackground={definitionBlock.print_background}
- includeTimestamp={definitionBlock.include_timestamp}
- customFilename={definitionBlock.custom_filename}
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.HumanInteraction ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Human Interaction Block</h1>
- <HumanInteractionBlockParameters
- instructions={
- activeBlock.instructions ??
- (isBlockOfType(
- definitionBlock,
- WorkflowBlockTypes.HumanInteraction,
- )
- ? definitionBlock.instructions
- : null)
- }
- positiveDescriptor={
- activeBlock.positive_descriptor ??
- (isBlockOfType(
- definitionBlock,
- WorkflowBlockTypes.HumanInteraction,
- )
- ? definitionBlock.positive_descriptor
- : null)
- }
- negativeDescriptor={
- activeBlock.negative_descriptor ??
- (isBlockOfType(
- definitionBlock,
- WorkflowBlockTypes.HumanInteraction,
- )
- ? definitionBlock.negative_descriptor
- : null)
- }
- timeoutSeconds={
- isBlockOfType(
- definitionBlock,
- WorkflowBlockTypes.HumanInteraction,
- )
- ? definitionBlock.timeout_seconds
- : null
- }
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.Conditional ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Conditional Block</h1>
- <ConditionalBlockParameters
- branchConditions={
- isBlockOfType(definitionBlock, WorkflowBlockTypes.Conditional)
- ? definitionBlock.branch_conditions
- : null
- }
- executedBranchId={activeBlock.executed_branch_id ?? null}
- executedBranchExpression={
- activeBlock.executed_branch_expression ?? null
- }
- executedBranchResult={activeBlock.executed_branch_result ?? null}
- executedBranchNextBlock={
- activeBlock.executed_branch_next_block ?? null
- }
- />
- </div>
- </div>
- ) : null}
- {activeBlock &&
- activeBlock.block_type === WorkflowBlockTypes.Taskv2 &&
- isBlockOfType(definitionBlock, WorkflowBlockTypes.Taskv2) ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Task v2 Block</h1>
- <Taskv2BlockParameters
- prompt={activeBlock.prompt ?? definitionBlock.prompt ?? ""}
- url={activeBlock.url ?? definitionBlock.url}
- maxSteps={definitionBlock.max_steps}
- totpVerificationUrl={definitionBlock.totp_verification_url}
- totpIdentifier={definitionBlock.totp_identifier}
- disableCache={definitionBlock.disable_cache}
- />
- </div>
- </div>
- ) : null}
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Agent Inputs</h1>
- {getOrderedRunParameters(
- workflow?.workflow_definition.parameters,
- parameters,
- ).map(([key, value]) => {
- return (
- <div key={key} className="flex gap-16">
- <span className="w-80 truncate text-lg" title={key}>
- {key}
- </span>
- {typeof value === "string" ||
- typeof value === "number" ||
- typeof value === "boolean" ? (
- <AutoResizingTextarea value={String(value)} readOnly />
- ) : (
- <CodeEditor
- value={JSON.stringify(value, null, 2)}
- readOnly
- language="json"
- minHeight="96px"
- maxHeight="200px"
- className="w-full"
- />
- )}
- </div>
- );
- })}
- {Object.keys(parameters).length === 0 ? (
- <div>No inputs found for this agent</div>
- ) : null}
- <h1 className="text-lg font-bold">Other Agent Inputs</h1>
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-lg">Webhook Callback URL</h1>
- </div>
- <Input value={webhookCallbackUrl ?? ""} readOnly />
- </div>
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-lg">Proxy Location</h1>
- </div>
- <ProxySelector
- value={proxyLocation ?? ProxyLocation.Residential}
- onChange={() => {
- // TODO
- }}
- />
- </div>
- {showExtraHttpHeaders ? (
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-lg">Extra HTTP Headers</h1>
- </div>
- <div className="w-full">
- <KeyValueInput
- value={JSON.stringify(extraHttpHeaders)}
- readOnly={true}
- onChange={() => {}}
- />
- </div>
- </div>
- ) : null}
- {workflowRun.browser_session_id ? (
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-lg">Browser Session ID</h1>
- </div>
- <Input value={workflowRun.browser_session_id} readOnly />
- </div>
- ) : null}
- {(workflowRun.run_with ?? workflow?.run_with) ? (
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-lg">Run With</h1>
- <h2 className="text-base text-muted-foreground">
- Execution mode for this run
- </h2>
- </div>
- <Input
- value={workflowRun.run_with ?? workflow?.run_with ?? ""}
- readOnly
- />
- </div>
- ) : null}
- {workflowRun.max_screenshot_scrolls != null ? (
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-lg">Max Screenshot Scrolls</h1>
- </div>
- <Input
- value={workflowRun.max_screenshot_scrolls.toString()}
- readOnly
- />
- </div>
- ) : null}
- </div>
- </div>
- {workflowRun.task_v2 ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Task 2.0 Inputs</h1>
- <div className="flex gap-16">
- <div className="w-80">
- <h1 className="text-lg">Task 2.0 Prompt</h1>
- <h2 className="text-base text-muted-foreground">
- The original prompt for the task
- </h2>
- </div>
- <AutoResizingTextarea
- value={workflowRun.task_v2.prompt ?? ""}
- readOnly
- />
- </div>
- </div>
- </div>
- ) : null}
- </div>
- );
-}
-
-export { WorkflowPostRunParameters };
diff --git a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunOutput.test.tsx b/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunOutput.test.tsx
deleted file mode 100644
index 61c998e3d..000000000
--- a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunOutput.test.tsx
+++ /dev/null
@@ -1,354 +0,0 @@
-// @vitest-environment jsdom
-
-import { QueryClient, QueryClientProvider } from "@tanstack/react-query";
-import { cleanup, render, screen } from "@testing-library/react";
-import { type ReactNode } from "react";
-import { MemoryRouter } from "react-router-dom";
-import { afterEach, beforeEach, describe, expect, it, vi } from "vitest";
-
-import { Status } from "@/api/types";
-import { WorkflowRun } from "../WorkflowRun";
-import type {
- WorkflowRunBlock,
- WorkflowRunTimelineBlockItem,
-} from "../types/workflowRunTypes";
-import { WorkflowRunOutput } from "./WorkflowRunOutput";
-
-const mocks = vi.hoisted(() => ({
- workflowRun: null as unknown,
- timeline: [] as unknown,
- timelineIsPlaceholder: false,
-}));
-
-vi.mock("@/api/AxiosClient", () => ({ getClient: vi.fn() }));
-vi.mock("@/hooks/useCredentialGetter", () => ({
- useCredentialGetter: () => null,
-}));
-
-vi.mock("../hooks/useWorkflowRunWithWorkflowQuery", () => ({
- useWorkflowRunWithWorkflowQuery: () => ({
- data: mocks.workflowRun,
- isLoading: false,
- }),
-}));
-
-vi.mock("../hooks/useWorkflowRunTimelineQuery", () => ({
- useWorkflowRunTimelineQuery: () => ({
- data: mocks.timeline,
- isLoading: false,
- isPlaceholderData: mocks.timelineIsPlaceholder,
- }),
-}));
-
-vi.mock("../components/CodeEditor", () => ({
- CodeEditor: ({ value }: { value: string }) => (
- <pre data-testid="code-editor">{value}</pre>
- ),
-}));
-
-vi.mock("@/components/SummarizeOutput", () => ({
- SummarizeOutput: () => null,
-}));
-
-vi.mock("@/components/ui/scroll-area", () => ({
- ScrollArea: ({ children }: { children?: ReactNode }) => <div>{children}</div>,
- ScrollAreaViewport: ({ children }: { children?: ReactNode }) => (
- <div>{children}</div>
- ),
-}));
-
-vi.mock("@/hooks/useApiCredential", () => ({
- useApiCredential: () => null,
-}));
-vi.mock("@/hooks/useWorkflowStudioEnabled", () => ({
- useWorkflowStudioEnabled: () => false,
-}));
-vi.mock("../hooks/useCacheKeyValuesQuery", () => ({
- useCacheKeyValuesQuery: () => ({ data: undefined }),
-}));
-vi.mock("../hooks/useBlockScriptsQuery", () => ({
- useBlockScriptsQuery: () => ({ data: undefined }),
-}));
-vi.mock("../hooks/useFallbackEpisodesQuery", () => ({
- useFallbackEpisodesQuery: () => ({ data: undefined }),
-}));
-vi.mock("../hooks/useRefreshOnboardingOnRunCompletion", () => ({
- useRefreshOnboardingOnRunCompletion: () => undefined,
-}));
-vi.mock("./useRunCompletionToast", () => ({
- useRunCompletionToast: () => undefined,
-}));
-vi.mock("./WorkflowRunTimeline", () => ({
- WorkflowRunTimeline: () => null,
-}));
-vi.mock("./WorkflowRunBlockDetail", () => ({
- WorkflowRunBlockDetail: () => null,
-}));
-
-function buildBlock(
- overrides: Partial<WorkflowRunBlock> = {},
-): WorkflowRunBlock {
- return {
- workflow_run_block_id: "wrb_default",
- workflow_run_id: "wr_default",
- parent_workflow_run_block_id: null,
- block_type: "http_request",
- label: null,
- description: null,
- title: null,
- status: Status.Completed,
- failure_reason: null,
- output: null,
- continue_on_failure: false,
- task_id: null,
- url: null,
- navigation_goal: null,
- navigation_payload: null,
- data_extraction_goal: null,
- data_schema: null,
- terminate_criterion: null,
- complete_criterion: null,
- include_action_history_in_verification: null,
- engine: null,
- actions: null,
- created_at: "2026-01-01T00:00:00Z",
- modified_at: "2026-01-01T00:00:00Z",
- duration: null,
- loop_values: null,
- current_value: null,
- current_index: null,
- ...overrides,
- };
-}
-
-function buildBlockItem(block: WorkflowRunBlock): WorkflowRunTimelineBlockItem {
- return {
- type: "block",
- block,
- children: [],
- thought: null,
- created_at: block.created_at,
- modified_at: block.modified_at,
- };
-}
-
-function renderWorkflowRunOutput(activeBlock: WorkflowRunBlock) {
- mocks.timeline = [buildBlockItem(activeBlock)];
- mocks.workflowRun = {
- workflow_run_id: activeBlock.workflow_run_id,
- workflow_title: "Demo workflow",
- outputs: {},
- downloaded_file_urls: [],
- downloaded_files: [],
- };
-
- return render(
- <MemoryRouter
- initialEntries={[
- `/runs/demo?active=${activeBlock.workflow_run_block_id}`,
- ]}
- >
- <WorkflowRunOutput />
- </MemoryRouter>,
- );
-}
-
-beforeEach(() => {
- mocks.workflowRun = null;
- mocks.timeline = [];
- mocks.timelineIsPlaceholder = false;
-});
-
-afterEach(() => {
- cleanup();
-});
-
-describe("WorkflowRunOutput", () => {
- it.each([
- { active: "", hasCurrentBlocks: true, expected: '"current output"' },
- { active: "", hasCurrentBlocks: false, expected: null },
- {
- active: "?active=wrb_cleanup_1",
- hasCurrentBlocks: true,
- expected: '"historical output"',
- },
- ])(
- "scopes default output to the current attempt ($active, $hasCurrentBlocks)",
- ({ active, hasCurrentBlocks, expected }) => {
- const historical = {
- ...buildBlockItem(
- buildBlock({
- workflow_run_block_id: "wrb_cleanup_1",
- label: "cleanup",
- block_type: "code",
- output: { extracted_information: "historical output" },
- }),
- ),
- attempt: 1,
- };
- const current = {
- ...buildBlockItem(
- buildBlock({
- workflow_run_block_id: "wrb_cleanup_2",
- label: "cleanup",
- block_type: "code",
- output: { extracted_information: "current output" },
- }),
- ),
- attempt: 2,
- };
- mocks.timeline = hasCurrentBlocks ? [historical, current] : [historical];
- mocks.workflowRun = {
- workflow_run_id: "wr_default",
- status: Status.Completed,
- attempt: 2,
- workflow: { workflow_definition: { finally_block_label: "cleanup" } },
- };
-
- render(
- <MemoryRouter initialEntries={[`/runs/wr_default${active}`]}>
- <WorkflowRunOutput />
- </MemoryRouter>,
- );
-
- expect(
- screen
- .queryAllByTestId("code-editor")
- .map((editor) => editor.textContent)
- .filter(Boolean),
- ).toEqual(expected ? [expected] : []);
- },
- );
-
- it("renders code block extracted information without the raw output wrapper", () => {
- const block = buildBlock({
- workflow_run_block_id: "wrb_code",
- block_type: "code",
- output: {
- extracted_information: {
- order_id: "ord_123",
- status: "shipped",
- },
- raw_code_output: "debug payload",
- },
- });
-
- renderWorkflowRunOutput(block);
-
- expect(screen.getByText("Extracted Information")).toBeDefined();
- expect(screen.getByText("order_id")).toBeDefined();
- expect(screen.getByText('"ord_123"')).toBeDefined();
- expect(screen.queryByText("raw_code_output")).toBeNull();
- expect(screen.queryByText('"debug payload"')).toBeNull();
- });
-
- it("keeps null code block extracted information in the extraction section", () => {
- const block = buildBlock({
- workflow_run_block_id: "wrb_code_null_extraction",
- block_type: "code",
- output: {
- extracted_information: null,
- raw_code_output: "debug payload",
- },
- });
-
- renderWorkflowRunOutput(block);
-
- expect(screen.getByText("Extracted Information")).toBeDefined();
- expect(screen.getByText("null")).toBeDefined();
- expect(screen.queryByText("raw_code_output")).toBeNull();
- expect(screen.queryByText('"debug payload"')).toBeNull();
- });
-});
-
-describe("WorkflowRunOutput with a retained timeline", () => {
- it("renders no block output while the timeline still belongs to the previous run", () => {
- mocks.timelineIsPlaceholder = true;
- const block = buildBlock({
- workflow_run_block_id: "wrb_retained",
- block_type: "code",
- output: { extracted_information: { order_id: "ord_123" } },
- });
-
- const { container } = renderWorkflowRunOutput(block);
-
- expect(container.textContent).toBe("");
- expect(screen.queryByText('"ord_123"')).toBeNull();
- });
-});
-
-describe("legacy WorkflowRun failure card", () => {
- it("uses the final attempt's code details and finally-block outcome", () => {
- const reason = "CodeBlock failed with ValueError at line 8";
- const body = buildBlock({
- block_type: "code",
- label: "body",
- status: Status.Failed,
- failure_reason: reason,
- });
- const finallyBlock = buildBlock({ label: "cleanup" });
- mocks.workflowRun = {
- workflow_run_id: "wr_retry",
- status: Status.Failed,
- attempt: 2,
- failure_reason: reason,
- task_v2: null,
- workflow: {
- title: "Retry workflow",
- workflow_permanent_id: "wpid_retry",
- workflow_definition: { blocks: [], finally_block_label: "cleanup" },
- },
- };
- mocks.timeline = [
- {
- ...buildBlockItem({
- ...body,
- workflow_run_block_id: "wrb_body_1",
- error_codes: ["runner_unavailable"],
- }),
- attempt: 1,
- },
- {
- ...buildBlockItem({
- ...finallyBlock,
- workflow_run_block_id: "wrb_cleanup_1",
- status: Status.Completed,
- }),
- attempt: 1,
- },
- {
- ...buildBlockItem({
- ...body,
- workflow_run_block_id: "wrb_body_2",
- error_codes: ["user_code_error"],
- }),
- attempt: 2,
- },
- {
- ...buildBlockItem({
- ...finallyBlock,
- workflow_run_block_id: "wrb_cleanup_2",
- status: Status.Failed,
- }),
- attempt: 2,
- },
- ];
-
- render(
- <QueryClientProvider client={new QueryClient()}>
- <MemoryRouter initialEntries={["/runs/wr_retry?embed=true"]}>
- <WorkflowRun />
- </MemoryRouter>
- </QueryClientProvider>,
- );
-
- expect(
- screen.getByText('"Execute on any outcome" block (cleanup) failed.'),
- ).toBeDefined();
- expect(
- screen.getByText("The block's code raised ValueError"),
- ).toBeDefined();
- expect(screen.queryByText(/completed successfully/)).toBeNull();
- expect(screen.queryByText("runner_unavailable")).toBeNull();
- });
-});
diff --git a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunOutput.tsx b/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunOutput.tsx
deleted file mode 100644
index 6ddf485c7..000000000
--- a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunOutput.tsx
+++ /dev/null
@@ -1,393 +0,0 @@
-import { FileIcon } from "@radix-ui/react-icons";
-import { useEffect, useMemo, useRef, useState } from "react";
-import { useSearchParams } from "react-router-dom";
-import { ArtifactDownloadLink } from "@/components/ArtifactDownloadLink";
-import { Button } from "@/components/ui/button";
-import { CodeEditor } from "../components/CodeEditor";
-import { useWorkflowRunWithWorkflowQuery } from "../hooks/useWorkflowRunWithWorkflowQuery";
-import { useActiveWorkflowRunItem } from "./useActiveWorkflowRunItem";
-import {
- getBlockOutputDisplayValue,
- getExtractedInformationDisplayValue,
- isAction,
- isWorkflowRunBlock,
- shouldShowExtractedInformation,
-} from "../types/workflowRunTypes";
-import { JsonExplorer } from "./blockDetail/BlockInspector";
-import { findBlockSurroundingAction } from "./workflowTimelineUtils";
-import { useWorkflowRunTimelineQuery } from "../hooks/useWorkflowRunTimelineQuery";
-import { Status } from "@/api/types";
-import { AutoResizingTextarea } from "@/components/AutoResizingTextarea/AutoResizingTextarea";
-import { SummarizeOutput } from "@/components/SummarizeOutput";
-import { statusIsAFailureType } from "@/routes/tasks/types";
-import {
- pickDownloadedFileFilename,
- getBlockDownloadedFileUrls,
-} from "./blockDownloadedFiles";
-
-function SummaryDisplay({
- summary,
- isStale,
-}: {
- summary: string;
- isStale: boolean;
-}) {
- return (
- <div
- role="status"
- aria-live="polite"
- className="space-y-1 rounded bg-slate-elevation3 p-4"
- >
- {isStale && (
- <p className="text-xs text-muted-foreground">
- Out of date - re-summarize for the current output.
- </p>
- )}
- <p
- className={
- isStale
- ? "whitespace-pre-wrap text-sm text-muted-foreground"
- : "whitespace-pre-wrap text-sm"
- }
- >
- {summary}
- </p>
- </div>
- );
-}
-
-function WorkflowRunOutput() {
- const [searchParams] = useSearchParams();
- const hasExplicitSelection = searchParams.has("active");
- const {
- data: retainedTimeline,
- isLoading: workflowRunTimelineIsLoading,
- isPlaceholderData: timelineIsPlaceholder,
- } = useWorkflowRunTimelineQuery();
- const workflowRunTimeline = timelineIsPlaceholder
- ? undefined
- : retainedTimeline;
- const [activeItem] = useActiveWorkflowRunItem();
- const { data: workflowRun, isPlaceholderData: runIsWithheld } =
- useWorkflowRunWithWorkflowQuery();
- const [blockSummaries, setBlockSummaries] = useState<
- Record<string, { summary: string; identity: string }>
- >({});
- const [outputSummary, setOutputSummary] = useState<{
- summary: string;
- identity: string;
- } | null>(null);
-
- const currentWorkflowRunId = workflowRun?.workflow_run_id;
- const prevRunIdRef = useRef(currentWorkflowRunId);
- useEffect(() => {
- const prev = prevRunIdRef.current;
- prevRunIdRef.current = currentWorkflowRunId;
- if (prev && prev !== currentWorkflowRunId) {
- setBlockSummaries({});
- setOutputSummary(null);
- }
- }, [currentWorkflowRunId]);
-
- // Local override so users on finalized runs (where there is no neutral
- // timeline selection to return to) can still view the full workflow-level
- // file list without clearing their block selection elsewhere on the page.
- const [showAllFilesOverride, setShowAllFilesOverride] = useState(false);
-
- const activeBlock = useMemo(() => {
- if (!workflowRunTimeline) {
- return undefined;
- }
- if (isWorkflowRunBlock(activeItem)) {
- return activeItem;
- }
- if (isAction(activeItem)) {
- return findBlockSurroundingAction(
- workflowRunTimeline,
- activeItem.action_id,
- );
- }
- return undefined;
- }, [workflowRunTimeline, activeItem]);
- const activeBlockId = activeBlock?.workflow_run_block_id;
- useEffect(() => {
- setShowAllFilesOverride(false);
- }, [activeBlockId]);
-
- const outputs = workflowRun?.outputs;
-
- const workflowOutputJsonCompact = useMemo(
- () => (outputs ? JSON.stringify(outputs) : ""),
- [outputs],
- );
-
- const workflowOutputJson = useMemo(
- () => (outputs ? JSON.stringify(outputs, null, 2) : ""),
- [outputs],
- );
-
- const blockOutputJsonCompact = useMemo(() => {
- if (
- !activeBlock ||
- activeBlock.output === undefined ||
- activeBlock.output === null
- ) {
- return "";
- }
- return JSON.stringify(getBlockOutputDisplayValue(activeBlock));
- }, [activeBlock]);
-
- if (workflowRunTimelineIsLoading || runIsWithheld) {
- return <div>Loading...</div>;
- }
-
- if (!workflowRunTimeline) {
- return null;
- }
-
- const showExtractedInformation =
- activeBlock && shouldShowExtractedInformation(activeBlock);
- const extractedInformationValue = activeBlock
- ? getExtractedInformationDisplayValue(activeBlock)
- : null;
- // Only flat/nested objects benefit from the structured tree; scalars, arrays,
- // and null read better as formatted JSON, so keep the CodeEditor for those.
- const extractedInformationIsRecord =
- typeof extractedInformationValue === "object" &&
- extractedInformationValue !== null &&
- !Array.isArray(extractedInformationValue);
-
- const showFailureReason =
- activeBlock &&
- activeBlock.status !== null &&
- (statusIsAFailureType({ status: activeBlock.status }) ||
- activeBlock.status === Status.Canceled);
-
- const allFileUrls = workflowRun?.downloaded_file_urls ?? [];
-
- // Prefer the rich downloaded_files array (carries filename) when the backend
- // sends it; falls back to URL parsing otherwise.
- const filenameByUrl = new Map<string, string>();
- for (const file of workflowRun?.downloaded_files ?? []) {
- if (file.filename) {
- filenameByUrl.set(file.url, file.filename);
- }
- }
-
- // Scope to the surrounding block whenever the user explicitly selected a
- // block or any action inside it, so drilling from a block into its child
- // steps keeps the files section anchored to that block. The helper swaps
- // expired block URLs for matching fresh run-level ones when available.
- const userSelectedBlock = Boolean(hasExplicitSelection && activeBlock);
- const showBlockFiles = userSelectedBlock && !showAllFilesOverride;
- const fileUrls =
- showBlockFiles && activeBlock
- ? getBlockDownloadedFileUrls(activeBlock.output, allFileUrls)
- : allFileUrls;
- const observerOutput = workflowRun?.task_v2?.output;
- const webhookFailureReasonData =
- workflowRun?.task_v2?.webhook_failure_reason ??
- workflowRun?.webhook_failure_reason;
-
- const blockContextKey = activeBlock
- ? `block:${activeBlock.workflow_run_id}:${activeBlock.workflow_run_block_id}`
- : "";
- const blockIdentity = `${blockContextKey}|${blockOutputJsonCompact}`;
- const storedBlockSummary = activeBlock
- ? blockSummaries[activeBlock.workflow_run_block_id]
- : undefined;
- const blockSummaryIsStale =
- !!storedBlockSummary && storedBlockSummary.identity !== blockIdentity;
-
- const runContextKey = `run:${currentWorkflowRunId ?? ""}`;
- const runIdentity = `${runContextKey}|${workflowOutputJsonCompact}`;
- const outputSummaryIsStale =
- !!outputSummary && outputSummary.identity !== runIdentity;
-
- return (
- <div className="space-y-5">
- {webhookFailureReasonData ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Webhook Failure Reason</h1>
- <div className="space-y-2 text-yellow-600">
- {webhookFailureReasonData}
- </div>
- </div>
- </div>
- ) : null}
- {activeBlock ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <div className="flex items-center justify-between">
- <h1 className="text-lg font-bold">Block Outputs</h1>
- {activeBlock.output !== null && !showFailureReason && (
- <SummarizeOutput
- key={blockContextKey}
- contextKey={blockContextKey}
- outputJson={blockOutputJsonCompact}
- workflowTitle={workflowRun?.workflow_title}
- blockLabel={activeBlock.label}
- hasSummary={!!storedBlockSummary}
- onSummary={(summary) =>
- setBlockSummaries((prev) => ({
- ...prev,
- [activeBlock.workflow_run_block_id]: {
- summary,
- identity: blockIdentity,
- },
- }))
- }
- />
- )}
- </div>
- {storedBlockSummary && (
- <SummaryDisplay
- summary={storedBlockSummary.summary}
- isStale={blockSummaryIsStale}
- />
- )}
- {showFailureReason ? (
- <div className="space-y-2">
- <h2>Failure Reason</h2>
- <AutoResizingTextarea
- value={
- activeBlock.status === "canceled"
- ? "This block was cancelled"
- : (activeBlock.failure_reason ?? "")
- }
- readOnly
- />
- </div>
- ) : showExtractedInformation ? (
- <div className="space-y-2">
- <h2>Extracted Information</h2>
- {extractedInformationIsRecord ? (
- <JsonExplorer
- value={extractedInformationValue}
- rootLabel="extracted_information"
- />
- ) : (
- <CodeEditor
- language="json"
- value={JSON.stringify(extractedInformationValue, null, 2)}
- minHeight="96px"
- maxHeight="200px"
- readOnly
- />
- )}
- </div>
- ) : activeBlock.output !== null ? (
- <div className="space-y-2">
- <h2>Output</h2>
- <CodeEditor
- language="json"
- value={JSON.stringify(activeBlock.output, null, 2)}
- minHeight="96px"
- maxHeight="200px"
- readOnly
- />
- </div>
- ) : (
- <div>This block has no outputs</div>
- )}
- </div>
- </div>
- ) : null}
- {observerOutput ? (
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <h1 className="text-lg font-bold">Task 2.0 Output</h1>
- <CodeEditor
- language="json"
- value={JSON.stringify(observerOutput, null, 2)}
- readOnly
- minHeight="96px"
- maxHeight="200px"
- />
- </div>
- </div>
- ) : null}
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <div className="flex items-center justify-between">
- <h1 className="text-lg font-bold">Agent Run Outputs</h1>
- {outputs && (
- <SummarizeOutput
- key={runContextKey}
- contextKey={runContextKey}
- outputJson={workflowOutputJsonCompact}
- workflowTitle={workflowRun?.workflow_title}
- hasSummary={!!outputSummary}
- onSummary={(summary) =>
- setOutputSummary({
- summary,
- identity: runIdentity,
- })
- }
- />
- )}
- </div>
- {outputSummary && (
- <SummaryDisplay
- summary={outputSummary.summary}
- isStale={outputSummaryIsStale}
- />
- )}
- <CodeEditor
- language="json"
- value={workflowOutputJson}
- readOnly
- minHeight="96px"
- maxHeight="200px"
- />
- </div>
- </div>
- <div className="rounded bg-slate-elevation2 p-6">
- <div className="space-y-4">
- <div className="flex items-center justify-between">
- <h1 className="text-lg font-bold">
- {showBlockFiles
- ? "Block Downloaded Files"
- : "Agent Run Downloaded Files"}
- </h1>
- {userSelectedBlock ? (
- <Button
- variant="ghost"
- size="sm"
- className="h-7 px-2.5 text-xs"
- onClick={() => setShowAllFilesOverride((v) => !v)}
- >
- {showAllFilesOverride
- ? "Show block files"
- : "Show all agent files"}
- </Button>
- ) : null}
- </div>
- <div className="space-y-2">
- {fileUrls.length > 0 ? (
- fileUrls.map((url) => {
- const filename = pickDownloadedFileFilename(url, filenameByUrl);
- return (
- <div key={url} title={url} className="flex gap-2">
- <FileIcon className="size-6" />
- <ArtifactDownloadLink
- href={url}
- className="underline underline-offset-4"
- >
- <span>{filename}</span>
- </ArtifactDownloadLink>
- </div>
- );
- })
- ) : (
- <div>No files downloaded</div>
- )}
- </div>
- </div>
- </div>
- </div>
- );
-}
-
-export { WorkflowRunOutput };
diff --git a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunRecording.test.tsx b/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunRecording.test.tsx
deleted file mode 100644
index 88cf7b3aa..000000000
--- a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunRecording.test.tsx
+++ /dev/null
@@ -1,61 +0,0 @@
-// @vitest-environment jsdom
-
-import { cleanup, render } from "@testing-library/react";
-import { MemoryRouter, Route, Routes } from "react-router-dom";
-import { afterEach, describe, expect, it, vi } from "vitest";
-
-import { WorkflowRunRecording } from "./WorkflowRunRecording";
-
-// The component only ever renders under the run-detail route, and the run id it
-// finds there is what separates "still resolving" from "no run in view".
-function renderUnderRunRoute() {
- return render(
- <MemoryRouter initialEntries={["/runs/wr_1"]}>
- <Routes>
- <Route path="/runs/:runId" element={<WorkflowRunRecording />} />
- </Routes>
- </MemoryRouter>,
- );
-}
-
-const mocks = vi.hoisted(() => ({
- workflowRun: undefined as unknown,
- isLoading: false,
- isPlaceholderData: false,
-}));
-
-vi.mock("posthog-js/react", () => ({
- usePostHog: () => ({ capture: vi.fn() }),
-}));
-vi.mock("../hooks/useWorkflowRunWithWorkflowQuery", () => ({
- useWorkflowRunWithWorkflowQuery: () => ({
- data: mocks.workflowRun,
- isLoading: mocks.isLoading,
- isPlaceholderData: mocks.isPlaceholderData,
- }),
-}));
-
-afterEach(() => {
- cleanup();
- mocks.workflowRun = undefined;
- mocks.isLoading = false;
- mocks.isPlaceholderData = false;
-});
-
-describe("WorkflowRunRecording", () => {
- it("does not claim a run has no recording while its payload is still withheld", () => {
- mocks.isPlaceholderData = true;
-
- const { container } = renderUnderRunRoute();
-
- expect(container.textContent).not.toContain("No recording available");
- });
-
- it("states that a resolved run has no recording", () => {
- mocks.workflowRun = { workflow_run_id: "wr_1", recording_urls: [] };
-
- const { container } = renderUnderRunRoute();
-
- expect(container.textContent).toContain("No recording available");
- });
-});
diff --git a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunRecording.tsx b/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunRecording.tsx
deleted file mode 100644
index f0d9c7673..000000000
--- a/skyvern-frontend/src/routes/workflows/workflowRun/WorkflowRunRecording.tsx
+++ /dev/null
@@ -1,90 +0,0 @@
-import { usePostHog } from "posthog-js/react";
-import { ArtifactVideo } from "@/components/ArtifactVideo";
-import { AspectRatio } from "@/components/ui/aspect-ratio";
-import { Skeleton } from "@/components/ui/skeleton";
-import { useFirstParam } from "@/hooks/useFirstParam";
-import { useWorkflowRunWithWorkflowQuery } from "../hooks/useWorkflowRunWithWorkflowQuery";
-import { getRecordingUrls } from "./recordingUrls";
-
-function WorkflowRunRecording() {
- const postHog = usePostHog();
- const workflowRunId = useFirstParam("workflowRunId", "runId");
- const {
- data: workflowRun,
- isLoading,
- isPlaceholderData: runIsPlaceholder,
- } = useWorkflowRunWithWorkflowQuery();
-
- const recordingUrls = getRecordingUrls(workflowRun);
-
- // isPlaceholderData stays true for as long as the query is disabled, so without
- // the run-id term a cleared run leaves the skeleton up permanently.
- if (isLoading || (Boolean(workflowRunId) && runIsPlaceholder)) {
- return (
- <AspectRatio ratio={16 / 9}>
- <Skeleton className="h-full w-full" />
- </AspectRatio>
- );
- }
-
- if (!workflowRun || recordingUrls.length === 0) {
- if (workflowRun?.recording_archived) {
- return (
- <div className="text-muted-foreground">
- This recording has been archived. To request restoration, please
- {/* TODO: add a "Request Restore" button */}
- </div>
- );
- }
- return <div>No recording available for this agent</div>;
- }
-
- const run = workflowRun;
-
- function handlePlay(index: number) {
- postHog.capture("run.recording.viewed", {
- org_id: run.workflow.organization_id,
- run_id: run.workflow_run_id,
- recording_index: index,
- recording_count: recordingUrls.length,
- });
- }
-
- const singleUrl = recordingUrls.length === 1 ? recordingUrls[0] : undefined;
- if (singleUrl) {
- return (
- <ArtifactVideo
- src={singleUrl}
- controls
- preload="metadata"
- className="w-full rounded-md"
- onPlay={() => handlePlay(0)}
- />
- );
- }
-
- return (
- <div className="flex flex-col gap-4">
- {recordingUrls.map((url, index) => (
- <div
- key={index} // presigned URLs change on refetch; list order is stable
- className="flex flex-col gap-2"
- >
- <div className="text-sm text-muted-foreground">
- Recording {index + 1} of {recordingUrls.length}
- </div>
- <ArtifactVideo
- src={url}
- controls
- preload="metadata"
- className="w-full rounded-md"
- onPlay={() => handlePlay(index)}
- />
- </div>
- ))}
- </div>
- );
-}
-
-export { WorkflowRunRecording };
diff --git a/skyvern-frontend/src/store/WorkflowHasChangesStore.ts b/skyvern-frontend/src/store/WorkflowHasChangesStore.ts
index e2fed4aae..b077ff0b1 100644
--- a/skyvern-frontend/src/store/WorkflowHasChangesStore.ts
+++ b/skyvern-frontend/src/store/WorkflowHasChangesStore.ts
@@ -271,6 +271,7 @@ const useWorkflowSave = (opts?: WorkflowSaveOpts) => {
extra_http_headers: extraHttpHeaders,
cdp_connect_headers: cdpConnectHeaders,
run_with: saveData.settings.runWith,
+ browser_type: saveData.settings.browserType ?? null,
cache_key: normalizedKey,
ai_fallback: saveData.settings.aiFallback ?? true,
enable_self_healing: saveData.settings.enableSelfHealing ?? false,
diff --git a/skyvern-frontend/src/util/featureFlags.ts b/skyvern-frontend/src/util/featureFlags.ts
index fd349b816..9a431f20c 100644
--- a/skyvern-frontend/src/util/featureFlags.ts
+++ b/skyvern-frontend/src/util/featureFlags.ts
@@ -25,10 +25,6 @@ export const WORKFLOW_TAGGING_FLAG = "WORKFLOW_TAGGING";
// Not enrolled reads as disabled, so the default stays the flat folders/list.
export const WORKFLOWS_DIRECTORY_TREE_FLAG = "WORKFLOWS_DIRECTORY_TREE";
-// Opt-in (0% base rollout) preview gating the redesigned workflow studio.
-// Not enrolled reads as disabled, so the default stays the legacy editor.
-export const WORKFLOW_STUDIO_FLAG = "workflow_studio_v2";
-
// Gates the optional, server-confirmed onboarding details step.
export const ONBOARDING_QUESTIONNAIRE_FLAG = "onboarding_questionnaire_v1";
@@ -38,9 +34,6 @@ export const ONBOARDING_QUESTIONNAIRE_FLAG = "onboarding_questionnaire_v1";
// /customer (see FeatureFlagProvider), keyed on organization_id like the backend gate.
export const CREDENTIAL_FALLBACK_RETRY_FLAG = "CREDENTIAL_FALLBACK_RETRY";
-// Opt-in preview (0% base): recordings synthesize code blocks instead of agent blocks.
-export const RECORD_BROWSER_CODE_FIRST_FLAG = "record_browser_code_first";
-
export const ONBOARDING_PROGRESS_FLAG = "onboarding_progress_v1";
export const ONBOARDING_TRACK_FLAG = "onboarding_track_v1";
diff --git a/skyvern/cli/mcpb/claude_desktop/package-lock.json b/skyvern/cli/mcpb/claude_desktop/package-lock.json
index ab0df1bf9..d7b8b80cb 100644
--- a/skyvern/cli/mcpb/claude_desktop/package-lock.json
+++ b/skyvern/cli/mcpb/claude_desktop/package-lock.json
@@ -716,9 +716,9 @@
}
},
"node_modules/qs": {
- "version": "6.15.3",
- "resolved": "https://registry.npmjs.org/qs/-/qs-6.15.3.tgz",
- "integrity": "sha512-O9gl3zCl5h5blw1KGUzQKhA5oUXSl8rwUIM5o0S3nCXMliSvy5Dzx7/DJcI+SwgICv+IneSZwhBh1oSyEHA71A==",
+ "version": "6.16.0",
+ "resolved": "https://registry.npmjs.org/qs/-/qs-6.16.0.tgz",
+ "integrity": "sha512-h6fhOIaRrID2CbEY2fqs+7t+UXZo+MLAnU5gRIq85uFtdiUPCdsApMlHhXogKVM4HM2DVbIjGNTTYH2OcmP1vA==",
"license": "BSD-3-Clause",
"dependencies": {
"es-define-property": "^1.0.1",
diff --git a/skyvern/exceptions.py b/skyvern/exceptions.py
index b656203d4..5eb9a7af4 100644
--- a/skyvern/exceptions.py
+++ b/skyvern/exceptions.py
@@ -780,6 +780,29 @@ class BlockedNavigationDestination(FailedToNavigateToUrl):
super().__init__(url=url, error_message=f"blocked navigation destination: {reason}")
+# Leads the error_message below, and so the run's failure_reason. A consumer that only ever sees
+# that string matches on this instead of restating the prose, since the egress-attributable browser
+# error code stays in the message and would otherwise classify the failure as ours.
+NO_ADDRESS_RECORD_NAV_ERROR_MARKER = "has no DNS address record"
+
+
+class UnresolvableNavigationHost(FailedToNavigateToUrl):
+ """The navigation target's host has no DNS address record, so no egress of ours can reach it.
+
+ A browser behind a proxy hands the hostname to the proxy (HTTP CONNECT, SOCKS remote DNS), so a
+ target whose record is gone surfaces as the proxy failing to open a tunnel rather than as
+ ERR_NAME_NOT_RESOLVED. The distinct type is what keeps that borrowed error code from being read
+ as our own egress failing; it subclasses FailedToNavigateToUrl so the failure stays terminal.
+ """
+
+ def __init__(self, url: str, host: str, error_message: str) -> None:
+ self.host = host
+ super().__init__(
+ url=url,
+ error_message=f"{host} {NO_ADDRESS_RECORD_NAV_ERROR_MARKER}: {error_message}",
+ )
+
+
class FailedToReloadPage(SkyvernException):
def __init__(self, url: str, error_message: str) -> None:
self.url = url
diff --git a/skyvern/forge/agent.py b/skyvern/forge/agent.py
index 3c5d61b10..e3d6b69f4 100644
--- a/skyvern/forge/agent.py
+++ b/skyvern/forge/agent.py
@@ -2298,6 +2298,20 @@ class ForgeAgent:
)
await browser_state.reload_page(page=pinned)
+ async def _restore_page_url(page: Any, url: str) -> None:
+ # Not `reload_page`: the tab is blanked in place, so reloading it reloads `about:blank`.
+ await browser_state.navigate_to_url(page=page, url=url)
+
+ def _download_attempts() -> int | None:
+ # The interceptor counts an attempt before it checks for a byte-identical file already on
+ # disk, so this still moves for a deduplicated download -- which leaves no filesystem
+ # trace at all. Attached dynamically to the Playwright context, hence the getattr chain
+ # (the same access `real_browser_state` uses); absent interceptor simply yields None.
+ context = browser_state.browser_context
+ interceptor = getattr(context, "_skyvern_cdp_download_interceptor", None) if context else None
+ attempts = getattr(interceptor, "_download_index", None)
+ return attempts if isinstance(attempts, int) else None
+
# Whether the control the run CLICKED is still in flight. Scoped to that one control on
# purpose: "is anything on this page busy?" strands a finished run on an unrelated upload
# widget or a stale modal the app left in the DOM. Narrow for the same reason -- every
@@ -2363,6 +2377,8 @@ class ForgeAgent:
page_fingerprint=_page_fingerprint,
page_probe=_page_probe,
reload_page=_reload_page,
+ restore_page_url=_restore_page_url,
+ download_attempts=_download_attempts,
block_type=str(task_block.block_type) if task_block is not None else None,
# Unfenced across both populations, unlike the settle probe above: that fence exists
# to keep a RENDERING wait off the bare arm, and this asks a different question. The
diff --git a/skyvern/forge/sdk/copilot/agent.py b/skyvern/forge/sdk/copilot/agent.py
index 6c6dc9613..7159c300f 100644
--- a/skyvern/forge/sdk/copilot/agent.py
+++ b/skyvern/forge/sdk/copilot/agent.py
@@ -35,6 +35,7 @@ from agents.items import ToolCallItem
from litellm.exceptions import NotFoundError as LiteLLMNotFoundError
from pydantic import JsonValue, TypeAdapter, ValidationError
+from skyvern.exceptions import NO_ADDRESS_RECORD_NAV_ERROR_MARKER
from skyvern.forge import app
from skyvern.forge.prompts import prompt_engine
from skyvern.forge.sdk.copilot.blocker_signal import (
@@ -294,7 +295,7 @@ def _render_code_only_browser_authoring_prompt() -> str:
return (
_code_only_browser_authoring_prompt()
- + "\n\nWhen a SYNTHESIZED CODE BLOCK is offered to you, it already encodes the page\n"
+ + "\n\nWhen a scouting tool offers a SYNTHESIZED CODE BLOCK, it already encodes the page\n"
"interactions you scouted as deterministic Playwright. Persist that block VERBATIM\n"
"via update_workflow / update_and_run_blocks — do not rewrite, reorder, or\n"
"re-derive its locators. Only hand-author the steps it does not cover, such as the\n"
@@ -3374,6 +3375,13 @@ _PROXY_TRANSPORT_NAV_ERROR_CODES = (
def _non_retriable_nav_reply(error_message: str) -> str:
_, separator, machine_error = error_message.rpartition(". Error message: ")
classification_source = machine_error if separator else error_message
+ # Ahead of the proxy-transport branch: a dead domain reaches us wearing a proxy error code, and
+ # answering it with "contact Skyvern Support" sends the user to us over a URL only they can fix.
+ if NO_ADDRESS_RECORD_NAV_ERROR_MARKER in classification_source:
+ return (
+ f"The site's domain has no DNS record, so it cannot be reached from any network. "
+ f"Error: {error_message}. Please check the URL for a typo, or confirm the site is still online."
+ )
if any(code in classification_source for code in _PROXY_TRANSPORT_NAV_ERROR_CODES):
return (
f"The site could not be reached through Skyvern's browser network. Error: {error_message}. "
diff --git a/skyvern/forge/sdk/copilot/code_block_synthesis.py b/skyvern/forge/sdk/copilot/code_block_synthesis.py
index c5de76ca0..fc9d345a2 100644
--- a/skyvern/forge/sdk/copilot/code_block_synthesis.py
+++ b/skyvern/forge/sdk/copilot/code_block_synthesis.py
@@ -284,8 +284,12 @@ def credential_scout_gap(
return ScoutGap(missing_fields=missing_fields, missing_submit=missing_submit)
-_ENTRY_TARGET_TOOLS = frozenset({"click", "type_text", CREDENTIAL_FILL_TOOL_NAME, "select_option", "press_key"})
-_DURABLE_FALLBACK_ENTRY_TARGET_TOOLS = frozenset({"type_text", CREDENTIAL_FILL_TOOL_NAME, "select_option"})
+_ENTRY_TARGET_TOOLS = frozenset(
+ {"click", "type_text", CREDENTIAL_FILL_TOOL_NAME, "select_option", "upload_file", "press_key"}
+)
+_DURABLE_FALLBACK_ENTRY_TARGET_TOOLS = frozenset(
+ {"type_text", CREDENTIAL_FILL_TOOL_NAME, "select_option", "upload_file"}
+)
_OPTIONAL_DISMISSAL_NAME_PATTERN = re.compile(
r"\b(?:accept|agree|allow|consent|cookies?|decline|reject|refuse|dismiss|got it|no thanks)\b|^(?:ok|okay)$",
re.I,
@@ -2664,6 +2668,26 @@ def synthesize_code_block(
lines.append(f"{action_indent}await page.wait_for_load_state({_py_str(_DOMCONTENTLOADED)})")
record_emission(trajectory_index, tool_name, "select_option", locator, line_start=line_start)
append_step(f"Select {value} in {_step_target(interaction)}", "select_option", line_start)
+ elif tool_name == "upload_file":
+ selector = str(interaction.get("selector") or "").strip()
+ if not selector:
+ notes.append("dropped a file upload with no durable selector")
+ diagnostics.dropped_interactions.append(
+ {"trajectory_index": trajectory_index, "tool_name": tool_name, "reason_code": "missing_selector"}
+ )
+ continue
+ parameter_name = str(interaction.get("parameter_name") or "upload_file")
+ param_key = _unique_key(_safe_param_base(parameter_name), used_param_keys)
+ parameters.append({"key": param_key, "workflow_parameter_type": "file_url"})
+ lines.append(f"{action_indent}await attach_authorized_file(page, {param_key}, {_py_str(selector)})")
+ record_emission(
+ trajectory_index,
+ tool_name,
+ "attach_authorized_file",
+ locator,
+ line_start=line_start,
+ )
+ append_step(f"Upload a file to {_step_target(interaction)}", "upload_file", line_start)
elif tool_name == "hover" and not strict_selectors:
# Non-strict only: recording trajectories carry deliberate hovers; the
# strict-imposition envelope keeps treating hover as unsupported.
diff --git a/skyvern/forge/sdk/db/repositories/workflow_parameters.py b/skyvern/forge/sdk/db/repositories/workflow_parameters.py
index 6fa5866e4..11bbce413 100644
--- a/skyvern/forge/sdk/db/repositories/workflow_parameters.py
+++ b/skyvern/forge/sdk/db/repositories/workflow_parameters.py
@@ -6,7 +6,7 @@ from datetime import datetime, timedelta, timezone
from typing import Any
import structlog
-from sqlalchemy import cast, func, select
+from sqlalchemy import cast, func, or_, select
from sqlalchemy.dialects.postgresql import JSONB, JSONPATH
from sqlalchemy.dialects.postgresql import insert as pg_insert
from sqlalchemy.ext.asyncio import AsyncSession
@@ -1487,15 +1487,35 @@ class WorkflowParametersRepository(BaseRepository):
self,
organization_id: str,
workflow_permanent_id: str,
+ request_cancel_token: str | None = None,
) -> WorkflowCopilotChat | None:
async with self.Session() as session:
query = (
select(WorkflowCopilotChatModel)
.filter(WorkflowCopilotChatModel.organization_id == organization_id)
.filter(WorkflowCopilotChatModel.workflow_permanent_id == workflow_permanent_id)
- .order_by(WorkflowCopilotChatModel.created_at.desc())
- .limit(1)
)
+ if request_cancel_token is not None:
+ completed_turn = (
+ select(WorkflowCopilotChatMessageModel.workflow_copilot_chat_message_id)
+ .where(
+ WorkflowCopilotChatMessageModel.workflow_copilot_chat_id
+ == WorkflowCopilotChatModel.workflow_copilot_chat_id
+ )
+ .where(WorkflowCopilotChatMessageModel.organization_id == organization_id)
+ .where(
+ WorkflowCopilotChatMessageModel.turn_outcome["request_cancel_token"].as_string()
+ == request_cancel_token
+ )
+ .exists()
+ )
+ pending_turn = func.jsonb_path_exists(
+ cast(WorkflowCopilotChatModel.pending_turns, JSONB),
+ cast("$.* ? (@.cancel_token == $token)", JSONPATH),
+ func.jsonb_build_object("token", request_cancel_token),
+ )
+ query = query.filter(or_(pending_turn, completed_turn))
+ query = query.order_by(WorkflowCopilotChatModel.created_at.desc()).limit(1)
chat = (await session.scalars(query)).first()
if not chat:
return None
diff --git a/skyvern/forge/sdk/routes/browser_sessions.py b/skyvern/forge/sdk/routes/browser_sessions.py
index e6caa33d4..97c5d346d 100644
--- a/skyvern/forge/sdk/routes/browser_sessions.py
+++ b/skyvern/forge/sdk/routes/browser_sessions.py
@@ -56,6 +56,7 @@ from skyvern.schemas.browser_sessions import (
)
from skyvern.schemas.proxy_pinning import should_generate_proxy_session_id
from skyvern.schemas.runs import ProxyLocation
+from skyvern.services.browser_recording.session_registry import interpretation_registry
from skyvern.webeye.schemas import BrowserSessionResponse
LOG = structlog.get_logger(__name__)
@@ -673,17 +674,40 @@ async def process_recording(
if not browser_session:
raise HTTPException(status_code=404, detail=f"Browser session {browser_session_id} not found")
+ # Record Browser now emits Code blocks exclusively. Reject processing before
+ # producing a workflow that this organization cannot execute.
+ await app.AGENT_FUNCTION.validate_code_block(current_org.organization_id)
+
+ recorded_actions = interpretation_registry.get_finalized_actions(
+ interpretation_session_id=recording_request.interpretation_session_id,
+ browser_session_id=browser_session_id,
+ organization_id=current_org.organization_id,
+ workflow_permanent_id=recording_request.workflow_permanent_id,
+ )
+ if recorded_actions is None and recording_request.interpretation_session_id is not None:
+ # Done can reach this route while the websocket is still flushing its
+ # final snapshot. Join that stop task so a successful chunk fallback
+ # cannot leave a late finalized-action cache entry behind.
+ await interpretation_registry.stop_session(browser_session_id)
+ recorded_actions = interpretation_registry.get_finalized_actions(
+ interpretation_session_id=recording_request.interpretation_session_id,
+ browser_session_id=browser_session_id,
+ organization_id=current_org.organization_id,
+ workflow_permanent_id=recording_request.workflow_permanent_id,
+ )
blocks, parameters, recording_id, evidence = await app.BROWSER_SESSION_RECORDING_SERVICE.process_recording(
organization_id=current_org.organization_id,
browser_session_id=browser_session_id,
compressed_chunks=recording_request.compressed_chunks,
workflow_permanent_id=recording_request.workflow_permanent_id,
draft_steps=recording_request.draft_steps,
- code_first=recording_request.code_first,
+ recorded_actions=recorded_actions,
supports_credential_tokens=recording_request.supports_credential_tokens,
recording_attempt_id=recording_request.recording_attempt_id,
interpretation_session_id=recording_request.interpretation_session_id,
)
+ if recorded_actions is not None:
+ interpretation_registry.discard_finalized_actions(recording_request.interpretation_session_id)
return ProcessBrowserSessionRecordingResponse(
recording_id=recording_id,
diff --git a/skyvern/forge/sdk/routes/streaming/cdp_input.py b/skyvern/forge/sdk/routes/streaming/cdp_input.py
index 9de5b834a..887bd35ad 100644
--- a/skyvern/forge/sdk/routes/streaming/cdp_input.py
+++ b/skyvern/forge/sdk/routes/streaming/cdp_input.py
@@ -3,10 +3,13 @@ CDP input channel for interactive browser control via Chrome DevTools Protocol.
"""
import asyncio
+import base64
+import binascii
import dataclasses
import json
import time
import typing as t
+from pathlib import PurePath
import structlog
from fastapi import WebSocket, WebSocketDisconnect
@@ -39,7 +42,16 @@ from skyvern.webeye.navigation import revalidate_redirect_chain, validate_naviga
LOG = structlog.get_logger()
_INPUT_KIND_LABELS = frozenset(
- {"mouseEvent", "keyEvent", "wheelEvent", "navigateEvent", "goBackEvent", "goForwardEvent", "reloadEvent"}
+ {
+ "mouseEvent",
+ "keyEvent",
+ "wheelEvent",
+ "fileUpload",
+ "navigateEvent",
+ "goBackEvent",
+ "goForwardEvent",
+ "reloadEvent",
+ }
)
_LATENCY_BUCKETS_SECONDS = [0.001, 0.002, 0.005, 0.01, 0.02, 0.03, 0.045, 0.06, 0.1, 0.25, 0.5, 1.0, 2.5, 5.0]
_meter = metrics.get_meter("skyvern.live_view")
@@ -72,6 +84,7 @@ _NAVIGATION_RESET_TIMEOUT_MS = 5_000
_TARGET_CLOSED_ERROR_TYPES = frozenset({"TargetClosedError", "CdpTargetClosedError"})
_PIPELINED_INPUT_KINDS = frozenset({"mouseEvent", "wheelEvent", "keyEvent"})
_MAX_IN_FLIGHT_INPUT_DISPATCHES = 32
+_MAX_FILE_UPLOAD_BYTES = 10 * 1024 * 1024
def _input_kind_label(kind: object) -> str:
@@ -450,6 +463,42 @@ async def _dispatch_history_event(
await cdp_session.send("Page.navigateToHistoryEntry", {"entryId": entry_id})
+async def _dispatch_file_upload_event(page: object, msg: dict, websocket: WebSocket) -> None:
+ file_name = msg.get("fileName")
+ mime_type = msg.get("mimeType")
+ encoded = msg.get("content")
+ if (
+ not isinstance(file_name, str)
+ or not file_name
+ or len(file_name) > 255
+ or PurePath(file_name).name != file_name
+ or "/" in file_name
+ or "\\" in file_name
+ or not isinstance(mime_type, str)
+ or not mime_type
+ or len(mime_type) > 255
+ or not isinstance(encoded, str)
+ or len(encoded) > ((_MAX_FILE_UPLOAD_BYTES + 2) // 3) * 4
+ ):
+ await websocket.send_json({"kind": "file-upload-error", "reason": "invalid_file"})
+ return
+ try:
+ content = base64.b64decode(encoded, validate=True)
+ except (ValueError, binascii.Error):
+ await websocket.send_json({"kind": "file-upload-error", "reason": "invalid_file"})
+ return
+ if len(content) > _MAX_FILE_UPLOAD_BYTES:
+ await websocket.send_json({"kind": "file-upload-error", "reason": "file_too_large"})
+ return
+
+ locator = page.locator('input[type="file"]') # type: ignore[attr-defined]
+ if await locator.count() != 1:
+ await websocket.send_json({"kind": "file-upload-error", "reason": "expected_one_file_input"})
+ return
+ await locator.set_input_files({"name": file_name, "mimeType": mime_type, "buffer": content})
+ await websocket.send_json({"kind": "file-upload-complete"})
+
+
async def _dispatch_event(
cdp_session: CDPSession,
page: object,
@@ -459,6 +508,10 @@ async def _dispatch_event(
log_id_value: str,
websocket: WebSocket,
) -> None:
+ if kind == "fileUpload":
+ await _dispatch_file_upload_event(page, msg, websocket)
+ return
+
if kind == "navigateEvent":
await _dispatch_navigate_event(page, msg, log_id_key, log_id_value, websocket)
return
diff --git a/skyvern/forge/sdk/routes/streaming/channels/message.py b/skyvern/forge/sdk/routes/streaming/channels/message.py
index 994464a30..d9a381e6b 100644
--- a/skyvern/forge/sdk/routes/streaming/channels/message.py
+++ b/skyvern/forge/sdk/routes/streaming/channels/message.py
@@ -147,6 +147,9 @@ class MessageInBeginExfiltration(Message):
@dataclasses.dataclass
class MessageInEndExfiltration(Message):
kind: t.Literal[MessageKind.END_EXFILTRATION] = MessageKind.END_EXFILTRATION
+ discard: bool = False
+ interpretation_session_id: str | None = None
+ workflow_permanent_id: str | None = None
@dataclasses.dataclass
@@ -352,7 +355,15 @@ def reify_channel_message(data: dict) -> MessageIn:
text = data.get("text") or ""
return MessageInClipboardPaste(text=text)
case MessageKind.END_EXFILTRATION:
- return MessageInEndExfiltration()
+ interpretation_session_id = data.get("interpretation_session_id")
+ workflow_permanent_id = data.get("workflow_permanent_id")
+ return MessageInEndExfiltration(
+ discard=data.get("discard") is True,
+ interpretation_session_id=(
+ interpretation_session_id if isinstance(interpretation_session_id, str) else None
+ ),
+ workflow_permanent_id=workflow_permanent_id if isinstance(workflow_permanent_id, str) else None,
+ )
case MessageKind.RECORDING_CAPTURE_PAUSE:
return MessageInRecordingCapturePause()
case MessageKind.RECORDING_CAPTURE_RESUME:
@@ -844,8 +855,26 @@ async def loop_stream_messages(message_channel: MessageChannel) -> None:
exfiltration_channel = None
if live_interpretation_browser_session_id:
- await interpretation_registry.stop_session(live_interpretation_browser_session_id)
+ if message.discard:
+ await interpretation_registry.stop_session(
+ live_interpretation_browser_session_id,
+ retain_finalized_actions=False,
+ )
+ else:
+ await interpretation_registry.stop_session(live_interpretation_browser_session_id)
live_interpretation_browser_session_id = None
+ elif (
+ message.discard
+ and message.interpretation_session_id
+ and message.workflow_permanent_id
+ and message_channel.browser_session
+ ):
+ interpretation_registry.discard_finalized_actions_if_owned(
+ interpretation_session_id=message.interpretation_session_id,
+ browser_session_id=message_channel.browser_session.persistent_browser_session_id,
+ organization_id=message_channel.organization_id,
+ workflow_permanent_id=message.workflow_permanent_id,
+ )
case MessageKind.RECORDING_CAPTURE_PAUSE:
if exfiltration_channel is not None:
diff --git a/skyvern/forge/sdk/routes/workflow_copilot.py b/skyvern/forge/sdk/routes/workflow_copilot.py
index c51cdd111..63be79bd2 100644
--- a/skyvern/forge/sdk/routes/workflow_copilot.py
+++ b/skyvern/forge/sdk/routes/workflow_copilot.py
@@ -192,18 +192,20 @@ DIAGNOSE_RUN_TURN_MESSAGE = "Diagnose run {run_id} and repair the workflow."
# WorkflowCopilotChat.tsx echoes it so a history reload does not reword the turn.
REFINE_RECORDING_RECEIPT = "Refine the recording ({action_count} actions) into a reusable workflow."
REFINE_RECORDING_INSTRUCTION = (
- "From the recorded browser evidence (provided separately as untrusted observations), infer the user's objective "
- "and refine the current literal replay draft into the simplest reusable workflow that reliably achieves it.\n\n"
- "Treat recorded actions as evidence, not a mandatory checklist. Remove an action only if it is clearly "
- "exploratory, corrective, duplicated, or incidental and no later interaction plausibly depends on the state it "
- "creates. Otherwise preserve its position in the sequence, especially when it navigates, reveals controls, opens "
- "menus or dialogs, switches context, submits forms, or advances a multi-step flow. You may replace it with a "
- "simpler equivalent that produces the same required state. If the dependency is uncertain and the workflow cannot "
- "be tested in this turn, retain the action.\n\n"
- "Generalize concrete user-entered values into workflow parameters. Prefer selectors supported by the recorded "
- "evidence and add appropriate validation and extraction logic.\n\n"
- "The program must fail closed: raise when the objective is not verifiably achieved and return output only "
- "on success. Write the workflow with the workflow tool in this turn; never answer in prose."
+ "From recorded browser evidence, infer the user's objective and refine the literal replay into the simplest "
+ "reusable workflow that achieves it.\n\n"
+ "Treat recorded actions as evidence, not a checklist. Remove exploratory, corrective, duplicated, or incidental "
+ "actions only when neither the objective nor any retained interaction depends on the state they create. You may "
+ "collapse a multi-action path into a direct action only when the replacement reaches the same required browser "
+ "state and does not depend on state created by removed actions.\n\n"
+ "For every required interaction, preserve its recorded semantics and actionable target by default. Do not "
+ "change the interaction primitive or retarget to an associated, nested, or semantically similar element unless "
+ "the replacement is successfully tested and produces the same observable state. If a dependency or equivalence "
+ "is uncertain and cannot be tested in this turn, retain the recorded interaction.\n\n"
+ "Generalize user-entered values into workflow parameters. Improve selectors only while targeting the same "
+ "actionable surface. Add validation and any extraction required by the objective.\n\n"
+ "Fail closed: raise unless the objective is verifiably achieved, and return output only on success. Write the "
+ "workflow with the workflow tool before replying; do not substitute a prose-only answer."
)
@@ -1002,6 +1004,7 @@ async def _persist_turn_messages(
narrative_payload: TurnNarrativePayload | None,
sender: WorkflowCopilotChatSender,
attached_files: list[CopilotAttachedFile] | None = None,
+ request_cancel_token: str | None = None,
) -> WorkflowCopilotChatMessage | None:
"""The only writer of a copilot turn's chat rows; idempotent per ``turn_id``.
@@ -1011,21 +1014,33 @@ async def _persist_turn_messages(
if turn_id is not None and turn_outcome is not None and not turn_outcome.copilot_turn_id:
turn_outcome = turn_outcome.model_copy(update={"copilot_turn_id": turn_id})
+ pending_user_message_id: str | None = None
if turn_id is not None:
stored_chat = await app.DATABASE.workflow_params.get_workflow_copilot_chat_by_id(
chat.organization_id, chat.workflow_copilot_chat_id
)
if isinstance(stored_chat, WorkflowCopilotChat):
pending = stored_chat.pending_turns.get(turn_id)
- if pending is not None and pending.question_interactions:
- if narrative_payload is None:
- narrative_payload = _make_error_narrative_payload(turn_id, None, assistant_content)
- narrative_payload["questionInteractions"] = [
- item.model_dump(mode="json") for item in pending.question_interactions
- ]
+ if pending is not None:
+ pending_user_message_id = pending.user_message_id
+ request_cancel_token = pending.cancel_token or request_cancel_token
+ if pending.question_interactions:
+ if narrative_payload is None:
+ narrative_payload = _make_error_narrative_payload(turn_id, None, assistant_content)
+ narrative_payload["questionInteractions"] = [
+ item.model_dump(mode="json") for item in pending.question_interactions
+ ]
+
+ if turn_outcome is not None:
+ turn_outcome = turn_outcome.model_copy(
+ update={
+ **({"user_message_id": pending_user_message_id} if pending_user_message_id is not None else {}),
+ **({"request_cancel_token": request_cancel_token} if request_cancel_token is not None else {}),
+ }
+ )
if not user_row_already_persisted:
- await asyncio.shield(
+ user_row = await asyncio.shield(
app.DATABASE.workflow_params.create_workflow_copilot_chat_message(
organization_id=chat.organization_id,
workflow_copilot_chat_id=chat.workflow_copilot_chat_id,
@@ -1035,6 +1050,9 @@ async def _persist_turn_messages(
attached_files=attached_files,
)
)
+ user_message_id = getattr(user_row, "workflow_copilot_chat_message_id", None)
+ if turn_outcome is not None and turn_outcome.user_message_id is None and user_message_id is not None:
+ turn_outcome = turn_outcome.model_copy(update={"user_message_id": user_message_id})
assistant_message: WorkflowCopilotChatMessage | None = None
existing = await _assistant_row_for_turn(chat, turn_id) if turn_id is not None else None
@@ -1044,7 +1062,22 @@ async def _persist_turn_messages(
and turn_outcome is not None
and turn_outcome.terminal_reason != INTERRUPTED_TERMINAL_REASON
)
- if existing is not None and superseding_recovery:
+ if existing is not None and superseding_recovery and turn_outcome is not None:
+ replacement_updates: dict[str, str] = {}
+ if (
+ turn_outcome.user_message_id is None
+ and existing.turn_outcome is not None
+ and existing.turn_outcome.user_message_id is not None
+ ):
+ replacement_updates["user_message_id"] = existing.turn_outcome.user_message_id
+ if (
+ turn_outcome.request_cancel_token is None
+ and existing.turn_outcome is not None
+ and existing.turn_outcome.request_cancel_token is not None
+ ):
+ replacement_updates["request_cancel_token"] = existing.turn_outcome.request_cancel_token
+ if replacement_updates:
+ turn_outcome = turn_outcome.model_copy(update=replacement_updates)
# Recovery reached this turn first and wrote an interrupted row. The turn then finished,
# so its real reply is the truth and replaces that row rather than being dropped.
LOG.info(
@@ -1163,6 +1196,7 @@ async def _persist_interrupted_turn(
sender: WorkflowCopilotChatSender = WorkflowCopilotChatSender.USER,
effective_mode: PersistedCopilotComposerMode | None = None,
code_available: bool | None = None,
+ request_cancel_token: str | None = None,
) -> None:
"""Write the assistant row for a turn that stopped before it finished.
@@ -1200,6 +1234,7 @@ async def _persist_interrupted_turn(
code_available=code_available,
),
narrative_payload=narrative_payload,
+ request_cancel_token=request_cancel_token,
)
@@ -1220,6 +1255,7 @@ async def _persist_cancel_turn(
sender: WorkflowCopilotChatSender = WorkflowCopilotChatSender.USER,
effective_mode: PersistedCopilotComposerMode | None = None,
code_available: bool | None = None,
+ request_cancel_token: str | None = None,
) -> None:
"""Persist a cancelled turn and emit a terminal SSE response frame.
@@ -1381,6 +1417,7 @@ async def _persist_cancel_turn(
global_llm_context=updated_global_llm_context,
turn_outcome=turn_outcome,
narrative_payload=narrative_payload,
+ request_cancel_token=request_cancel_token,
)
response_time = assistant_message.created_at if assistant_message else datetime.now(UTC)
try:
@@ -1520,6 +1557,7 @@ async def _finalise_normal_turn(
turn_outcome=agent_result.turn_outcome,
narrative_payload=narrative_payload,
attached_files=attached_files,
+ request_cancel_token=chat_request.cancel_token,
)
response_data = {
@@ -2000,6 +2038,11 @@ async def _new_copilot_chat_post(
code_available=current_code_available,
turn_id=turn_id,
)
+ if agent_result is not None:
+ # Without this turn's candidate token the recovered result reads as owning whatever
+ # the row holds now, so its clear would take a candidate another turn published.
+ recovered_result.proposal_owner_turn_id = agent_result.proposal_owner_turn_id
+ recovered_result.proposal_revision = agent_result.proposal_revision
LOG.error(
translated_log_message,
organization_id=organization.organization_id,
@@ -2374,6 +2417,7 @@ async def _new_copilot_chat_post(
cancel_source=user_cancel_source[0],
effective_mode=effective_mode,
code_available=current_code_available,
+ request_cancel_token=chat_request.cancel_token,
)
terminal_frame_emitted = True
capture_code_mode_opt_out_after_persist()
@@ -2470,6 +2514,7 @@ async def _new_copilot_chat_post(
sender=_turn_opener_sender(chat_request),
effective_mode=effective_mode,
code_available=current_code_available,
+ request_cancel_token=chat_request.cancel_token,
)
)
except Exception as writer_exc:
@@ -2525,6 +2570,7 @@ async def _new_copilot_chat_post(
sender=_turn_opener_sender(chat_request),
effective_mode=effective_mode,
code_available=current_code_available,
+ request_cancel_token=chat_request.cancel_token,
)
)
raise
@@ -2931,6 +2977,7 @@ async def _recover_interrupted_copilot_turn(
idempotency_digest=entry.idempotency_digest,
effective_mode=entry.copilot_effective_mode,
code_available=entry.copilot_code_available,
+ request_cancel_token=entry.cancel_token,
)
LOG.info(
"Recovered an interrupted copilot turn on read",
@@ -3069,6 +3116,7 @@ async def workflow_copilot_chat_history(
credential_recovery_token: str | None = Header(None, alias="X-Copilot-Credential-Recovery-Token"),
workflow_permanent_id: str | None = None,
workflow_copilot_chat_id: str | None = None,
+ request_cancel_token: str | None = None,
organization: Organization = Depends(org_auth_service.get_current_org),
) -> WorkflowCopilotChatHistoryResponse:
if workflow_copilot_chat_id:
@@ -3080,6 +3128,7 @@ async def workflow_copilot_chat_history(
chat = await app.DATABASE.workflow_params.get_latest_workflow_copilot_chat(
organization_id=organization.organization_id,
workflow_permanent_id=workflow_permanent_id,
+ request_cancel_token=request_cancel_token,
)
else:
raise HTTPException(
@@ -3102,6 +3151,22 @@ async def workflow_copilot_chat_history(
proposed_workflow, proposed_workflow_metadata, proposed_workflow_run = await _history_proposal_state(
chat, organization.organization_id
)
+ request_turn_id = None
+ if chat is not None and request_cancel_token is not None:
+ request_turn_id = next(
+ (turn_id for turn_id, entry in chat.pending_turns.items() if entry.cancel_token == request_cancel_token),
+ None,
+ )
+ if request_turn_id is None:
+ request_turn_id = next(
+ (
+ message.turn_outcome.copilot_turn_id
+ for message in reversed(chat_messages)
+ if message.turn_outcome is not None
+ and message.turn_outcome.request_cancel_token == request_cancel_token
+ ),
+ None,
+ )
return WorkflowCopilotChatHistoryResponse(
pending_credential_requests=await pending_credential_requests(
organization.organization_id,
@@ -3112,6 +3177,7 @@ async def workflow_copilot_chat_history(
if chat
else [],
workflow_copilot_chat_id=chat.workflow_copilot_chat_id if chat else None,
+ request_turn_id=request_turn_id,
question_interactions=list(
{
item.interaction_id: item
@@ -3138,7 +3204,11 @@ async def workflow_copilot_chat_history(
),
None,
),
- chat_history=await _history_with_resolved_attachments(chat_messages, organization.organization_id),
+ chat_history=await _history_with_resolved_attachments(
+ chat_messages,
+ organization.organization_id,
+ chat.pending_turns if chat else None,
+ ),
proposed_workflow=proposed_workflow,
proposed_workflow_metadata=proposed_workflow_metadata,
proposed_workflow_run=proposed_workflow_run,
@@ -3513,14 +3583,16 @@ async def workflow_copilot_apply_proposed_workflow(
async def _history_with_resolved_attachments(
- messages: list[WorkflowCopilotChatMessage], organization_id: str
+ messages: list[WorkflowCopilotChatMessage],
+ organization_id: str,
+ pending_turns: dict[str, CopilotPendingTurn] | None = None,
) -> list[WorkflowCopilotChatHistoryMessage]:
"""Re-resolve every attachment this chat carries before serving it.
The row records identity, never availability: a file that expired since it was attached must
render as missing on reload rather than as a reference the user can still act on.
"""
- history = convert_to_history_messages(messages)
+ history = convert_to_history_messages(messages, pending_turns)
resolved = {
attached.file_id: attached
for attached in await _resolve_copilot_attached_files(
@@ -3543,15 +3615,39 @@ async def _history_with_resolved_attachments(
def convert_to_history_messages(
messages: list[WorkflowCopilotChatMessage],
+ pending_turns: dict[str, CopilotPendingTurn] | None = None,
) -> list[WorkflowCopilotChatHistoryMessage]:
+ opener_turn_ids = {
+ outcome.user_message_id: outcome.copilot_turn_id
+ for message in messages
+ if (outcome := message.turn_outcome) is not None
+ and outcome.user_message_id is not None
+ and outcome.copilot_turn_id is not None
+ }
+ opener_turn_ids.update(
+ {
+ pending.user_message_id: turn_id
+ for turn_id, pending in (pending_turns or {}).items()
+ if pending.user_message_id is not None
+ }
+ )
+
+ def resolve_turn_id(message: WorkflowCopilotChatMessage) -> str | None:
+ if message.turn_outcome is not None:
+ return message.turn_outcome.copilot_turn_id
+ message_id = getattr(message, "workflow_copilot_chat_message_id", None)
+ return opener_turn_ids.get(message_id) if isinstance(message_id, str) else None
+
return [
WorkflowCopilotChatHistoryMessage(
sender=message.sender,
content=message.content,
+ turn_id=resolve_turn_id(message),
audio_artifact_id=message.audio_artifact_id,
attached_files=message.attached_files,
turn_outcome=message.turn_outcome,
created_at=message.created_at,
+ modified_at=getattr(message, "modified_at", message.created_at),
narrative_payload=message.narrative_payload,
)
for message in messages
diff --git a/skyvern/forge/sdk/schemas/copilot_turn_outcome.py b/skyvern/forge/sdk/schemas/copilot_turn_outcome.py
index 1172610c4..e63e6b07f 100644
--- a/skyvern/forge/sdk/schemas/copilot_turn_outcome.py
+++ b/skyvern/forge/sdk/schemas/copilot_turn_outcome.py
@@ -104,6 +104,12 @@ class TurnOutcome(BaseModel):
copilot_last_code_build_failed: bool = False
copilot_pending_capability: str | None = None
copilot_turn_id: str | None = None
+ # Durable link back to this turn's opener row. This keeps history
+ # correlation correct when concurrent turns finish out of order.
+ user_message_id: str | None = None
+ # Client-generated per-request correlation value. Retained after the pending
+ # marker is cleared so a disconnected client can recover the exact chat.
+ request_cancel_token: str | None = None
idempotency_digest: str | None = None
unresolved_runtime_failure: UnresolvedRuntimeFailure | None = None
connected_account_choices: list[ConnectedAccountChoice] | None = None
diff --git a/skyvern/forge/sdk/schemas/workflow_copilot.py b/skyvern/forge/sdk/schemas/workflow_copilot.py
index fca842080..0dc30009e 100644
--- a/skyvern/forge/sdk/schemas/workflow_copilot.py
+++ b/skyvern/forge/sdk/schemas/workflow_copilot.py
@@ -415,6 +415,7 @@ class WorkflowCopilotApplyProposedWorkflowRequest(BaseModel):
class WorkflowCopilotChatHistoryMessage(BaseModel):
sender: WorkflowCopilotChatSender = Field(..., description="Message sender")
content: str = Field(..., description="Message content")
+ turn_id: str | None = Field(None, description="Turn that owns this row")
audio_artifact_id: str | None = Field(None, description="Artifact ID for captured dictation audio")
attached_files: Annotated[list[CopilotAttachedFile], BeforeValidator(_attached_files_or_empty)] = Field(
default_factory=list, description="Uploaded files the user attached to this message"
@@ -425,6 +426,10 @@ class WorkflowCopilotChatHistoryMessage(BaseModel):
description="Persisted narrative bubble snapshot; lets a reload re-render per-block cards.",
)
created_at: datetime = Field(..., description="When the message was created")
+ modified_at: datetime | None = Field(
+ None,
+ description="When this version was persisted; legacy in-process constructors may omit it",
+ )
class WorkflowCopilotChatSummary(BaseModel):
@@ -798,6 +803,9 @@ class WorkflowCopilotChatHistoryResponse(BaseModel):
question_interactions: list[QuestionInteraction] = Field(default_factory=list)
pending_question_cancel_token: str | None = None
workflow_copilot_chat_id: str | None = Field(None, description="Latest chat ID for the workflow")
+ request_turn_id: str | None = Field(
+ None, description="Turn matched by request_cancel_token when recovery requests provide one"
+ )
chat_history: list[WorkflowCopilotChatHistoryMessage] = Field(default_factory=list, description="Chat messages")
proposed_workflow: dict | None = Field(None, description="Latest workflow proposed by the copilot")
proposed_workflow_metadata: CopilotProposalMetadata | None = None
diff --git a/skyvern/forge/sdk/workflow/models/block.py b/skyvern/forge/sdk/workflow/models/block.py
index ecca50a9f..a47755e3d 100644
--- a/skyvern/forge/sdk/workflow/models/block.py
+++ b/skyvern/forge/sdk/workflow/models/block.py
@@ -310,7 +310,7 @@ from skyvern.webeye.browser_state import BrowserState, get_browser_state_diagnos
from skyvern.webeye.cdp_download_interceptor import normalize_download_filename, settle_browser_downloads_for_context
from skyvern.webeye.navigation import (
default_navigation_settle,
- is_egress_attributable_navigation_error,
+ is_egress_attributable_navigation_failure,
navigate_with_retry,
redact_url_secrets,
)
@@ -2259,7 +2259,7 @@ class BaseTaskBlock(Block):
if (
isinstance(e, FailedToNavigateToUrl)
and not isinstance(e, BlockedNavigationDestination)
- and not is_egress_attributable_navigation_error(e.error_message)
+ and not is_egress_attributable_navigation_failure(e)
):
# The target site did not load. Site-caused, and already surfaced on the run
# through the failure_reason recorded below. Two classes stay at error
@@ -2267,7 +2267,9 @@ class BaseTaskBlock(Block):
# is the SSRF guard tripping, and EGRESS_ATTRIBUTABLE_NAV_ERRORS is our own
# egress failing, which get_or_create_page may recover from on a different
# proxy node. The block-execution handler above applies only the first of
- # those two carve-outs; see the PR discussion.
+ # those two carve-outs; see the PR discussion. Classify on the exception, not
+ # its message: a target with no address record borrows an egress-attributable
+ # error code from the proxy while being as site-caused as a site that 404s.
LOG.warning(
"Failed to get browser state for first task",
task_id=task.task_id,
diff --git a/skyvern/forge/sdk/workflow/models/code_block_recorder.py b/skyvern/forge/sdk/workflow/models/code_block_recorder.py
index 2b6b46c25..19bbf66a0 100644
--- a/skyvern/forge/sdk/workflow/models/code_block_recorder.py
+++ b/skyvern/forge/sdk/workflow/models/code_block_recorder.py
@@ -93,6 +93,12 @@ _PAGE_ACTION_MAP: dict[str, ActionType] = {
"reload": ActionType.RELOAD_PAGE,
"evaluate": ActionType.EXECUTE_JS,
}
+# Effects a code block has on a tab other than its own. Those pages are raw Playwright objects
+# below this proxy, so the broker hands the call back here to keep them on the action timeline.
+_LISTED_PAGE_ACTION_MAP: dict[str, ActionType] = {
+ "close": ActionType.CLOSE_PAGE,
+ "bring_to_front": ActionType.SWITCH_TAB,
+}
_LOCATOR_ACTION_MAP: dict[str, ActionType] = {
"click": ActionType.CLICK,
"dblclick": ActionType.CLICK,
@@ -270,6 +276,11 @@ def _recorded_action_fields(
download_url = _string_value(kwargs.get("download_url", _arg(args, 1)))
if download_url is not None:
fields["download_url"] = download_url
+ elif action_type in (ActionType.SWITCH_TAB, ActionType.CLOSE_PAGE):
+ # Required on SwitchTabAction, so without it the typed action degrades to a base Action.
+ tab_index = kwargs.get("tab_index")
+ if isinstance(tab_index, int):
+ fields["tab_index"] = tab_index
elif action_type == ActionType.SELECT_OPTION:
option = _select_option(kwargs.get("value", _arg(args, value_index)), kwargs)
if option is not None:
@@ -795,6 +806,20 @@ class RecordingPage:
),
)
+ async def _record_listed_page_effect(
+ self,
+ name: str,
+ target: str | None,
+ call: Callable[[], Awaitable[None]],
+ tab_index: int | None = None,
+ ) -> None:
+ """Record a code block's effect on a sibling tab, which is a raw Page this proxy does not wrap."""
+ action_type = _LISTED_PAGE_ACTION_MAP.get(name)
+ if action_type is None:
+ await call()
+ return
+ await self.__recorder.record(action_type, name, target, call, (), {"tab_index": tab_index})
+
def _brokered_default_timeout(self, scope: Literal["page", "context"]) -> float | None:
"""Return the trusted timeout that a secure-runner block must restore."""
if scope == "page":
diff --git a/skyvern/forge/taskv3/engine.py b/skyvern/forge/taskv3/engine.py
index e3824e2eb..896ce2d4f 100644
--- a/skyvern/forge/taskv3/engine.py
+++ b/skyvern/forge/taskv3/engine.py
@@ -52,7 +52,12 @@ from skyvern.forge.taskv3.loop import (
run_agent_tool_loop,
)
from skyvern.forge.taskv3.opaque_refs import OpaqueUrlRefs, mask_opaque_urls
-from skyvern.forge.taskv3.tools import PageProvider, build_browser_tools
+from skyvern.forge.taskv3.tools import (
+ BlankWorkingPageGuard,
+ PageProvider,
+ apply_blank_page_guard,
+ build_browser_tools,
+)
LOG = structlog.get_logger()
@@ -203,6 +208,8 @@ async def run_task_v3_agent_loop(
initial_navigation_status: int | None = None,
page_probe: Callable[[], Awaitable[str | None]] | None = None,
reload_page: Callable[[], Awaitable[None]] | None = None,
+ restore_page_url: Callable[[Any, str], Awaitable[None]] | None = None,
+ download_attempts: Callable[[], int | None] | None = None,
block_type: str | None = None,
) -> LoopOutcome:
"""Run one Task V3 task to completion against `page`, returning the loop outcome.
@@ -256,6 +263,27 @@ async def run_task_v3_agent_loop(
# contract have to agree: with the switch off, a live object would report 0 opportunities for a
# tier that is turned off rather than saying it was not there.
semantic_commit_stats = None if page_free or not settings.TASK_V3_SEMANTIC_COMMIT_VERIFY else SemanticCommitStats()
+
+ # Defaulted BEFORE the guard is built: the guard holds this set by reference to exclude files
+ # `file_upload` staged, so it has to be the same object the loop writes into.
+ if staged_downloads is None:
+ staged_downloads = set()
+
+ def _deadline_remaining() -> float | None:
+ return None if deadline_seconds is None else (loop_started_at + deadline_seconds) - time.monotonic()
+
+ blank_page_guard = (
+ None
+ if page_free or restore_page_url is None
+ else BlankWorkingPageGuard(
+ page_provider,
+ restore_page_url,
+ downloads_dir=downloads_dir,
+ remaining_seconds=_deadline_remaining,
+ download_attempts=download_attempts,
+ staged_downloads=staged_downloads,
+ )
+ )
browser_tools = (
[]
if page_free
@@ -309,8 +337,6 @@ async def run_task_v3_agent_loop(
# run's deadline and cancellation so probing cannot overrun either. Page-free runs never probe.
activity = ActivityRecency()
submit_watch = SubmitWatch()
- if staged_downloads is None:
- staged_downloads = set()
# A page-free run has no tool that could trigger a download, so a blocker would refuse every
# completed verdict forever; it has nothing to type a verification code into either.
if page_free:
@@ -331,6 +357,9 @@ async def run_task_v3_agent_loop(
verification_blocker=verification_blocker,
)
tools = browser_tools + (extra_tools or []) + [finish_tool]
+ # The COMPLETE dispatch list, not just the browser tools: auth / captcha / code tools and finish
+ # are appended here and would otherwise be able to inspect and act on a blank page.
+ apply_blank_page_guard(tools, blank_page_guard)
base_system_prompt = PAGE_FREE_SYSTEM_PROMPT if page_free else SYSTEM_PROMPT
# Keyed on which hooks are present, not completion_probe alone: an extraction blocker-only
# case needs the model told it ends the run itself; a wait-only probe has nothing to explain.
@@ -383,6 +412,22 @@ async def run_task_v3_agent_loop(
_exit_ctx = skyvern_context.current()
if _exit_ctx is not None and _exit_ctx.refresh_working_page:
_exit_ctx.refresh_working_page = False
+ # Backstop for the handoff the ticket rides on: the loop can end without a further tool call
+ # (budget, cancellation, a raise), leaving the page blank for the next url-less block. In the
+ # `finally` because a raising block can still be followed by another one
+ # (continue-on-failure); `ensure_live` swallows everything, so it cannot turn a success into
+ # a failure or mask the exception on its way out. Bounded by what is left of the run's
+ # deadline, and skipped outright once cancelled -- finalization must not outlive either.
+ if blank_page_guard is not None:
+ _cancelled = False
+ if should_cancel is not None:
+ try:
+ _cancelled = await should_cancel()
+ except Exception:
+ _cancelled = False
+ if not _cancelled:
+ # The guard bounds itself by `_deadline_remaining`; no second computation here.
+ await blank_page_guard.ensure_live()
if refs.refs:
outcome.reason = refs.resolve(outcome.reason)
outcome.extracted_output = refs.resolve_deep(outcome.extracted_output)
diff --git a/skyvern/forge/taskv3/tools.py b/skyvern/forge/taskv3/tools.py
index 4200316c6..726cc792b 100644
--- a/skyvern/forge/taskv3/tools.py
+++ b/skyvern/forge/taskv3/tools.py
@@ -28,6 +28,7 @@ from contextvars import ContextVar
from datetime import datetime
from enum import Enum
from typing import TYPE_CHECKING, Any, Awaitable, Callable, NamedTuple
+from urllib.parse import urlparse
import structlog
from PIL import Image, ImageDraw
@@ -62,6 +63,7 @@ from skyvern.forge.taskv3.loop import (
from skyvern.forge.taskv3.preflight import PREFLIGHT_TOOL_NAMES, preflight_tool_action
from skyvern.forge.taskv3.target_label import TARGET_KIND_TOKENS, TARGET_NAME_CAP
from skyvern.webeye.browser_driver_errors import is_driver_timeout_error
+from skyvern.webeye.browser_state import BLANK_PAGE_URLS
from skyvern.webeye.utils.page import OTP_INPUT_PRIVACY_JS, OTP_SAFE_FRAGMENT_HTML_JS, mask_otp_values_in_html
if TYPE_CHECKING:
@@ -4784,6 +4786,11 @@ _OBSERVE_OWN_IDENTITY_ATTRS = ("id", "name", "data-testid")
# is kept on the control's own: an ancestor with id="row" and one with data-testid="row" are two
# different statements about the page, and printing both as `within='row'` erases that.
_OBSERVE_WITHIN_KINDS = tuple(f"within.{attr}" for attr in _OBSERVE_OWN_IDENTITY_ATTRS)
+# The label the page bound to the one form control enclosed with this one -- the field it belongs to.
+# Ranked under the control's own identity and over both `within` and `section`: a wrapper id narrows
+# to the same row a field does, but the task names the field, never `lookupRow1`, so a set told apart
+# only by a framework-minted wrapper id leaves the model exactly where it started.
+_OBSERVE_FIELD_KIND = "field"
# Each pass appends at most one qualifier per line, so a set that only splits in stages -- a tier
# that separates a group of three into a pair and a single, then a deeper tier that separates the
# pair -- still converges. This bounds work, not correctness: whatever is left after the last pass
@@ -4799,10 +4806,11 @@ def _disambiguate_digest_bodies(
that control sits -- its own identity, the identity of what encloses it, the heading it is filed
under -- that tells the set apart. Returns the bodies and how many lines are still identical to
another afterwards, which is what production has to be able to see."""
- parsed: list[tuple[list[tuple[str, str]], list[tuple[str, str]], str | None]] = []
+ parsed: list[tuple[list[tuple[str, str]], list[tuple[str, str]], str | None, str | None]] = []
for e in elements:
owns: list[tuple[str, str]] = []
within: list[tuple[str, str]] = []
+ field_label: str | None = None
section: str | None = None
for entry in e.get("placement") or []:
if not isinstance(entry, (list, tuple)) or len(entry) != 2:
@@ -4820,21 +4828,26 @@ def _disambiguate_digest_bodies(
# compared at different real depths. Both qualifiers are still true of their own
# element, which is all the line claims.
within.append((kind, value))
+ elif kind == _OBSERVE_FIELD_KIND and field_label is None:
+ field_label = value
elif kind == "section" and section is None:
section = value
- parsed.append((owns, within, section))
+ parsed.append((owns, within, field_label, section))
- tiers: list[tuple[str, int]] = [("own", i) for i in range(max((len(o) for o, _, _ in parsed), default=1) or 1)]
- tiers += [("within", i) for i in range(max((len(w) for _, w, _ in parsed), default=0))]
+ tiers: list[tuple[str, int]] = [("own", i) for i in range(max((len(o) for o, _, _, _ in parsed), default=1) or 1)]
+ tiers.append((_OBSERVE_FIELD_KIND, 0))
+ tiers += [("within", i) for i in range(max((len(w) for _, w, _, _ in parsed), default=0))]
tiers.append(("section", 0))
def candidate(index: int, tier: tuple[str, int]) -> tuple[str, str] | None:
- owns, within, section = parsed[index]
+ owns, within, field_label, section = parsed[index]
kind, depth = tier
if kind == "own":
return owns[depth] if depth < len(owns) else None
if kind == "within":
return within[depth] if depth < len(within) else None
+ if kind == _OBSERVE_FIELD_KIND:
+ return (_OBSERVE_FIELD_KIND, field_label) if field_label else None
return ("section", section) if section else None
def suffix(index: int, tier: tuple[str, int]) -> str:
@@ -5111,6 +5124,58 @@ async () => {
}
const out = [];
const labelOfControl = new Map();
+ // Only the name the page DECLARED for a control -- aria-label, a bound <label>, aria-labelledby.
+ // Deliberately not rec.label: that falls through to the placeholder and then to the user's typed
+ // value, so a qualifier read from it would promote a shared template hint to an address, and would
+ // change every turn as the agent fills the row.
+ const boundLabelOfControl = new Map();
+ // Element -> every listed element beneath it that COULD be a field. Built after the sweep, from the
+ // records that survived and from their live parents, so a field the page moved during the observer
+ // drain is counted where it now is rather than where it was when its record was made.
+ const fieldsUnder = new Map();
+ // Boxes whose subtree gained or lost a child while observe was running. The census can only count
+ // what it enumerated; a node the page INSERTS during the observer drain is in no collection at all,
+ // so no amount of re-reading the survivors will see it. Rather than enumerate live candidates -- the
+ // fifth patch to the same class, with a sixth behind it -- a box the page changed underneath us is
+ // one whose field count we cannot establish, so the qualifier is withheld there. This subsumes
+ // inserted, moved and removed in one predicate, at the level the leaks actually occur.
+ const boxChangedUnderUs = new Set();
+ // Definitively a command control, decided from the element AS IT STANDS. Read live for the same
+ // reason the ancestry is: a page can flip role="button" to role="combobox", or switch on
+ // contenteditable, in response to a marker write, and the fingerprint tracks neither -- so a
+ // classification captured before the drain survives re-resolution while being wrong.
+ //
+ // "Definitively" is the whole point, and it is the complement of a much smaller, more stable list:
+ // everything else counts as a possible field. The qualifier is DISCARDABLE -- emitting nothing
+ // costs one hint and the lines stay honestly identical, while emitting wrongly names a control
+ // after a field it does not belong to. So this fails CLOSED: anything unrecognised, and anything
+ // that throws, counts as a field and suppresses the qualifier rather than going uncounted and
+ // letting a two-field box read as one.
+ const _isCommandControl = (el) => {
+ try {
+ const tag = String(el.tagName || '').toLowerCase();
+ if (tag === 'button' || tag === 'a' || tag === 'summary') return true;
+ if (tag === 'input' && /^(?:submit|button|reset|image)$/.test(String(el.type || ''))) return true;
+ const role = String(_attr(el, 'role') || '').trim().toLowerCase();
+ return /^(?:button|link|tab|menuitem|menuitemcheckbox|menuitemradio)$/.test(role);
+ } catch (e) { return false; }
+ };
+ // Affirmatively a field: something a value is entered into. This gates NAMING only, never counting.
+ // The two directions are not symmetric, and that is the whole point: widening the census SUPPRESSES
+ // more qualifiers (fail closed), but widening the naming source makes more NON-fields eligible to
+ // lend their text to a neighbour (fail open). So counting stays the complement above -- an
+ // unrecognised widget counts and suppresses -- while naming is this affirmative list, and a
+ // `[role=option]` or a bare `[tabindex]` is neither: it counts, and it never names.
+ const _isFieldControl = (el) => {
+ try {
+ const tag = String(el.tagName || '').toLowerCase();
+ if (tag === 'select' || tag === 'textarea') return true;
+ if (tag === 'input') return !/^(?:submit|button|reset|image)$/.test(String(el.type || ''));
+ if (el.isContentEditable === true) return true;
+ const role = String(_attr(el, 'role') || '').trim().toLowerCase();
+ return /^(?:textbox|combobox|searchbox|spinbutton)$/.test(role);
+ } catch (e) { return false; }
+ };
// Monotonic across observe() calls (persisted on window), and never reassigned on an element that
// already has one, so a data-tv3 marker always denotes the same element. Resetting the counter per
// call let a selector remembered from an earlier observe silently resolve to a different node.
@@ -5293,6 +5358,20 @@ async () => {
// page, and reporting the shared one alone throws that away.
for (const a of _identitiesOf(n)) { if (kept < 8) { out.push(['within.' + a[0], a[1]]); kept++; } }
}
+ // The one named field control the page put in the SAME box as this one. The claim is
+ // CO-LOCATION, not ownership: `label[for]` binds a name to a control, but nothing binds that
+ // control to a neighbouring one, so the box is the direct parent and a box holding a second
+ // field at any depth names neither. Reaching wider would hand every row button of a panel the
+ // name of that panel's filter box -- confidently and wrongly, which is worse than leaving the
+ // lines identical, because duplicate_digest_lines reports identical lines honestly.
+ if (_isCommandControl(el)) {
+ const box = _parentOf.call(el);
+ const under = box && !boxChangedUnderUs.has(box) ? fieldsUnder.get(box) : null;
+ if (under && under.length === 1 && _parentOf.call(under[0]) === box
+ && _isFieldControl(under[0]) && boundLabelOfControl.has(under[0])) {
+ out.push(['field', boundLabelOfControl.get(under[0])]);
+ }
+ }
const h = _headingFor(el);
if (h) out.push(['section', h]);
} catch (e) { /* fail open: fewer qualifiers, never a dropped element */ }
@@ -5864,6 +5943,7 @@ async () => {
for (const l of el.labels) { strongLabel = _labelText(l); if (strongLabel) break; }
}
if (!strongLabel) strongLabel = byId('aria-labelledby');
+ const boundName = strongLabel;
let slottedName = false;
if (!strongLabel) strongLabel = (el.innerText || '').trim();
if (!strongLabel && host) { strongLabel = slottedText(el, host); slottedName = !!strongLabel; }
@@ -6013,7 +6093,14 @@ async () => {
// A submit or button input is named by its caption, and a caption is what a refusal beside it
// repeats; a field's own control is the only thing a wrapper holds.
const captioned = rec.tag === 'input' && /^(?:submit|button|reset|image)$/.test(rec.type || '');
- if ((rec.tag === 'input' && !captioned) || rec.tag === 'select' || rec.tag === 'textarea') labelOfControl.set(el, rec.label.slice(0, 140).replace(/\s+/g, ' ').trim());
+ const isField = (rec.tag === 'input' && !captioned) || rec.tag === 'select' || rec.tag === 'textarea';
+ if (isField) labelOfControl.set(el, rec.label.slice(0, 140).replace(/\s+/g, ' ').trim());
+ // Retained at _RETAIN_WIDTH, not cut to a display width: this value is PRINTED as a qualifier,
+ // and masking is by provenance over the WHOLE minted URL, so a page-side cut leaves a fragment
+ // the masker cannot recognise -- keeping the prefix, where the credential sits. Python masks,
+ // then caps. labelOfControl above is compared, never printed, so its 140 stays. Stored for every
+ // listed element, because what counts as a field is decided later, from the live element.
+ if (boundName) boundLabelOfControl.set(el, boundName.slice(0, _RETAIN_WIDTH).replace(/\s+/g, ' ').trim());
// The element rides on its OWN record, under a name generated for this call in Python. There is
// then no second structure to misalign: every manipulation of `out` below carries each element
// with its own record. A rec->element lookup, or a parallel array, would each be an INDEPENDENT
@@ -6071,14 +6158,14 @@ async () => {
try { connected = _isConnected.call(c.el); } catch (e) { connected = false; }
if (!lost && connected) continue;
const at = c.rec === null ? -1 : out.indexOf(c.rec);
- if (at !== -1) { out.splice(at, 1); labelOfControl.delete(c.el); dropped++; }
+ if (at !== -1) { out.splice(at, 1); labelOfControl.delete(c.el); boundLabelOfControl.delete(c.el); dropped++; }
}
if (lost && !rem.shared) { if (rem.fresh) markersWritten--; else markersReused--; }
continue;
}
if (rem.rec === null || still !== rem.m) {
const at = rem.rec === null ? -1 : out.indexOf(rem.rec);
- if (at !== -1) { out.splice(at, 1); labelOfControl.delete(rem.el); dropped++; }
+ if (at !== -1) { out.splice(at, 1); labelOfControl.delete(rem.el); boundLabelOfControl.delete(rem.el); dropped++; }
if (rem.fresh) markersWritten--; else markersReused--;
}
}
@@ -6111,6 +6198,7 @@ async () => {
}
} catch (e) { mutated = true; }
}
+ const droppedStillLive = [];
for (let k = out.length - 1; k >= 0; k--) {
const rec = out[k];
const el = elOfRec.get(rec);
@@ -6129,7 +6217,14 @@ async () => {
}
} else { checkInconclusive = false; ok = resolvesTo(rec.selector, el) || checkInconclusive; }
}
- if (!ok) { labelOfControl.delete(el); out.splice(k, 1); dropped++; }
+ if (!ok) {
+ // Dropped from the digest, but the control can still BE there -- a page that changes a second
+ // field's label mid-run fails the fingerprint without detaching it. It must not name anything
+ // (it has no line for the model to read), and it must still COUNT, or the box it sits in reads
+ // as holding one field when it holds two.
+ if (el && connected) droppedStillLive.push(el);
+ labelOfControl.delete(el); boundLabelOfControl.delete(el); out.splice(k, 1); dropped++;
+ }
else if (el) {
// Recomputed after the drain rather than added to the fingerprint above: a page's own observer
// can change either attribute in response to a marker write, and the line has to describe the
@@ -6141,10 +6236,43 @@ async () => {
try { now = _a11yRemoved(el); } catch (e) { now = ''; }
if (now) rec.a11yRemoved = now; else delete rec.a11yRemoved;
}
- const place = _placement(el);
- if (place.length) rec.placement = place;
}
}
+ // The field census, built here and not earlier: from the records that SURVIVED the sweep, and from
+ // each one's parent read now. A registry built while records were made answers about the page as it
+ // was before the observer drain, so a field the page moved into a box afterwards goes uncounted and
+ // the box reads as holding one. No depth bound -- the box is read by "holds no second field at any
+ // depth", so a bound would not bound work, it would make the count silently wrong.
+ for (const m of _witnessed) {
+ if (m.type !== 'childList') continue;
+ try {
+ let n = m.target;
+ while (n && !boxChangedUnderUs.has(n)) { boxChangedUnderUs.add(n); n = _parentOf.call(n); }
+ } catch (e) { /* fail open on the walk; the box itself is already marked */ }
+ }
+ const census = [];
+ for (const rec of out) { const el = elOfRec.get(rec); if (el) census.push(el); }
+ for (const el of droppedStillLive) { if (_isConnected.call(el)) census.push(el); }
+ for (const el of census) {
+ if (_isCommandControl(el)) continue;
+ let anc = el;
+ for (;;) {
+ anc = _parentOf.call(anc);
+ if (!anc) break;
+ const held = fieldsUnder.get(anc);
+ if (held) held.push(el); else fieldsUnder.set(anc, [el]);
+ }
+ }
+ // Placement is read only after the sweep above has finished, because it is the one thing here that
+ // reads OTHER records' controls. Computed inside that loop, a record at a lower index had not been
+ // swept yet, so a field about to be dropped could still name a survivor -- printing `field='X'`
+ // with no line for X anywhere in the digest.
+ for (const rec of out) {
+ const el = elOfRec.get(rec);
+ if (!el) continue;
+ const place = _placement(el);
+ if (place.length) rec.placement = place;
+ }
// Page-text digest: outcome states (submission confirmations, rejection banners, validation
// summaries) live in non-interactive nodes the element list can never carry. Three sources in
// priority order — ARIA status channels (uncapped within the 900 total), class/id-named message
@@ -12346,6 +12474,255 @@ def _human_download_size(num_bytes: int) -> str:
return f"{size:.1f} TB"
+# v1's equivalent restore is unbounded, but it runs once per download action while this guard runs
+# before EVERY tool call, so an unbounded navigation here can stall a run in a way v1 cannot. One
+# full page-load attempt plus settle, and then a small number of retries before latching off -- an
+# unreachable URL must not re-burn the timeout on every remaining tool call.
+_BLANK_PAGE_RESTORE_TIMEOUT_SECONDS = 90.0
+_BLANK_PAGE_RESTORE_MAX_FAILURES = 3
+# How many further tool calls a download stays eligible to have opened the tab it is blamed for.
+# Fleet-wide 91.3% of downloads open no second page, so arming that never expires would hand the
+# next legitimately-blank tab -- one the model opened -- to the close path.
+_DOWNLOAD_ARM_TOOL_CALLS = 2
+
+
+def _url_origin(url: str) -> str:
+ """Scheme + host + port, and nothing else. A page URL reaches structured logs here, and its PATH
+ and query routinely carry the secret itself -- a password-reset token, `/verify/<token>`, a
+ signed object path. Built from `hostname`, not `netloc`, so `user:pass@host` credentials cannot
+ ride along. Not a general masker: the provenance masker (`opaque_refs.mask`) rewrites only URLs
+ the payload masker minted, so it is a no-op on a live page URL and cannot cover this (SKY-16337).
+ """
+ # `parsed.port` is lazily evaluated and raises ValueError on an out-of-range port, so it has to
+ # be inside the try: both call sites are in paths documented as unable to raise.
+ try:
+ parsed = urlparse(url)
+ if not parsed.scheme or not parsed.hostname:
+ return "unparseable"
+ port = f":{parsed.port}" if parsed.port else ""
+ except Exception:
+ return "unparseable"
+ return f"{parsed.scheme}://{parsed.hostname}{port}"
+
+
+class BlankWorkingPageGuard:
+ """Keeps the working page off a blank document a download left behind.
+
+ Two shapes, because the browser produces two and v1 repairs both:
+
+ * **The download opens a second tab** and that tab takes the working-page slot (it is newest, and
+ `list_valid_pages` counts `about:blank` as valid). This is what actually happens in production:
+ on the workflow SKY-16322 was filed against, 1545 of 1546 downloads went from 1 page to 2, and
+ v1 repaired every one of them by CLOSING the extra page, never by navigating. Closing is gated
+ on a download having been seen, and consumes that arming, exactly as v1 scopes its close to one
+ download action -- the gate is what separates a download popup from the model opening a tab,
+ which by URL alone are the same thing.
+ * **The tab is blanked in place**, which v1 repairs by navigating back. Kept because v1 still
+ carries the code, but note it is close to dead: its log fired once in 30 days fleet-wide.
+
+ Either way the next url-less block would otherwise inherit `about:blank` and
+ `resolve_inherited_workflow_task_page` raises `InvalidWorkflowTaskURLState`.
+
+ Checked as a condition before every tool call rather than fired on the download-detection edge:
+ a download is counted once, so an edge-triggered repair can be spent on a call where the page is
+ still fine and then never fire again.
+ """
+
+ def __init__(
+ self,
+ page_provider: PageProvider,
+ restore_page_url: Callable[[Any, str], Awaitable[None]],
+ downloads_dir: str | None = None,
+ remaining_seconds: Callable[[], float | None] | None = None,
+ download_attempts: Callable[[], int | None] | None = None,
+ staged_downloads: set[str] | None = None,
+ ) -> None:
+ self._page_provider = page_provider
+ self._restore_page_url = restore_page_url
+ self._downloads_dir = downloads_dir
+ self._remaining_seconds = remaining_seconds
+ self._download_attempts = download_attempts
+ self._staged_downloads = staged_downloads
+ self._attempts_seen: int | None = None
+ self._live_page: Any = None
+ self._live_url = ""
+ self._failures = 0
+ self._arm_budget = 0
+ self._seen_downloads: set[str] = set()
+ self._baselined = False
+
+ def _bound(self, timeout: float | None = None) -> float:
+ """Every repair is bounded by what is left of the run, not just the post-loop one: this runs
+ before EVERY tool call, so work near the deadline would otherwise push past it."""
+ limits = [_BLANK_PAGE_RESTORE_TIMEOUT_SECONDS]
+ if timeout is not None:
+ limits.append(timeout)
+ if self._remaining_seconds is not None:
+ left = self._remaining_seconds()
+ if left is not None:
+ limits.append(left)
+ return min(limits)
+
+ def _scan_for_new_download(self) -> bool:
+ """Arm from signals this guard reads itself, not from the download-signal wrapper.
+
+ That wrapper only covers the tools `build_browser_tools` builds, but this guard runs over the
+ COMPLETE dispatch list, and the file routinely lands after the click handler has returned. A
+ download followed by `finish` would otherwise arm nothing and the tab would survive -- the
+ original bug, reintroduced through the arming path.
+
+ Two sources, because neither alone is sufficient:
+
+ * The interceptor's attempt counter, which increments in `_resolve_save_path` BEFORE the
+ duplicate check, so it still moves when a download is deduplicated against a byte-identical
+ file already on disk -- a case that leaves the directory completely unchanged and is
+ therefore invisible to any filesystem check.
+ * The directory itself, for download paths that do not run through that interceptor.
+
+ Files `file_upload` staged into the same directory are excluded: they are this run writing its
+ own upload, not the browser downloading anything, and arming on one would hand a tab the model
+ opened to the close path. The first call only baselines -- `downloads_dir` is per run, so a
+ file a previous block downloaded is not this block's.
+ """
+ armed = False
+ if self._download_attempts is not None:
+ try:
+ attempts = self._download_attempts()
+ except Exception:
+ attempts = None
+ if attempts is not None:
+ if self._attempts_seen is not None and attempts > self._attempts_seen:
+ armed = True
+ self._attempts_seen = attempts
+ if self._downloads_dir:
+ try:
+ names = set(os.listdir(self._downloads_dir))
+ except OSError:
+ names = None
+ if names is not None:
+ if self._staged_downloads:
+ names -= self._staged_downloads
+ # Through `_download_signal_identity`, or ONE download arms twice: once when
+ # `report.pdf.crdownload` appears and again when it is renamed to `report.pdf`. The
+ # second arming has no tab-opening event behind it and would grant a fresh close
+ # window at an arbitrary later point -- exactly the misfire the decay bound exists to
+ # stop. The invariant is one close per DOWNLOAD, which means per identity.
+ identities = {_download_signal_identity(name) for name in names}
+ new = identities - self._seen_downloads
+ self._seen_downloads = identities
+ if self._baselined and new:
+ armed = True
+ if not self._baselined:
+ self._baselined = True
+ return False
+ return armed
+
+ async def ensure_live(self, timeout: float | None = None) -> None:
+ armed_now = self._scan_for_new_download()
+ if armed_now:
+ self._arm_budget = _DOWNLOAD_ARM_TOOL_CALLS
+ # Two passes at most: closing an interloper can expose a page that itself needs restoring.
+ for _ in range(2):
+ try:
+ working_page = await self._page_provider()
+ except Exception:
+ # The page is gone, not blank. The handler re-resolves and raises the real error;
+ # logging a "restore failed" traceback on every call would only bury it.
+ return
+ if working_page is None or working_page.is_closed():
+ return
+ url = working_page.url
+ if url and url not in BLANK_PAGE_URLS:
+ # A live working page means this download opened no tab that took the slot. Let the
+ # arming decay rather than saving it for a blank tab the model opens much later.
+ if not armed_now and self._arm_budget > 0:
+ self._arm_budget -= 1
+ # Track every navigation, not just the first: the repair has to return the tab to
+ # where the download was triggered from, rarely where the block started.
+ if working_page is not self._live_page or url != self._live_url:
+ self._failures = 0
+ self._live_page = working_page
+ self._live_url = url
+ return
+ if self._live_page is None or not self._live_url:
+ return
+ if working_page is self._live_page:
+ await self._restore_in_place(working_page, timeout)
+ return
+ if not await self._close_interloper(working_page):
+ return
+
+ async def _close_interloper(self, blank_page: Any) -> bool:
+ """Close a blank page that a download opened, so the working page falls back to the live one
+ (`get_working_page` re-derives from the context and takes the newest valid page). Returns
+ whether it closed anything."""
+ if self._arm_budget <= 0:
+ # Without a recent download to account for it, a blank page that is not ours is a tab the
+ # model legitimately opened and has not navigated yet. Closing it would destroy real work.
+ return False
+ if self._live_page.is_closed() or self._live_page.url in BLANK_PAGE_URLS:
+ return False
+ self._arm_budget = 0
+ close_bound = self._bound()
+ if close_bound <= 0:
+ return False
+ LOG.info("taskv3 closing the blank page a download opened", origin=_url_origin(self._live_url))
+ try:
+ # A popup on a stalled CDP connection can leave close() pending forever, and this runs
+ # both before dispatch and during final cleanup -- unbounded, it would block cancellation.
+ async with asyncio.timeout(close_bound):
+ await blank_page.close()
+ except Exception:
+ LOG.warning("taskv3 failed to close the blank page a download opened", exc_info=True)
+ return False
+ return True
+
+ async def _restore_in_place(self, working_page: Any, timeout: float | None) -> None:
+ if self._failures >= _BLANK_PAGE_RESTORE_MAX_FAILURES:
+ # Latched off: this runs before EVERY tool call, so a URL that cannot be reloaded would
+ # otherwise burn the full timeout again on each one for the rest of the block.
+ return
+ bound = self._bound(timeout)
+ if bound <= 0:
+ return
+ LOG.warning("taskv3 restoring a working page left blank", origin=_url_origin(self._live_url))
+ try:
+ async with asyncio.timeout(bound):
+ await self._restore_page_url(working_page, self._live_url)
+ self._failures = 0
+ except Exception:
+ # A run on a blank page is degraded, not broken: never fail a tool call over the repair.
+ self._failures += 1
+ LOG.warning("taskv3 blank working page restore failed", attempts=self._failures, exc_info=True)
+
+
+def apply_blank_page_guard(tools: list[ToolSpec], guard: BlankWorkingPageGuard | None) -> None:
+ """Repair before the handler runs, so the tool acts on -- and the model perceives -- a live
+ document. Applied by the engine to the COMPLETE dispatch list, not just the browser tools: the
+ auth / captcha / code tools and `finish` are appended afterwards and would otherwise be able to
+ inspect and act on `about:blank`.
+
+ No page-change flag is set on the tool result: the loop reads `page_state_changed` as progress
+ evidence (it clears the retry ledger and scores hard progress), and a repair is not progress.
+ That closes the result channel only -- a repair landing mid-batch still moves the loop's own
+ before/after page fingerprint, which the loop scores on the run's behalf. Closing that too means
+ re-baselining the loop's samplers, a change to shared v3 loop behaviour, deliberately not here.
+ Stale observe refs fail closed with a re-observe error, which is the safe degrade.
+ """
+ if guard is None:
+ return
+ for tool_spec in tools:
+
+ async def wrapped(
+ args: dict[str, Any],
+ _handler: Callable[[dict[str, Any]], Awaitable[ToolResult]] = tool_spec.handler,
+ ) -> ToolResult:
+ await guard.ensure_live()
+ return await _handler(args)
+
+ tool_spec.handler = wrapped
+
+
def _apply_download_signal(tools: list[ToolSpec], downloads_dir: str | None) -> None:
"""Wrap every tool in the given list so a file landing in `downloads_dir` during (or between)
calls is reported in the next tool result, without a dedicated tool call. Tools assembled later
diff --git a/skyvern/schemas/browser_sessions.py b/skyvern/schemas/browser_sessions.py
index 879de76cc..6245af413 100644
--- a/skyvern/schemas/browser_sessions.py
+++ b/skyvern/schemas/browser_sessions.py
@@ -130,10 +130,6 @@ class ProcessBrowserSessionRecordingRequest(BaseModel):
default=None,
description="Optional live interpretation drafts to commit instead of reprocessing the compressed recording.",
)
- code_first: bool = Field(
- default=False,
- description="When true, synthesize deterministic code blocks from the recording instead of agent blocks.",
- )
supports_credential_tokens: bool = Field(
default=False,
description=(
diff --git a/skyvern/services/browser_recording/code_first.py b/skyvern/services/browser_recording/code_first.py
index aed386cec..642e3acae 100644
--- a/skyvern/services/browser_recording/code_first.py
+++ b/skyvern/services/browser_recording/code_first.py
@@ -55,11 +55,11 @@ DRAFT_TIMESTAMP_TOLERANCE_MS = 1.0
ActionDraftPair = tuple[Action, RecordingDraftStep | None]
# Interactions that silently vanish from a workflow if the synthesizer cannot
-# locate them; any unlocatable one forces the legacy (LLM agent block) fallback.
+# locate them; any unlocatable one rejects the recording result.
# Dropped hovers are tolerated: they are usually incidental to a located click.
-_REQUIRED_LOCATOR_TOOLS = frozenset({"click", "type_text", "select_option", CREDENTIAL_FILL_TOOL_NAME})
+_REQUIRED_LOCATOR_TOOLS = frozenset({"click", "type_text", "select_option", "upload_file", CREDENTIAL_FILL_TOOL_NAME})
-# The credential field each recorded kind fills. `secret` mirrors the legacy path's single
+# The credential field each recorded kind fills. A secret credential exposes one
# `secret_value`; `magic_link` resolves at runtime rather than as a field, and `credit_card`
# spreads over six fields the recorder cannot tell apart, so both stay plain parameters.
CREDENTIAL_FILL_FIELD_BY_KIND: dict[CredentialKind, str] = {
@@ -137,8 +137,8 @@ def transfer_focus_click_credentials(pairs: list[ActionDraftPair]) -> list[Actio
The recording panel offers the credential prompt on every step it classifies, and the
mouse-driven login clicks the password box before typing into it, so the user can bind on
- the click card - after which the panel dismisses the prompt on the fill. The legacy path
- reads the credential off whichever step carries it; a code block has to fill it.
+ the click card - after which the panel dismisses the prompt on the fill. Move that binding
+ to the fill that the code block executes.
"""
transferred = list(pairs)
for index, (action, draft) in enumerate(pairs):
@@ -243,12 +243,16 @@ def _interaction_for_action(action: Action, draft: RecordingDraftStep | None = N
base["role"] = target.role
if target.accessible_name:
base["accessible_name"] = target.accessible_name
+ if target.selector or (target.role and target.accessible_name):
+ base["observed_hidden"] = True
if isinstance(action, ActionClick):
return {"tool_name": "click", **base}
if isinstance(action, ActionHover):
return {"tool_name": "hover", **base}
if isinstance(action, ActionInputText):
+ if (target.input_type or "").lower() == "file":
+ return {"tool_name": "upload_file", "parameter_name": "upload_file", **base}
credential_fill = _credential_fill(draft)
if credential_fill is not None:
return {**credential_fill, **base}
@@ -410,11 +414,22 @@ def _segment_summary(segment: RecordingSegment, emitted_indices: set[int] | None
if emitted_indices is None or trajectory_index in emitted_indices:
actions.append(action)
trajectory_index += 1
- typed = [action for action in actions if isinstance(action, ActionInputText)]
+ uploads = [
+ action
+ for action in actions
+ if isinstance(action, ActionInputText) and (action.target.input_type or "").lower() == "file"
+ ]
+ typed = [
+ action
+ for action in actions
+ if isinstance(action, ActionInputText) and (action.target.input_type or "").lower() != "file"
+ ]
clicks = [action for action in actions if isinstance(action, ActionClick)]
if any((action.target.input_type or "").lower() == "password" for action in typed):
return f"log_in_to_{site}" if site else "log_in"
+ if uploads:
+ return "upload_a_file"
if any(not _SEARCH_FIELD_HINTS.isdisjoint(_target_name(action.target).split("_")) for action in typed):
return f"search_{site}" if site else "search"
if typed and all((action.target.tag_name or "").upper() == "SELECT" for action in typed):
@@ -459,7 +474,7 @@ def actions_to_code_first_blocks(
*,
bind_credentials: bool = True,
) -> CodeFirstResult | None:
- """Convert recorded actions into code blocks and parameters; None means fall back to legacy blocks."""
+ """Convert recorded actions into code blocks and parameters; None means synthesis was rejected."""
if draft_steps is not None and not draft_steps:
# The user deleted every interpreted step; commit an empty workflow rather
# than falling back to blocks derived from the raw actions.
@@ -472,6 +487,7 @@ def actions_to_code_first_blocks(
# (selector, role, accessible name). The same field re-filled in a later
# segment reuses its key; a different same-labeled field gets a fresh key.
parameter_identities: dict[str, tuple[str, str, str] | None] = {}
+ parameter_types: dict[str, str] = {}
# Credential parameter key (the credential id, a token the editor re-keys) -> credential id.
credential_ids_by_key: dict[str, str] = {}
used_labels: set[str] = set()
@@ -540,19 +556,20 @@ def actions_to_code_first_blocks(
break
renames[key] = resolved_key
parameter_identities.setdefault(resolved_key, identity)
+ parameter_types.setdefault(resolved_key, str(parameter.get("workflow_parameter_type") or "string"))
block_parameter_keys.append(resolved_key)
for original_key, resolved_key in renames.items():
- # A typed-text fill is the only emission that reads a string parameter,
- # always as `str(<key>)`, so this rename cannot touch selectors. Rename
- # targets never collide with this block's own keys, so one rename cannot
- # cascade onto the fill of another parameter.
code = code.replace(f"str({original_key})", f"str({resolved_key})")
+ code = code.replace(
+ f"attach_authorized_file(page, {original_key}, ",
+ f"attach_authorized_file(page, {resolved_key}, ",
+ )
try:
CodeBlock.is_safe_code(code)
# ast.parse raises ValueError on e.g. null bytes; any failure inside the
- # safety gate must fall back to legacy blocks, never surface as a 500.
+ # safety gate rejects the recording result rather than surfacing as a 500.
except (SyntaxError, ValueError, InsecureCodeDetected):
LOG.warning(
"record_browser.code_first_safety_rejected",
@@ -587,9 +604,9 @@ def actions_to_code_first_blocks(
parameters: list[WorkflowDefinitionYamlParametersItem] = [
WorkflowDefinitionYamlParametersItem_Workflow(
key=key,
- workflow_parameter_type="string",
- # Like the legacy path, recorded values never persist as defaults: a
- # secret typed into any field (not just type=password) must not land
+ workflow_parameter_type=parameter_types.get(key, "string"),
+ # Recorded values never persist as defaults: a secret typed into any
+ # field (not just type=password) must not land
# in a DB-stored, API-exposed default_value. The user binds values.
default_value="",
description="",
diff --git a/skyvern/services/browser_recording/interpretation.py b/skyvern/services/browser_recording/interpretation.py
index 5eae9f378..fcd991090 100644
--- a/skyvern/services/browser_recording/interpretation.py
+++ b/skyvern/services/browser_recording/interpretation.py
@@ -295,28 +295,20 @@ class RecordingInterpretationSession:
# followed by the keypress, not replace it.
sm.PressKey(),
sm.UrlChange(),
- sm.Wait(),
]
self._all_actions: list[Action] = []
self._processed_event_count = 0
self._capture_paused = False
- def reset_wait_capture(self) -> None:
- for machine in self._action_machines:
- if isinstance(machine, sm.Wait):
- machine.reset()
-
def set_deltas_enabled(self, enabled: bool) -> None:
# Deltas require both the client capability and the server kill switch.
self._deltas_enabled = enabled and settings.RECORDING_INTERPRETATION_DELTAS_ENABLED
def pause_capture(self) -> None:
self._capture_paused = True
- self.reset_wait_capture()
def resume_capture(self) -> None:
self._capture_paused = False
- self.reset_wait_capture()
# Enrichment deltas that landed while paused were dropped client-side, so
# send an authoritative snapshot to resync on resume.
if self.session_revision > 0:
@@ -362,6 +354,9 @@ class RecordingInterpretationSession:
await self._interpret(finalized=True)
return self.steps
+ def recorded_actions(self) -> list[Action]:
+ return list(self._all_actions)
+
def _cancel_debounce(self) -> None:
if self._debounce_task and not self._debounce_task.done():
self._debounce_task.cancel()
diff --git a/skyvern/services/browser_recording/service.py b/skyvern/services/browser_recording/service.py
index c26129303..20fac7208 100644
--- a/skyvern/services/browser_recording/service.py
+++ b/skyvern/services/browser_recording/service.py
@@ -1,8 +1,6 @@
-import asyncio
import base64
import functools
import json
-import pathlib
import re
import typing as t
import zlib
@@ -14,15 +12,11 @@ import structlog
import skyvern.services.browser_recording.state_machines as sm
from skyvern.client.types.workflow_definition_yaml_blocks_item import (
WorkflowDefinitionYamlBlocksItem_Action,
+ WorkflowDefinitionYamlBlocksItem_Code,
WorkflowDefinitionYamlBlocksItem_GotoUrl,
- WorkflowDefinitionYamlBlocksItem_Login,
WorkflowDefinitionYamlBlocksItem_Wait,
)
-from skyvern.client.types.workflow_definition_yaml_parameters_item import (
- WorkflowDefinitionYamlParametersItem,
- WorkflowDefinitionYamlParametersItem_Credential,
- WorkflowDefinitionYamlParametersItem_Workflow,
-)
+from skyvern.client.types.workflow_definition_yaml_parameters_item import WorkflowDefinitionYamlParametersItem
from skyvern.config import settings
from skyvern.forge import app
from skyvern.forge.prompts import prompt_engine
@@ -39,43 +33,13 @@ from skyvern.services.browser_recording.types import (
ActionKind,
ActionUrlChange,
ActionWait,
- CredentialKind,
ExfiltratedCdpEvent,
ExfiltratedConsoleEvent,
ExfiltratedEvent,
- OutputBlock,
- ProcessedBlock,
RecordingDraftStep,
)
-def summarize_exfiltrated_recording_events(events: list[ExfiltratedEvent]) -> dict[str, t.Any]:
- cdp_by_event_name: dict[str, int] = {}
- console_by_dom_type: dict[str, int] = {}
- console_by_exfil_event_name: dict[str, int] = {}
- cdp_total = 0
- console_total = 0
-
- for ev in events:
- if isinstance(ev, ExfiltratedCdpEvent):
- cdp_total += 1
- cdp_by_event_name[ev.event_name] = cdp_by_event_name.get(ev.event_name, 0) + 1
- elif isinstance(ev, ExfiltratedConsoleEvent):
- console_total += 1
- dom_type = ev.params.type
- console_by_dom_type[dom_type] = console_by_dom_type.get(dom_type, 0) + 1
- console_by_exfil_event_name[ev.event_name] = console_by_exfil_event_name.get(ev.event_name, 0) + 1
-
- return {
- "recording_exfil_total_events": len(events),
- "recording_exfil_cdp_event_count": cdp_total,
- "recording_exfil_console_event_count": console_total,
- "recording_exfil_cdp_event_name_counts": cdp_by_event_name,
- "recording_exfil_console_dom_type_counts": console_by_dom_type,
- "recording_exfil_console_exfil_event_name_counts": console_by_exfil_event_name,
- }
-
-
def _durable_recording_url(url: str) -> str:
parsed = urlparse(url)
if parsed.scheme not in {"http", "https"} or not parsed.hostname:
@@ -256,13 +220,12 @@ def build_durable_recording_evidence(actions: list[Action]) -> list[dict[str, t.
def build_durable_recording_metadata(
*,
draft_steps: list[RecordingDraftStep] | None,
- blocks: list[ProcessedBlock],
+ blocks: list[WorkflowDefinitionYamlBlocksItem_Code],
parameters: list[WorkflowDefinitionYamlParametersItem],
- code_first: bool,
interpretation_session_id: str | None,
) -> dict[str, t.Any]:
return {
- "code_first": code_first,
+ "code_first": True,
"interpretation_session_id": interpretation_session_id,
"draft_steps": [
{
@@ -290,8 +253,6 @@ MAX_BASE64_SIZE = 14 * 1024 * 1024 # ~10MB compressed + base64 overhead
# allows a generous ~10x expansion for legitimate recordings while rejecting bombs that would
# otherwise inflate to gigabytes and exhaust process memory on a shared host.
MAX_DECOMPRESSED_SIZE = 100 * 1024 * 1024 # 100MB
-DEFAULT_DRAFT_ACTION_TITLE = "Browser Action"
-DEFAULT_LOGIN_BLOCK_TITLE = "Log in"
def _gunzip_bounded(compressed_data: bytes, max_output_size: int) -> bytes | None:
@@ -398,134 +359,6 @@ def deterministic_goto_url_label(url: str) -> str:
return normalize_recording_block_label(f"goto_{host}" if host else None, fallback="goto_url")
-LOGIN_CREDENTIAL_KINDS: frozenset[CredentialKind] = frozenset({"password", "totp", "magic_link"})
-
-# A credential parameter resolves to a dict of fields, so binding one to a recorded fill means
-# naming the field that fill typed. Only `secret` is unambiguous: SecretCredential holds a single
-# value. A credit card spreads over card_number / card_cvv / card_exp_month / card_exp_year /
-# card_holder_name, and the recorder classifies the field only as "credit_card" - it does not
-# capture which one, so there is nothing to point the fill at. Card fills stay unbound.
-CREDENTIAL_FIELD_BY_KIND: dict[CredentialKind, str] = {"secret": "secret_value"}
-
-
-def _credential_id(step: RecordingDraftStep) -> str:
- return (step.credential_id or "").strip()
-
-
-def _is_login_credential_step(step: RecordingDraftStep) -> bool:
- return bool(_credential_id(step)) and step.credential_kind in LOGIN_CREDENTIAL_KINDS
-
-
-def _is_secret_fill(step: RecordingDraftStep) -> bool:
- return bool(_credential_id(step)) and step.credential_kind in CREDENTIAL_FIELD_BY_KIND
-
-
-def _is_typed_form_field(step: RecordingDraftStep) -> bool:
- return step.block_type == "action" and step.action_kind == ActionKind.INPUT_TEXT
-
-
-def _is_submit_click(step: RecordingDraftStep) -> bool:
- return step.block_type == "action" and step.action_kind == ActionKind.CLICK
-
-
-def bound_credential_ids(draft_steps: list[RecordingDraftStep]) -> set[str]:
- """Credentials the generated blocks reference.
-
- Emitted blocks carry each credential under its id, as a parameter-key *token*: the recorder
- cannot see the target workflow, so it cannot pick a key that is free there. The editor
- allocates the real key when it applies the blocks and substitutes it for the token.
- """
- return {_credential_id(step) for step in draft_steps if _is_login_credential_step(step) or _is_secret_fill(step)}
-
-
-def bind_credential_to_action_goal(navigation_goal: str, placeholder_keys: list[str], field_reference: str) -> str:
- """Repoint the fill instruction's `{{ placeholder }}` at the credential field.
-
- The recorder wrote the placeholder itself (`Type 'API token' with {{ api_token }}.`) as a
- workflow parameter with an empty default; left alone it renders empty and the credential
- goes unused.
- """
- goal = navigation_goal
- for key in placeholder_keys:
- goal = re.sub(r"\{\{\s*" + re.escape(key) + r"\s*\}\}", "{{ " + field_reference + " }}", goal)
- return goal
-
-
-def _is_login_fill_of(draft_steps: list[RecordingDraftStep], at: int, credential_id: str) -> bool:
- return (
- at < len(draft_steps)
- and _is_login_credential_step(draft_steps[at])
- and _credential_id(draft_steps[at]) == credential_id
- )
-
-
-class LoginRun(t.NamedTuple):
- """The steps one login block replaces, plus the credential step it is built from."""
-
- anchor: int
- indices: list[int]
-
-
-def login_block_runs(draft_steps: list[RecordingDraftStep]) -> dict[int, LoginRun]:
- """Run-start index -> the run one login block replaces.
-
- A login block fills and submits the form itself from the credential, so it replaces the
- steps that typed that form and nothing more: the credential fill, the one field typed just
- before it (the identifier), further fills of the same credential with at most the submit
- click between them (password, then a TOTP code), and one trailing submit click.
-
- Every bound is tight because each loosening has a demonstrated way to eat recorded steps.
- Reaching over same-URL steps swallows the session on a modal or single-page-app login;
- absorbing every preceding field eats a signup form's name and email; grouping every use of
- one credential erases the work between a login and a later re-auth. Clicks and text entry
- are what ordinary work looks like, so only adjacency separates a second field of the same
- form from that work.
-
- The cost is under-collapsing: a login split over two pages (identifier, then password on a
- new URL) leaves its first page a separate action block that still executes, typing an
- empty parameter into the identifier before the login block runs. That is visible and
- editable on the canvas; a step absorbed by a wrong guess is neither.
- """
- runs: dict[int, LoginRun] = {}
- claimed: set[int] = set()
- total = len(draft_steps)
- index = 0
-
- while index < total:
- step = draft_steps[index]
- if index in claimed or not _is_login_credential_step(step):
- index += 1
- continue
-
- credential_id = _credential_id(step)
- start = end = index
-
- if start - 1 >= 0 and start - 1 not in claimed and _is_typed_form_field(draft_steps[start - 1]):
- start -= 1
-
- while True:
- if _is_login_fill_of(draft_steps, end + 1, credential_id):
- end += 1
- elif (
- end + 1 < total
- and _is_submit_click(draft_steps[end + 1])
- and _is_login_fill_of(draft_steps, end + 2, credential_id)
- ):
- end += 2
- else:
- break
-
- if end + 1 < total and _is_submit_click(draft_steps[end + 1]):
- end += 1
-
- indices = list(range(start, end + 1))
- runs[start] = LoginRun(anchor=index, indices=indices)
- claimed.update(indices)
- index = end + 1
-
- return runs
-
-
def deterministic_wait_seconds(duration_ms: int) -> int:
return int(max(duration_ms / 1000.0, ActionWait.MIN_DURATION_THRESHOLD_MS / 1000.0))
@@ -726,12 +559,12 @@ class Processor:
sm.Click(),
sm.Hover(),
sm.InputText(),
+ sm.FileUpload(),
sm.Select(),
# After InputText: an Enter that submits a field must record as the fill
# followed by the keypress, not replace it.
sm.PressKey(),
sm.UrlChange(),
- sm.Wait(),
]
for event in events:
@@ -769,246 +602,9 @@ class Processor:
# NOTE: append-only — the live interpreter calls this each iteration and
# tracks emitted actions by index, so collapsing here would shrink the list
- # and drop a later wait. Collapsing happens in the raw process() path only.
+ # and drop a later wait.
return actions
- @staticmethod
- def _collapse_consecutive_waits(actions: list[Action]) -> list[Action]:
- collapsed: list[Action] = []
-
- for action in actions:
- previous = collapsed[-1] if collapsed else None
- if isinstance(action, ActionWait) and isinstance(previous, ActionWait):
- collapsed[-1] = ActionWait(
- kind=ActionKind.WAIT.value,
- target=previous.target,
- timestamp_start=previous.timestamp_start,
- timestamp_end=action.timestamp_end,
- url=action.url,
- duration_ms=previous.duration_ms + action.duration_ms,
- )
- continue
- collapsed.append(action)
-
- return collapsed
-
- def dedupe_block_labels(self, suspects: list[OutputBlock]) -> list[OutputBlock]:
- """
- Detect if any block labels are duplicated, and, if so, rename them for
- uniqueness.
- """
-
- blocks: list[OutputBlock] = []
- labels: set[str] = set()
-
- for block in suspects:
- if block.label not in labels:
- labels.add(block.label)
- blocks.append(block)
- continue
- else:
- original_label = block.label
- count = 0
- while True:
- new_label = f"{original_label}_{count}"
- if new_label not in labels:
- cls = block.__class__
- data = block.model_dump() | {"label": new_label}
- new_block = cls(**data)
- blocks.append(new_block)
- labels.add(new_label)
- break
- count += 1
-
- return blocks
-
- async def actions_to_blocks(self, actions: list[Action]) -> list[OutputBlock]:
- """
- Convert a list of `Action` objects into workflow definition (YAML) blocks.
- """
- tasks: list[asyncio.Task] = []
-
- for action in actions:
- action_kind = action.kind.value
-
- match action.kind:
- case ActionKind.CLICK | ActionKind.HOVER | ActionKind.INPUT_TEXT | ActionKind.PRESS_KEY:
- task = asyncio.create_task(self.create_action_block(action))
- tasks.append(task)
- case ActionKind.URL_CHANGE:
- task = asyncio.create_task(self.create_url_block(action))
- tasks.append(task)
- case ActionKind.WAIT:
- task = asyncio.create_task(self.create_wait_block(action))
- tasks.append(task)
- case _:
- LOG.warning(
- f"{self.class_name} Unknown action kind: {action_kind}",
- action=action,
- **self.identity,
- )
- continue
-
- blocks: list[OutputBlock] = await asyncio.gather(*tasks)
-
- blocks = self.dedupe_block_labels(blocks)
-
- return blocks
-
- def blocks_to_parameters(
- self,
- blocks: list[OutputBlock],
- credential_ids: set[str] | None = None,
- ) -> list[WorkflowDefinitionYamlParametersItem]:
- """
- Convert a list of workflow definition (YAML) blocks into workflow definition (YAML) parameters.
-
- Keys in `credential_ids` are credential tokens (see `bound_credential_ids`); they become
- credential parameters, not empty string workflow parameters.
- """
- credential_ids = credential_ids or set()
- parameter_names: set[str] = set()
-
- for block in blocks:
- if isinstance(block, WorkflowDefinitionYamlBlocksItem_Action):
- for param_name in block.parameter_keys or []:
- if param_name not in credential_ids:
- parameter_names.add(param_name)
-
- parameters: list[WorkflowDefinitionYamlParametersItem] = []
-
- for param_name in parameter_names:
- parameter = WorkflowDefinitionYamlParametersItem_Workflow(
- key=param_name,
- workflow_parameter_type="string",
- default_value="",
- description="",
- )
- parameters.append(parameter)
-
- for credential_id in sorted(credential_ids):
- parameters.append(
- WorkflowDefinitionYamlParametersItem_Credential(
- key=credential_id,
- credential_id=credential_id,
- description="",
- )
- )
-
- return parameters
-
- def login_run_to_block(
- self,
- draft_steps: list[RecordingDraftStep],
- run: LoginRun,
- ) -> WorkflowDefinitionYamlBlocksItem_Login:
- """
- Collapse a recorded login form interaction into a single login block.
- """
- credential_step = draft_steps[run.anchor]
- tokens = list(dict.fromkeys(_credential_id(draft_steps[index]) for index in run.indices))
- tokens = [token for token in tokens if token]
-
- return WorkflowDefinitionYamlBlocksItem_Login(
- label=normalize_recording_block_label(credential_step.label, fallback="login"),
- title=credential_step.title or DEFAULT_LOGIN_BLOCK_TITLE,
- url=(credential_step.url or "").strip() or None,
- parameter_keys=tokens,
- # Editor's convertToNode reads block.parameters.map(p => p.key); mirror the action block.
- parameters=[{"key": token} for token in tokens],
- )
-
- def drafts_to_blocks(self, draft_steps: list[RecordingDraftStep]) -> list[OutputBlock]:
- """
- Convert user-editable live recording drafts into workflow definition blocks.
- """
- blocks: list[OutputBlock] = []
- login_runs = login_block_runs(draft_steps)
- absorbed_indices = {index for run in login_runs.values() for index in run.indices}
-
- for index, draft_step in enumerate(draft_steps):
- if index in login_runs:
- blocks.append(self.login_run_to_block(draft_steps, login_runs[index]))
- continue
-
- if index in absorbed_indices:
- continue
-
- match draft_step.block_type:
- case "action":
- # A secret fill's placeholder parameter is replaced by the credential outright:
- # the credential field supplies the value, so the empty string parameter and the
- # instruction that referenced it both go.
- navigation_goal = draft_step.navigation_goal or ""
- parameters = list(draft_step.parameters)
- parameter_keys = list(draft_step.parameter_keys)
-
- if _is_secret_fill(draft_step) and draft_step.credential_kind:
- token = _credential_id(draft_step)
- field = CREDENTIAL_FIELD_BY_KIND[draft_step.credential_kind]
- navigation_goal = bind_credential_to_action_goal(
- navigation_goal, parameter_keys, f"{token}.{field}"
- )
- parameters = [{"key": token}]
- parameter_keys = [token]
-
- block = WorkflowDefinitionYamlBlocksItem_Action(
- label=normalize_recording_block_label(draft_step.label, fallback="act"),
- title=draft_step.title or DEFAULT_DRAFT_ACTION_TITLE,
- navigation_goal=navigation_goal,
- error_code_mapping=None,
- parameters=parameters,
- parameter_keys=parameter_keys,
- )
- case "goto_url":
- url = (draft_step.url or "").strip()
- if not url:
- LOG.warning(
- "skipping draft goto_url block with empty URL",
- draft_step=draft_step.model_dump(mode="json"),
- **self.identity,
- )
- continue
- fallback_label = deterministic_goto_url_label(url)
- title_candidate = (draft_step.title or "").strip()
- label_candidate = (draft_step.label or "").strip()
- if title_candidate:
- goto_label = normalize_recording_block_label(
- title_candidate,
- fallback=fallback_label,
- )
- elif label_candidate:
- goto_label = normalize_recording_block_label(
- label_candidate,
- fallback=fallback_label,
- )
- else:
- goto_label = fallback_label
- block = WorkflowDefinitionYamlBlocksItem_GotoUrl(
- label=goto_label,
- url=url,
- )
- case "wait":
- wait_sec = max(
- int(draft_step.wait_sec or 0),
- int(ActionWait.MIN_DURATION_THRESHOLD_MS / 1000.0),
- )
- block = WorkflowDefinitionYamlBlocksItem_Wait(
- label=normalize_recording_block_label(draft_step.label, fallback="wait"),
- wait_sec=wait_sec,
- )
- case _:
- LOG.warning(
- "skipping unsupported draft block type",
- draft_step=draft_step.model_dump(mode="json"),
- **self.identity,
- )
- continue
-
- blocks.append(block)
-
- return self.dedupe_block_labels(blocks)
-
async def create_action_block(self, action: ActionBlockable) -> WorkflowDefinitionYamlBlocksItem_Action:
"""
Create a YAML action block from an `ActionBlockable`.
@@ -1099,16 +695,20 @@ class Processor:
self,
compressed_chunks: list[str],
draft_steps: list[RecordingDraftStep] | None = None,
- code_first: bool = False,
+ recorded_actions: list[Action] | None = None,
supports_credential_tokens: bool = False,
- ) -> tuple[list[ProcessedBlock], list[WorkflowDefinitionYamlParametersItem], RecordingEvidencePacket | None]:
+ ) -> tuple[
+ list[WorkflowDefinitionYamlBlocksItem_Code],
+ list[WorkflowDefinitionYamlParametersItem],
+ RecordingEvidencePacket,
+ ]:
"""
Process the compressed browser session recording into workflow definition blocks.
"""
blocks, parameters, _, _, evidence = await self.process_with_evidence(
compressed_chunks,
draft_steps=draft_steps,
- code_first=code_first,
+ recorded_actions=recorded_actions,
supports_credential_tokens=supports_credential_tokens,
)
return blocks, parameters, evidence
@@ -1117,106 +717,54 @@ class Processor:
self,
compressed_chunks: list[str],
draft_steps: list[RecordingDraftStep] | None = None,
- code_first: bool = False,
+ recorded_actions: list[Action] | None = None,
supports_credential_tokens: bool = False,
) -> tuple[
- list[ProcessedBlock],
+ list[WorkflowDefinitionYamlBlocksItem_Code],
list[WorkflowDefinitionYamlParametersItem],
list[dict[str, t.Any]],
dict[str, t.Any],
- RecordingEvidencePacket | None,
+ RecordingEvidencePacket,
]:
- events = self.compressed_chunks_to_events(compressed_chunks)
- actions = self.events_to_actions(events)
- refinement_evidence: RecordingEvidencePacket | None = None
-
- if code_first:
- # Code-first always re-derives selector-bearing actions from raw events;
- # draft steps carry no locators and act only as an edit overlay.
- # Built from the actions this pass already derived so refinement never re-uploads
- # or re-derives the recording.
- refinement_evidence = build_recording_evidence(
- actions,
- draft_steps,
- browser_session_id=self.browser_session_id,
- workflow_permanent_id=self.workflow_permanent_id,
- recording_attempt_id=self.recording_attempt_id or f"rra_{uuid4().hex}",
- )
- code_first_result = actions_to_code_first_blocks(
- actions, draft_steps, bind_credentials=supports_credential_tokens
- )
- if code_first_result is not None:
- code_blocks, code_parameters = code_first_result
- LOG.info(
- "record_browser.process_recording_code_first",
- recording_code_block_count=len(code_blocks),
- recording_code_parameter_count=len(code_parameters),
- recording_action_count=len(actions),
- **self.identity,
- )
- blocks: list[ProcessedBlock] = list(code_blocks)
- return (
- blocks,
- code_parameters,
- build_durable_recording_evidence(actions),
- build_durable_recording_metadata(
- draft_steps=draft_steps,
- blocks=blocks,
- parameters=code_parameters,
- code_first=True,
- interpretation_session_id=self.interpretation_session_id,
- ),
- refinement_evidence,
- )
+ if recorded_actions is None:
+ events = self.compressed_chunks_to_events(compressed_chunks)
+ actions = self.events_to_actions(events)
+ else:
+ actions = recorded_actions
+ refinement_evidence = build_recording_evidence(
+ actions,
+ draft_steps,
+ browser_session_id=self.browser_session_id,
+ workflow_permanent_id=self.workflow_permanent_id,
+ recording_attempt_id=self.recording_attempt_id or f"rra_{uuid4().hex}",
+ )
+ result = actions_to_code_first_blocks(actions, draft_steps, bind_credentials=supports_credential_tokens)
+ if result is None:
+ blocks: list[WorkflowDefinitionYamlBlocksItem_Code] = []
+ parameters: list[WorkflowDefinitionYamlParametersItem] = []
LOG.warning(
- "record_browser.code_first_fallback_to_legacy",
+ "record_browser.process_recording_no_code_blocks",
recording_action_count=len(actions),
**self.identity,
)
-
- # `is not None` (not truthiness): an empty list means the user deleted every
- # live-interpreted step, which must not fall back to re-processing raw events.
- if draft_steps is not None:
+ else:
+ blocks, parameters = result
LOG.info(
- "record_browser.process_recording_drafts",
- recording_draft_step_count=len(draft_steps),
+ "record_browser.process_recording_code_first",
+ recording_code_block_count=len(blocks),
+ recording_code_parameter_count=len(parameters),
+ recording_action_count=len(actions),
**self.identity,
)
- blocks = self.drafts_to_blocks(draft_steps)
- parameters = self.blocks_to_parameters(blocks, bound_credential_ids(draft_steps))
- return (
- blocks,
- parameters,
- build_durable_recording_evidence(actions),
- build_durable_recording_metadata(
- draft_steps=draft_steps,
- blocks=blocks,
- parameters=parameters,
- code_first=code_first,
- interpretation_session_id=self.interpretation_session_id,
- ),
- refinement_evidence,
- )
-
- LOG.info(
- "record_browser.process_recording_payload",
- recording_compressed_chunk_count=len(compressed_chunks),
- **summarize_exfiltrated_recording_events(events),
- **self.identity,
- )
- actions = self._collapse_consecutive_waits(actions)
- blocks = await self.actions_to_blocks(actions)
- parameters = self.blocks_to_parameters(blocks)
return (
blocks,
parameters,
build_durable_recording_evidence(actions),
build_durable_recording_metadata(
- draft_steps=None,
+ draft_steps=draft_steps,
blocks=blocks,
parameters=parameters,
- code_first=code_first,
interpretation_session_id=self.interpretation_session_id,
),
refinement_evidence,
@@ -1231,15 +779,15 @@ class BrowserSessionRecordingService:
workflow_permanent_id: str,
compressed_chunks: list[str],
draft_steps: list[RecordingDraftStep] | None = None,
- code_first: bool = False,
+ recorded_actions: list[Action] | None = None,
supports_credential_tokens: bool = False,
recording_attempt_id: str | None = None,
interpretation_session_id: str | None = None,
) -> tuple[
- list[ProcessedBlock],
+ list[WorkflowDefinitionYamlBlocksItem_Code],
list[WorkflowDefinitionYamlParametersItem],
str | None,
- RecordingEvidencePacket | None,
+ RecordingEvidencePacket,
]:
"""
Process compressed browser session recording events into workflow definition blocks.
@@ -1255,7 +803,7 @@ class BrowserSessionRecordingService:
blocks, parameters, evidence, metadata, refinement_evidence = await processor.process_with_evidence(
compressed_chunks,
draft_steps=draft_steps,
- code_first=code_first,
+ recorded_actions=recorded_actions,
supports_credential_tokens=supports_credential_tokens,
)
if not blocks:
@@ -1270,64 +818,3 @@ class BrowserSessionRecordingService:
metadata=metadata,
)
return blocks, parameters, recording.recording_id, refinement_evidence
-
-
-async def smoke() -> None:
- with open(pathlib.Path("/path/to/uncompressed/events.json")) as f:
- raw_events: list[dict] = json.load(f)
-
- events: list[ExfiltratedEvent] = []
-
- for i, raw_event in enumerate(raw_events):
- if not isinstance(raw_event, dict):
- LOG.debug(f"~ skipping non-dict event: {raw_event}")
- continue
- if raw_event.get("source") == "cdp":
- try:
- event = ExfiltratedCdpEvent(**raw_event)
- except Exception:
- LOG.exception(f"{i} Failed to parse exfiltrated CDP event")
- LOG.debug(f"~ raw event: {json.dumps(raw_event, sort_keys=True, indent=2)}")
- continue
- events.append(event)
- elif raw_event.get("source") == "console":
- event = ExfiltratedConsoleEvent(**raw_event)
- events.append(event)
-
- LOG.debug(f"{len(events)} events.")
-
- my_local_org_id = "o_389844905020748346"
- processor = Processor("pbs_123", my_local_org_id, "wpid_123")
- actions = processor.events_to_actions(events)
-
- LOG.debug(f"{len(actions)} actions:")
-
- for action in actions:
- id = action.target.sky_id if action.target.sky_id else action.target.id
- text = ",".join(action.target.texts or [])
- LOG.debug(f" {action.kind} [{id}] [{text}] @ {action.url}")
-
- blocks = await processor.actions_to_blocks(actions)
-
- LOG.debug(f"{len(blocks)} blocks:")
-
- for block in blocks:
- LOG.debug(f" {block.label}")
-
- if isinstance(block, WorkflowDefinitionYamlBlocksItem_Action):
- LOG.debug(f" title: {block.title}")
- LOG.debug(f" nav goal: {block.navigation_goal}")
-
- if isinstance(block, WorkflowDefinitionYamlBlocksItem_GotoUrl):
- LOG.debug(f" url: {block.url}")
-
- if isinstance(block, WorkflowDefinitionYamlBlocksItem_Wait):
- LOG.debug(f" wait sec: {block.wait_sec}")
-
-
-# if __name__ == "__main__":
-# from skyvern.forge.forge_app_initializer import start_forge_app
-
-# start_forge_app()
-
-# asyncio.run(smoke())
diff --git a/skyvern/services/browser_recording/session_registry.py b/skyvern/services/browser_recording/session_registry.py
index 4a1db2321..f10a70697 100644
--- a/skyvern/services/browser_recording/session_registry.py
+++ b/skyvern/services/browser_recording/session_registry.py
@@ -1,5 +1,6 @@
import asyncio
import time
+from dataclasses import dataclass
import structlog
@@ -8,7 +9,7 @@ from skyvern.services.browser_recording.interpretation import (
OnRecordingInterpretationUpdate,
RecordingInterpretationSession,
)
-from skyvern.services.browser_recording.types import RecordingDraftStep
+from skyvern.services.browser_recording.types import Action, RecordingDraftStep
LOG = structlog.get_logger(__name__)
@@ -17,10 +18,21 @@ LOG = structlog.get_logger(__name__)
SESSION_TTL_SECONDS = 60 * 30
+@dataclass(frozen=True)
+class FinalizedRecordingActions:
+ browser_session_id: str
+ organization_id: str
+ workflow_permanent_id: str
+ actions: list[Action]
+
+
class RecordingInterpretationSessionRegistry:
def __init__(self) -> None:
self._sessions: dict[str, RecordingInterpretationSession] = {}
self._last_seen: dict[str, float] = {}
+ self._finalized_actions: dict[str, FinalizedRecordingActions] = {}
+ self._finalized_last_seen: dict[str, float] = {}
+ self._stopping_sessions: dict[str, asyncio.Future[list[RecordingDraftStep]]] = {}
def start_session(
self,
@@ -102,17 +114,122 @@ class RecordingInterpretationSessionRegistry:
session.resume_capture()
- async def stop_session(self, browser_session_id: str) -> list[RecordingDraftStep]:
+ async def stop_session(
+ self,
+ browser_session_id: str,
+ *,
+ retain_finalized_actions: bool = True,
+ ) -> list[RecordingDraftStep]:
+ stopping = self._stopping_sessions.get(browser_session_id)
+ if stopping is not None:
+ return await asyncio.shield(stopping)
+
+ stopping = asyncio.get_running_loop().create_future()
+ self._stopping_sessions[browser_session_id] = stopping
+ try:
+ drafts = await self._stop_session(
+ browser_session_id,
+ retain_finalized_actions=retain_finalized_actions,
+ )
+ except asyncio.CancelledError:
+ stopping.cancel()
+ raise
+ except Exception as exc:
+ stopping.set_exception(exc)
+ # The initiating caller receives the exception directly. Mark the
+ # future observed too in case there were no concurrent waiters.
+ stopping.exception()
+ raise
+ else:
+ stopping.set_result(drafts)
+ return drafts
+ finally:
+ if self._stopping_sessions.get(browser_session_id) is stopping:
+ self._stopping_sessions.pop(browser_session_id, None)
+
+ async def _stop_session(
+ self,
+ browser_session_id: str,
+ *,
+ retain_finalized_actions: bool,
+ ) -> list[RecordingDraftStep]:
session = self._sessions.pop(browser_session_id, None)
self._last_seen.pop(browser_session_id, None)
if not session:
return []
try:
- return await session.flush()
+ drafts = await session.flush()
+ if retain_finalized_actions:
+ self._finalized_actions[session.interpretation_session_id] = FinalizedRecordingActions(
+ browser_session_id=session.browser_session_id,
+ organization_id=session.organization_id,
+ workflow_permanent_id=session.workflow_permanent_id,
+ actions=session.recorded_actions(),
+ )
+ self._finalized_last_seen[session.interpretation_session_id] = time.monotonic()
+ return drafts
finally:
session.cancel()
+ def get_finalized_actions(
+ self,
+ *,
+ interpretation_session_id: str | None,
+ browser_session_id: str,
+ organization_id: str,
+ workflow_permanent_id: str,
+ ) -> list[Action] | None:
+ self._prune_expired_sessions()
+ if interpretation_session_id is None:
+ return None
+
+ finalized = self._finalized_actions.get(interpretation_session_id)
+ if finalized is None:
+ return None
+ if (
+ finalized.browser_session_id != browser_session_id
+ or finalized.organization_id != organization_id
+ or finalized.workflow_permanent_id != workflow_permanent_id
+ ):
+ LOG.warning(
+ "Rejected finalized recording actions with mismatched ownership",
+ interpretation_session_id=interpretation_session_id,
+ browser_session_id=browser_session_id,
+ organization_id=organization_id,
+ workflow_permanent_id=workflow_permanent_id,
+ )
+ return None
+
+ self._finalized_last_seen[interpretation_session_id] = time.monotonic()
+ return list(finalized.actions)
+
+ def discard_finalized_actions(self, interpretation_session_id: str | None) -> None:
+ if interpretation_session_id is None:
+ return
+
+ self._finalized_actions.pop(interpretation_session_id, None)
+ self._finalized_last_seen.pop(interpretation_session_id, None)
+
+ def discard_finalized_actions_if_owned(
+ self,
+ *,
+ interpretation_session_id: str,
+ browser_session_id: str,
+ organization_id: str,
+ workflow_permanent_id: str,
+ ) -> bool:
+ actions = self.get_finalized_actions(
+ interpretation_session_id=interpretation_session_id,
+ browser_session_id=browser_session_id,
+ organization_id=organization_id,
+ workflow_permanent_id=workflow_permanent_id,
+ )
+ if actions is None:
+ return False
+ self.discard_finalized_actions(interpretation_session_id)
+ return True
+
def discard_session(self, browser_session_id: str) -> None:
session = self._sessions.pop(browser_session_id, None)
self._last_seen.pop(browser_session_id, None)
@@ -130,11 +247,20 @@ class RecordingInterpretationSessionRegistry:
LOG.info("Pruning stale recording interpretation session", browser_session_id=browser_session_id)
self.discard_session(browser_session_id)
+ expired_finalized_ids = [
+ interpretation_session_id
+ for interpretation_session_id, last_seen in self._finalized_last_seen.items()
+ if now - last_seen > SESSION_TTL_SECONDS
+ ]
+ for interpretation_session_id in expired_finalized_ids:
+ self._finalized_last_seen.pop(interpretation_session_id, None)
+ self._finalized_actions.pop(interpretation_session_id, None)
+
async def stop_all(self) -> None:
await asyncio.gather(*(self.stop_session(browser_session_id) for browser_session_id in list(self._sessions)))
# Process-local singleton. Requires sticky routing (or a single worker) so the message
-# WebSocket and interpretation session stay on the same API instance. Multi-pod deployments
-# without affinity need shared session storage (e.g. Redis).
+# WebSocket, final action handoff, and interpretation session stay on the same API instance.
+# Multi-pod deployments without affinity need shared session storage (e.g. Redis).
interpretation_registry = RecordingInterpretationSessionRegistry()
diff --git a/skyvern/services/browser_recording/state_machines/__init__.py b/skyvern/services/browser_recording/state_machines/__init__.py
index 90a15c570..e9ad7b4b7 100644
--- a/skyvern/services/browser_recording/state_machines/__init__.py
+++ b/skyvern/services/browser_recording/state_machines/__init__.py
@@ -1,19 +1,19 @@
from .click import StateMachineClick as Click
+from .file_upload import StateMachineFileUpload as FileUpload
from .hover import StateMachineHover as Hover
from .input_text import StateMachineInputText as InputText
from .press_key import StateMachinePressKey as PressKey
from .select import StateMachineSelect as Select
from .state_machine import StateMachine
from .url_change import StateMachineUrlChange as UrlChange
-from .wait import StateMachineWait as Wait
__all__ = [
"Click",
+ "FileUpload",
"Hover",
"InputText",
"PressKey",
"Select",
"StateMachine",
"UrlChange",
- "Wait",
]
diff --git a/skyvern/services/browser_recording/state_machines/file_upload.py b/skyvern/services/browser_recording/state_machines/file_upload.py
new file mode 100644
index 000000000..ae2703e0f
--- /dev/null
+++ b/skyvern/services/browser_recording/state_machines/file_upload.py
@@ -0,0 +1,47 @@
+import structlog
+
+from skyvern.services.browser_recording.types import (
+ Action,
+ ActionInputText,
+ ActionKind,
+ ActionTarget,
+ ExfiltratedEvent,
+ Mouse,
+)
+
+from .state_machine import StateMachine
+
+LOG = structlog.get_logger()
+
+
+class StateMachineFileUpload(StateMachine):
+ def tick(self, event: ExfiltratedEvent, current_actions: list[Action]) -> ActionInputText | None:
+ if event.source != "console" or event.params.type != "change":
+ return None
+ target = event.params.target
+ if (target.tagName or "").upper() != "INPUT" or (target.inputType or "").lower() != "file":
+ return None
+ if not target.value:
+ return None
+
+ LOG.debug(f"~ emitting file upload action [{target.skyId or target.id}]")
+ return ActionInputText(
+ kind=ActionKind.INPUT_TEXT.value,
+ target=ActionTarget(
+ class_name=target.className,
+ id=target.id,
+ mouse=Mouse(xp=None, yp=None),
+ sky_id=target.skyId,
+ tag_name=target.tagName,
+ texts=target.text,
+ selector=target.selector,
+ role=target.role,
+ accessible_name=target.accessibleName,
+ input_type=target.inputType,
+ autocomplete=target.autocomplete,
+ ),
+ timestamp_start=event.params.timestamp,
+ timestamp_end=event.params.timestamp,
+ url=event.params.url,
+ input_value="",
+ )
diff --git a/skyvern/services/browser_recording/state_machines/select.py b/skyvern/services/browser_recording/state_machines/select.py
index b3b2577e7..b34c375c6 100644
--- a/skyvern/services/browser_recording/state_machines/select.py
+++ b/skyvern/services/browser_recording/state_machines/select.py
@@ -42,8 +42,8 @@ class StateMachineSelect(StateMachine):
value = target.value
# A card-expiry month is usually a <select autocomplete="cc-exp-month">, so a select can
# classify as secret and reach here with its value already nulled by ingest redaction.
- # Keep the step blank rather than dropping it, as StateMachineInputText does, so the
- # legacy path still shows the field; code-first drops an empty select_option regardless.
+ # Keep the live draft blank rather than dropping it, as StateMachineInputText does;
+ # final synthesis drops an empty select_option regardless.
secret = is_secret_field(
target.inputType,
target.autocomplete,
diff --git a/skyvern/services/browser_recording/state_machines/wait.py b/skyvern/services/browser_recording/state_machines/wait.py
deleted file mode 100644
index febf4a015..000000000
--- a/skyvern/services/browser_recording/state_machines/wait.py
+++ /dev/null
@@ -1,122 +0,0 @@
-import collections
-import typing as t
-
-import structlog
-
-from skyvern.services.browser_recording.types import (
- Action,
- ActionKind,
- ActionTarget,
- ActionWait,
- ExfiltratedEvent,
- Mouse,
-)
-
-from .state_machine import StateMachine
-
-LOG = structlog.get_logger()
-
-PAGE_ACTIVITY_EVENT_NAME = "net:activity"
-PAGE_ACTIVITY_EVENT_NAME_PREFIX = "nav:"
-
-MAX_PAGE_ACTIVITY_TIMESTAMPS = 256
-
-
-class StateMachineWait(StateMachine):
- state: t.Literal["void"] = "void"
- last_event_timestamp: float | None = None
- threshold_ms: int = ActionWait.MIN_DURATION_THRESHOLD_MS
-
- def __init__(self, threshold_ms: int | None = None) -> None:
- self.threshold_ms = max(
- threshold_ms or ActionWait.MIN_DURATION_THRESHOLD_MS,
- ActionWait.MIN_DURATION_THRESHOLD_MS,
- )
-
- self.page_activity_timestamps_ms: collections.deque[float] = collections.deque(
- maxlen=MAX_PAGE_ACTIVITY_TIMESTAMPS
- )
-
- # client_ms - server_ms, measured from console events (which carry both
- # clocks). Used to project server-stamped CDP activity timestamps into the
- # client clock so the gap and the busy-span are compared in one domain.
- # A property of the connection, not the wait window — never cleared on reset.
- self.client_server_offset_ms: float | None = None
-
- self.reset()
-
- def tick(self, event: ExfiltratedEvent, current_actions: list[Action]) -> ActionWait | None:
- if event.source == "cdp":
- if event.event_name == PAGE_ACTIVITY_EVENT_NAME or event.event_name.startswith(
- PAGE_ACTIVITY_EVENT_NAME_PREFIX
- ):
- offset_ms = self.client_server_offset_ms or 0.0
- self.page_activity_timestamps_ms.append(event.timestamp * 1000.0 + offset_ms)
- return None
-
- if event.source != "console":
- return None
-
- self.client_server_offset_ms = event.params.timestamp - event.timestamp * 1000.0
-
- if self.last_event_timestamp is not None:
- gap_ms = int(event.params.timestamp - self.last_event_timestamp)
-
- if gap_ms >= self.threshold_ms:
- # Size the wait to the page-busy span, not the full gap (drops the idle/think/lag tail).
- last_activity_ms = self._last_activity_in_gap_ms(self.last_event_timestamp, event.params.timestamp)
- busy_ms = int(last_activity_ms - self.last_event_timestamp) if last_activity_ms is not None else 0
-
- if last_activity_ms is None or busy_ms < self.threshold_ms:
- LOG.debug(
- "~ suppressing wait action: page settled before the wait threshold",
- gap_ms=gap_ms,
- busy_ms=busy_ms,
- )
- else:
- LOG.debug("~ emitting wait action", gap_ms=gap_ms, busy_ms=busy_ms)
-
- action_target = ActionTarget(
- class_name=None,
- id=None,
- mouse=Mouse(xp=None, yp=None),
- sky_id=None,
- tag_name=None,
- texts=[],
- )
-
- action = ActionWait(
- kind=ActionKind.WAIT.value,
- target=action_target,
- timestamp_start=self.last_event_timestamp,
- timestamp_end=last_activity_ms,
- url=event.params.url,
- duration_ms=busy_ms,
- )
-
- self.reset()
-
- return action
-
- self.last_event_timestamp = event.params.timestamp
-
- return None
-
- def _last_activity_in_gap_ms(self, gap_start_ms: float, gap_end_ms: float) -> float | None:
- while self.page_activity_timestamps_ms and self.page_activity_timestamps_ms[0] < gap_start_ms:
- self.page_activity_timestamps_ms.popleft()
-
- in_gap = [timestamp for timestamp in self.page_activity_timestamps_ms if timestamp <= gap_end_ms]
- return max(in_gap) if in_gap else None
-
- def on_action(self, action: Action, current_actions: list[Action]) -> bool:
- if action.kind == ActionKind.HOVER:
- return True
-
- self.reset()
-
- return True
-
- def reset(self) -> None:
- self.state = "void"
- self.last_event_timestamp = None
diff --git a/skyvern/services/browser_recording/types.py b/skyvern/services/browser_recording/types.py
index 40e5b910e..365f1b08c 100644
--- a/skyvern/services/browser_recording/types.py
+++ b/skyvern/services/browser_recording/types.py
@@ -10,7 +10,6 @@ from pydantic import BaseModel, ConfigDict, Field
from skyvern.client.types.workflow_definition_yaml_blocks_item import (
WorkflowDefinitionYamlBlocksItem_Action,
- WorkflowDefinitionYamlBlocksItem_Code,
WorkflowDefinitionYamlBlocksItem_GotoUrl,
WorkflowDefinitionYamlBlocksItem_Login,
WorkflowDefinitionYamlBlocksItem_Wait,
@@ -77,6 +76,8 @@ class ActionUrlChange(ActionBase):
class ActionWait(ActionBase):
+ """Compatibility model for recordings captured before inferred waits were disabled."""
+
kind: t.Literal[ActionKind.WAIT]
# --
duration_ms: int
@@ -124,10 +125,6 @@ OutputBlock = t.Union[
WorkflowDefinitionYamlBlocksItem_Wait,
]
-# What process_recording can return: legacy interpretable blocks, plus code blocks
-# when code-first mode is enabled. The live interpreter only ever sees OutputBlock.
-ProcessedBlock = t.Union[OutputBlock, WorkflowDefinitionYamlBlocksItem_Code]
-
class RecordingDraftStep(BaseModel):
step_id: str
diff --git a/skyvern/utils/url_validators.py b/skyvern/utils/url_validators.py
index 94cdbd9c5..995982f45 100644
--- a/skyvern/utils/url_validators.py
+++ b/skyvern/utils/url_validators.py
@@ -13,6 +13,13 @@ from skyvern.exceptions import BlockedHost, InvalidUrl, SkyvernHTTPException, Un
SAFE_REDIRECT_STATUS_CODES = {301, 302, 303, 307, 308}
MAX_SAFE_REDIRECTS = 10
+# getaddrinfo codes that mean the resolver answered "this name has no address", as opposed to
+# EAI_AGAIN/EAI_FAIL, which mean the resolver could not answer at all. EAI_NODATA is absent on
+# some platforms and folded into EAI_NONAME on others.
+_NO_SUCH_HOST_DNS_ERRNOS = frozenset(
+ getattr(socket, name) for name in ("EAI_NONAME", "EAI_NODATA") if hasattr(socket, name)
+)
+
_BLOCKED_INTERNAL_HOSTNAMES = frozenset({"localhost", "metadata.google.internal", "kubernetes.default.svc"})
_BLOCKED_INTERNAL_SUFFIXES = (".local", ".localhost", ".internal", ".cluster.local")
_LOCAL_BROWSER_HOSTNAMES = frozenset({"localhost", "host.docker.internal"})
@@ -277,6 +284,37 @@ def resolve_fetch_host_ips(host: str) -> tuple[str, ...]:
return tuple(resolved_ips)
+def host_has_no_address_record(host: str) -> bool:
+ """True only when the resolver answers definitively that ``host`` has no address.
+
+ Deliberately narrower than ``UnresolvableHost``: a timeout or SERVFAIL means the resolver could
+ not answer, and reading that as a dead host would misattribute every navigation failure during
+ a resolver outage. Callers use this to attribute a navigation that already failed, never to
+ decide whether one is allowed -- ``is_blocked_host`` ignores worker-side resolution failures
+ for that reason, since the browser resolves through the run proxy.
+ """
+ normalized = _normalize_host(host)
+ if not normalized:
+ return False
+
+ try:
+ ipaddress.ip_address(normalized)
+ except ValueError:
+ pass
+ else:
+ # A literal address needs no resolution, so DNS can say nothing about it.
+ return False
+
+ try:
+ socket.getaddrinfo(normalized, None, type=socket.SOCK_STREAM)
+ except socket.gaierror as error:
+ # gaierror subclasses OSError, so it must be caught first.
+ return error.errno in _NO_SUCH_HOST_DNS_ERRNOS
+ except (OSError, UnicodeError):
+ return False
+ return False
+
+
def _raise_if_best_effort_fetch_host_is_blocked(url: str) -> None:
# Browsers treat a backslash in the authority as a separator; urlsplit does not, which would
# otherwise let "http://<blocked-ip>\.example.com" read as an unrelated host here while the
diff --git a/skyvern/webeye/browser_state.py b/skyvern/webeye/browser_state.py
index c709f7a99..b8ac6b0e5 100644
--- a/skyvern/webeye/browser_state.py
+++ b/skyvern/webeye/browser_state.py
@@ -78,7 +78,7 @@ class BrowserState(Protocol):
async def set_working_page(self, page: Page | None, index: int = 0) -> None: ...
- async def set_active_page(self, page: Page) -> None: ...
+ async def set_active_page(self, page: Page, *, prune_excess_pages: bool = True) -> None: ...
async def navigate_to_url(
self,
diff --git a/skyvern/webeye/navigation.py b/skyvern/webeye/navigation.py
index 5eaff361a..ac00160ca 100644
--- a/skyvern/webeye/navigation.py
+++ b/skyvern/webeye/navigation.py
@@ -14,8 +14,14 @@ from skyvern.constants import (
PERMANENT_NAV_ERRORS,
SKIP_INNER_NAV_RETRY_ERRORS,
)
-from skyvern.exceptions import BlockedHost, BlockedNavigationDestination, FailedToNavigateToUrl, InvalidUrl
-from skyvern.utils.url_validators import canonical_navigation_host, is_blocked_host
+from skyvern.exceptions import (
+ BlockedHost,
+ BlockedNavigationDestination,
+ FailedToNavigateToUrl,
+ InvalidUrl,
+ UnresolvableNavigationHost,
+)
+from skyvern.utils.url_validators import canonical_navigation_host, host_has_no_address_record, is_blocked_host
LOG = structlog.get_logger()
@@ -126,6 +132,38 @@ def is_egress_attributable_navigation_error(error_message: str) -> bool:
return any(pattern in error_message for pattern in EGRESS_ATTRIBUTABLE_NAV_ERRORS)
+def is_egress_attributable_navigation_failure(error: BaseException) -> bool:
+ """Whether a navigation failure is attributable to our own egress rather than the target.
+
+ The error code alone does not settle it: a target with no DNS address record makes the run
+ proxy report a tunnel failure, and ``UnresolvableNavigationHost`` is that case already
+ distinguished. Prefer this over the message-only form wherever the exception is in hand.
+ """
+ if isinstance(error, UnresolvableNavigationHost):
+ return False
+ message = error.error_message if isinstance(error, FailedToNavigateToUrl) else str(error)
+ return is_egress_attributable_navigation_error(message)
+
+
+async def _unresolvable_navigation_host(url: str, error_message: str) -> str | None:
+ """The host of ``url`` when an egress-attributable failure was really a dead target.
+
+ A proxied browser delegates hostname resolution to the proxy, so a target whose address record
+ is gone comes back as the proxy failing to open a tunnel. Corroborating with our own resolver is
+ what separates "the pool cannot reach this host" from "this host no longer exists" -- the latter
+ must not redraw a proxy node, page our egress, or tell the customer our network is at fault.
+ """
+ if not is_egress_attributable_navigation_error(error_message):
+ return None
+ try:
+ host = canonical_navigation_host(url)
+ except InvalidUrl:
+ return None
+ if not host:
+ return None
+ return host if await asyncio.to_thread(host_has_no_address_record, host) else None
+
+
def redact_url_secrets(url: str) -> str:
"""Reduce a URL to scheme and host for display.
@@ -210,6 +248,20 @@ async def navigate_with_retry(
safe_error_str = error_str.replace(url, display_url) if log_url is not None else error_str
if is_skip_inner_retry_error(error_str):
+ # The host is the one part of a secret URL that redaction keeps, so naming it is
+ # safe for a redacting caller too.
+ dead_host = await _unresolvable_navigation_host(url, error_str)
+ if dead_host is not None:
+ LOG.warning(
+ "Navigation target has no DNS address record, failing without a proxy retry",
+ url=display_url,
+ host=dead_host,
+ error=safe_error_str,
+ )
+ raise UnresolvableNavigationHost(
+ url=display_url, host=dead_host, error_message=safe_error_str
+ ) from (None if redacting else error)
+
LOG.warning(
"Non-retriable navigation error, failing immediately",
url=display_url,
diff --git a/skyvern/webeye/real_browser_state.py b/skyvern/webeye/real_browser_state.py
index 8dc32ad2c..1deaf88e4 100644
--- a/skyvern/webeye/real_browser_state.py
+++ b/skyvern/webeye/real_browser_state.py
@@ -27,6 +27,7 @@ from skyvern.exceptions import (
FailedToReloadPage,
FailedToStopLoadingPage,
MissingBrowserStatePage,
+ UnresolvableNavigationHost,
)
from skyvern.forge import app
from skyvern.forge.sdk.core import skyvern_context
@@ -57,6 +58,34 @@ class _BrowserConnectionProbeFailed(str):
"""Diagnostic reason whose stale driver may be stopped before replacement."""
+def browser_context_stopped_reason(context: BrowserContext | None) -> str | None:
+ """Why this context's driver is unusable, or None while it still looks live.
+
+ A reused browser state (e.g. a persistent debug session) can have a stopped driver after a
+ prior owner's cleanup; page.goto then raises "Connection closed while reading from the
+ driver". A bare pw.stop() leaves browser.is_connected() stale and never flips
+ _close_was_called, so the shared driver Connection's closed-error is inspected too.
+ """
+ if context is None:
+ return "browser_context_missing"
+ impl = getattr(context, "_impl_obj", None)
+ if getattr(impl, "_close_was_called", False) is True:
+ return "browser_context_close_called"
+ if getattr(impl, "_closed", False) is True:
+ return "browser_context_closed"
+ connection = getattr(impl, "_connection", None)
+ if getattr(connection, "_closed_error", None) is not None:
+ return "playwright_driver_connection_closed"
+ browser = getattr(context, "browser", None)
+ if browser is None:
+ return None
+ try:
+ connected = bool(browser.is_connected())
+ except Exception as exc:
+ return _BrowserConnectionProbeFailed(f"browser_connection_probe_failed:{type(exc).__name__}")
+ return None if connected else "browser_context_disconnected"
+
+
def _same_page_ignoring_fragment(left: str | None, right: str | None) -> bool:
if not left or not right:
return False
@@ -505,9 +534,19 @@ class RealBrowserState(BrowserState):
if max_pages <= 0 or len(pages) <= max_pages:
return pages
- reserved_pages = pages[-max_pages:]
-
- closing_pages = pages[: len(pages) - max_pages]
+ # Oldest first, skipping the selected tab: a code block can switch to an older tab, and
+ # closing it here would leave get_working_page resuming on the newest one. Skipping it costs
+ # the next-oldest page instead, so the list still comes back within the cap.
+ active_page = self.__active_page
+ excess = len(pages) - max_pages
+ closing_pages: list[Page] = []
+ for page in pages:
+ if len(closing_pages) == excess:
+ break
+ if page is not active_page:
+ closing_pages.append(page)
+ closing_ids = {id(page) for page in closing_pages}
+ reserved_pages = [page for page in pages if id(page) not in closing_ids]
LOG.warning(
"The page number exceeds the limit, closing the oldest pages. It might cause the video missing",
closing_pages=closing_pages,
@@ -554,10 +593,14 @@ class RealBrowserState(BrowserState):
self.__active_page = None
self.__active_page_known_pages = set()
- async def set_active_page(self, page: Page) -> None:
+ async def set_active_page(self, page: Page, *, prune_excess_pages: bool = True) -> None:
self.__active_page = page
self.__page = page
- self.__active_page_known_pages = set(await self.list_valid_pages())
+ # list_valid_pages closes the oldest tabs past the cap, and the snapshot below only needs to
+ # know which pages existed. A caller that is pinning a tab on someone else's behalf passes
+ # False so pinning never closes a tab; max_pages <= 0 lists without pruning.
+ max_pages = settings.BROWSER_MAX_PAGES_NUMBER if prune_excess_pages else 0
+ self.__active_page_known_pages = set(await self.list_valid_pages(max_pages))
async def get_or_create_page(
self,
@@ -594,6 +637,10 @@ class RealBrowserState(BrowserState):
browser_session_id=browser_session_id,
)
except Exception as e:
+ # Recreating the context draws a different proxy node, and no proxy node can invent an
+ # address record for a host that has none.
+ if isinstance(e, UnresolvableNavigationHost):
+ raise
error_message = e.error_message if isinstance(e, FailedToNavigateToUrl) else str(e)
if is_permanent_navigation_error(error_message):
raise
@@ -803,29 +850,8 @@ class RealBrowserState(BrowserState):
return self._browser_state_diagnostic
def _connection_status(self) -> tuple[bool, str | None]:
- # A reused browser state (e.g. a persistent debug session) can have a stopped driver
- # after a prior owner's cleanup; page.goto then raises "Connection closed while reading
- # from the driver". A bare pw.stop() leaves browser.is_connected() stale and never flips
- # _close_was_called, so also inspect the shared driver Connection's closed-error.
- context = self.browser_context
- if context is None:
- return False, "browser_context_missing"
- impl = getattr(context, "_impl_obj", None)
- if getattr(impl, "_close_was_called", False) is True:
- return False, "browser_context_close_called"
- if getattr(impl, "_closed", False) is True:
- return False, "browser_context_closed"
- connection = getattr(impl, "_connection", None)
- if getattr(connection, "_closed_error", None) is not None:
- return False, "playwright_driver_connection_closed"
- browser = getattr(context, "browser", None)
- if browser is None:
- return True, None
- try:
- connected = bool(browser.is_connected())
- except Exception as exc:
- return False, _BrowserConnectionProbeFailed(f"browser_connection_probe_failed:{type(exc).__name__}")
- return connected, None if connected else "browser_context_disconnected"
+ stopped = browser_context_stopped_reason(self.browser_context)
+ return stopped is None, stopped
def is_connected(self) -> bool:
connected, reason = self._connection_status()
diff --git a/skyvern/webeye/skycdp/facade/locator.py b/skyvern/webeye/skycdp/facade/locator.py
index a8e1c66ff..a238f070f 100644
--- a/skyvern/webeye/skycdp/facade/locator.py
+++ b/skyvern/webeye/skycdp/facade/locator.py
@@ -7,6 +7,7 @@ offers, and the reason callers use locators instead of handles.
from __future__ import annotations
+import re
from dataclasses import dataclass
from typing import TYPE_CHECKING, Any
@@ -288,6 +289,11 @@ const __queryAll = (root, selector) => {
"""
+# Playwright's parseSelectorString: a quote inside an unquoted `text=` body is literal, so `text=Men's >> nth=0`
+# still splits at `>>`.
+_UNQUOTED_TEXT_BODY = re.compile(r"^\s*text\s*=.")
+
+
def _split_selector_chain(selector: str) -> list[str]:
"""Split a Playwright `a >> b` chain into steps at the top level, leaving quoted `>>` alone."""
parts: list[str] = []
@@ -296,11 +302,16 @@ def _split_selector_chain(selector: str) -> list[str]:
i = 0
while i < len(selector):
ch = selector[i]
+ if ch == "\\" and i + 1 < len(selector):
+ current.append(ch)
+ current.append(selector[i + 1])
+ i += 2
+ continue
if quote is not None:
if ch == quote:
quote = None
current.append(ch)
- elif ch in "\"'":
+ elif ch in "\"'" and not _UNQUOTED_TEXT_BODY.match("".join(current)):
quote = ch
current.append(ch)
elif selector.startswith(">>", i):
@@ -402,7 +413,17 @@ class _LocatorOperation:
def _query_operations(selector: str) -> list[_LocatorOperation]:
if not isinstance(selector, str):
raise TypeError("selector must be a string")
- return [_LocatorOperation(kind="query", selector=part) for part in _split_selector_chain(selector)]
+ operations: list[_LocatorOperation] = []
+ for part in _split_selector_chain(selector):
+ if part.startswith("nth="):
+ try:
+ index = int(part[len("nth=") :])
+ except ValueError:
+ raise CdpError(f"invalid nth selector step {part!r}") from None
+ operations.append(_LocatorOperation(kind="nth", index=index))
+ else:
+ operations.append(_LocatorOperation(kind="query", selector=part))
+ return operations
def _validate_text_filter(name: str, value: str | None) -> None:
diff --git a/skyvern/webeye/skycdp/facade/page.py b/skyvern/webeye/skycdp/facade/page.py
index 35ea57d2e..1cd4102cc 100644
--- a/skyvern/webeye/skycdp/facade/page.py
+++ b/skyvern/webeye/skycdp/facade/page.py
@@ -294,6 +294,8 @@ class EventContextManager:
class Page:
+ _skyvern_brokerable_handle = True
+
def __init__(self, context: BrowserContext, session: CdpSession) -> None:
self._context = context
self._session = session
diff --git a/tests/unit/forge/sdk/routes/streaming/test_cdp_input_session.py b/tests/unit/forge/sdk/routes/streaming/test_cdp_input_session.py
index bde5731f7..d49609355 100644
--- a/tests/unit/forge/sdk/routes/streaming/test_cdp_input_session.py
+++ b/tests/unit/forge/sdk/routes/streaming/test_cdp_input_session.py
@@ -92,6 +92,59 @@ class _FakePage:
self.url = url
+class _FakeFileInput:
+ def __init__(self) -> None:
+ self.files: dict[str, str | bytes] | None = None
+
+ async def count(self) -> int:
+ return 1
+
+ async def set_input_files(self, files: dict[str, str | bytes]) -> None:
+ self.files = files
+
+
+class _FakeUploadPage:
+ def __init__(self) -> None:
+ self.input = _FakeFileInput()
+ self.selector = ""
+
+ def locator(self, selector: str) -> _FakeFileInput:
+ self.selector = selector
+ return self.input
+
+
+class _FakeUploadWebSocket:
+ def __init__(self) -> None:
+ self.sent: list[dict[str, Any]] = []
+
+ async def send_json(self, value: dict[str, Any]) -> None:
+ self.sent.append(value)
+
+
+async def test_file_upload_dispatches_to_the_pages_only_file_input() -> None:
+ page = _FakeUploadPage()
+ websocket = _FakeUploadWebSocket()
+
+ await cdp_input._dispatch_file_upload_event(
+ page,
+ {
+ "fileName": "equipment_checklist.txt",
+ "mimeType": "text/plain",
+ "content": "Y2hlY2tsaXN0",
+ },
+ websocket, # type: ignore[arg-type]
+ )
+
+ assert page.selector == 'input[type="file"]'
+ assert page.input.files == {
+ "name": "equipment_checklist.txt",
+ "mimeType": "text/plain",
+ "buffer": b"checklist",
+ }
+ assert websocket.sent == [{"kind": "file-upload-complete"}]
+
+
def _build(monkeypatch: pytest.MonkeyPatch, page: Any) -> tuple[cdp_input.ActivePageCdpInputSession, list[Any]]:
resolved: list[Any] = [page]
diff --git a/tests/unit/services/test_browser_recording.py b/tests/unit/services/test_browser_recording.py
index 7ee09860c..103b72809 100644
--- a/tests/unit/services/test_browser_recording.py
+++ b/tests/unit/services/test_browser_recording.py
@@ -34,7 +34,6 @@ from skyvern.services.browser_recording.service import (
_resolve_enrichment_handler,
build_durable_recording_evidence,
deterministic_input_text_parameter_key,
- summarize_exfiltrated_recording_events,
)
from skyvern.services.browser_recording.state_machines.press_key import playwright_key
from skyvern.services.browser_recording.types import (
@@ -746,7 +745,7 @@ def make_focus_event(target: dict[str, t.Any], timestamp: float) -> ExfiltratedC
return make_console_event(params=params, timestamp=timestamp)
-def test_wait_suppressed_when_page_idle() -> None:
+def test_elapsed_time_does_not_emit_wait() -> None:
target = dict(id="button-1", skyId="sky-123", tagName="BUTTON", text=["Click me"])
events = [
@@ -760,7 +759,7 @@ def test_wait_suppressed_when_page_idle() -> None:
assert [action.kind for action in actions] == [ActionKind.CLICK]
-def test_wait_emitted_and_sized_to_page_busy_span() -> None:
+def test_page_activity_does_not_emit_wait() -> None:
target = dict(id="button-1", skyId="sky-123", tagName="BUTTON", text=["Click me"])
events = [
@@ -774,167 +773,9 @@ def test_wait_emitted_and_sized_to_page_busy_span() -> None:
processor = Processor(PBS_ID, ORG_ID, WP_ID)
actions = processor.events_to_actions(events)
- assert [action.kind for action in actions] == [ActionKind.CLICK, ActionKind.WAIT]
- wait_action = actions[1]
- assert isinstance(wait_action, ActionWait)
- # busy span = last activity (7000) - click (1000); the 2s idle tail is excluded.
- assert wait_action.duration_ms == 6000
-
-
-def test_wait_suppressed_when_page_settles_before_threshold() -> None:
- target = dict(id="button-1", skyId="sky-123", tagName="BUTTON", text=["Click me"])
-
- events = [
- make_click_event(target=target, timestamp=1000.0),
- # page settles ~1.5s in, then a long idle tail — below the wait threshold
- make_cdp_event("net:activity", timestamp_seconds=2.5, params={"count": 4}),
- make_focus_event(target=target, timestamp=12000.0),
- ]
-
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- actions = processor.events_to_actions(events)
-
assert [action.kind for action in actions] == [ActionKind.CLICK]
-def _two_consecutive_wait_events() -> list[t.Any]:
- # Two busy stretches (focus events produce no action) yield two adjacent waits;
- # a wait resets the timer, so each stretch needs its own pair of focus events.
- target = dict(id="button-1", skyId="sky-123", tagName="BUTTON", text=["Click me"])
- return [
- make_focus_event(target=target, timestamp=1000.0),
- make_cdp_event("net:activity", timestamp_seconds=6.5, params={"count": 9}),
- make_focus_event(target=target, timestamp=7000.0),
- make_focus_event(target=target, timestamp=13000.0),
- make_cdp_event("net:activity", timestamp_seconds=18.5, params={"count": 9}),
- make_focus_event(target=target, timestamp=19000.0),
- ]
-
-
-def test_events_to_actions_keeps_waits_separate_for_live_path() -> None:
- # events_to_actions feeds the incremental live interpreter, which tracks
- # actions by index, so it must stay append-only (no collapsing here).
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- actions = processor.events_to_actions(_two_consecutive_wait_events())
-
- assert [action.kind for action in actions] == [ActionKind.WAIT, ActionKind.WAIT]
-
-
-def test_collapse_consecutive_waits_merges_durations() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- actions = processor.events_to_actions(_two_consecutive_wait_events())
-
- collapsed = Processor._collapse_consecutive_waits(actions)
-
- assert [action.kind for action in collapsed] == [ActionKind.WAIT]
- wait_action = collapsed[0]
- assert isinstance(wait_action, ActionWait)
- # 5500ms + 5500ms busy spans summed into a single wait.
- assert wait_action.duration_ms == 11000
-
-
-def make_skewed_console_event(
- event_type: str,
- target: dict[str, t.Any],
- client_ms: float,
- server_skew_seconds: float,
-) -> ExfiltratedConsoleEvent:
- """A console event whose server clock is offset from the client clock."""
- params: dict[str, t.Any] = {
- "type": event_type,
- "target": target,
- "timestamp": client_ms,
- "url": "https://example.com",
- "activeElement": {"tagName": "BUTTON"},
- "window": {"height": 800, "width": 1200, "scrollX": 0, "scrollY": 0},
- "mousePosition": {"xp": 0.5, "yp": 0.5},
- }
- return ExfiltratedConsoleEvent(
- kind="exfiltrated-event",
- source="console",
- event_name="user_interaction",
- params=params,
- timestamp=client_ms / 1000.0 + server_skew_seconds,
- )
-
-
-def test_wait_offset_projection_cancels_client_server_clock_skew() -> None:
- # Server clock runs 60s ahead of the client clock. The Wait machine must
- # project the server-stamped CDP activity back into the client clock so the
- # busy span is measured correctly; otherwise the activity falls outside the
- # client-clock gap and the (real) wait is wrongly suppressed.
- skew = 60.0
- target = dict(id="button-1", skyId="sky-123", tagName="BUTTON", text=["Click me"])
-
- events = [
- make_skewed_console_event("click", target, client_ms=1000.0, server_skew_seconds=skew),
- # real activity at client 4s/7s -> server-stamped 64s/67s
- make_cdp_event("net:activity", timestamp_seconds=4.0 + skew, params={"count": 12}),
- make_cdp_event("net:activity", timestamp_seconds=7.0 + skew, params={"count": 3}),
- make_skewed_console_event("focus", target, client_ms=9000.0, server_skew_seconds=skew),
- ]
-
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- actions = processor.events_to_actions(events)
-
- assert [action.kind for action in actions] == [ActionKind.CLICK, ActionKind.WAIT]
- wait_action = actions[1]
- assert isinstance(wait_action, ActionWait)
- assert wait_action.duration_ms == 6000
-
-
-def test_wait_ignores_activity_outside_the_idle_gap() -> None:
- target = dict(id="button-1", skyId="sky-123", tagName="BUTTON", text=["Click me"])
-
- events = [
- # activity happened before the gap even started
- make_cdp_event("net:activity", timestamp_seconds=0.5, params={"count": 3}),
- make_click_event(target=target, timestamp=1000.0),
- make_focus_event(target=target, timestamp=8000.0),
- ]
-
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- actions = processor.events_to_actions(events)
-
- assert [action.kind for action in actions] == [ActionKind.CLICK]
-
-
-def test_summarize_exfiltrated_recording_events_mixed() -> None:
- target = dict(id="button-1", skyId="sky-123", tagName="BUTTON", text=["Click me"])
- click = make_click_event(target=target, timestamp=1000.0)
- keypress = make_console_event(
- params={
- "type": "keypress",
- "target": target,
- "timestamp": 1001.0,
- },
- timestamp=1001.0,
- )
- cdp_nav = ExfiltratedCdpEvent(
- kind="exfiltrated-event",
- event_name="nav:frame_navigated",
- params=ExfiltratedEventCdpParams(),
- source="cdp",
- timestamp=999.0,
- )
- cdp_nav_2 = ExfiltratedCdpEvent(
- kind="exfiltrated-event",
- event_name="nav:frame_navigated",
- params=ExfiltratedEventCdpParams(),
- source="cdp",
- timestamp=1002.0,
- )
-
- summary = summarize_exfiltrated_recording_events([cdp_nav, click, keypress, cdp_nav_2])
-
- assert summary["recording_exfil_total_events"] == 4
- assert summary["recording_exfil_cdp_event_count"] == 2
- assert summary["recording_exfil_console_event_count"] == 2
- assert summary["recording_exfil_cdp_event_name_counts"] == {"nav:frame_navigated": 2}
- assert summary["recording_exfil_console_dom_type_counts"] == {"click": 1, "keypress": 1}
- assert summary["recording_exfil_console_exfil_event_name_counts"] == {"user_interaction": 2}
-
-
def test_gunzip_bounded_returns_full_payload_under_limit() -> None:
payload = b"hello world" * 500
compressed = gzip.compress(payload)
@@ -1540,6 +1381,38 @@ def test_change_on_a_checkbox_records_only_the_click() -> None:
assert [a.kind for a in actions] == [ActionKind.CLICK]
+def test_change_on_a_file_input_records_an_upload_without_the_local_path() -> None:
+ target = {
+ "id": "fileInput",
+ "skyId": "sky-file",
+ "tagName": "INPUT",
+ "inputType": "file",
+ "selector": "#fileInput",
+ "value": r"C:\fakepath\equipment_checklist.txt",
+ }
+
+ actions = Processor(PBS_ID, ORG_ID, WP_ID).events_to_actions([make_change_event(target=target, timestamp=1000.0)])
+
+ assert len(actions) == 1
+ assert isinstance(actions[0], ActionInputText)
+ assert actions[0].target.input_type == "file"
+ assert actions[0].input_value == ""
+
+
+def test_clearing_a_file_input_does_not_record_an_upload() -> None:
+ target = {
+ "id": "fileInput",
+ "tagName": "INPUT",
+ "inputType": "file",
+ "selector": "#fileInput",
+ "value": "",
+ }
+
+ assert (
+ Processor(PBS_ID, ORG_ID, WP_ID).events_to_actions([make_change_event(target=target, timestamp=1000.0)]) == []
+ )
+
+
def test_select_placeholder_row_records_nothing() -> None:
target = {**SELECT_TARGET, "value": ""}
diff --git a/tests/unit/services/test_browser_recording_code_first.py b/tests/unit/services/test_browser_recording_code_first.py
index 159969f56..fe3a15d34 100644
--- a/tests/unit/services/test_browser_recording_code_first.py
+++ b/tests/unit/services/test_browser_recording_code_first.py
@@ -183,6 +183,22 @@ def test_password_input_becomes_parameter_without_default() -> None:
assert parameters[0].default_value == ""
+def test_recorded_targets_get_extended_visibility_waits() -> None:
+ actions: list[Action] = [
+ make_click(1000, selector="#generate"),
+ make_click(40000, selector="#download"),
+ ]
+
+ result = actions_to_code_first_blocks(actions, None)
+
+ assert result is not None
+ blocks, _ = result
+ readiness = 'await page.locator("#download").wait_for(state="visible", timeout=120000)'
+ click = 'await page.locator("#download").click()'
+ assert readiness in blocks[0].code
+ assert blocks[0].code.index(readiness) < blocks[0].code.index(click)
+
+
def test_select_element_maps_to_select_option() -> None:
actions: list[Action] = [
make_input(1000, "CA", selector="#state", tag_name="SELECT", accessible_name="State"),
@@ -195,6 +211,48 @@ def test_select_element_maps_to_select_option() -> None:
assert 'await page.locator("#state").select_option("CA")' in blocks[0].code
+def test_file_input_maps_to_an_authorized_file_attachment() -> None:
+ actions: list[Action] = [
+ make_input(1000, "", selector="#fileInput", tag_name="INPUT", input_type="file"),
+ make_click(2000, selector="#submit", role="button", accessible_name="Submit checklist"),
+ ]
+
+ result = actions_to_code_first_blocks(actions, None)
+
+ assert result is not None
+ blocks, parameters = result
+ assert 'await attach_authorized_file(page, upload_file, "#fileInput")' in blocks[0].code
+ assert "set_input_files" not in blocks[0].code
+ assert blocks[0].parameter_keys == ["upload_file"]
+ assert len(parameters) == 1
+ assert parameters[0].key == "upload_file"
+ assert parameters[0].workflow_parameter_type == "file_url"
+ assert parameters[0].default_value == ""
+
+
+def test_file_inputs_in_separate_blocks_keep_distinct_parameter_references() -> None:
+ actions: list[Action] = [
+ make_input(1000, "", selector="#firstFile", tag_name="INPUT", input_type="file"),
+ make_url_change(5000, "https://example.com/second"),
+ make_input(
+ 6000,
+ "",
+ url="https://example.com/second",
+ selector="#secondFile",
+ tag_name="INPUT",
+ input_type="file",
+ ),
+ ]
+
+ result = actions_to_code_first_blocks(actions, None)
+
+ assert result is not None
+ blocks, parameters = result
+ assert 'attach_authorized_file(page, upload_file, "#firstFile")' in blocks[0].code
+ assert 'attach_authorized_file(page, upload_file_2, "#secondFile")' in blocks[1].code
+ assert [parameter.key for parameter in parameters] == ["upload_file", "upload_file_2"]
+
+
def test_wait_and_hover_emit_deterministic_lines() -> None:
actions: list[Action] = [
make_hover(1000, selector="#menu", accessible_name="Menu"),
@@ -428,19 +486,17 @@ async def test_process_binds_credentials_only_for_a_caller_that_substitutes_toke
)
processor = Processor(PBS_ID, ORG_ID, WP_ID)
- unsupported_blocks, _, _ = await processor.process(["chunk"], draft_steps=drafts, code_first=True)
- supported_blocks, _, _ = await processor.process(
- ["chunk"], draft_steps=drafts, code_first=True, supports_credential_tokens=True
- )
+ unsupported_blocks, _, _ = await processor.process(["chunk"], draft_steps=drafts)
+ supported_blocks, _, _ = await processor.process(["chunk"], draft_steps=drafts, supports_credential_tokens=True)
- # The route defaults to off, so an old frontend never receives code reading a token it
+ # Token substitution defaults to off, so an old frontend never receives code reading a token it
# cannot rename - which would persist a block that fails at run time with a NameError.
assert "cred_123" not in unsupported_blocks[0].code
assert "cred_123.password" in supported_blocks[0].code
@pytest.mark.asyncio
-async def test_process_code_first_returns_code_blocks_and_evidence(monkeypatch: pytest.MonkeyPatch) -> None:
+async def test_process_returns_code_blocks_and_evidence(monkeypatch: pytest.MonkeyPatch) -> None:
typed_value = "hunter2-distinct"
actions: list[Action] = [
make_input(1000, typed_value, selector="#query", input_type="text"),
@@ -454,7 +510,7 @@ async def test_process_code_first_returns_code_blocks_and_evidence(monkeypatch:
)
processor = Processor(PBS_ID, ORG_ID, WP_ID, recording_attempt_id="rra_test")
- blocks, _, evidence = await processor.process(["chunk"], code_first=True)
+ blocks, _, evidence = await processor.process(["chunk"])
assert len(blocks) == 1
assert blocks[0].block_type == "code"
@@ -468,20 +524,32 @@ async def test_process_code_first_returns_code_blocks_and_evidence(monkeypatch:
@pytest.mark.asyncio
-async def test_process_code_first_falls_back_to_legacy_when_synthesis_empty(
+async def test_process_does_not_fall_back_to_legacy_blocks_when_synthesis_is_rejected(
monkeypatch: pytest.MonkeyPatch,
) -> None:
monkeypatch.setattr(Processor, "compressed_chunks_to_events", lambda self, chunks: [])
+ action = make_click(1000)
+ monkeypatch.setattr(
+ Processor,
+ "events_to_actions",
+ lambda self, events, machines=None, initial_actions=None: [action],
+ )
+
+ async def fail_legacy_conversion(*args: object, **kwargs: object) -> object:
+ raise AssertionError("legacy draft conversion must not run")
+
+ monkeypatch.setattr(Processor, "create_action_block", fail_legacy_conversion)
processor = Processor(PBS_ID, ORG_ID, WP_ID)
- blocks, parameters, _ = await processor.process(["chunk"], code_first=True)
+ blocks, parameters, evidence = await processor.process(["chunk"], draft_steps=[draft_for(action)])
assert blocks == []
assert parameters == []
+ assert evidence is not None
@pytest.mark.asyncio
-async def test_process_code_first_prefers_draft_overlay_over_drafts_to_blocks(
+async def test_process_applies_draft_overlay_before_code_synthesis(
monkeypatch: pytest.MonkeyPatch,
) -> None:
kept = make_click(1000, selector="#keep", accessible_name="Keep")
@@ -494,7 +562,7 @@ async def test_process_code_first_prefers_draft_overlay_over_drafts_to_blocks(
)
processor = Processor(PBS_ID, ORG_ID, WP_ID)
- blocks, _, _ = await processor.process(["chunk"], draft_steps=[draft_for(kept)], code_first=True)
+ blocks, _, _ = await processor.process(["chunk"], draft_steps=[draft_for(kept)])
assert len(blocks) == 1
assert blocks[0].block_type == "code"
diff --git a/tests/unit/test_browser_recording_interpretation.py b/tests/unit/test_browser_recording_interpretation.py
index 7e934dd90..352e2ea3f 100644
--- a/tests/unit/test_browser_recording_interpretation.py
+++ b/tests/unit/test_browser_recording_interpretation.py
@@ -1,7 +1,6 @@
from __future__ import annotations
import asyncio
-import typing as t
from unittest.mock import AsyncMock, MagicMock
import pytest
@@ -16,7 +15,7 @@ from skyvern.services.browser_recording.interpretation import (
RecordingInterpretationSession,
streaming_events_to_recording_events,
)
-from skyvern.services.browser_recording.service import Processor, bound_credential_ids
+from skyvern.services.browser_recording.service import Processor
from skyvern.services.browser_recording.types import (
ActionKind,
ActionTarget,
@@ -66,140 +65,6 @@ def test_streaming_console_event_reifies_for_recording_processor() -> None:
assert reified[0].params.target.skyId == "sky-1"
-def test_drafts_to_blocks_preserves_action_parameters_and_sanitizes_duplicate_labels() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- parameter = {
- "key": "customer_name",
- "workflow_parameter_type": "string",
- "default_value": "",
- "description": "",
- }
- drafts = [
- RecordingDraftStep(
- step_id="step-1",
- action_kind=ActionKind.CLICK,
- block_type="action",
- label="123 Submit!",
- title="Submit form",
- navigation_goal="Click submit",
- parameters=[parameter],
- parameter_keys=["customer_name"],
- ),
- RecordingDraftStep(
- step_id="step-2",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="123 Submit!",
- title="Type name",
- navigation_goal="Type the customer name",
- parameters=[parameter],
- parameter_keys=["customer_name"],
- ),
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
- parameters = processor.blocks_to_parameters(blocks)
-
- assert [block.label for block in blocks] == ["act_123_Submit", "act_123_Submit_0"]
- assert blocks[0].parameters == [parameter]
- assert blocks[0].parameter_keys == ["customer_name"]
- assert [parameter.key for parameter in parameters] == ["customer_name"]
-
-
-def test_drafts_to_blocks_skips_empty_goto_url() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- RecordingDraftStep(
- step_id="step-1",
- action_kind=ActionKind.URL_CHANGE,
- block_type="goto_url",
- label="visit",
- url="",
- )
- ]
-
- assert processor.drafts_to_blocks(drafts) == []
-
-
-def test_drafts_to_blocks_goto_url_label_follows_edited_title_and_url() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- RecordingDraftStep(
- step_id="step-1",
- action_kind=ActionKind.URL_CHANGE,
- block_type="goto_url",
- label="goto_wikipedia_com",
- title="Go to wikipedia.org",
- url="https://wikipedia.org/wiki/Foo",
- )
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
-
- assert len(blocks) == 1
- assert blocks[0].label == "Go_to_wikipedia_org"
- assert blocks[0].url == "https://wikipedia.org/wiki/Foo"
-
-
-def test_drafts_to_blocks_goto_url_label_derives_from_url_without_title_or_label() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- RecordingDraftStep(
- step_id="step-1",
- action_kind=ActionKind.URL_CHANGE,
- block_type="goto_url",
- label="",
- url="https://www.wikipedia.org/wiki/Foo",
- )
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
-
- assert len(blocks) == 1
- assert blocks[0].label == "goto_www_wikipedia_org"
- assert blocks[0].url == "https://www.wikipedia.org/wiki/Foo"
-
-
-def test_drafts_to_blocks_goto_url_label_preserves_edited_label_without_title() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- RecordingDraftStep(
- step_id="step-1",
- action_kind=ActionKind.URL_CHANGE,
- block_type="goto_url",
- label="Open Wikipedia",
- url="https://www.wikipedia.org/wiki/Foo",
- )
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
-
- assert len(blocks) == 1
- assert blocks[0].label == "Open_Wikipedia"
- assert blocks[0].url == "https://www.wikipedia.org/wiki/Foo"
-
-
-async def test_processor_process_uses_draft_steps_without_compressed_chunks() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- RecordingDraftStep(
- step_id="step-1",
- action_kind=ActionKind.WAIT,
- block_type="wait",
- label="wait",
- wait_sec=2,
- )
- ]
-
- blocks, parameters, _ = await processor.process([], draft_steps=drafts)
-
- assert len(blocks) == 1
- assert blocks[0].block_type == "wait"
- assert blocks[0].wait_sec == 5
- assert parameters == []
-
-
def _click_streaming_event(
*,
timestamp: float = 1234.0,
@@ -221,6 +86,8 @@ def _click_streaming_event(
"id": target_id,
"text": ["Submit"],
"skyId": sky_id,
+ "selector": f"#{target_id}",
+ "accessibleName": target_id,
},
"mousePosition": {"xp": 0.5, "yp": 0.5},
"activeElement": {"tagName": "BUTTON"},
@@ -234,6 +101,40 @@ def _click_streaming_event(
)
+async def test_live_interpretation_drops_inferred_waits() -> None:
+ session = RecordingInterpretationSession(
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ on_update=lambda _: None,
+ )
+ first_focus = _click_streaming_event(timestamp=1000.0, capture_seq=0)
+ first_focus.params["type"] = "focus"
+ first_focus.timestamp = 1.0
+ second_focus = _click_streaming_event(timestamp=7000.0, capture_seq=2)
+ second_focus.params["type"] = "focus"
+ second_focus.timestamp = 7.0
+ session.ingest_events(
+ [
+ first_focus,
+ StreamingExfiltratedEvent(
+ event_name="net:activity",
+ source=StreamingExfiltratedEventSource.CDP,
+ timestamp=6.5,
+ capture_seq=1,
+ params={"count": 3},
+ ),
+ second_focus,
+ ]
+ )
+
+ steps = await session.flush()
+
+ assert steps == []
+ assert session.recorded_actions() == []
+
+
@pytest.mark.asyncio
async def test_jittered_reclick_yields_single_draft_step(monkeypatch: pytest.MonkeyPatch) -> None:
async def fake_llm(*args: object, **kwargs: object) -> dict[str, object]:
@@ -829,7 +730,7 @@ async def test_new_attempt_id_mid_recording_continues_session_and_keeps_drafts(
recording when it lost its in-memory state (e.g. page reload). The registry
must continue the populated session: resync the panel with the accumulated
drafts instead of blanking it, keep interpreting new events, and let the
- finished recording build blocks from everything captured.
+ finished recording retain everything captured.
"""
from skyvern.services.browser_recording.session_registry import RecordingInterpretationSessionRegistry
@@ -863,6 +764,7 @@ async def test_new_attempt_id_mid_recording_continues_session_and_keeps_drafts(
populated_snapshots = [u for u in panel if u.is_snapshot and u.steps]
assert populated_snapshots, "expected the panel to display the interpreted drafts"
accumulated_step_count = len(session_one.steps)
+ accumulated_step_ids = {step.step_id for step in session_one.steps}
assert accumulated_step_count >= 1
# A reconnect arrives with a NEW attempt id (same browser session, not finalized).
@@ -891,344 +793,210 @@ async def test_new_attempt_id_mid_recording_continues_session_and_keeps_drafts(
await session_two._interpret(finalized=False)
assert len(session_two.steps) > accumulated_step_count
- # Finishing builds blocks from everything captured across the reconnect.
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- blocks = processor.drafts_to_blocks(session_two.steps)
- assert len(blocks) == len(session_two.steps)
+ assert accumulated_step_ids < {step.step_id for step in session_two.steps}
+
+ drafts = await registry.stop_session(PBS_ID)
+ assert (
+ registry.get_finalized_actions(
+ interpretation_session_id=session_two.interpretation_session_id,
+ browser_session_id=PBS_ID,
+ organization_id="other-org",
+ workflow_permanent_id=WP_ID,
+ )
+ is None
+ )
+ actions = registry.get_finalized_actions(
+ interpretation_session_id=session_two.interpretation_session_id,
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ )
+ blocks, _, _ = await Processor(PBS_ID, ORG_ID, WP_ID).process(
+ [],
+ draft_steps=drafts,
+ recorded_actions=actions,
+ )
+ code = "\n".join(block.code for block in blocks)
+ assert "#a" in code
+ assert "#b" in code
+ assert "#c" in code
+
+ registry.discard_finalized_actions(session_two.interpretation_session_id)
+ assert (
+ registry.get_finalized_actions(
+ interpretation_session_id=session_two.interpretation_session_id,
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ )
+ is None
+ )
registry.discard_session(PBS_ID)
-def test_drafts_to_blocks_collapses_credentialed_login_form_into_one_login_block() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- login_url = "https://example.com/login"
- drafts = [
- RecordingDraftStep(
- step_id="step-goto",
- action_kind=ActionKind.URL_CHANGE,
- block_type="goto_url",
- label="goto",
- url=login_url,
- ),
- RecordingDraftStep(
- step_id="step-email",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="type_email",
- url=login_url,
- parameters=[{"key": "email"}],
- parameter_keys=["email"],
- ),
- RecordingDraftStep(
- step_id="step-password",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="type_password",
- title="Log into example",
- url=login_url,
- credential_kind="password",
- credential_id="cred_abc",
- ),
- RecordingDraftStep(
- step_id="step-submit",
- action_kind=ActionKind.CLICK,
- block_type="action",
- label="click_sign_in",
- url=login_url,
- ),
- RecordingDraftStep(
- step_id="step-after",
- action_kind=ActionKind.CLICK,
- block_type="action",
- label="click_dashboard",
- url="https://example.com/dashboard",
+async def test_process_recording_discards_finalized_actions_after_success(monkeypatch: pytest.MonkeyPatch) -> None:
+ from skyvern.forge.sdk.routes import browser_sessions as browser_sessions_routes
+ from skyvern.schemas.browser_sessions import ProcessBrowserSessionRecordingRequest
+
+ registry = MagicMock()
+ registry.get_finalized_actions.return_value = [MagicMock()]
+ monkeypatch.setattr(browser_sessions_routes, "interpretation_registry", registry)
+
+ persistent_sessions_manager = MagicMock()
+ persistent_sessions_manager.get_session = AsyncMock(return_value=MagicMock())
+ recording_service = MagicMock()
+ recording_service.process_recording = AsyncMock(return_value=([], [], None, None))
+ route_app = MagicMock(
+ PERSISTENT_SESSIONS_MANAGER=persistent_sessions_manager,
+ BROWSER_SESSION_RECORDING_SERVICE=recording_service,
+ AGENT_FUNCTION=MagicMock(validate_code_block=AsyncMock()),
+ )
+ monkeypatch.setattr(browser_sessions_routes, "app", route_app)
+
+ await browser_sessions_routes.process_recording(
+ browser_session_id=PBS_ID,
+ recording_request=ProcessBrowserSessionRecordingRequest(
+ workflow_permanent_id=WP_ID,
+ interpretation_session_id="interpretation-1",
),
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
- parameters = processor.blocks_to_parameters(blocks, bound_credential_ids(drafts))
-
- assert [block.block_type for block in blocks] == ["goto_url", "login", "action"]
- login_block = blocks[1]
- assert login_block.url == login_url
- assert login_block.title == "Log into example"
- # Keyed by the credential id: a token the editor swaps for a real key when it applies the blocks.
- assert login_block.parameter_keys == ["cred_abc"]
- assert login_block.parameters == [{"key": "cred_abc"}]
- # The email typing and the submit click are the login block's job now.
- assert blocks[2].label == "click_dashboard"
-
- credential_parameters = [p for p in parameters if p.parameter_type == "credential"]
- assert [(p.key, p.credential_id) for p in credential_parameters] == [("cred_abc", "cred_abc")]
- # The absorbed email step's placeholder parameter goes with it.
- assert [p.key for p in parameters if p.parameter_type == "workflow"] == []
-
+ current_org=MagicMock(organization_id=ORG_ID),
+ )
-def test_drafts_to_blocks_points_a_secret_fill_at_the_credential_field() -> None:
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- RecordingDraftStep(
- step_id="step-secret",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="type_api_token",
- url="https://example.com/settings",
- navigation_goal="Type 'API token' with {{ api_token }}.",
- parameters=[{"key": "api_token"}],
- parameter_keys=["api_token"],
- credential_kind="secret",
- credential_id="cred_secret",
- ),
- ]
+ registry.discard_finalized_actions.assert_called_once_with("interpretation-1")
- blocks = processor.drafts_to_blocks(drafts)
- parameters = processor.blocks_to_parameters(blocks, bound_credential_ids(drafts))
- assert [block.block_type for block in blocks] == ["action"]
- # The instruction reads the credential field, not the empty placeholder it replaced.
- assert blocks[0].navigation_goal == "Type 'API token' with {{ cred_secret.secret_value }}."
- assert blocks[0].parameter_keys == ["cred_secret"]
- assert blocks[0].parameters == [{"key": "cred_secret"}]
- assert [(p.parameter_type, p.key) for p in parameters] == [("credential", "cred_secret")]
+async def test_process_recording_retains_finalized_actions_after_failure(monkeypatch: pytest.MonkeyPatch) -> None:
+ from skyvern.forge.sdk.routes import browser_sessions as browser_sessions_routes
+ from skyvern.schemas.browser_sessions import ProcessBrowserSessionRecordingRequest
+
+ registry = MagicMock()
+ registry.get_finalized_actions.return_value = [MagicMock()]
+ monkeypatch.setattr(browser_sessions_routes, "interpretation_registry", registry)
+
+ persistent_sessions_manager = MagicMock()
+ persistent_sessions_manager.get_session = AsyncMock(return_value=MagicMock())
+ recording_service = MagicMock()
+ recording_service.process_recording = AsyncMock(side_effect=RuntimeError("processing failed"))
+ route_app = MagicMock(
+ PERSISTENT_SESSIONS_MANAGER=persistent_sessions_manager,
+ BROWSER_SESSION_RECORDING_SERVICE=recording_service,
+ AGENT_FUNCTION=MagicMock(validate_code_block=AsyncMock()),
+ )
+ monkeypatch.setattr(browser_sessions_routes, "app", route_app)
+
+ with pytest.raises(RuntimeError, match="processing failed"):
+ await browser_sessions_routes.process_recording(
+ browser_session_id=PBS_ID,
+ recording_request=ProcessBrowserSessionRecordingRequest(
+ workflow_permanent_id=WP_ID,
+ interpretation_session_id="interpretation-1",
+ ),
+ current_org=MagicMock(organization_id=ORG_ID),
+ )
+ registry.discard_finalized_actions.assert_not_called()
-def test_drafts_to_blocks_leaves_a_card_fill_unbound() -> None:
- """A card credential spans several fields and the recorder does not say which one was typed.
- Attaching it anyway would emit a credential parameter nothing references while the
- instruction still rendered the empty placeholder.
- """
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- RecordingDraftStep(
- step_id="step-card",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="type_card_number",
- url="https://example.com/checkout",
- navigation_goal="Type 'Card number' with {{ cardnumber }}.",
- parameters=[{"key": "cardnumber"}],
- parameter_keys=["cardnumber"],
- credential_kind="credit_card",
- credential_id="cred_card",
+async def test_process_recording_requires_code_block_access(monkeypatch: pytest.MonkeyPatch) -> None:
+ from skyvern.exceptions import DisabledBlockExecutionError
+ from skyvern.forge.sdk.routes import browser_sessions as browser_sessions_routes
+ from skyvern.schemas.browser_sessions import ProcessBrowserSessionRecordingRequest
+
+ persistent_sessions_manager = MagicMock()
+ persistent_sessions_manager.get_session = AsyncMock(return_value=MagicMock())
+ recording_service = MagicMock()
+ recording_service.process_recording = AsyncMock()
+ agent_function = MagicMock()
+ agent_function.validate_code_block = AsyncMock(side_effect=DisabledBlockExecutionError("CodeBlock is disabled"))
+ monkeypatch.setattr(
+ browser_sessions_routes,
+ "app",
+ MagicMock(
+ PERSISTENT_SESSIONS_MANAGER=persistent_sessions_manager,
+ BROWSER_SESSION_RECORDING_SERVICE=recording_service,
+ AGENT_FUNCTION=agent_function,
),
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
- parameters = processor.blocks_to_parameters(blocks, bound_credential_ids(drafts))
+ )
- assert [block.block_type for block in blocks] == ["action"]
- assert blocks[0].navigation_goal == "Type 'Card number' with {{ cardnumber }}."
- assert blocks[0].parameter_keys == ["cardnumber"]
- assert [(p.parameter_type, p.key) for p in parameters] == [("workflow", "cardnumber")]
+ with pytest.raises(DisabledBlockExecutionError, match="CodeBlock is disabled"):
+ await browser_sessions_routes.process_recording(
+ browser_session_id=PBS_ID,
+ recording_request=ProcessBrowserSessionRecordingRequest(workflow_permanent_id=WP_ID),
+ current_org=MagicMock(organization_id=ORG_ID),
+ )
+ recording_service.process_recording.assert_not_awaited()
-def test_drafts_to_blocks_keeps_post_login_steps_when_the_url_never_changes() -> None:
- """A modal/SPA login must not swallow the rest of the session.
- The whole recording sits on one URL, so a same-URL span reaches to the last step.
- """
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- app_url = "https://example.com/app"
- drafts = [
- RecordingDraftStep(
- step_id="step-goto",
- action_kind=ActionKind.URL_CHANGE,
- block_type="goto_url",
- label="goto",
- url=app_url,
- ),
- RecordingDraftStep(
- step_id="step-open-modal",
- action_kind=ActionKind.CLICK,
- block_type="action",
- label="click_sign_in_link",
- url=app_url,
- ),
- RecordingDraftStep(
- step_id="step-email",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="type_email",
- url=app_url,
- parameters=[{"key": "email"}],
- parameter_keys=["email"],
- ),
- RecordingDraftStep(
- step_id="step-password",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="type_password",
- url=app_url,
- credential_kind="password",
- credential_id="cred_abc",
- ),
- RecordingDraftStep(
- step_id="step-submit",
- action_kind=ActionKind.CLICK,
- block_type="action",
- label="click_submit",
- url=app_url,
- ),
- RecordingDraftStep(
- step_id="step-after-1",
- action_kind=ActionKind.CLICK,
- block_type="action",
- label="click_new_invoice",
- url=app_url,
- ),
- RecordingDraftStep(
- step_id="step-after-2",
- action_kind=ActionKind.INPUT_TEXT,
- block_type="action",
- label="type_amount",
- url=app_url,
- parameters=[{"key": "amount"}],
- parameter_keys=["amount"],
- ),
- RecordingDraftStep(
- step_id="step-after-3",
- action_kind=ActionKind.CLICK,
- block_type="action",
- label="click_save",
- url=app_url,
+async def test_process_recording_waits_for_late_finalized_actions(monkeypatch: pytest.MonkeyPatch) -> None:
+ from skyvern.forge.sdk.routes import browser_sessions as browser_sessions_routes
+ from skyvern.schemas.browser_sessions import ProcessBrowserSessionRecordingRequest
+
+ finalized_actions = [MagicMock()]
+ registry = MagicMock()
+ registry.get_finalized_actions.side_effect = [None, finalized_actions]
+ registry.stop_session = AsyncMock(return_value=[])
+ monkeypatch.setattr(browser_sessions_routes, "interpretation_registry", registry)
+
+ recording_service = MagicMock()
+ recording_service.process_recording = AsyncMock(return_value=([], [], None, None))
+ monkeypatch.setattr(
+ browser_sessions_routes,
+ "app",
+ MagicMock(
+ PERSISTENT_SESSIONS_MANAGER=MagicMock(get_session=AsyncMock(return_value=MagicMock())),
+ BROWSER_SESSION_RECORDING_SERVICE=recording_service,
+ AGENT_FUNCTION=MagicMock(validate_code_block=AsyncMock()),
),
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
- parameters = processor.blocks_to_parameters(blocks, bound_credential_ids(drafts))
-
- assert [block.label for block in blocks] == [
- "goto",
- "click_sign_in_link",
- "type_password",
- "click_new_invoice",
- "type_amount",
- "click_save",
- ]
- assert [block.block_type for block in blocks] == [
- "goto_url",
- "action",
- "login",
- "action",
- "action",
- "action",
- ]
- # The post-login work keeps its own parameters.
- assert sorted(p.key for p in parameters if p.parameter_type == "workflow") == ["amount"]
-
-
-def _step(step_id: str, kind: ActionKind, label: str, **kwargs: t.Any) -> RecordingDraftStep:
- return RecordingDraftStep(
- step_id=step_id,
- action_kind=kind,
- block_type=kwargs.pop("block_type", "action"),
- label=label,
- **kwargs,
)
+ await browser_sessions_routes.process_recording(
+ browser_session_id=PBS_ID,
+ recording_request=ProcessBrowserSessionRecordingRequest(
+ workflow_permanent_id=WP_ID,
+ interpretation_session_id="interpretation-1",
+ ),
+ current_org=MagicMock(organization_id=ORG_ID),
+ )
-def test_drafts_to_blocks_keeps_the_work_between_a_login_and_a_later_reauth() -> None:
- """The same credential used twice is two logins, not one span over everything between."""
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- _step("s1", ActionKind.INPUT_TEXT, "type_password_1", credential_kind="password", credential_id="cred_abc"),
- _step("s2", ActionKind.CLICK, "click_submit_1"),
- _step("s3", ActionKind.URL_CHANGE, "goto_dashboard", block_type="goto_url", url="https://example.com/app"),
- _step("s4", ActionKind.CLICK, "click_billing"),
- _step("s5", ActionKind.INPUT_TEXT, "type_amount", parameters=[{"key": "amount"}], parameter_keys=["amount"]),
- _step("s6", ActionKind.URL_CHANGE, "goto_reauth", block_type="goto_url", url="https://example.com/reauth"),
- _step("s7", ActionKind.INPUT_TEXT, "type_password_2", credential_kind="password", credential_id="cred_abc"),
- _step("s8", ActionKind.CLICK, "click_submit_2"),
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
-
- assert [block.label for block in blocks] == [
- "type_password_1",
- "goto_dashboard",
- "click_billing",
- "type_amount",
- "goto_reauth",
- "type_password_2",
- ]
- assert [block.block_type for block in blocks] == [
- "login",
- "goto_url",
- "action",
- "action",
- "goto_url",
- "login",
- ]
-
-
-def test_drafts_to_blocks_absorbs_only_the_field_next_to_the_credential() -> None:
- """A signup-style form's earlier fields are not the login block's to type."""
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- _step("s1", ActionKind.INPUT_TEXT, "type_first_name", parameters=[{"key": "first"}], parameter_keys=["first"]),
- _step("s2", ActionKind.INPUT_TEXT, "type_last_name", parameters=[{"key": "last"}], parameter_keys=["last"]),
- _step("s3", ActionKind.INPUT_TEXT, "type_email", parameters=[{"key": "email"}], parameter_keys=["email"]),
- _step("s4", ActionKind.INPUT_TEXT, "type_password", credential_kind="password", credential_id="cred_abc"),
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
- parameters = processor.blocks_to_parameters(blocks, bound_credential_ids(drafts))
-
- # Only the email field next to the password is absorbed; the name fields survive.
- assert [block.label for block in blocks] == ["type_first_name", "type_last_name", "type_password"]
- assert [block.block_type for block in blocks] == ["action", "action", "login"]
- assert sorted(p.key for p in parameters if p.parameter_type == "workflow") == ["first", "last"]
-
-
-def test_drafts_to_blocks_keeps_in_app_work_between_two_uses_of_one_credential() -> None:
- """A re-auth on a single-page app is a second login, not one span over the work between.
-
- Nothing navigates and every step in between is a click or a text entry, which is exactly
- what ordinary app work looks like - so adjacency, not step kind, has to bound the group.
- """
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- _step("s1", ActionKind.INPUT_TEXT, "type_password_1", credential_kind="password", credential_id="cred_abc"),
- _step("s2", ActionKind.CLICK, "click_submit_1"),
- _step("s3", ActionKind.CLICK, "click_invoices"),
- _step("s4", ActionKind.INPUT_TEXT, "type_amount", parameters=[{"key": "amount"}], parameter_keys=["amount"]),
- _step("s5", ActionKind.CLICK, "click_save"),
- _step("s6", ActionKind.CLICK, "click_reauth_prompt"),
- _step("s7", ActionKind.INPUT_TEXT, "type_password_2", credential_kind="password", credential_id="cred_abc"),
- ]
-
- blocks = processor.drafts_to_blocks(drafts)
-
- assert [block.label for block in blocks] == [
- "type_password_1",
- "click_invoices",
- "type_amount",
- "click_save",
- "click_reauth_prompt",
- "type_password_2",
- ]
- assert [block.block_type for block in blocks] == [
- "login",
- "action",
- "action",
- "action",
- "action",
- "login",
- ]
+ registry.stop_session.assert_awaited_once_with(PBS_ID)
+ assert recording_service.process_recording.await_args.kwargs["recorded_actions"] == finalized_actions
+ registry.discard_finalized_actions.assert_called_once_with("interpretation-1")
-def test_drafts_to_blocks_merges_a_totp_fill_across_the_submit_click() -> None:
- """One step may separate two fills of a credential: the submit between password and code."""
- processor = Processor(PBS_ID, ORG_ID, WP_ID)
- drafts = [
- _step("s1", ActionKind.INPUT_TEXT, "type_password", credential_kind="password", credential_id="cred_abc"),
- _step("s2", ActionKind.CLICK, "click_next"),
- _step("s3", ActionKind.INPUT_TEXT, "type_code", credential_kind="totp", credential_id="cred_abc"),
- _step("s4", ActionKind.CLICK, "click_verify"),
- ]
+async def test_concurrent_stop_session_waits_for_the_same_flush() -> None:
+ from skyvern.services.browser_recording.session_registry import RecordingInterpretationSessionRegistry
- blocks = processor.drafts_to_blocks(drafts)
+ registry = RecordingInterpretationSessionRegistry()
+ registry.start_session(
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ on_update=lambda _update: None,
+ )
+ session = registry._sessions[PBS_ID]
+ flush_started = asyncio.Event()
+ release_flush = asyncio.Event()
+
+ async def delayed_flush() -> list[RecordingDraftStep]:
+ flush_started.set()
+ await release_flush.wait()
+ return []
+
+ session.flush = AsyncMock(side_effect=delayed_flush)
+ first = asyncio.create_task(registry.stop_session(PBS_ID))
+ await flush_started.wait()
+ second = asyncio.create_task(registry.stop_session(PBS_ID))
+ await asyncio.sleep(0)
- assert [block.block_type for block in blocks] == ["login"]
- assert blocks[0].parameter_keys == ["cred_abc"]
+ assert not second.done()
+ release_flush.set()
+ assert await first == []
+ assert await second == []
+ session.flush.assert_awaited_once()
diff --git a/tests/unit/test_copilot_non_retriable_nav.py b/tests/unit/test_copilot_non_retriable_nav.py
index dc7b7b393..deb4490f0 100644
--- a/tests/unit/test_copilot_non_retriable_nav.py
+++ b/tests/unit/test_copilot_non_retriable_nav.py
@@ -56,6 +56,16 @@ _PROXY_FAILURE_REPLY = (
"Please try again later, or contact Skyvern Support if the problem continues."
)
_NON_PROXY_FAILURE_REPLY = "The target URL could not be reached. Error: {reason}. Please verify the URL and try again."
+# A host with no address record still carries the proxy's error code, because a proxied browser
+# delegates resolution to the proxy. UnresolvableNavigationHost prefixes the real cause.
+_NO_ADDRESS_RECORD_FAILURE_REASON = (
+ "Failed to navigate to url https://gone.example/. Error message: gone.example has no DNS "
+ "address record: net::ERR_TUNNEL_CONNECTION_FAILED"
+)
+_NO_ADDRESS_RECORD_REPLY = (
+ "The site's domain has no DNS record, so it cannot be reached from any network. "
+ "Error: {reason}. Please check the URL for a typo, or confirm the site is still online."
+)
def _fresh_context() -> CopilotContext:
@@ -111,6 +121,7 @@ def _record_and_render_terminal_reply(reason: str, *, block_type: str) -> tuple[
pytest.param("SSL error: net::ERR_SSL_PROTOCOL_ERROR", id="ssl_prefixed"),
pytest.param(_TUNNEL_FAILURE_REASON, id="tunnel_connection_failed"),
pytest.param(_SOCKS_FAILURE_REASON, id="socks_connection_failed"),
+ pytest.param(_NO_ADDRESS_RECORD_FAILURE_REASON, id="no_dns_address_record"),
pytest.param(
"Failed to navigate to url https://x.test. Error message: net::ERR_SOCKS_CONNECTION_HOST_UNREACHABLE",
id="socks_host_unreachable",
@@ -433,6 +444,17 @@ def test_proxy_transport_full_flow_uses_browser_network_reply(reason: str, block
assert ctx.last_test_non_retriable_nav_error == reason
+def test_a_host_with_no_address_record_is_not_answered_with_contact_support() -> None:
+ # The proxy error code is still in the reason, so without the marker check this renders the
+ # browser-network reply and sends the user to Support over a URL only they can fix.
+ _, result = _record_and_render_terminal_reply(_NO_ADDRESS_RECORD_FAILURE_REASON, block_type="navigation")
+
+ assert result.user_response == _NO_ADDRESS_RECORD_REPLY.format(reason=_NO_ADDRESS_RECORD_FAILURE_REASON)
+ assert "Skyvern Support" not in result.user_response
+ assert result.turn_outcome is not None
+ assert result.turn_outcome.terminal_reason == "non_retriable_nav"
+
+
@pytest.mark.parametrize(
"reason",
[
diff --git a/tests/unit/test_real_browser_state_active_page.py b/tests/unit/test_real_browser_state_active_page.py
index a4a76a1ba..a86b15b91 100644
--- a/tests/unit/test_real_browser_state_active_page.py
+++ b/tests/unit/test_real_browser_state_active_page.py
@@ -5,6 +5,7 @@ from unittest.mock import AsyncMock, MagicMock
import pytest
+from skyvern.config import settings
from skyvern.exceptions import MissingBrowserStatePage
from skyvern.webeye.real_browser_state import RealBrowserState
@@ -170,3 +171,51 @@ async def test_must_get_working_page_raises_when_context_is_gone() -> None:
with pytest.raises(MissingBrowserStatePage):
await state.must_get_working_page()
state.browser_context.new_page.assert_not_awaited()
+
+
+def _pages_over_the_cap() -> list[MagicMock]:
+ pages = [_mock_page(f"https://example.test/{index}") for index in range(settings.BROWSER_MAX_PAGES_NUMBER + 2)]
+ for page in pages:
+ page.close = AsyncMock()
+ return pages
+
+
+async def test_pinning_a_tab_without_pruning_closes_nothing() -> None:
+ pages = _pages_over_the_cap()
+ state = _make_state()
+ state.browser_context.pages = pages
+
+ await state.set_active_page(pages[-1], prune_excess_pages=False)
+
+ assert [page.close.await_count for page in pages] == [0] * len(pages)
+ assert await state.get_working_page() is pages[-1]
+
+
+async def test_pinning_a_tab_prunes_by_default() -> None:
+ pages = _pages_over_the_cap()
+ state = _make_state()
+ state.browser_context.pages = pages
+
+ await state.set_active_page(pages[-1])
+
+ assert [page.close.await_count for page in pages[:2]] == [1, 1]
+
+
+async def test_the_cap_keeps_the_selected_tab_even_when_it_is_oldest() -> None:
+ pages = _pages_over_the_cap()
+ state = _make_state()
+ state.browser_context.pages = pages
+ await state.set_active_page(pages[0], prune_excess_pages=False)
+
+ remaining = await state.list_valid_pages()
+
+ assert pages[0].close.await_count == 0
+ assert remaining[0] is pages[0]
+ # Reserving the selected tab costs the next-oldest one rather than the cap.
+ assert len(remaining) == settings.BROWSER_MAX_PAGES_NUMBER
+ assert pages[1].close.await_count == 1
+ assert pages[2].close.await_count == 1
+ assert await state.get_working_page() is pages[0]
diff --git a/tests/unit/test_record_browser_teardown_integration.py b/tests/unit/test_record_browser_teardown_integration.py
index b5e2fb086..bc6141133 100644
--- a/tests/unit/test_record_browser_teardown_integration.py
+++ b/tests/unit/test_record_browser_teardown_integration.py
@@ -24,8 +24,11 @@ from skyvern.forge.sdk.routes.streaming.channels.exfiltration import (
from skyvern.forge.sdk.routes.streaming.channels.exfiltration import (
ExfiltrationChannel,
)
-from skyvern.forge.sdk.routes.streaming.channels.message import MessageChannelContext
-from skyvern.services.browser_recording.service import Processor
+from skyvern.forge.sdk.routes.streaming.channels.message import (
+ MessageChannelContext,
+ MessageInEndExfiltration,
+ reify_channel_message,
+)
from skyvern.services.browser_recording.session_registry import RecordingInterpretationSessionRegistry
ORG_ID = "org_123"
@@ -126,19 +129,94 @@ async def _assert_closed_target_preserves_drafts(
await channel.stop() # message.py:727 — must not raise on a closed target
drafts = await registry.stop_session(PBS_ID) # message.py:731 — only reached if stop() didn't crash
- # The drafts survived teardown and convert into real workflow blocks.
+ # The drafts survived teardown for final code synthesis.
assert drafts, "teardown dropped the recorded drafts"
- blocks = Processor(PBS_ID, ORG_ID, WP_ID).drafts_to_blocks(drafts)
- assert blocks, "surviving drafts did not produce workflow blocks"
@pytest.mark.asyncio
-async def test_end_exfiltration_on_closed_target_preserves_drafts_into_blocks(monkeypatch: pytest.MonkeyPatch) -> None:
+async def test_end_exfiltration_on_closed_target_preserves_drafts(monkeypatch: pytest.MonkeyPatch) -> None:
await _assert_closed_target_preserves_drafts(monkeypatch, _vnc_context())
@pytest.mark.asyncio
-async def test_end_exfiltration_without_vnc_on_closed_target_preserves_drafts_into_blocks(
+async def test_end_exfiltration_without_vnc_on_closed_target_preserves_drafts(
monkeypatch: pytest.MonkeyPatch,
) -> None:
await _assert_closed_target_preserves_drafts(monkeypatch, _message_context())
+
+
+async def test_discard_end_exfiltration_does_not_cache_finalized_actions() -> None:
+ registry = RecordingInterpretationSessionRegistry()
+ registry.start_session(
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ on_update=lambda _update: None,
+ recording_attempt_id="attempt-1",
+ )
+ interpretation_session_id = registry._sessions[PBS_ID].interpretation_session_id
+
+ message = reify_channel_message({"kind": "end-exfiltration", "discard": True})
+ assert isinstance(message, MessageInEndExfiltration)
+ assert message.discard is True
+
+ await registry.stop_session(PBS_ID, retain_finalized_actions=not message.discard)
+
+ assert (
+ registry.get_finalized_actions(
+ interpretation_session_id=interpretation_session_id,
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ )
+ is None
+ )
+
+
+async def test_discard_after_finalization_purges_only_owned_actions() -> None:
+ registry = RecordingInterpretationSessionRegistry()
+ registry.start_session(
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ on_update=lambda _update: None,
+ recording_attempt_id="attempt-1",
+ )
+ interpretation_session_id = registry._sessions[PBS_ID].interpretation_session_id
+ await registry.stop_session(PBS_ID)
+
+ message = reify_channel_message(
+ {
+ "kind": "end-exfiltration",
+ "discard": True,
+ "interpretation_session_id": interpretation_session_id,
+ "workflow_permanent_id": WP_ID,
+ }
+ )
+ assert isinstance(message, MessageInEndExfiltration)
+ assert message.interpretation_session_id == interpretation_session_id
+ assert message.workflow_permanent_id == WP_ID
+
+ assert not registry.discard_finalized_actions_if_owned(
+ interpretation_session_id=interpretation_session_id,
+ browser_session_id=PBS_ID,
+ organization_id="another-org",
+ workflow_permanent_id=WP_ID,
+ )
+ assert registry.discard_finalized_actions_if_owned(
+ interpretation_session_id=interpretation_session_id,
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ )
+ assert (
+ registry.get_finalized_actions(
+ interpretation_session_id=interpretation_session_id,
+ browser_session_id=PBS_ID,
+ organization_id=ORG_ID,
+ workflow_permanent_id=WP_ID,
+ )
+ is None
+ )
diff --git a/tests/unit/test_skycdp_locator_chain.py b/tests/unit/test_skycdp_locator_chain.py
new file mode 100644
index 000000000..84407fcb3
--- /dev/null
+++ b/tests/unit/test_skycdp_locator_chain.py
@@ -0,0 +1,26 @@
+import pytest
+
+from skyvern.webeye.skycdp.errors import CdpError
+from skyvern.webeye.skycdp.facade.locator import _query_operations, _split_selector_chain
+
+
+def test_nth_selector_step_becomes_an_index_operation() -> None:
+ kinds = [(op.kind, op.selector, op.index) for op in _query_operations("ul#list >> nth=-1 >> li")]
+ assert kinds == [("query", "ul#list", None), ("nth", None, -1), ("query", "li", None)]
+
+
+def test_malformed_nth_selector_step_is_a_cdp_error() -> None:
+ with pytest.raises(CdpError, match="nth=abc"):
+ _query_operations("div >> nth=abc")
+
+
+def test_apostrophe_in_an_unquoted_text_body_does_not_swallow_the_next_step() -> None:
+ assert _split_selector_chain("text=Men's >> nth=0") == ["text=Men's", "nth=0"]
+ assert _split_selector_chain("text='Men's shoes' >> nth=0") == ["text='Men's shoes'", "nth=0"]
+ assert _split_selector_chain('li:has-text("a >> b") >> nth=0') == ['li:has-text("a >> b")', "nth=0"]
+
+
+def test_an_escaped_quote_does_not_swallow_the_next_step() -> None:
+ assert _split_selector_chain(r"div[title='John\'s'] >> nth=0") == [r"div[title='John\'s']", "nth=0"]
+ assert _split_selector_chain(r'div[title="a\"b"] >> nth=0') == [r'div[title="a\"b"]', "nth=0"]
+ assert _split_selector_chain(r"li:has-text('a >> b') >> nth=0") == [r"li:has-text('a >> b')", "nth=0"]
diff --git a/tests/unit/test_taskv3_engine.py b/tests/unit/test_taskv3_engine.py
index 33728ee45..3c4bf68d1 100644
--- a/tests/unit/test_taskv3_engine.py
+++ b/tests/unit/test_taskv3_engine.py
@@ -8,6 +8,7 @@ from __future__ import annotations
import asyncio
import json
+from pathlib import Path
from typing import Any
from unittest.mock import AsyncMock
@@ -44,7 +45,12 @@ from skyvern.forge.taskv3.loop import (
from skyvern.forge.taskv3.opaque_refs import mask_opaque_urls
from skyvern.forge.taskv3.tools import PAGE_UNAVAILABLE_ERROR
from tests.unit.test_taskv3_loop import _ScriptedCaller
-from tests.unit.test_taskv3_tools import _SURFACE_OFF_TOOL_NAMES, _FakePage, _fixed_page_provider
+from tests.unit.test_taskv3_tools import (
+ _SURFACE_OFF_TOOL_NAMES,
+ _DownloadFakePage,
+ _FakePage,
+ _fixed_page_provider,
+)
@pytest.mark.asyncio
@@ -1735,3 +1741,64 @@ def test_combobox_bullet_is_pinned_so_every_edit_is_re_derived_against_the_error
bullet = next(line for line in SYSTEM_PROMPT.splitlines() if "select_combobox` tool" in line)
assert bullet == _COMBOBOX_BULLET
+
+
+async def test_engine_restores_a_blank_working_page_before_the_block_hands_it_on(tmp_path: Path) -> None:
+ # `finish` is assembled outside `build_browser_tools` (engine: browser_tools + extras + finish)
+ # and is therefore never wrapped, so a block shaped `click(download) -> finish` reaches the end
+ # of the loop with the tab still on `about:blank`. The next url-less block inherits it and
+ # `resolve_inherited_workflow_task_page` raises InvalidWorkflowTaskURLState (SKY-16322). Only the
+ # post-loop backstop can repair this shape -- the per-call guard has no later call to run on.
+ page = _DownloadFakePage(tmp_path)
+ page._click_blanks = True
+ before = page.url
+ restored: list[tuple[Any, str]] = []
+
+ async def _restore(target: Any, url: str) -> None:
+ restored.append((target, url))
+ target.url = url
+
+ script = [
+ [("click", {"selector": "#dl"})],
+ [("finish", {"status": "completed", "reason": "downloaded the statement"})],
+ ]
+ outcome = await run_task_v3_agent_loop(
+ page_provider=_fixed_page_provider(page),
+ llm_caller=_ScriptedCaller(script),
+ goal="Download the statement.",
+ restore_page_url=_restore,
+ )
+
+ assert outcome.status == "completed"
+ assert restored == [(page, before)]
+ assert page.url == before
+
+
+async def test_engine_repairs_a_blank_page_when_the_loop_ends_without_another_tool_call(
+ tmp_path: Path,
+) -> None:
+ # The per-call guard repairs BEFORE a tool runs, so it cannot help when the blanking click is the
+ # last thing that happens -- the turn budget runs out and the loop returns with the page still
+ # blank. Only the post-loop backstop covers that, and the next url-less block is what pays.
+ page = _DownloadFakePage(tmp_path)
+ page._click_blanks = True
+ before = page.url
+ restored: list[tuple[Any, str]] = []
+
+ async def _restore(target: Any, url: str) -> None:
+ restored.append((target, url))
+ target.url = url
+
+ outcome = await run_task_v3_agent_loop(
+ page_provider=_fixed_page_provider(page),
+ llm_caller=_ScriptedCaller([[("click", {"selector": "#dl"})]]),
+ goal="Download the statement.",
+ restore_page_url=_restore,
+ max_turns=1,
+ )
+
+ assert outcome.status != "completed" # ran out of turns rather than finishing
+ assert restored == [(page, before)]
+ assert page.url == before
diff --git a/tests/unit/test_taskv3_tools.py b/tests/unit/test_taskv3_tools.py
index dc92da0e6..e3ee37176 100644
--- a/tests/unit/test_taskv3_tools.py
+++ b/tests/unit/test_taskv3_tools.py
@@ -45,6 +45,7 @@ from skyvern.forge.taskv3.tools import (
_SEMANTIC_COMMIT_STATE_JS,
NAVIGATION_DEAD_END_STATUSES,
PAGE_UNAVAILABLE_ERROR,
+ BlankWorkingPageGuard,
_annotate_screenshot,
_invalid_selector_result,
_is_host_anchored_selector,
@@ -52,6 +53,7 @@ from skyvern.forge.taskv3.tools import (
)
from skyvern.forge.taskv3.tools import _upload_submit_delay as _REAL_UPLOAD_SUBMIT_DELAY
from skyvern.forge.taskv3.tools import (
+ apply_blank_page_guard,
build_browser_tools,
pending_marker,
)
@@ -297,6 +299,13 @@ class _FakePage:
self.calls: list[tuple[str, dict[str, Any]]] = []
self.element = _FakeElement(self)
self._request_listeners: list[Any] = []
+ self._closed = False
+
+ def is_closed(self) -> bool:
+ return self._closed
+
+ async def close(self) -> None:
+ self._closed = True
def on(self, event: str, callback: Any) -> None:
if event == "request":
@@ -4345,10 +4354,15 @@ class _DownloadFakePage(_FakePage):
self._downloads_dir = downloads_dir
self._click_writes: str | None = None
self._observe_writes: str | None = None
+ # A download-intent click strands the tab on a blank document (SKY-16322): Chrome converts
+ # the navigation into a download and the tab it started from is left on `about:blank`.
+ self._click_blanks = False
async def click(self, selector: str, timeout: int | None = None) -> None:
if self._click_writes:
(self._downloads_dir / self._click_writes).write_bytes(b"x" * 500)
+ if self._click_blanks:
+ self.url = "about:blank"
await super().click(selector, timeout=timeout)
async def evaluate(self, js: str) -> str:
@@ -16044,6 +16058,52 @@ async def test_observe_masks_a_minted_url_longer_than_its_display_caps(carrier:
assert "opaque_url_" in (r.content if carrier.startswith("alert") else line)
+# The same URL with a signature carrying no long run of hex characters. The `field=` qualifier is
+# screened by `_OPAQUE_ID_RUN_RE` before it is ever printed, so a signature like the one above is
+# dropped by that screen and a masking test written with it goes green without reaching the masker
+# at all. This one survives the screen, which is what puts the masker under test.
+_LONG_SIGNED_REF_URL_NO_HEX_RUN = (
+ "https://files.example.test/uploads/report.pdf"
+ "?X-Amz-Algorithm=AWS4-HMAC-SHA256&X-Amz-Credential=AKIAEXAMPLE%2F20260827%2Fus-east-1%2Fs3%2Faws4_request"
+ "&X-Amz-Date=20260827T000000Z&X-Amz-Expires=3600&X-Amz-SignedHeaders=host"
+ "&X-Amz-Signature=QUJDREVGR0hJSktMTU5PUFFSU1RVVldYWVowMTIzNDU2Nzg5QUJDREVGR0hJSkxNTk9Q"
+)
+
+
+@_skip_no_browser
+async def test_observe_masks_a_minted_url_carried_into_a_field_qualifier() -> None:
+ # Masking is by provenance over the WHOLE URL, so any cap applied before the masker runs leaves a
+ # fragment it cannot recognise -- and `X-Amz-Credential` sits in the PREFIX, which is the part a
+ # truncation keeps. The qualifier is a print path like any other, so it has to retain wide and let
+ # Python mask before capping. This is the same defect #16858 fixed one qualifier earlier.
+ assert len(_LONG_SIGNED_REF_URL_NO_HEX_RUN) > 200
+ assert not _OPAQUE_ID_RUN_RE.search(_LONG_SIGNED_REF_URL_NO_HEX_RUN), "screened before the masker"
+ # Two boxes, identical buttons: the qualifier only prints where a reading is actually ambiguous.
+ row = (
+ '<div><label for="f{i}">{url}</label>'
+ '<input id="f{i}" type="text" style="width:140px;height:24px">'
+ '<button style="width:60px;height:28px">Search</button></div>'
+ )
+ html = (
+ "<!doctype html><html><body>"
+ + row.format(i=0, url=html_escape(_LONG_SIGNED_REF_URL_NO_HEX_RUN))
+ + row.format(i=1, url="Plain second field")
+ + "</body></html>"
+ )
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page), opaque_refs=_refs_for(_LONG_SIGNED_REF_URL_NO_HEX_RUN))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [ln for ln in _lines(r.content) if "'Search'" in ln]
+ assert len(icons) == 2, r.content
+ named = [ln for ln in icons if "field=" in ln]
+ assert len(named) == 2, r.content
+ # The leak check is over the whole digest, not just the qualifier: a truncated URL keeps its head.
+ assert _LONG_SIGNED_REF_ARTIFACT not in r.content, r.content
+ assert "opaque_url_" in "".join(named), named
+
+
# A design system's question component: the label lives in the component's root and wraps a <slot>
# for the light-DOM question text, so `.labels` finds it but its innerText is only the required
# marker -- the model pattern-completed values from its neighbours. The two ancestor-root shapes in
@@ -22742,6 +22802,409 @@ async def test_observe_says_nothing_past_the_last_heading_a_capped_scan_saw() ->
assert "section=" not in past, r.content
+def _lookup_rows(names: tuple[str, ...], label_last: bool = False, noise: str = "") -> str:
+ """A form row repeating one identical icon control beside a labelled field control."""
+ control = (
+ '<div>{noise}<input id="f{i}" type="text" disabled="true" style="width:160px;height:24px">'
+ '<button style="width:24px;height:24px"><span>Search</span></button></div>'
+ )
+ label = '<label for="f{i}">{name}</label>'
+ order = (control + label) if label_last else (label + control)
+ return "".join("<div>" + order.format(i=i, name=n, noise=noise.format(i=i)) + "</div>" for i, n in enumerate(names))
+
+
+@_skip_no_browser
[email protected]("label_last", [False, True], ids=["label-first", "label-last"])
+ "noise",
+ [
+ "",
+ '<input type="hidden" name="key{i}">',
+ '<input type="submit" value="Go" style="width:40px;height:20px">',
+ ],
+ ids=["bare", "hidden-key", "submit-input"],
+)
+async def test_observe_names_a_repeated_icon_control_by_the_field_it_belongs_to(label_last: bool, noise: str) -> None:
+ # SKY-16320. A form repeating one icon control across several labelled fields: the buttons are
+ # strongly AND identically named ('Search'), carry no identity of their own, sit in wrappers that
+ # carry none, and share one heading -- so every address tier found nothing and all three lines
+ # rendered the same bytes. The page bound a label to the field control beside each button, and
+ # that binding is what separates them. Run both DOM orders because a reading that depends on the
+ # label PRECEDING the control answers this shape by accident and the mirrored one not at all.
+ # The noise cases are markup a real lookup widget carries: a hidden input holding the resolved
+ # record id, and a submit control in the row. Neither names a field, so neither may make the row
+ # read as a section holding several -- counting them at all put the bug straight back.
+ names = ("Alpha Contact", "Beta Contact", "Gamma Contact")
+ html = f"""<!doctype html><html><body><h2>Contacts</h2>{_lookup_rows(names, label_last, noise)}</body></html>"""
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 3, r.content
+ # Each icon must answer with the field the page bound beside IT -- three distinct strings
+ # alone would let the icons be told apart while still naming the wrong field.
+ for expected in names:
+ assert sum(f"field={expected!r}" in line for line in icons) == 1, r.content
+ assert r.data["summary"]["duplicate_digest_lines"] == 0, r.data
+
+
+@_skip_no_browser
+async def test_observe_ranks_the_field_over_a_wrapper_id_and_a_heading_but_under_its_own() -> None:
+ # Every set here can be split by BOTH tiers that bracket it unaided, which is what makes rank the
+ # only thing the assertions can be reading: a tier that could not split would fall through
+ # whatever the order and prove nothing. Mutation-check by reordering tiers, not by reverting.
+ rows = "".join(
+ f'<div id="lookupRow{i}"><h3>Section {i}</h3><label for="g{i}">{n}</label><div>'
+ f'<input id="g{i}" type="text" style="width:160px;height:24px">'
+ f'<button style="width:24px;height:24px"><span>Search</span></button></div></div>'
+ for i, n in enumerate(("Alpha Contact", "Beta Contact"))
+ )
+ owned = "".join(
+ f'<h3>{h}</h3><div><label for="d{i}">{h} note</label>'
+ f'<input id="d{i}" type="text" style="width:160px;height:24px">'
+ f'<button id="del-{i}" style="width:70px;height:30px">Delete</button></div>'
+ for i, h in enumerate(("Work", "Study"))
+ )
+ html = f"""<!doctype html><html><body>{rows}{owned}</body></html>"""
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 2, r.content
+ # A per-row wrapper id and a per-row heading would each split these on their own. Both are
+ # framework bookkeeping; only the field names what the task names, so the field must win.
+ for expected in ("Alpha Contact", "Beta Contact"):
+ assert sum(f"field={expected!r}" in line for line in icons) == 1, r.content
+ assert all("within." not in line and "section=" not in line for line in icons), r.content
+ dels = [line for line in _lines(r.content) if "'Delete'" in line]
+ assert len(dels) == 2, r.content
+ # These carry their own ids AND a field AND a heading -- the control's own address outranks
+ # both, so neither of the other two may be reached.
+ assert "id='del-0'" in dels[0] and "id='del-1'" in dels[1], r.content
+ assert all("field=" not in line and "section=" not in line for line in dels), r.content
+
+
+@_skip_no_browser
[email protected]("filter_nested", [False, True], ids=["rows-nested", "filter-nested"])
+async def test_observe_does_not_name_a_control_by_a_field_it_shares_no_box_with(filter_nested: bool) -> None:
+ # A panel holding one filter box and a list of row buttons. The page bound a name to the box; it
+ # bound nothing between that box and these buttons. Reaching for it anyway hands every row the
+ # filter's name -- confidently, and wrongly -- which is worse than leaving the lines identical,
+ # because the model acts on a qualifier it reads as a semantic claim.
+ #
+ # The filter is a DIRECT CHILD of the panel on purpose. Nested one level deeper, widening the box
+ # by a single level still passes, so the fixture would not discriminate the rule from a looser
+ # one -- and no counter can catch that class, because the wrong name is SHARED and the lines go
+ # on colliding. Only this shape can.
+ # Two arrangements, because the two halves of the rule fail to different shapes. rows-nested puts
+ # the filter beside the row blocks and needs the box NOT to widen outward; filter-nested puts the
+ # buttons beside the filter's wrapper and needs the field to be a direct child of the box. A
+ # fixture with only one of them leaves the other half of the rule free to be deleted.
+ if filter_nested:
+ panel = (
+ '<div><div><label for="q{i}">{name}</label>'
+ '<input id="q{i}" type="text" style="width:160px;height:24px"></div>'
+ '<button style="width:70px;height:30px">Select</button>'
+ '<button style="width:70px;height:30px">Select</button></div>'
+ )
+ else:
+ panel = (
+ '<div><label for="q{i}">{name}</label>'
+ '<input id="q{i}" type="text" style="width:160px;height:24px">'
+ '<div><span>Row one</span><button style="width:70px;height:30px">Select</button></div>'
+ '<div><span>Row two</span><button style="width:70px;height:30px">Select</button></div></div>'
+ )
+ html = (
+ "<!doctype html><html><body>"
+ + panel.format(i=0, name="Search company")
+ + panel.format(i=1, name="Search wholesaler")
+ + "</body></html>"
+ )
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ picks = [line for line in _lines(r.content) if "'Select'" in line]
+ assert len(picks) == 4, r.content
+ assert all("field=" not in line for line in picks), r.content
+ # And the count must still say the reading is unresolved rather than hide it.
+ assert r.data["summary"]["duplicate_digest_lines"] == 4, r.data
+
+
+@_skip_no_browser
+async def test_observe_will_not_name_a_box_that_holds_a_second_field_after_the_unnamed_one() -> None:
+ # Whether a box holds one field is a fact about the BOX. A second field that happens to carry no
+ # declared name is still a second field, so the button could belong to either and the page has
+ # not said which -- counting only the named ones would let the survivor name the whole row.
+ # The second field is wrapped in the first row and bare in the second: a validation-state wrapper
+ # appears around exactly the errored field, so a rule that only counted direct children would let
+ # the qualifier blink in and out as the user fixes the form.
+ row = (
+ '<div><label for="a{i}">{name}</label><input id="a{i}" style="width:140px;height:24px">'
+ '{open}<input id="b{i}" placeholder="Account number" style="width:140px;height:24px">{close}'
+ '<button style="width:60px;height:28px">Search</button></div>'
+ )
+ html = (
+ "<!doctype html><html><body>"
+ + row.format(i=0, name="Contact name", open='<div class="has-error">', close="</div>")
+ + row.format(i=1, name="Owner name", open="", close="")
+ + "</body></html>"
+ )
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 2, r.content
+ assert all("field=" not in line for line in icons), r.content
+ assert r.data["summary"]["duplicate_digest_lines"] == 2, r.data
+
+
+@_skip_no_browser
+ "source",
+ [
+ "contenteditable-sibling",
+ "moved-in-by-observer",
+ "reclassified-by-observer",
+ "dropped-but-still-live",
+ "inserted-during-drain",
+ ],
+)
+async def test_observe_withholds_the_field_when_the_box_holds_a_second_one_it_almost_missed(
+ source: str,
+) -> None:
+ # The census decides whether the box holds exactly ONE field, so every way it can undercount is a
+ # way to emit a confident wrong name. Three sources that each defeated an earlier version: a field
+ # that is not an <input>; a field the page MOVES into the box while observe is running; a widget
+ # the page RECLASSIFIES from a command to a field mid-run, which the fingerprint does not track,
+ # so it survives re-resolution while its stored classification is stale; and a field whose record
+ # the sweep DROPS while the control stays on the page. Ancestry and classification are read from
+ # the live element after the sweep, the census includes dropped-but-live candidates, a box whose
+ # children changed underneath us is withheld outright, and anything not definitively a command
+ # control counts -- so all five fail closed.
+ if source == "contenteditable-sibling":
+ rows = "".join(
+ f'<div><label for="i{i}">{n}</label><input id="i{i}" style="width:140px;height:24px">'
+ f'<div contenteditable="true" aria-label="Notes {i}" style="width:120px;height:24px">y</div>'
+ f'<button style="width:60px;height:28px">Search</button></div>'
+ for i, n in enumerate(("Alpha Contact", "Beta Contact"))
+ )
+ html = f"<!doctype html><html><body>{rows}</body></html>"
+ elif source == "inserted-during-drain":
+ # The page APPENDS a field to the box on the marker write. It is in no collection the census
+ # can read -- not a survivor, not a dropped original -- so no amount of re-reading what was
+ # enumerated will see it. A box whose children changed underneath us is one whose field count
+ # cannot be established, so the qualifier is withheld there rather than guessed at.
+ rows = "".join(
+ f'<div class="row" id="r{i}"><label for="a{i}">{n}</label>'
+ f'<input id="a{i}" style="width:120px;height:26px">'
+ f'<button style="width:50px;height:26px">Search</button></div>'
+ for i, n in enumerate(("Account", "Policy"))
+ )
+ html = f"""<!doctype html><html><body>{rows}
+ <script>
+ new MutationObserver(function() {{
+ ['r0', 'r1'].forEach(function(id) {{
+ var row = document.getElementById(id);
+ if (row && !row.querySelector('.late')) {{
+ var i = document.createElement('input');
+ i.className = 'late';
+ i.setAttribute('aria-label', 'Late field ' + id);
+ i.style.width = '120px';
+ i.style.height = '26px';
+ row.appendChild(i);
+ }}
+ }});
+ }}).observe(document.documentElement, {{attributes: true, subtree: true,
+ attributeFilter: ['data-tv3']}});
+ </script></body></html>"""
+ elif source == "dropped-but-still-live":
+ # The second field's label changes on the marker write, so it fails the fingerprint and its
+ # record is dropped -- but the control is still in the box. It must not name anything, having
+ # no line for the model to read, and it must still count.
+ rows = "".join(
+ f'<div><label for="a{i}">{n}</label><input id="a{i}" style="width:120px;height:26px">'
+ f'<input id="b{i}" aria-label="Second {i}" style="width:120px;height:26px">'
+ f'<button style="width:50px;height:26px">Search</button></div>'
+ for i, n in enumerate(("Account", "Policy"))
+ )
+ html = f"""<!doctype html><html><body>{rows}
+ <script>
+ new MutationObserver(function() {{
+ ['b0', 'b1'].forEach(function(id) {{
+ var n = document.getElementById(id);
+ if (n && n.getAttribute('aria-label').indexOf('changed') === -1) {{
+ n.setAttribute('aria-label', 'changed ' + id);
+ }}
+ }});
+ }}).observe(document.documentElement, {{attributes: true, subtree: true,
+ attributeFilter: ['data-tv3']}});
+ </script></body></html>"""
+ elif source == "reclassified-by-observer":
+ # role=button at record time, role=combobox by the time the census runs.
+ rows = "".join(
+ f'<div><label for="a{i}">{n}</label><input id="a{i}" style="width:120px;height:26px">'
+ f'<div id="w{i}" role="button" tabindex="0" aria-label="{n} picker"'
+ f' style="width:100px;height:26px">pick</div>'
+ f'<button style="width:50px;height:26px">Search</button></div>'
+ for i, n in enumerate(("Account", "Policy"))
+ )
+ html = f"""<!doctype html><html><body>{rows}
+ <script>
+ new MutationObserver(function() {{
+ ['w0', 'w1'].forEach(function(id) {{
+ var n = document.getElementById(id);
+ if (n && n.getAttribute('role') === 'button') n.setAttribute('role', 'combobox');
+ }});
+ }}).observe(document.documentElement, {{attributes: true, subtree: true,
+ attributeFilter: ['data-tv3']}});
+ </script></body></html>"""
+ else:
+ # The observer fires on observe's own marker write, after the records were made.
+ html = """<!doctype html><html><body>
+ <div class="row" id="r0"><label for="a0">Account</label>
+ <input id="a0" style="width:120px;height:26px">
+ <button style="width:50px;height:26px">Search</button></div>
+ <div class="row" id="r1"><label for="a1">Policy</label>
+ <input id="a1" style="width:120px;height:26px">
+ <button style="width:50px;height:26px">Search</button></div>
+ <div id="park"><label for="x0">Extra one</label>
+ <input id="x0" style="width:120px;height:26px">
+ <label for="x1">Extra two</label>
+ <input id="x1" style="width:120px;height:26px"></div>
+ <script>
+ new MutationObserver(function() {
+ var a = document.getElementById('x0'), b = document.getElementById('x1');
+ if (a && a.parentElement.id === 'park') document.getElementById('r0').appendChild(a);
+ if (b && b.parentElement.id === 'park') document.getElementById('r1').appendChild(b);
+ }).observe(document.documentElement, {attributes: true, subtree: true,
+ attributeFilter: ['data-tv3']});
+ </script></body></html>"""
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 2, r.content
+ assert all("field=" not in line for line in icons), r.content
+
+
+@_skip_no_browser
+async def test_observe_will_not_name_an_icon_after_something_that_is_not_a_field() -> None:
+ # The census is deliberately WIDE -- everything not definitively a command control counts -- and
+ # that is fail-closed for COUNTING, because a wider census suppresses more. It is NOT fail-closed
+ # for NAMING: a wider source makes more non-fields eligible to lend their text to a neighbour.
+ # A labelled [role=option] is listed, is not a command control, and is not a field, so it must
+ # count against the box and must never name the icon beside it.
+ rows = "".join(
+ f'<div><div role="option" aria-label="{n}" tabindex="0" style="width:140px;height:26px">x</div>'
+ f'<button style="width:60px;height:26px">Search</button></div>'
+ for n in ("Alpha plan", "Beta plan")
+ )
+ html = f"<!doctype html><html><body>{rows}</body></html>"
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 2, r.content
+ assert all("field=" not in line for line in icons), r.content
+ # And the reading stays honestly unresolved rather than being told apart by a false name.
+ assert r.data["summary"]["duplicate_digest_lines"] == 2, r.data
+
+
+@_skip_no_browser
+async def test_observe_names_an_icon_by_a_contenteditable_field_beside_it() -> None:
+ # A field is not only an <input>. The same module lists contenteditable elements and types into
+ # them, so a box whose one field is a labelled contenteditable names its icon like any other --
+ # the complement of the test above, and what stops "fails closed" from meaning "never fires".
+ rows = "".join(
+ f'<div><div contenteditable="true" aria-label="{n}" style="width:160px;height:24px">x</div>'
+ f'<button style="width:60px;height:28px">Search</button></div>'
+ for n in ("Alpha Contact", "Beta Contact")
+ )
+ html = f"<!doctype html><html><body>{rows}</body></html>"
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 2, r.content
+ for expected in ("Alpha Contact", "Beta Contact"):
+ assert sum(f"field={expected!r}" in line for line in icons) == 1, r.content
+ assert r.data["summary"]["duplicate_digest_lines"] == 0, r.data
+
+
+@_skip_no_browser
+async def test_observe_sees_a_second_field_in_the_box_however_deeply_it_is_nested() -> None:
+ # The box is read by "holds no second field at any depth", so the registration that answers that
+ # question must not be depth-bounded. A bound does not bound the work here -- it makes the count
+ # WRONG, and silently: a box whose second field sits one level past it reports as holding one and
+ # names a control after it. That turns a refusal to guess back into a confident wrong answer,
+ # which is the failure this tier is built to avoid. Nested well past any plausible bound.
+ deep_open, deep_close = "<div>" * 8, "</div>" * 8
+ row = (
+ '<div><label for="a{i}">{near}</label><input id="a{i}" style="width:140px;height:24px">'
+ '<button style="width:60px;height:28px">Search</button>'
+ + deep_open
+ + '<label for="b{i}">{deep}</label><input id="b{i}" style="width:140px;height:24px">'
+ + deep_close
+ + "</div>"
+ )
+ html = (
+ "<!doctype html><html><body>"
+ + row.format(i=0, near="Contact name", deep="Deep account")
+ + row.format(i=1, near="Owner name", deep="Deep owner")
+ + "</body></html>"
+ )
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 2, r.content
+ assert all("field=" not in line for line in icons), r.content
+ assert r.data["summary"]["duplicate_digest_lines"] == 2, r.data
+
+
+@_skip_no_browser
+ "attrs",
+ ['placeholder="Search {name}..."', 'value="{name}"'],
+ ids=["placeholder", "typed-value"],
+)
+async def test_observe_names_a_field_only_from_a_name_the_page_bound(attrs: str) -> None:
+ # The qualifier must come from a DECLARED name. A placeholder is a hint every field of a template
+ # shares, and a typed value changes as the agent fills the row -- either one would make an
+ # unchanged page render a different digest each turn, on exactly these repeated-row forms.
+ rows = "".join(
+ f'<div><div><input id="p{i}" type="text" {attrs.format(name=n)} style="width:160px;height:24px">'
+ f'<button style="width:24px;height:24px"><span>Search</span></button></div></div>'
+ for i, n in enumerate(("Alpha Contact", "Beta Contact"))
+ )
+ html = f"""<!doctype html><html><body><h2>Contacts</h2>{rows}</body></html>"""
+ async with _content_page(html) as page:
+ tools = build_browser_tools(_fixed_page_provider(page))
+ r = await _tool(tools, "observe").handler({})
+ assert r.status == "ok", r.content
+ icons = [line for line in _lines(r.content) if "'Search'" in line]
+ assert len(icons) == 2, r.content
+ assert all("field=" not in line for line in icons), r.content
+ assert r.data["summary"]["duplicate_digest_lines"] == 2, r.data
+
+
def _surface_oracle(
pre: str | None, post: str, chosen: str, rows: list[str] | None
) -> taskv3_tools.CommitStatus | None:
@@ -23529,3 +23992,528 @@ async def test_the_isolated_leg_is_recorded_as_isolated(monkeypatch: pytest.Monk
await _tool(build_browser_tools(_fixed_page_provider(_FakePage())), "click").handler({"selector": "#submit"})
assert (taskv3_loop._HIT_CLASS.get() or {}).get("isolated") is True
taskv3_loop._HIT_CLASS.set(None)
+
+
+# --- Blank working-page guard (SKY-16322). A download leaves the working page on a blank document
+# and the next url-less block raises InvalidWorkflowTaskURLState. Production shape, measured: on the
+# affected workflow 1545 of 1546 downloads went 1 page -> 2 (the download opens a tab), and v1
+# repaired every one by CLOSING it. Fleet-wide the reverse dominates -- 91.3% of downloads open no
+# second page at all -- so the no-interloper case is the one the close path must stay off. ---
+
+
+def _guarded(page_provider: Any, tmp_path: Path) -> tuple[list[Any], Any, list[tuple[Any, str]]]:
+ calls: list[tuple[Any, str]] = []
+
+ async def _restore(page: Any, url: str) -> None:
+ calls.append((page, url))
+ page.url = url
+
+ guard = BlankWorkingPageGuard(page_provider, _restore, downloads_dir=str(tmp_path))
+ tools = build_browser_tools(page_provider, downloads_dir=str(tmp_path))
+ apply_blank_page_guard(tools, guard)
+ return tools, guard, calls
+
+
+def _newest_open(pages: list[Any]) -> Callable[[], Awaitable[Any]]:
+ """Mirrors get_working_page: newest page that is still open (a closed page leaves context.pages)."""
+
+ async def _provider() -> Any:
+ live = [p for p in pages if not p.is_closed()]
+ return live[-1] if live else None
+
+ return _provider
+
+
+class _PopupDownloadPage(_DownloadFakePage):
+ """A download click that opens a second tab and leaves it blank -- the measured production shape."""
+
+ def __init__(self, downloads_dir: Path, pages: list[Any], popup: Any) -> None:
+ super().__init__(downloads_dir)
+ self._pages = pages
+ self._popup = popup
+
+ async def click(self, selector: str, timeout: int | None = None) -> None:
+ await super().click(selector, timeout=timeout)
+ self._pages.append(self._popup)
+
+
+async def test_the_blank_tab_a_download_opens_is_closed_so_the_working_page_falls_back(
+ tmp_path: Path,
+) -> None:
+ # THE production case. Without this the popup is newest, takes the working-page slot, and the
+ # next url-less block inherits about:blank.
+ pages: list[Any] = []
+ popup = _DownloadFakePage(tmp_path)
+ popup.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, popup)
+ opener._click_writes = "statement.pdf"
+ pages.append(opener)
+ tools, _guard, calls = _guarded(_newest_open(pages), tmp_path)
+
+ r = await _tool(tools, "click").handler({"selector": "#dl"})
+ assert "Downloaded: statement.pdf" in r.content
+ await _tool(tools, "get_html").handler({})
+
+ assert popup.is_closed()
+ assert not opener.is_closed()
+ assert await _newest_open(pages)() is opener
+ assert calls == [] # closing is the repair here; nothing is navigated
+
+
+async def test_a_download_that_opens_no_second_page_closes_nothing(tmp_path: Path) -> None:
+ # Fleet-wide this is 91.3% of downloads. The close path must stay completely off.
+ page = _DownloadFakePage(tmp_path)
+ page._click_writes = "statement.pdf"
+ before = page.url
+ tools, _guard, calls = _guarded(_fixed_page_provider(page), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ await _tool(tools, "get_html").handler({})
+
+ assert not page.is_closed()
+ assert page.url == before
+ assert calls == []
+
+
+async def test_a_second_tab_the_model_opened_is_never_closed(tmp_path: Path) -> None:
+ # The regression this fix could cause: closing a tab the model deliberately opened. A real tab
+ # navigates, so it is not blank when the guard next looks.
+ pages: list[Any] = []
+ real_tab = _DownloadFakePage(tmp_path)
+ real_tab.url = "https://example.test/report"
+ opener = _PopupDownloadPage(tmp_path, pages, real_tab)
+ opener._click_writes = "statement.pdf"
+ pages.append(opener)
+ tools, _guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ await _tool(tools, "get_html").handler({})
+
+ assert not real_tab.is_closed()
+ assert not opener.is_closed()
+
+
+async def test_a_blank_tab_with_no_download_to_account_for_it_is_never_closed(tmp_path: Path) -> None:
+ # No download observed => the blank tab is one the model opened and has not navigated yet.
+ # v1 gates its close on a download for exactly this reason; the gate is the whole safety margin.
+ pages: list[Any] = []
+ fresh_tab = _DownloadFakePage(tmp_path)
+ fresh_tab.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, fresh_tab)
+ pages.append(opener)
+ tools, _guard, calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#link"}) # no file lands
+ await _tool(tools, "get_html").handler({})
+
+ assert not fresh_tab.is_closed()
+ assert calls == []
+
+
+async def test_one_download_arms_exactly_one_close(tmp_path: Path) -> None:
+ # v1 scopes its close to a single download action. A second blank tab appearing later, with no
+ # new download, must survive.
+ pages: list[Any] = []
+ popup = _DownloadFakePage(tmp_path)
+ popup.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, popup)
+ opener._click_writes = "statement.pdf"
+ pages.append(opener)
+ tools, _guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ await _tool(tools, "get_html").handler({})
+ assert popup.is_closed()
+
+ later_tab = _DownloadFakePage(tmp_path)
+ later_tab.url = "about:blank"
+ pages.append(later_tab)
+ await _tool(tools, "get_html").handler({})
+
+ assert not later_tab.is_closed()
+
+
+async def test_a_tab_blanked_in_place_is_navigated_back(tmp_path: Path) -> None:
+ # The other shape. v1 still carries code for it, though its log fired once in 30 days fleet-wide.
+ page = _DownloadFakePage(tmp_path)
+ page._click_writes = "statement.pdf"
+ page._click_blanks = True
+ before = page.url
+ tools, _guard, calls = _guarded(_fixed_page_provider(page), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ r2 = await _tool(tools, "get_html").handler({})
+
+ assert calls == [(page, before)]
+ assert page.url == before
+ assert "about:blank" not in r2.content
+
+
+async def test_repair_targets_where_the_download_fired_not_where_the_block_started(tmp_path: Path) -> None:
+ # A guard that remembered only the run's FIRST url would pass the other tests and drop the run
+ # back onto the wrong page in production.
+ page = _DownloadFakePage(tmp_path)
+ started_on = page.url
+ tools, _guard, calls = _guarded(_fixed_page_provider(page), tmp_path)
+
+ await _tool(tools, "get_html").handler({})
+ page.url = "https://example.test/statements"
+ page._click_writes = "statement.pdf"
+ page._click_blanks = True
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ await _tool(tools, "get_html").handler({})
+
+ assert calls == [(page, "https://example.test/statements")]
+ assert started_on not in [url for _, url in calls]
+
+
+async def test_a_live_working_page_is_never_re_navigated(tmp_path: Path) -> None:
+ # Kills the mutant that drops the blank-URL check: it would re-navigate the live page after every
+ # download, wiping form state mid-run.
+ page = _DownloadFakePage(tmp_path)
+ page._click_writes = "statement.pdf"
+ tools, _guard, calls = _guarded(_fixed_page_provider(page), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ page.url = "https://example.test/apply?step=2"
+ await _tool(tools, "get_html").handler({})
+ await _tool(tools, "click").handler({"selector": "#next"})
+
+ assert calls == []
+ assert page.url == "https://example.test/apply?step=2"
+
+
+async def test_a_restore_that_hangs_is_bounded_and_gives_up_rather_than_retrying_forever(
+ tmp_path: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+ import skyvern.forge.taskv3.tools as tools_mod
+
+ monkeypatch.setattr(tools_mod, "_BLANK_PAGE_RESTORE_TIMEOUT_SECONDS", 0.01)
+ page = _DownloadFakePage(tmp_path)
+ page._click_blanks = True
+ attempts: list[int] = []
+
+ async def _hang(_page: Any, _url: str) -> None:
+ attempts.append(1)
+ await asyncio.sleep(10)
+
+ guard = BlankWorkingPageGuard(_fixed_page_provider(page), _hang, downloads_dir=str(tmp_path))
+ tools = build_browser_tools(_fixed_page_provider(page), downloads_dir=str(tmp_path))
+ apply_blank_page_guard(tools, guard)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ for _ in range(6):
+ assert (await _tool(tools, "get_html").handler({})).status == "ok"
+
+ assert len(attempts) == tools_mod._BLANK_PAGE_RESTORE_MAX_FAILURES
+ assert page.url == "about:blank"
+
+
+async def test_a_closed_page_is_not_navigated(tmp_path: Path) -> None:
+ page = _DownloadFakePage(tmp_path)
+ page._click_blanks = True
+ tools, _guard, calls = _guarded(_fixed_page_provider(page), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ page._closed = True
+ await _tool(tools, "get_html").handler({})
+
+ assert calls == []
+
+
+def test_only_the_origin_of_a_page_url_can_reach_a_log() -> None:
+ # P1: a page URL reaching structlog can carry an OAuth code, a signed token, or customer data in
+ # its PATH or QUERY, and `user:pass@` credentials in its netloc.
+ from skyvern.forge.taskv3.tools import _url_origin
+
+ assert _url_origin("https://site.test/verify/SECRET-TOKEN?code=abc123") == "https://site.test"
+ assert _url_origin("https://user:[email protected]:8443/a") == "https://site.test:8443"
+ for blank in ("about:blank", ":", ""):
+ assert _url_origin(blank) == "unparseable"
+
+
+async def test_the_interloper_is_kept_when_there_is_no_live_page_to_fall_back_to(tmp_path: Path) -> None:
+ # Closing is only a repair because the working page falls back to a live one. If the remembered
+ # page is itself gone or blank there is nothing to fall back TO, and closing would leave the run
+ # with no usable page at all -- strictly worse than the blank page it was fixing.
+ pages: list[Any] = []
+ popup = _DownloadFakePage(tmp_path)
+ popup.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, popup)
+ opener._click_writes = "statement.pdf"
+ pages.append(opener)
+ tools, _guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ opener._closed = True # the opener dies before the guard next looks
+ await _tool(tools, "get_html").handler({})
+
+ assert not popup.is_closed()
+
+
+async def test_a_download_landing_after_the_click_still_arms_the_close(tmp_path: Path) -> None:
+ # The arming must not depend on the download-signal wrapper, which covers only the tools
+ # build_browser_tools builds. The file routinely lands after the click handler returns, and if the
+ # next call is `finish` -- which that wrapper never sees -- nothing would arm and the tab would
+ # survive: the original bug, reintroduced through the arming path.
+ pages: list[Any] = []
+ popup = _DownloadFakePage(tmp_path)
+ popup.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, popup)
+ pages.append(opener)
+ tools, guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"}) # opens the tab, writes nothing yet
+ (tmp_path / "statement.pdf").write_bytes(b"x" * 500) # lands after the handler returned
+
+ # A tool the download-signal wrapper does not wrap at all.
+ await guard.ensure_live()
+
+ assert popup.is_closed()
+ assert not opener.is_closed()
+
+
+async def test_download_arming_expires_so_a_later_blank_tab_survives(tmp_path: Path) -> None:
+ # Fleet-wide 91.3% of downloads open no second page, so arming that never expires would be live
+ # for the rest of the block and would close the next tab the model opened before it navigated.
+ page = _DownloadFakePage(tmp_path)
+ page._click_writes = "statement.pdf"
+ pages: list[Any] = [page]
+ tools, guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#dl"}) # download, no popup
+ for _ in range(4):
+ await _tool(tools, "get_html").handler({}) # live page each time: arming decays
+
+ later_tab = _DownloadFakePage(tmp_path)
+ later_tab.url = "about:blank"
+ pages.append(later_tab)
+ await guard.ensure_live()
+
+ assert not later_tab.is_closed()
+
+
+async def test_a_restore_is_bounded_by_what_is_left_of_the_run(tmp_path: Path) -> None:
+ # The per-call wrapper gets no explicit timeout, so without the run-level bound a restore near the
+ # deadline could spend the full allowance before the handler it precedes even starts.
+ page = _DownloadFakePage(tmp_path)
+ page._click_blanks = True
+ attempts: list[float] = []
+
+ async def _record(_page: Any, _url: str) -> None:
+ attempts.append(1.0)
+ await asyncio.sleep(10)
+
+ guard = BlankWorkingPageGuard(
+ _fixed_page_provider(page), _record, downloads_dir=str(tmp_path), remaining_seconds=lambda: 0.01
+ )
+ tools = build_browser_tools(_fixed_page_provider(page), downloads_dir=str(tmp_path))
+ apply_blank_page_guard(tools, guard)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ assert (await _tool(tools, "get_html").handler({})).status == "ok"
+
+ assert attempts # it tried
+ # and gave up on the run's remaining time, not on the 90s allowance
+ guard._remaining_seconds = lambda: 0.0
+ await guard.ensure_live()
+ assert len(attempts) == 1
+
+
+async def test_a_file_a_previous_block_downloaded_does_not_arm_this_one(tmp_path: Path) -> None:
+ # downloads_dir persists across the blocks of a workflow run. Without a baselining first scan,
+ # every block would start already armed by its predecessor's file and would close the first tab
+ # the model opened.
+ (tmp_path / "statement-from-the-previous-block.pdf").write_bytes(b"x" * 500)
+
+ pages: list[Any] = []
+ fresh_tab = _DownloadFakePage(tmp_path)
+ fresh_tab.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, fresh_tab)
+ pages.append(opener)
+ tools, guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#link"}) # opens a tab, downloads nothing
+ await guard.ensure_live()
+
+ assert not fresh_tab.is_closed()
+
+
+async def test_a_deduplicated_download_still_arms_the_close(tmp_path: Path) -> None:
+ # The CDP interceptor discards a download whose bytes already match a file on disk
+ # (`_existing_file_is_identical` -> `_discard_duplicate_download` -> return), so the directory is
+ # COMPLETELY unchanged. No filesystem check can see it. Its attempt counter still moves, because
+ # it increments in `_resolve_save_path` before the duplicate check.
+ (tmp_path / "statement.pdf").write_bytes(b"x" * 500) # the earlier, identical copy
+
+ pages: list[Any] = []
+ popup = _DownloadFakePage(tmp_path)
+ popup.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, popup)
+ pages.append(opener)
+
+ attempts = [7]
+ calls: list[tuple[Any, str]] = []
+
+ async def _restore(page: Any, url: str) -> None:
+ calls.append((page, url))
+
+ provider = _newest_open(pages)
+ guard = BlankWorkingPageGuard(
+ provider, _restore, downloads_dir=str(tmp_path), download_attempts=lambda: attempts[0]
+ )
+ tools = build_browser_tools(provider, downloads_dir=str(tmp_path))
+ apply_blank_page_guard(tools, guard)
+
+ await _tool(tools, "click").handler({"selector": "#dl"}) # opens the tab; writes no new file
+ attempts[0] += 1 # the interceptor counted the attempt, then deduplicated it
+ await guard.ensure_live()
+
+ assert popup.is_closed()
+
+
+async def test_a_file_staged_by_file_upload_does_not_arm_the_close(tmp_path: Path) -> None:
+ # file_upload writes its own source into the SAME directory. That is this run staging an upload,
+ # not the browser downloading anything, and arming on it would hand a tab the model opened to the
+ # close path.
+ staged: set[str] = set()
+ pages: list[Any] = []
+ fresh_tab = _DownloadFakePage(tmp_path)
+ fresh_tab.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, fresh_tab)
+ pages.append(opener)
+
+ async def _restore(page: Any, url: str) -> None:
+ raise AssertionError("must not restore")
+
+ provider = _newest_open(pages)
+ guard = BlankWorkingPageGuard(provider, _restore, downloads_dir=str(tmp_path), staged_downloads=staged)
+ tools = build_browser_tools(provider, downloads_dir=str(tmp_path))
+ apply_blank_page_guard(tools, guard)
+
+ await _tool(tools, "click").handler({"selector": "#link"}) # opens a blank tab
+ (tmp_path / "resume.pdf").write_bytes(b"x" * 500)
+ staged.add("resume.pdf") # the loop records it as this run's own staged upload
+ await guard.ensure_live()
+
+ assert not fresh_tab.is_closed()
+
+
+async def test_a_popup_close_that_hangs_does_not_block_the_run(tmp_path: Path) -> None:
+ # A popup on a stalled CDP connection can leave close() pending forever, and this path runs both
+ # before dispatch and during final cleanup.
+ pages: list[Any] = []
+
+ class _HangingClosePage(_DownloadFakePage):
+ async def close(self) -> None:
+ await asyncio.sleep(10)
+
+ popup = _HangingClosePage(tmp_path)
+ popup.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, popup)
+ opener._click_writes = "statement.pdf"
+ pages.append(opener)
+
+ async def _restore(page: Any, url: str) -> None:
+ return None
+
+ provider = _newest_open(pages)
+ guard = BlankWorkingPageGuard(provider, _restore, downloads_dir=str(tmp_path), remaining_seconds=lambda: 0.01)
+ tools = build_browser_tools(provider, downloads_dir=str(tmp_path))
+ apply_blank_page_guard(tools, guard)
+
+ await _tool(tools, "click").handler({"selector": "#dl"})
+ started = time.monotonic()
+ assert (await _tool(tools, "get_html").handler({})).status == "ok"
+ elapsed = time.monotonic() - started
+
+ # The discriminator is the BOUND, not the return: an unbounded close returns too, ten seconds
+ # later, having held up dispatch (and, at the post-loop site, cancellation) the whole time. The
+ # threshold sits between the two outcomes with margin either side, not at the bound itself.
+ assert elapsed < 4.0, f"the hung close was not bounded: {elapsed:.2f}s"
+ assert not popup.is_closed()
+
+
+async def test_one_download_arms_once_across_its_temp_name_and_its_rename(tmp_path: Path) -> None:
+ # A download appears twice in the directory: `report.pdf.<uuid>.crdownload`, then `report.pdf`.
+ # On raw names that is two arming events, and the second has no tab-opening behind it -- it would
+ # grant a fresh close window at an arbitrary later point and close a tab the model opened. The
+ # invariant is one close per DOWNLOAD, so the scan normalises through _download_signal_identity.
+ pages: list[Any] = []
+ popup = _DownloadFakePage(tmp_path)
+ popup.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, popup)
+ pages.append(opener)
+ tools, guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "get_html").handler({}) # baseline on an empty directory
+
+ temp = "report.pdf." + "a1b2c3d4e5f6a7b8c9d0e1f2a3b4c5d6" + ".crdownload"
+ (tmp_path / temp).write_bytes(b"partial")
+ await _tool(tools, "click").handler({"selector": "#dl"}) # sees the temp file, arms once
+ await _tool(tools, "get_html").handler({})
+ assert popup.is_closed()
+
+ # The rename lands. It must NOT arm again.
+ (tmp_path / temp).unlink()
+ (tmp_path / "report.pdf").write_bytes(b"x" * 500)
+
+ later_tab = _DownloadFakePage(tmp_path)
+ later_tab.url = "about:blank"
+ pages.append(later_tab)
+ await guard.ensure_live()
+
+ assert not later_tab.is_closed()
+
+
+async def test_the_baseline_stores_identities_so_a_previous_blocks_temp_name_does_not_arm(
+ tmp_path: Path,
+) -> None:
+ # The one place the identity substitution is not mechanical: the first scan baselines IDENTITIES,
+ # not raw names. A predecessor block that left `report.pdf.<uuid>.crdownload` behind must not arm
+ # this block when that file is renamed to `report.pdf` under it.
+ temp = "report.pdf." + "f0e1d2c3b4a5968778695a4b3c2d1e0f" + ".crdownload"
+ (tmp_path / temp).write_bytes(b"partial")
+
+ pages: list[Any] = []
+ fresh_tab = _DownloadFakePage(tmp_path)
+ fresh_tab.url = "about:blank"
+ opener = _PopupDownloadPage(tmp_path, pages, fresh_tab)
+ pages.append(opener)
+ tools, guard, _calls = _guarded(_newest_open(pages), tmp_path)
+
+ await _tool(tools, "click").handler({"selector": "#link"}) # baselines on the identity, opens a tab
+ (tmp_path / temp).unlink()
+ (tmp_path / "report.pdf").write_bytes(b"x" * 500) # the predecessor's file, renamed
+ await guard.ensure_live()
+
+ assert not fresh_tab.is_closed()
+
+
+def test_an_out_of_range_port_does_not_escape_the_origin_helper() -> None:
+ # `parsed.port` is lazy and raises ValueError, and both call sites sit in paths documented as
+ # unable to raise -- including the engine `finally`.
+ from skyvern.forge.taskv3.tools import _url_origin
+
+ assert _url_origin("https://site.test:99999/a") == "unparseable"
+ assert _url_origin("https://site.test:8443/a") == "https://site.test:8443"
diff --git a/tests/unit/test_workflow_copilot_prompt_rewrite.py b/tests/unit/test_workflow_copilot_prompt_rewrite.py
index 640ace1c3..0c41ae327 100644
--- a/tests/unit/test_workflow_copilot_prompt_rewrite.py
+++ b/tests/unit/test_workflow_copilot_prompt_rewrite.py
@@ -170,6 +170,12 @@ def test_rendered_prompt_keeps_security_ask_telemetry_and_workflow_wide_edit_sco
assert "ACTIVE BLOCK AUTHORING POLICY: CODE-ONLY BROWSER MODE" in new_workflow_prompt
+def test_rendered_prompt_scopes_verbatim_code_to_scouting_tool_output() -> None:
+ prompt = _render_production_prompt()
+
+ assert "scouting tool offers a SYNTHESIZED CODE BLOCK" in prompt
+
+
def test_ask_carve_out_gates_money_and_destruction_and_never_a_site_sent_message() -> None:
prompt = _render_production_prompt()
carve_out = _sentence_containing(prompt, "spends money")
diff --git a/tests/unit/test_workflow_copilot_route.py b/tests/unit/test_workflow_copilot_route.py
index 3b9168bd0..ab76c6e11 100644
--- a/tests/unit/test_workflow_copilot_route.py
+++ b/tests/unit/test_workflow_copilot_route.py
@@ -850,6 +850,88 @@ def test_chat_history_serves_a_saved_envelope_back_untouched() -> None:
assert served["terminalMessage"] == payload["terminalMessage"]
+def test_chat_history_correlates_interleaved_turn_openers() -> None:
+ now = datetime.now(timezone.utc)
+ messages = [
+ WorkflowCopilotChatMessage(
+ workflow_copilot_chat_message_id="message-refine",
+ workflow_copilot_chat_id="chat-1",
+ sender=WorkflowCopilotChatSender.PRODUCT,
+ content="Refine the recording (2 actions) into a reusable workflow.",
+ created_at=now,
+ modified_at=now,
+ ),
+ WorkflowCopilotChatMessage(
+ workflow_copilot_chat_message_id="message-later",
+ workflow_copilot_chat_id="chat-1",
+ sender=WorkflowCopilotChatSender.USER,
+ content="Add a confirmation step.",
+ created_at=now,
+ modified_at=now,
+ ),
+ WorkflowCopilotChatMessage(
+ workflow_copilot_chat_message_id="response-later",
+ workflow_copilot_chat_id="chat-1",
+ sender=WorkflowCopilotChatSender.AI,
+ content="Added it.",
+ turn_outcome=TurnOutcome(
+ response_kind=ResponseKind.BUILD,
+ copilot_turn_id="turn-later",
+ user_message_id="message-later",
+ ),
+ created_at=now,
+ modified_at=now,
+ ),
+ WorkflowCopilotChatMessage(
+ workflow_copilot_chat_message_id="response-refine",
+ workflow_copilot_chat_id="chat-1",
+ sender=WorkflowCopilotChatSender.AI,
+ content="Refinement ready.",
+ turn_outcome=TurnOutcome(
+ response_kind=ResponseKind.BUILD,
+ copilot_turn_id="turn-refine",
+ user_message_id="message-refine",
+ ),
+ created_at=now,
+ modified_at=now,
+ ),
+ ]
+
+ history = convert_to_history_messages(messages)
+
+ assert [message.turn_id for message in history] == [
+ "turn-refine",
+ "turn-later",
+ "turn-later",
+ "turn-refine",
+ ]
+
+
+def test_chat_history_correlates_a_pending_turn_opener() -> None:
+ now = datetime.now(timezone.utc)
+ message = WorkflowCopilotChatMessage(
+ workflow_copilot_chat_message_id="message-refine",
+ workflow_copilot_chat_id="chat-1",
+ sender=WorkflowCopilotChatSender.PRODUCT,
+ content="Refine the recording (2 actions) into a reusable workflow.",
+ created_at=now,
+ modified_at=now,
+ )
+
+ history = convert_to_history_messages(
+ [message],
+ {
+ "turn-refine": _make_pending_turn(
+ "turn-refine",
+ 10,
+ user_message_id="message-refine",
+ )
+ },
+ )
+
+ assert history[0].turn_id == "turn-refine"
+
+
@pytest.mark.asyncio
async def test_cancel_turn_never_applies_the_proposal(
monkeypatch: pytest.MonkeyPatch,
@@ -2358,26 +2440,45 @@ async def test_route_error_after_staged_commit_clears_stale_proposal_despite_kee
@pytest.mark.asyncio
@pytest.mark.parametrize(
- "newer_candidate_lands_during_commit",
- [False, True],
- ids=["assistant-row-write-fails", "candidate-superseded-then-assistant-row-write-fails"],
+ "scenario",
+ [
+ "assistant-row-write-fails",
+ "candidate-superseded-then-assistant-row-write-fails",
+ "agent-fails-after-newer-candidate-lands",
+ ],
)
-async def test_route_error_after_auto_accept_commit_keeps_the_committed_workflow(
+async def test_route_error_recovery_keeps_the_committed_workflow_and_other_turns_candidate(
monkeypatch: pytest.MonkeyPatch,
- newer_candidate_lands_during_commit: bool,
+ scenario: str,
api_key_request: MagicMock,
copilot_stream: MagicMock,
organization: SimpleNamespace,
) -> None:
- """Recovery after a successful auto-accept commit must not restore the pre-turn workflow over
- it. A newer candidate landing during the commit discards this turn's proposal, so recovery
- then sees a result that no longer carries the draft it committed."""
+ """Recovery must not restore the pre-turn workflow over a successful auto-accept commit, nor clear a
+ candidate another turn published, whether or not this turn produced a result."""
captured = install_fake_create(monkeypatch)
+ agent_fails = scenario == "agent-fails-after-newer-candidate-lands"
+ newer_candidate_lands_during_commit = scenario == "candidate-superseded-then-assistant-row-write-fails"
+
+ def candidate(owner_turn_id: str, *, unvalidated: bool = False) -> dict[str, Any]:
+ return {
+ "_copilot_yaml": "title: Committed draft",
+ **({"_copilot_unvalidated": True} if unvalidated else {}),
+ COPILOT_PROPOSAL_METADATA_KEY: {
+ "owner_turn_id": owner_turn_id,
+ "revision": 1,
+ "canonical_fingerprint": "fp",
+ "disposition": "review_untested" if unvalidated else "review_tested",
+ },
+ }
+
+ # An unvalidated pre-turn draft is what makes auto-accept recovery attempt a clear on the tokenless path.
+ chat_row: dict[str, Any] = {"proposed_workflow": candidate("turn-older", unvalidated=True) if agent_fails else None}
chat = SimpleNamespace(
workflow_copilot_chat_id="chat-1",
workflow_permanent_id="wpid-1",
organization_id="org-1",
- proposed_workflow=None,
+ proposed_workflow=chat_row["proposed_workflow"],
auto_accept=True,
)
original_workflow = _make_copilot_workflow("Saved workflow", _NOW)
@@ -2395,39 +2496,52 @@ async def test_route_error_after_auto_accept_commit_keeps_the_committed_workflow
real_restore = workflow_copilot_route._restore_workflow_definition
_, workflow_params = setup_new_copilot_mocks(monkeypatch, chat, original_workflow, agent_result)
monkeypatch.setattr(workflow_copilot_route, "_restore_workflow_definition", real_restore)
- workflow_params.clear_workflow_copilot_candidate = AsyncMock()
+ chat_reads: list[None] = []
- def candidate(owner_turn_id: str) -> dict[str, Any]:
- return {
- "_copilot_yaml": "title: Committed draft",
- COPILOT_PROPOSAL_METADATA_KEY: {
- "owner_turn_id": owner_turn_id,
- "revision": 1,
- "canonical_fingerprint": "fp",
- "disposition": "review_tested",
- },
- }
+ async def get_chat(*args: object, **kwargs: object) -> SimpleNamespace:
+ # The route's first read is its turn-start load; a candidate reaches its in-memory chat only through
+ # later reads of the row, as in production.
+ chat_reads.append(None)
+ if len(chat_reads) == 1:
+ return chat
+ return SimpleNamespace(**{**vars(chat), "proposed_workflow": chat_row["proposed_workflow"]})
+
+ workflow_params.get_workflow_copilot_chat_by_id = AsyncMock(side_effect=get_chat)
+
+ async def clear_candidate(
+ *, expected_owner_turn_id: str | None, expected_revision: int | None, **_: object
+ ) -> None:
+ token = (chat_row["proposed_workflow"] or {}).get(COPILOT_PROPOSAL_METADATA_KEY, {})
+ if (token.get("owner_turn_id"), token.get("revision")) != (expected_owner_turn_id, expected_revision):
+ raise CopilotProposalConflictError("Copilot proposal changed")
+ chat_row["proposed_workflow"] = None
+
+ workflow_params.clear_workflow_copilot_candidate = clear_candidate
if newer_candidate_lands_during_commit:
agent_result.proposal_owner_turn_id = "turn-this"
agent_result.proposal_revision = 1
- async def run_agent_publishing_candidate(*args: object, **kwargs: object) -> AgentResult:
- chat.proposed_workflow = candidate("turn-this")
- return agent_result
+ async def run_agent(*args: object, **kwargs: object) -> AgentResult:
+ if agent_fails:
+ chat_row["proposed_workflow"] = candidate("turn-newer")
+ raise RuntimeError("agent run failed")
+ if newer_candidate_lands_during_commit:
+ chat_row["proposed_workflow"] = candidate("turn-this")
+ return agent_result
- monkeypatch.setattr(workflow_copilot_route, "run_copilot_agent", run_agent_publishing_candidate)
+ monkeypatch.setattr(workflow_copilot_route, "run_copilot_agent", run_agent)
canonical: dict[str, str] = {original_workflow.workflow_id: original_workflow.title}
async def update_workflow_definition(**kwargs: Any) -> None:
canonical[kwargs["workflow_id"]] = kwargs["title"]
if newer_candidate_lands_during_commit and kwargs["title"] == committed.title:
- chat.proposed_workflow = candidate("turn-newer")
+ chat_row["proposed_workflow"] = candidate("turn-newer")
monkeypatch.setattr(app, "WORKFLOW_SERVICE", SimpleNamespace(update_workflow_definition=update_workflow_definition))
- first_assistant_write_failure = [RuntimeError("assistant row write failed")]
+ first_assistant_write_failure = [] if agent_fails else [RuntimeError("assistant row write failed")]
assistant_rows: list[dict[str, Any]] = []
async def create_message(**kwargs: Any) -> SimpleNamespace:
@@ -2445,7 +2559,10 @@ async def test_route_error_after_auto_accept_commit_keeps_the_committed_workflow
await handler(copilot_stream)
assert [row["turn_outcome"].response_kind for row in assistant_rows] == [ResponseKind.RECOVER]
- assert canonical == {original_workflow.workflow_id: committed.title}
+ assert canonical == {original_workflow.workflow_id: original_workflow.title if agent_fails else committed.title}
+ assert chat_row["proposed_workflow"] == (
+ None if scenario == "assistant-row-write-fails" else candidate("turn-newer")
+ )
@pytest.mark.asyncio
@@ -4299,13 +4416,14 @@ async def test_chat_history_shows_the_user_message_while_the_turn_is_still_runni
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""Reloading mid-turn — a tab refresh — shows the user's message without inventing an outcome."""
- chat = _make_persisted_chat([_make_pending_turn("turn-a", 30)])
+ chat = _make_persisted_chat([_make_pending_turn("turn-a", 30, user_message_id="wccm-0")])
store, _ = _install_reconcile_store(monkeypatch, chat)
store.add_message(WorkflowCopilotChatSender.USER, "build me a scraper")
response = await _load_history()
assert [m.content for m in response.chat_history] == ["build me a scraper"]
+ assert response.chat_history[0].turn_id == "turn-a"
assert all(m.turn_outcome is None for m in response.chat_history)
assert "turn-a" in chat.pending_turns
@@ -4599,7 +4717,9 @@ async def test_reconcile_drops_the_marker_of_a_turn_that_already_replied(
@pytest.mark.asyncio
async def test_persist_turn_messages_is_idempotent_for_one_turn(monkeypatch: pytest.MonkeyPatch) -> None:
"""A second finalizer pass for the same turn adds no duplicate user or assistant row."""
- chat = _make_persisted_chat([_make_pending_turn("turn-a", 10)])
+ chat = _make_persisted_chat(
+ [_make_pending_turn("turn-a", 10, user_message_id="wccm-0", cancel_token="cancel-turn-a")]
+ )
store, _ = _install_reconcile_store(monkeypatch, chat)
store.add_message(WorkflowCopilotChatSender.USER, "build me a scraper")
@@ -4621,6 +4741,8 @@ async def test_persist_turn_messages_is_idempotent_for_one_turn(monkeypatch: pyt
assert len(store.assistant_messages) == 1
assert store.assistant_messages[0].turn_outcome is not None
assert store.assistant_messages[0].turn_outcome.copilot_turn_id == "turn-a"
+ assert store.assistant_messages[0].turn_outcome.user_message_id == "wccm-0"
+ assert store.assistant_messages[0].turn_outcome.request_cancel_token == "cancel-turn-a"
@pytest.mark.asyncio
@@ -4644,9 +4766,16 @@ async def test_a_turn_that_never_started_keeps_its_attachments_on_the_recovery_r
turn_outcome=TurnOutcome(response_kind=ResponseKind.RECOVER),
narrative_payload=None,
attached_files=[CopilotAttachedFile(file_id="file_1", filename="targets.xlsx")],
+ request_cancel_token="cancel-turn-a",
)
assert [f.file_id for f in store.user_messages[0].attached_files] == ["file_1"]
+ assert store.assistant_messages[0].turn_outcome is not None
+ assert (
+ store.assistant_messages[0].turn_outcome.user_message_id
+ == store.user_messages[0].workflow_copilot_chat_message_id
+ )
+ assert store.assistant_messages[0].turn_outcome.request_cancel_token == "cancel-turn-a"
@pytest.mark.asyncio
@@ -4654,7 +4783,7 @@ async def test_a_finished_turn_replaces_the_interrupted_row_rather_than_dropping
monkeypatch: pytest.MonkeyPatch,
) -> None:
"""Recovery can reach a slow-but-live turn first; when that turn finishes, its reply is the truth."""
- chat = _make_persisted_chat([_make_pending_turn("turn-a", 10)])
+ chat = _make_persisted_chat([])
store, _ = _install_reconcile_store(monkeypatch, chat)
store.add_message(WorkflowCopilotChatSender.USER, "build me a scraper")
store.add_message(
@@ -4665,6 +4794,8 @@ async def test_a_finished_turn_replaces_the_interrupted_row_rather_than_dropping
reason_code=INTERRUPTED_TERMINAL_REASON,
terminal_reason=INTERRUPTED_TERMINAL_REASON,
copilot_turn_id="turn-a",
+ user_message_id="wccm-0",
+ request_cancel_token="cancel-turn-a",
),
)
@@ -4686,6 +4817,8 @@ async def test_a_finished_turn_replaces_the_interrupted_row_rather_than_dropping
outcome = store.assistant_messages[0].turn_outcome
assert outcome is not None
assert outcome.terminal_reason != INTERRUPTED_TERMINAL_REASON
+ assert outcome.user_message_id == "wccm-0"
+ assert outcome.request_cancel_token == "cancel-turn-a"
def test_reconcile_threshold_outlasts_the_turn_enforcement_ceiling() -> None:
@@ -5493,6 +5626,20 @@ def _refine_recording_mocks(monkeypatch: pytest.MonkeyPatch, *, session_found: b
return workflow_params
+def test_refine_recording_instruction_balances_simplification_and_interaction_fidelity() -> None:
+ instruction = workflow_copilot_route.REFINE_RECORDING_INSTRUCTION.lower()
+
+ assert len(instruction.split()) < 200
+ assert "collapse a multi-action path" in instruction
+ assert "neither the objective nor any retained interaction" in instruction
+ assert "interaction primitive" in instruction
+ assert "actionable target" in instruction
+ assert "successfully tested" in instruction
+ assert "extraction required by the objective" in instruction
+ assert "prose-only answer" in instruction
+ assert "never answer in prose" not in instruction
+
+
@pytest.mark.asyncio
async def test_refine_recording_dispatches_a_product_turn_with_the_packet_as_untrusted_evidence(
monkeypatch: pytest.MonkeyPatch,
diff --git a/tests/unit_tests/webeye/test_navigation_retry.py b/tests/unit_tests/webeye/test_navigation_retry.py
index 375d0f5ca..d1a9d6a31 100644
--- a/tests/unit_tests/webeye/test_navigation_retry.py
+++ b/tests/unit_tests/webeye/test_navigation_retry.py
@@ -8,8 +8,15 @@ from unittest.mock import AsyncMock, MagicMock
import pytest
import structlog.testing
-from skyvern.exceptions import BlockedHost, BlockedNavigationDestination, FailedToNavigateToUrl, UnresolvableHost
+from skyvern.exceptions import (
+ BlockedHost,
+ BlockedNavigationDestination,
+ FailedToNavigateToUrl,
+ UnresolvableHost,
+ UnresolvableNavigationHost,
+)
from skyvern.webeye.navigation import (
+ is_egress_attributable_navigation_failure,
navigate_with_retry,
redact_url_secrets,
revalidate_redirect_chain,
@@ -228,6 +235,102 @@ async def test_get_or_create_page_retries_retriable_failed_navigation() -> None:
browser_state.validate_browser_context.assert_awaited_once_with(page)
+_DEAD_URL = "https://gone.example.com/search"
+
+
+def _no_such_host(*args: object, **kwargs: object) -> list[object]:
+ raise socket.gaierror(socket.EAI_NONAME, "nodename nor servname provided, or not known")
+
+
+def _resolver_cannot_answer(*args: object, **kwargs: object) -> list[object]:
+ raise socket.gaierror(socket.EAI_AGAIN, "Temporary failure in name resolution")
+
+
+async def _navigate_failing_with(error_message: str) -> FailedToNavigateToUrl:
+ page = AsyncMock()
+ page.goto = AsyncMock(side_effect=Exception(error_message))
+
+ with pytest.raises(FailedToNavigateToUrl) as raised:
+ await navigate_with_retry(
+ navigate=lambda strategy: page.goto(_DEAD_URL, timeout=30000, wait_until=strategy),
+ url=_DEAD_URL,
+ retry_times=5,
+ settle=AsyncMock(),
+ sleep=AsyncMock(),
+ )
+
+ assert page.goto.call_count == 1
+ return raised.value
+
+
+ "error_message",
+ [
+ pytest.param("net::ERR_TUNNEL_CONNECTION_FAILED", id="tunnel"),
+ pytest.param("net::ERR_SOCKS_CONNECTION_FAILED", id="socks"),
+ ],
+)
+async def test_a_proxy_error_for_a_host_with_no_address_record_is_not_blamed_on_our_egress(
+ monkeypatch: pytest.MonkeyPatch, error_message: str
+) -> None:
+ # A proxied browser hands the hostname to the proxy, so a domain whose record is gone comes
+ # back as the proxy failing to open a tunnel rather than as ERR_NAME_NOT_RESOLVED.
+ monkeypatch.setattr("skyvern.utils.url_validators.socket.getaddrinfo", _no_such_host)
+
+ error = await _navigate_failing_with(error_message)
+
+ assert isinstance(error, UnresolvableNavigationHost)
+ assert error.host == "gone.example.com"
+ assert is_egress_attributable_navigation_failure(error) is False
+ # The browser's code is retained so downstream consumers still read the failure as terminal.
+ assert error_message in error.error_message
+
+
+async def test_a_proxy_error_for_a_resolvable_host_stays_attributed_to_our_egress() -> None:
+ # The autouse fixture resolves every host, so this is a genuine proxy fault and must keep the
+ # recovery path that retries it on a different proxy node.
+ error = await _navigate_failing_with("net::ERR_TUNNEL_CONNECTION_FAILED")
+
+ assert not isinstance(error, UnresolvableNavigationHost)
+ assert is_egress_attributable_navigation_failure(error) is True
+
+
+async def test_a_resolver_that_cannot_answer_does_not_reattribute_a_proxy_error(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ # EAI_AGAIN means the resolver never answered. Reading it as a dead host would reclassify every
+ # egress failure fleet-wide the moment worker DNS degrades.
+ monkeypatch.setattr("skyvern.utils.url_validators.socket.getaddrinfo", _resolver_cannot_answer)
+
+ error = await _navigate_failing_with("net::ERR_TUNNEL_CONNECTION_FAILED")
+
+ assert not isinstance(error, UnresolvableNavigationHost)
+ assert is_egress_attributable_navigation_failure(error) is True
+
+
+async def test_get_or_create_page_does_not_redraw_a_proxy_node_for_an_unresolvable_host() -> None:
+ browser_state = RealBrowserState(pw=AsyncMock())
+ browser_state.get_working_page = AsyncMock(return_value=None)
+ browser_state.check_and_fix_state = AsyncMock(
+ side_effect=UnresolvableNavigationHost(
+ url=_DEAD_URL,
+ host="gone.example.com",
+ error_message="net::ERR_TUNNEL_CONNECTION_FAILED",
+ )
+ )
+ browser_state.close_current_open_page = AsyncMock(return_value=True)
+
+ with pytest.raises(UnresolvableNavigationHost):
+ await browser_state.get_or_create_page(url=_DEAD_URL)
+
+ assert browser_state.check_and_fix_state.await_count == 1
+ browser_state.close_current_open_page.assert_not_awaited()
+
+
@pytest.mark.parametrize("url", BLOCKED_DESTINATIONS)
def test_validate_navigation_destination_rejects_internal_and_local_targets(url: str) -> None:
with pytest.raises(BlockedNavigationDestination):
diff --git a/uv.lock b/uv.lock
index abdab931a..20f2bfa6c 100644
--- a/uv.lock
+++ b/uv.lock
@@ -4964,11 +4964,11 @@ wheels = [
[[package]]
name = "pip"
-version = "26.1.2"
+version = "26.2.1"
source = { registry = "https://pypi.org/simple" }
-sdist = { url = "https://files.pythonhosted.org/packages/01/91/47e7d486260f618783899587af63ccf7980fb60245c3e63dd4571c6b57ad/pip-26.1.2.tar.gz", hash = "sha256:f49cd134c61cf2fd75e0ce2676db03e4054504a5a4986d00f8299ae632dc4605", size = 1840799, upload-time = "2026-05-31T17:33:58.56Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/ae/15/4500e320e6b101ec3b719ae85b697d9940b6cda672bc555bd6016fc60c6f/pip-26.2.1.tar.gz", hash = "sha256:f6ad667e89a1fe78046c8f13232b247200f5258d7828f3f7883d660878e0813f", size = 1848877, upload-time = "2026-08-04T22:51:14.148Z" }
wheels = [
- { url = "https://files.pythonhosted.org/packages/5d/95/6b5cb3461ea5673ba0995989746db58eb18b91b54dbf331e72f569540946/pip-26.1.2-py3-none-any.whl", hash = "sha256:382ff9f685ee3bc25864f820aa50505825f10f5458ffff07e30a6d96e5715cab", size = 1813144, upload-time = "2026-05-31T17:33:56.772Z" },
+ { url = "https://files.pythonhosted.org/packages/f3/6e/1736e5b4ae2b778ef2f81c47d797de9f891d4d8acb047a24ca37a60294dd/pip-26.2.1-py3-none-any.whl", hash = "sha256:71138adf1f4ca900cdb7d289c21b7494329f2332b6d85f0e1c42108c0384ed3e", size = 1816632, upload-time = "2026-08-04T22:51:12.472Z" },
]
[[package]]