Skip to content

Automate browser based workflows with AI

Show what the copilot actually did in its activity feed (SKY-13203) (#7869)

Andrew Neilson7d0370eaf0ba
diff --git a/skyvern-frontend/src/routes/workflows/copilot/NarrativeView.test.ts b/skyvern-frontend/src/routes/workflows/copilot/NarrativeView.test.ts
index a5e228cf2..1ca864cd0 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/NarrativeView.test.ts
+++ b/skyvern-frontend/src/routes/workflows/copilot/NarrativeView.test.ts
@@ -636,9 +636,22 @@ describe("condenseActivityEntries", () => {
it("leaves a still-pending retry visible, tagged with the attempt count", () => {
const s = reduceEvents([
- toolCall({ tool_call_id: "c1", tool_name: "extract" }),
- toolResult({ tool_call_id: "c1", tool_name: "extract", success: false }),
- toolCall({ tool_call_id: "c2", tool_name: "extract" }),
+ toolCall({
+ tool_call_id: "c1",
+ tool_name: "extract",
+ display_label: "Extracting",
+ }),
+ toolResult({
+ tool_call_id: "c1",
+ tool_name: "extract",
+ success: false,
+ display_label: "Extracting",
+ }),
+ toolCall({
+ tool_call_id: "c2",
+ tool_name: "extract",
+ display_label: "Extracting",
+ }),
]);
const condensed = condenseActivityEntries(s.designActivity);
expect(condensed).toHaveLength(1);
diff --git a/skyvern-frontend/src/routes/workflows/copilot/copilotPhases.test.ts b/skyvern-frontend/src/routes/workflows/copilot/copilotPhases.test.ts
index 3b7586b44..72130d0c9 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/copilotPhases.test.ts
+++ b/skyvern-frontend/src/routes/workflows/copilot/copilotPhases.test.ts
@@ -77,6 +77,12 @@ describe("toolActivityDisplayLabel — discovery tools (SKY-12385)", () => {
);
});
+ it("labels fill_credential_field without naming any credential", () => {
+ expect(toolActivityDisplayLabel("fill_credential_field")).toBe(
+ "Entering saved credentials",
+ );
+ });
+
it("still falls back to Working for unmapped tools", () => {
expect(toolActivityDisplayLabel("some_unmapped_tool")).toBe("Working");
});
@@ -156,6 +162,31 @@ describe("derivePhases — condenses each phase's entries (SKY-11971)", () => {
});
});
+ it("does not fold a failed edit of one block into a later edit of a different block", () => {
+ const t = turn({
+ designActivity: [
+ entry({
+ id: "tr-1",
+ kind: "tool_result",
+ toolName: "edit_block",
+ displayLabel: 'Editing block "Login Form"',
+ success: false,
+ }),
+ entry({
+ id: "tr-2",
+ kind: "tool_result",
+ toolName: "edit_block",
+ displayLabel: 'Editing block "Search Box"',
+ success: true,
+ }),
+ ],
+ });
+ const rows = derivePhases(t);
+ const exploreEntries = phase(rows, "explore").entries;
+ expect(exploreEntries).toHaveLength(2);
+ expect(exploreEntries.some((e) => e.attempts !== undefined)).toBe(false);
+ });
+
it("REGRESSION PIN: the explore 'N steps' stub matches the condensed row count, not the raw event count (Claude catch)", () => {
const t = turn({
terminal: "response",
diff --git a/skyvern-frontend/src/routes/workflows/copilot/narrativeState.phaseHint.test.ts b/skyvern-frontend/src/routes/workflows/copilot/narrativeState.phaseHint.test.ts
index 7e23c791c..9588704a7 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/narrativeState.phaseHint.test.ts
+++ b/skyvern-frontend/src/routes/workflows/copilot/narrativeState.phaseHint.test.ts
@@ -5,12 +5,14 @@ import {
EMPTY_NARRATIVE,
TurnNarrativeState,
applyNarrativeEvent,
+ hydrateNarrativeFromPayload,
} from "./narrativeState";
import {
WorkflowCopilotBlockProgressUpdate,
WorkflowCopilotRunOutcomeUpdate,
WorkflowCopilotStreamResponseUpdate,
WorkflowCopilotToolCallUpdate,
+ WorkflowCopilotToolResultUpdate,
WorkflowCopilotTurnStartUpdate,
} from "./workflowCopilotTypes";
@@ -135,7 +137,7 @@ describe("applyNarrativeEvent — lastActivityAtMs", () => {
it("stamps even when the tool is denylisted from the visible activity log", () => {
const s = applyNarrativeEvent(
EMPTY_NARRATIVE,
- toolCall({ tool_name: "list_credentials" }),
+ toolCall({ tool_name: "get_run_results" }),
777,
);
expect(s.lastActivityAtMs).toBe(777);
@@ -364,3 +366,145 @@ describe("applyNarrativeEvent — response hydration resets phase-hint fields",
expect(s.activitySeq).toBe(0);
});
});
+
+describe("applyNarrativeEvent — server-authored display labels", () => {
+ const toolResult = (
+ overrides: Partial<WorkflowCopilotToolResultUpdate> = {},
+ ): WorkflowCopilotToolResultUpdate => ({
+ type: "tool_result",
+ tool_name: "edit_block",
+ success: true,
+ summary: "",
+ iteration: 0,
+ tool_call_id: "call-1",
+ ...overrides,
+ });
+
+ it("renders a credential lookup label-only even when an older backend sends a count summary", () => {
+ const s = applyNarrativeEvent(
+ EMPTY_NARRATIVE,
+ toolResult({
+ tool_name: "list_credentials",
+ summary: "Found 4 credential(s)",
+ success: true,
+ }),
+ );
+ expect(s.designActivity).toHaveLength(1);
+ expect(s.designActivity[0]!.text).toBe("Checking saved credentials");
+ });
+
+ it("keeps a credential blocker explanation that arrives as a successful summary", () => {
+ const s = applyNarrativeEvent(
+ EMPTY_NARRATIVE,
+ toolResult({
+ tool_name: "list_credentials",
+ summary:
+ "Saved-credential scouting is not authorized for this request.",
+ success: true,
+ }),
+ );
+ expect(s.designActivity[0]!.text).toBe(
+ "Saved-credential scouting is not authorized for this request.",
+ );
+ });
+
+ it("still renders a credential lookup failure summary", () => {
+ const s = applyNarrativeEvent(
+ EMPTY_NARRATIVE,
+ toolResult({
+ tool_name: "list_credentials",
+ summary: "Failed: the credential store could not be reached",
+ success: false,
+ }),
+ );
+ expect(s.designActivity[0]!.text).toBe(
+ "Failed: the credential store could not be reached",
+ );
+ });
+
+ it("prefers the server display_label over the local map on tool_call and tool_result", () => {
+ let s: TurnNarrativeState = applyNarrativeEvent(
+ EMPTY_NARRATIVE,
+ toolCall({
+ tool_name: "edit_block",
+ display_label: 'Editing block "Log in"',
+ }),
+ );
+ s = applyNarrativeEvent(
+ s,
+ toolResult({ display_label: 'Editing block "Log in"' }),
+ );
+ expect(s.designActivity.map((e) => e.displayLabel)).toEqual([
+ 'Editing block "Log in"',
+ 'Editing block "Log in"',
+ ]);
+ expect(s.designActivity.map((e) => e.text)).toEqual([
+ 'Editing block "Log in"…',
+ 'Editing block "Log in"',
+ ]);
+ expect(s.designActivity.every((e) => !e.text.includes("Working"))).toBe(
+ true,
+ );
+ });
+
+ it("falls back to the local label map when the server sends no display_label", () => {
+ const s = applyNarrativeEvent(EMPTY_NARRATIVE, toolResult());
+ expect(s.designActivity[0]?.displayLabel).toBe("Editing block");
+ });
+
+ it("renders a credential-lookup row instead of suppressing it", () => {
+ let s: TurnNarrativeState = applyNarrativeEvent(
+ EMPTY_NARRATIVE,
+ toolCall({ tool_name: "list_credentials", tool_call_id: "call-9" }),
+ );
+ s = applyNarrativeEvent(
+ s,
+ toolResult({ tool_name: "list_credentials", tool_call_id: "call-9" }),
+ );
+ expect(s.designActivity.map((e) => e.text)).toEqual([
+ "Checking saved credentials…",
+ "Checking saved credentials",
+ ]);
+ });
+
+ it("keeps the persisted displayLabel when hydrating from narrative_payload", () => {
+ const s = hydrateNarrativeFromPayload({
+ turnId: "turn-1",
+ turnIndex: 0,
+ mode: "build",
+ designStarted: true,
+ designEnded: true,
+ draft: null,
+ blocks: [],
+ terminal: "response",
+ terminalMessage: "Done.",
+ narrativeSummary: "Done.",
+ priorBlockCount: null,
+ designActivity: [
+ {
+ kind: "tool_call",
+ text: 'Editing block "Log in"…',
+ iteration: 0,
+ toolName: "edit_block",
+ displayLabel: 'Editing block "Log in"',
+ id: "tc-call-1",
+ },
+ {
+ kind: "tool_result",
+ text: "Checking saved credentials",
+ iteration: 1,
+ toolName: "list_credentials",
+ displayLabel: "Checking saved credentials",
+ success: true,
+ id: "tr-call-9",
+ },
+ ],
+ startedAt: "2026-06-10T00:00:00Z",
+ endedAt: "2026-06-10T00:02:00Z",
+ });
+ expect(s?.designActivity.map((e) => e.displayLabel)).toEqual([
+ 'Editing block "Log in"',
+ "Checking saved credentials",
+ ]);
+ });
+});
diff --git a/skyvern-frontend/src/routes/workflows/copilot/narrativeState.ts b/skyvern-frontend/src/routes/workflows/copilot/narrativeState.ts
index 28c9595c3..d6a6dd924 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/narrativeState.ts
+++ b/skyvern-frontend/src/routes/workflows/copilot/narrativeState.ts
@@ -346,12 +346,12 @@ export const RUN_TOOLS = new Set([
// Tool names we never surface in the user-facing activity log. Internal
// observation/maintenance tools are not interesting context for the user.
const ACTIVITY_TOOL_DENYLIST = new Set([
- "list_credentials",
"get_run_results",
"get_browser_screenshot",
]);
-// Mirror of the backend _TOOL_ACTIVITY_DISPLAY_LABELS in narration.py.
+// Version-skew fallback for the server-authored display_label: mirror of the
+// backend _TOOL_ACTIVITY_DISPLAY_LABELS in narration.py.
const ACTIVITY_TOOL_DISPLAY_LABELS: Record<string, string> = {
update_workflow: "Updating workflow",
update_and_run_blocks: "Testing workflow",
@@ -367,6 +367,11 @@ const ACTIVITY_TOOL_DISPLAY_LABELS: Record<string, string> = {
inspect_current_workflow: "Inspecting workflow",
discover_workflow_entrypoint: "Finding the entry page",
inspect_page_for_composition: "Inspecting the page",
+ list_credentials: "Checking saved credentials",
+ fill_credential_field: "Entering saved credentials",
+ edit_block: "Editing block",
+ delete_block: "Deleting block",
+ synthesize_demonstrated_block: "Building a block from the recorded steps",
};
export function toolActivityDisplayLabel(toolName?: string | null): string {
@@ -392,16 +397,26 @@ function buildActivityFromToolCall(
};
}
+// summarize_tool_result's pre-SKY-13203 credential summary: "Found N credential(s)".
+const LEGACY_CREDENTIAL_COUNT_SUMMARY_RE = /^Found \d+ credential\(s\)$/;
+
function buildActivityFromToolResult(
event: WorkflowCopilotToolResultUpdate,
): ActivityEntry | null {
if (ACTIVITY_TOOL_DENYLIST.has(event.tool_name)) {
return null;
}
- const displayLabel = toolActivityDisplayLabel(event.tool_name);
+ const displayLabel =
+ event.display_label ?? toolActivityDisplayLabel(event.tool_name);
+ // Drop only the legacy enumeration count a pre-SKY-13203 backend still emits
+ // for credential lookups; a blocker explanation also arrives as a successful
+ // summary and must survive.
+ const suppressSummary =
+ event.tool_name === "list_credentials" &&
+ LEGACY_CREDENTIAL_COUNT_SUMMARY_RE.test(event.summary ?? "");
return {
kind: "tool_result",
- text: event.summary || displayLabel,
+ text: suppressSummary ? displayLabel : event.summary || displayLabel,
iteration: event.iteration,
toolName: event.tool_name,
displayLabel,
@@ -456,10 +471,11 @@ export function toolCallIdOf(entry: ActivityEntry): string | undefined {
// so a narration that streamed before the later attempt's result still
// reads as arriving before it, not after.
//
-// Known tradeoff: toolName is the only correlation signal available here
-// (no argument/target identity on the wire), so two independent same-tool
-// calls where only the first fails will also fold into one falsely-labeled
-// "retry" row. Accepted for this content-classification pass.
+// Correlation uses toolName plus displayLabel, which narrows the old
+// toolName-only rule: edits of two differently-named blocks no longer fold.
+// It is not full target identity — the label is humanized and length-capped,
+// so `login_v1`/`login_v2` and two labels sharing a 40-char prefix still
+// collide and fold. Raw-argument identity on the wire would close that.
//
// A tool without a dedicated backend summary falls back to a bare "OK"
// (summarize_tool_result in output_utils.py). That used to sit right below
@@ -524,6 +540,9 @@ export function condenseActivityEntries(
prevTool &&
entry.toolName !== undefined &&
prevTool.toolName === entry.toolName &&
+ (prevTool.displayLabel === undefined ||
+ entry.displayLabel === undefined ||
+ prevTool.displayLabel === entry.displayLabel) &&
prevTool.success === false
) {
condensed[lastToolIdx] = null;
diff --git a/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts b/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts
index afd64b047..bba8b174c 100644
--- a/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts
+++ b/skyvern-frontend/src/routes/workflows/copilot/workflowCopilotTypes.ts
@@ -207,6 +207,7 @@ export interface WorkflowCopilotToolCallUpdate {
export interface WorkflowCopilotToolResultUpdate {
type: "tool_result";
tool_name: string;
+ display_label?: string | null;
success: boolean;
summary: string;
iteration: number;
diff --git a/skyvern-frontend/src/routes/workflows/studio/runProjections.ts b/skyvern-frontend/src/routes/workflows/studio/runProjections.ts
index 41e417266..71e6f9800 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runProjections.ts
+++ b/skyvern-frontend/src/routes/workflows/studio/runProjections.ts
@@ -16,6 +16,10 @@ import { basicLocalTimeFormat, normalizeUtcTimestamp } from "@/util/timeFormat";
export type RunOutcome = "idle" | "running" | "failed" | "success";
+// Sentinel formatElapsed returns for a run with no usable start; callers that
+// hide the value instead of rendering a bare dash must compare against this.
+export const ELAPSED_NEVER_STARTED = "—";
+
// Run timestamps are naive ISO (no Z), so normalize to UTC before diffing — else
// a live run's (now - start) is skewed by the local timezone offset.
export function formatElapsed(
@@ -23,11 +27,11 @@ export function formatElapsed(
endIso: string | null,
): string {
if (!startIso) {
- return "—";
+ return ELAPSED_NEVER_STARTED;
}
const start = new Date(normalizeUtcTimestamp(startIso)).getTime();
if (Number.isNaN(start)) {
- return "—";
+ return ELAPSED_NEVER_STARTED;
}
const endMs = endIso
? new Date(normalizeUtcTimestamp(endIso)).getTime()
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 7e147f001..b5aff65b5 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runview/RunView.test.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/runview/RunView.test.tsx
@@ -737,6 +737,24 @@ describe("RunView live affordances", () => {
expect(scope.queryByText(/Run queued/)).not.toBeNull();
expect(scope.queryByRole("button", { name: "Live" })).toBeNull();
});
+
+ test("a queued run reports no elapsed time in the strip or the timeline", () => {
+ // created_at is always populated by the API, so only started_at can decide
+ // whether the run has actually accrued elapsed time.
+ seedCompletedRun({
+ status: Status.Queued,
+ created_at: "2026-06-30T23:59:00Z",
+ queued_at: "2026-06-30T23:59:30Z",
+ started_at: null,
+ finished_at: null,
+ });
+ const { container } = renderRunView();
+ const scope = within(container);
+
+ expect(scope.queryByText(/^Ran for /)).toBeNull();
+ expect(scope.queryByText(/·\s*\d+[ms]/)).toBeNull();
+ expect(scope.queryByText(/·\s*—/)).toBeNull();
+ });
});
describe("RunView iteration selection", () => {
diff --git a/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx b/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx
index bba45b58f..168ef8a8c 100644
--- a/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx
+++ b/skyvern-frontend/src/routes/workflows/studio/runview/RunView.tsx
@@ -27,6 +27,7 @@ import { findActiveItem } from "../../workflowRun/workflowTimelineUtils";
import { getOrderedRunParameters } from "../../utils";
import {
buildFilmstrip,
+ ELAPSED_NEVER_STARTED,
formatElapsed,
formatRunTimesTooltip,
runHasOutputs,
@@ -381,10 +382,16 @@ export function RunView({
return <RunPlaceholder loading={isLoading || runIdPending} />;
}
- const elapsed = formatElapsed(
- workflowRun.started_at ?? workflowRun.created_at ?? null,
+ // Same rule as the summary strip: created_at is always set, so falling back to
+ // it shows a queued run an elapsed time it never accrued. The never-started
+ // sentinel is dropped so the timeline omits the value entirely rather than
+ // rendering a bare dash.
+ const elapsedValue = formatElapsed(
+ workflowRun.started_at ?? null,
finalized ? (workflowRun.finished_at ?? null) : null,
);
+ const elapsed =
+ elapsedValue === ELAPSED_NEVER_STARTED ? undefined : elapsedValue;
const elapsedTitle = formatRunTimesTooltip(workflowRun);
const failureReason = formatFailureReason(
workflowRun.failure_reason ?? "The run failed.",
diff --git a/skyvern/forge/sdk/copilot/context.py b/skyvern/forge/sdk/copilot/context.py
index 328000ebe..fd43819b0 100644
--- a/skyvern/forge/sdk/copilot/context.py
+++ b/skyvern/forge/sdk/copilot/context.py
@@ -727,6 +727,7 @@ class InFlightStreamToolCall:
call_id: str
tool_name: str
iteration: int
+ display_label: str | None = None
@dataclass
diff --git a/skyvern/forge/sdk/copilot/narration.py b/skyvern/forge/sdk/copilot/narration.py
index d4eca57c4..b95bf97b1 100644
--- a/skyvern/forge/sdk/copilot/narration.py
+++ b/skyvern/forge/sdk/copilot/narration.py
@@ -28,6 +28,7 @@ from urllib.parse import urlparse
import structlog
from skyvern.forge.sdk.copilot.llm_config import get_fast_copilot_handler, resolve_fast_copilot_handler
+from skyvern.forge.sdk.copilot.output_utils import sanitize_block_label_for_display
from skyvern.forge.sdk.schemas.workflow_copilot import (
WorkflowCopilotBlockProgressUpdate,
WorkflowCopilotNarrationUpdate,
@@ -70,7 +71,7 @@ MAX_DESIGN_ACTIVITY_ENTRIES = 50
# Tools whose calls/results are never surfaced in the user-facing activity log.
# Mirror of the FE ACTIVITY_TOOL_DENYLIST in narrativeState.ts.
-ACTIVITY_TOOL_DENYLIST = frozenset({"list_credentials", "get_run_results", "get_browser_screenshot"})
+ACTIVITY_TOOL_DENYLIST = frozenset({"get_run_results", "get_browser_screenshot"})
# Tools that kick off a block run. Mirror of the FE RUN_TOOLS in narrativeState.ts.
# Their tool_call is recorded before the run flips running_block_label to the
@@ -100,18 +101,50 @@ _TOOL_ACTIVITY_DISPLAY_LABELS = {
"inspect_current_workflow": "Inspecting workflow",
"discover_workflow_entrypoint": "Finding the entry page",
"inspect_page_for_composition": "Inspecting the page",
+ "list_credentials": "Checking saved credentials",
+ "validate_block": "Checking the block",
+ "console_messages": "Reading the browser console",
+ "fill_credential_field": "Entering saved credentials",
+ "edit_block": "Editing block",
+ "delete_block": "Deleting block",
+ "synthesize_demonstrated_block": "Building a block from the recorded steps",
}
+# Tools whose label names the block they operate on, read from the tool's own
+# `label` argument.
+_BLOCK_TARGET_LABEL_TOOLS = frozenset({"edit_block", "delete_block"})
+_BLOCK_TARGET_VERSION_SUFFIX_RE = re.compile(r"_v\d+$", re.IGNORECASE)
-def tool_activity_display_label(tool_name: str) -> str:
- """Return a product-safe label for user-visible activity rows."""
- return _TOOL_ACTIVITY_DISPLAY_LABELS.get(tool_name, "Working")
+
+def _humanize_block_target(target: str) -> str:
+ # Mirror of the FE humanizeBlockLabel in blockLabel.ts, so the row matches
+ # the block card rendered beside it.
+ words = [w for w in re.split(r"[_\s]+", _BLOCK_TARGET_VERSION_SUFFIX_RE.sub("", target)) if w]
+ if not words:
+ return target
+ return " ".join(word[0].upper() + word[1:] for word in words)
-def build_tool_call_activity(tool_name: str, iteration: int, tool_call_id: str) -> NarrativeActivityEntry | None:
+def tool_activity_display_label(tool_name: str, tool_input: dict[str, Any] | None = None) -> str:
+ """Return a product-safe label for user-visible activity rows."""
+ label = _TOOL_ACTIVITY_DISPLAY_LABELS.get(tool_name, "Working")
+ if tool_name in _BLOCK_TARGET_LABEL_TOOLS and tool_input is not None:
+ target = tool_input.get("label")
+ if isinstance(target, str) and target.strip():
+ # The target is LLM-authored, so it goes through the same quote/length
+ # clamp the result-row summaries use before it is interpolated.
+ humanized = sanitize_block_label_for_display(_humanize_block_target(target))
+ if humanized:
+ return f'{label} "{humanized}"'
+ return label
+
+
+def build_tool_call_activity(
+ tool_name: str, iteration: int, tool_call_id: str, display_label: str | None = None
+) -> NarrativeActivityEntry | None:
if tool_name in ACTIVITY_TOOL_DENYLIST:
return None
- display_label = tool_activity_display_label(tool_name)
+ display_label = display_label or tool_activity_display_label(tool_name)
return {
"kind": "tool_call",
"text": f"{display_label}…",
@@ -123,11 +156,16 @@ def build_tool_call_activity(tool_name: str, iteration: int, tool_call_id: str)
def build_tool_result_activity(
- tool_name: str, summary: str, success: bool, iteration: int, tool_call_id: str
+ tool_name: str,
+ summary: str,
+ success: bool,
+ iteration: int,
+ tool_call_id: str,
+ display_label: str | None = None,
) -> NarrativeActivityEntry | None:
if tool_name in ACTIVITY_TOOL_DENYLIST:
return None
- display_label = tool_activity_display_label(tool_name)
+ display_label = display_label or tool_activity_display_label(tool_name)
return {
"kind": "tool_result",
"text": summary or display_label,
@@ -188,6 +226,9 @@ class NarratorState:
pending_activity: deque[_ToolActivityEntry] = field(default_factory=lambda: deque(maxlen=MAX_TOOL_ACTIVITY_BUFFER))
in_flight_task: asyncio.Task[None] | None = None
pending_transition: TransitionKind | None = None
+ # Which iteration recorded pending_transition, so suppressing iteration N cannot
+ # discard a transition an earlier typed-row-free iteration is still waiting on.
+ pending_transition_iteration: int | None = None
user_goal: str = ""
# Tool whose tool_called arrived but tool_output hasn't yet. Cleared on
# tool_output so post-tool transitions describe the finished action, not
@@ -209,10 +250,17 @@ class NarratorState:
run_tool_call_buckets: dict[str, str | None] = field(default_factory=dict)
# Per-turn (NarratorState lives one turn); collapses repeated code-repair progress to one entry.
emitted_progress_texts: set[str] = field(default_factory=set)
+ # Iterations already carrying a typed tool row; narrator prose is suppressed for
+ # those. Iteration numbers restart per stream_to_sse pass, so it is cleared on entry.
+ iterations_with_tool_activity: set[int] = field(default_factory=set)
def record_activity(self, entry: NarrativeActivityEntry | None) -> None:
if entry is None:
return
+ if entry.get("kind") in ("tool_call", "tool_result"):
+ iteration = entry.get("iteration")
+ if isinstance(iteration, int):
+ self.iterations_with_tool_activity.add(iteration)
label = self._activity_bucket_label(entry)
if label is None:
self.design_activity.append(entry)
@@ -267,6 +315,7 @@ class NarratorState:
or _TRANSITION_PRIORITY[kind] > _TRANSITION_PRIORITY[self.pending_transition]
):
self.pending_transition = kind
+ self.pending_transition_iteration = self.current_iteration
@dataclass(frozen=True)
@@ -316,9 +365,11 @@ class _NarratorPromptContext:
pending_tool_name: str | None = None
-def should_emit(state: NarratorState, now: float) -> bool:
+def should_emit(state: NarratorState, now: float, iteration: int | None = None) -> bool:
if state.pending_transition is None:
return False
+ if iteration is not None and iteration in state.iterations_with_tool_activity:
+ return False
if state.in_flight_task is not None and not state.in_flight_task.done():
return False
last_event = max(state.last_emitted_at or 0.0, state.last_attempted_at or 0.0)
@@ -335,12 +386,23 @@ def schedule_narration(
"""Kick off a background narration task if the gate allows. Fire-and-drop:
errors, timeouts, and empty responses are swallowed inside the task."""
now = time.monotonic()
- if not should_emit(state, now):
+ if not should_emit(state, now, iteration):
+ # A transition on an iteration that carries a typed row is already
+ # described by that row, so consume it here. Banking it would let it
+ # surface as stale prose against an unrelated later iteration.
+ if (
+ state.pending_transition is not None
+ and iteration in state.iterations_with_tool_activity
+ and state.pending_transition_iteration == iteration
+ ):
+ state.pending_transition = None
+ state.pending_transition_iteration = None
return
transition = state.pending_transition
assert transition is not None # guaranteed by should_emit
state.pending_transition = None
+ state.pending_transition_iteration = None
# Bound failure-path retries to the same gap window successes use; without
# this, a flaky narrator re-fires every poll tick.
state.last_attempted_at = now
@@ -397,6 +459,8 @@ async def _narration_task_body(
if not narration:
return
+ if iteration in state.iterations_with_tool_activity:
+ return
narration_ts = datetime.now(timezone.utc)
try:
@@ -642,9 +706,6 @@ def extract_tool_details(tool_name: str, parsed: dict[str, Any], *, success: boo
return "configuration invalid"
return ""
- if tool_name == "list_credentials":
- return _format_int_count(data, "credential")
-
if tool_name == "get_block_schema":
return _format_int_count(data, "step type")
diff --git a/skyvern/forge/sdk/copilot/output_utils.py b/skyvern/forge/sdk/copilot/output_utils.py
index 4bc9d4e71..6cf9dc33f 100644
--- a/skyvern/forge/sdk/copilot/output_utils.py
+++ b/skyvern/forge/sdk/copilot/output_utils.py
@@ -6,7 +6,8 @@ import base64
import binascii
import json
import re
-from collections.abc import Iterable, Iterator
+import unicodedata
+from collections.abc import Callable, Iterable, Iterator
from typing import TYPE_CHECKING, Any
from skyvern.forge.sdk.agents.context import sanitize_agent_tool_result_for_llm as sanitize_generic_tool_result_for_llm
@@ -332,7 +333,9 @@ def _result_data(result: dict[str, Any]) -> dict[str, Any]:
def _clean_structured_user_facing_text(value: Any, *, blocked_tool: str | None = None) -> str | None:
if not isinstance(value, str):
return None
- cleaned = " ".join(value.split())
+ # This branch returns straight to the caller without passing through
+ # _sanitize_failure_text, so it needs its own credential-id redaction.
+ cleaned = " ".join(_CREDENTIAL_ID_RE.sub("[credential]", value).split())
if not cleaned:
return None
try:
@@ -445,6 +448,43 @@ def user_facing_success(
_HEADERS_BLOB_RE = re.compile(r"\s*headers:\s*\{[^{}]*\}\s*", re.IGNORECASE)
_LARGE_DICT_BLOB_RE = re.compile(r"\{[^{}]{40,}\}")
+# Every credential-shaped id prefix in skyvern/forge/sdk/db/id.py, so a new one does
+# not quietly become renderable: cred, the vault/parameter kinds, folders, OAuth,
+# and run credential selections.
+_CREDENTIAL_ID_RE = re.compile(
+ r"`?\b(?:cred|cp|cfld|blc|bccd|bsi|opp|azcp|asp|goac|moac|wrcs)_[A-Za-z0-9][A-Za-z0-9_-]*`?"
+)
+# Punctuation an LLM-authored block label may keep. Everything outside this set and
+# the letter/digit categories becomes a space: a whitelist, because any quote-class
+# codepoint left in a label lets it visually close the quoting around it and append a
+# fabricated verdict to the row. Modifier letters (Lm) are excluded despite being
+# alphanumeric — U+02EE and friends are quote look-alikes.
+_LABEL_ALLOWED_PUNCTUATION = frozenset(" -_./")
+_LABEL_ALLOWED_CATEGORIES = frozenset({"Lu", "Ll", "Lt", "Lo", "Nd"})
+
+
+def sanitize_block_label_for_display(label: str, max_chars: int = 40) -> str:
+ """Make an LLM-authored block label safe to interpolate into a feed row."""
+ cleaned = "".join(
+ char if unicodedata.category(char) in _LABEL_ALLOWED_CATEGORIES or char in _LABEL_ALLOWED_PUNCTUATION else " "
+ for char in label
+ )
+ cleaned = " ".join(cleaned.split())
+ if len(cleaned) > max_chars:
+ cleaned = cleaned[: max_chars - 1].rstrip() + "…"
+ return cleaned
+
+
+_MAX_JOINED_BLOCK_LABELS = 5
+
+
+def _joined_block_labels(labels: list[Any], render: Callable[[object], str], *, cap: bool) -> str:
+ """Render a comma list, capped for display so a many-block run stays one row.
+ Uncapped for agent state, which reads the full set back."""
+ shown = [render(label) for label in (labels[:_MAX_JOINED_BLOCK_LABELS] if cap else labels)]
+ remaining = len(labels) - len(shown)
+ joined = ", ".join(label for label in shown if label)
+ return f"{joined} (+{remaining} more)" if remaining > 0 else joined
def _sanitize_failure_text(text: str, max_chars: int = 120) -> str:
@@ -456,6 +496,7 @@ def _sanitize_failure_text(text: str, max_chars: int = 120) -> str:
must pass through unchanged."""
text = _HEADERS_BLOB_RE.sub(" ", text)
text = _LARGE_DICT_BLOB_RE.sub("{...}", text)
+ text = _CREDENTIAL_ID_RE.sub("[credential]", text)
text = " ".join(text.split())
if not text:
return "(no details)"
@@ -485,8 +526,14 @@ def _describe_value_shape(value: Any) -> str:
return "value"
-def summarize_tool_result(tool_name: str, result: dict[str, Any]) -> str:
- """Create a brief human-readable summary of a tool result."""
+def summarize_tool_result(tool_name: str, result: dict[str, Any], *, for_display: bool = False) -> str:
+ """Summarize a tool result. ``for_display`` clamps LLM-authored block labels for
+ the activity feed; the default leaves them verbatim because this string is parsed
+ back into agent state by ``context.merge_turn_summary``."""
+
+ def block_label(value: object) -> str:
+ return sanitize_block_label_for_display(str(value)) if for_display else str(value)
+
if not result.get("ok", False):
return f"Failed: {_sanitize_failure_text(_extract_failure_message(result))}"
@@ -513,19 +560,22 @@ def summarize_tool_result(tool_name: str, result: dict[str, Any]) -> str:
return f"Schema for {data.get('block_type', '?')}"
if tool_name == "validate_block":
if data.get("valid"):
- return f"Block '{data.get('label', '?')}' is valid"
+ return f"Block '{block_label(data.get('label', '?'))}' is valid"
return "Block validation failed"
if tool_name == "run_blocks_and_collect_debug":
if not isinstance(raw_data, dict):
return "Run debug completed"
- executed = data.get("executed_block_labels") or [b.get("label", "?") for b in data.get("blocks", [])]
+ raw_executed = data.get("executed_block_labels") or [b.get("label", "?") for b in data.get("blocks", [])]
+ executed = _joined_block_labels(raw_executed, block_label, cap=for_display)
status = data.get("overall_status", "?")
requested = data.get("requested_block_labels") or []
- if requested and executed and list(executed) != list(requested):
- skipped = [label for label in requested if label not in set(executed)]
- suffix = f" (skipped prefix from cache: {', '.join(skipped)})" if skipped else ""
- return f"Run {', '.join(executed)}: {status}{suffix}"
- return f"Run {', '.join(executed)}: {status}"
+ if requested and raw_executed and list(raw_executed) != list(requested):
+ skipped = _joined_block_labels(
+ [label for label in requested if label not in set(raw_executed)], block_label, cap=for_display
+ )
+ suffix = f" (skipped prefix from cache: {skipped})" if skipped else ""
+ return f"Run {executed}: {status}{suffix}"
+ return f"Run {executed}: {status}"
if tool_name == "get_browser_screenshot":
url = data.get("url")
return f"Screenshot taken ({url[:80]})" if url else "Screenshot taken"
@@ -627,7 +677,20 @@ _USER_FACING_LOOP_MESSAGE = "The agent got stuck retrying the same step — movi
_USER_FACING_JINJA_MESSAGE = "A workflow parameter could not be filled in."
_USER_FACING_GENERIC_FAILURE = _STRUCTURED_UNSAFE_FALLBACK
-_USER_FACING_EMPTY_SUCCESS_TOOLS: frozenset[str] = frozenset({"click", "type_text", "evaluate", "select_option"})
+_USER_FACING_EMPTY_SUCCESS_TOOLS: frozenset[str] = frozenset(
+ {
+ "click",
+ "type_text",
+ "evaluate",
+ "select_option",
+ "list_credentials",
+ # The server-authored display label already names the operation and its
+ # target block; a bare "OK" summary would render instead of it.
+ "edit_block",
+ "delete_block",
+ "synthesize_demonstrated_block",
+ }
+)
def _translate_failure_for_user(error_text: str) -> str:
@@ -665,7 +728,7 @@ def format_tool_result_for_user(
return _translate_failure_for_user(_extract_failure_message(result))
if tool_name in _USER_FACING_EMPTY_SUCCESS_TOOLS:
return ""
- return summarize_tool_result(tool_name, result)
+ return summarize_tool_result(tool_name, result, for_display=True)
def truncate_output(output: Any, max_chars: int = 2000) -> str | None:
diff --git a/skyvern/forge/sdk/copilot/streaming_adapter.py b/skyvern/forge/sdk/copilot/streaming_adapter.py
index 2ae881c48..4ed2a7a8e 100644
--- a/skyvern/forge/sdk/copilot/streaming_adapter.py
+++ b/skyvern/forge/sdk/copilot/streaming_adapter.py
@@ -267,6 +267,9 @@ async def stream_to_sse(
asyncio's cancellation machinery still runs normally.
"""
call_id_to_name: dict[str, str] = {}
+ # Label authored once at tool_called so the matching result renders the same
+ # string; the result event carries no arguments to re-derive it from.
+ call_id_to_label: dict[str, str] = {}
# Counts completed tool round-trips (tool_called + tool_output pair), not
# raw stream events. Both TOOL_CALL and TOOL_RESULT for the same round
# carry the same iteration value; it advances after the matching result.
@@ -286,6 +289,12 @@ async def stream_to_sse(
getattr(ctx, "organization_id", None),
)
narrator_enabled = narrator_state.resolved_handler is not None
+ # Iteration numbers restart at 0 on every enforcement pass, so a set carried
+ # over from the previous pass would suppress this pass's clean iterations —
+ # and a transition tagged with a previous pass's iteration number could be
+ # consumed by this pass's same-numbered iteration by coincidence.
+ narrator_state.iterations_with_tool_activity.clear()
+ narrator_state.pending_transition_iteration = None
ctx.narrator_state = narrator_state
user_message = getattr(ctx, "user_message", "") or ""
if user_message and not narrator_state.user_goal:
@@ -319,35 +328,42 @@ async def stream_to_sse(
call_id = _get_raw_field(raw, "call_id") or _get_raw_field(raw, "id") or ""
tool_name = _get_raw_field(raw, "name") or "unknown"
call_id_to_name[call_id] = tool_name
+
+ raw_args = _get_raw_field(raw, "arguments")
+ tool_input: dict[str, Any] = {}
+ if isinstance(raw_args, str):
+ try:
+ tool_input = json.loads(raw_args)
+ except (json.JSONDecodeError, TypeError):
+ # A placeholder, not the raw text: _sanitize_input redacts by
+ # key name, which cannot see inside one opaque string.
+ tool_input = {"raw": "<unparsed arguments>"}
+ if not isinstance(tool_input, dict):
+ tool_input = {"raw": "<unparsed arguments>"}
+ elif isinstance(raw_args, dict):
+ tool_input = raw_args
+
+ display_label = tool_activity_display_label(tool_name, tool_input)
+ call_id_to_label[call_id] = display_label
ctx.in_flight_stream_tool_call = InFlightStreamToolCall(
- call_id=call_id, tool_name=tool_name, iteration=iteration
+ call_id=call_id, tool_name=tool_name, iteration=iteration, display_label=display_label
+ )
+ narrator_state.record_activity(
+ build_tool_call_activity(tool_name, iteration, call_id, display_label=display_label)
)
- narrator_state.record_activity(build_tool_call_activity(tool_name, iteration, call_id))
if not client_gone:
- raw_args = _get_raw_field(raw, "arguments")
- tool_input: dict[str, Any] = {}
- if isinstance(raw_args, str):
- try:
- tool_input = json.loads(raw_args)
- except (json.JSONDecodeError, TypeError):
- tool_input = {"raw": raw_args}
- elif isinstance(raw_args, dict):
- tool_input = raw_args
-
await stream.send(
WorkflowCopilotToolCallUpdate(
type=WorkflowCopilotStreamMessageType.TOOL_CALL,
tool_name=tool_name,
- display_label=tool_activity_display_label(tool_name),
+ display_label=display_label,
tool_input=_sanitize_input(tool_input),
iteration=iteration,
tool_call_id=call_id,
)
)
- # First narration lands here (~seconds after submit) rather
- # than waiting for tool_output of a long tool.
if narrator_enabled:
narrator_state.pending_tool_name = tool_name
narrator_state.current_iteration = iteration
@@ -384,8 +400,11 @@ async def stream_to_sse(
detail = summarize_tool_result_detail(
parsed, tool_name=tool_name, blocker_signal=blocker_signals, success=success
)
+ result_label = call_id_to_label.get(call_id) or tool_activity_display_label(tool_name)
narrator_state.record_activity(
- build_tool_result_activity(tool_name, summary, success, iteration, call_id)
+ build_tool_result_activity(
+ tool_name, summary, success, iteration, call_id, display_label=result_label
+ )
)
if not client_gone:
@@ -393,6 +412,7 @@ async def stream_to_sse(
WorkflowCopilotToolResultUpdate(
type=WorkflowCopilotStreamMessageType.TOOL_RESULT,
tool_name=tool_name,
+ display_label=result_label,
success=success,
summary=summary,
iteration=iteration,
@@ -522,10 +542,13 @@ async def flush_goal_satisfied_tool_result(stream: EventSourceStream, ctx: Copil
blocker_signals = _tool_blocker_signal_candidates(ctx)
summary = format_tool_result_for_user(pending.tool_name, parsed, blocker_signal=blocker_signals)
success = user_facing_success(parsed, blocker_signal=blocker_signals)
+ display_label = pending.display_label or tool_activity_display_label(pending.tool_name)
narrator_state = ctx.narrator_state
if narrator_state is not None:
narrator_state.record_activity(
- build_tool_result_activity(pending.tool_name, summary, success, pending.iteration, pending.call_id)
+ build_tool_result_activity(
+ pending.tool_name, summary, success, pending.iteration, pending.call_id, display_label=display_label
+ )
)
if await stream.is_disconnected():
return
@@ -533,6 +556,7 @@ async def flush_goal_satisfied_tool_result(stream: EventSourceStream, ctx: Copil
WorkflowCopilotToolResultUpdate(
type=WorkflowCopilotStreamMessageType.TOOL_RESULT,
tool_name=pending.tool_name,
+ display_label=display_label,
success=success,
summary=summary,
iteration=pending.iteration,
diff --git a/skyvern/forge/sdk/db/repositories/tasks.py b/skyvern/forge/sdk/db/repositories/tasks.py
index 8a6ab4c4c..048c434d9 100644
--- a/skyvern/forge/sdk/db/repositories/tasks.py
+++ b/skyvern/forge/sdk/db/repositories/tasks.py
@@ -86,10 +86,15 @@ class TasksRepository(BaseRepository):
terminate_criterion = _sanitize(terminate_criterion)
workflow_system_prompt = _sanitize(workflow_system_prompt)
+ # created_at is passed so an already-running task cannot start before it was created; None
+ # falls through to the column default.
+ started_at = naive_utc_now() if str(status) == TaskStatus.running.value else None
+
async with self.Session() as session:
new_task = TaskModel(
status=status,
- started_at=naive_utc_now() if str(status) == TaskStatus.running.value else None,
+ started_at=started_at,
+ created_at=started_at,
task_type=task_type,
url=url,
title=title,
diff --git a/skyvern/forge/sdk/schemas/sdk_actions.py b/skyvern/forge/sdk/schemas/sdk_actions.py
index 85b41c995..814b296c3 100644
--- a/skyvern/forge/sdk/schemas/sdk_actions.py
+++ b/skyvern/forge/sdk/schemas/sdk_actions.py
@@ -42,7 +42,12 @@ class ClickAction(SdkActionBase):
selector: str | None = Field(default="", description="CSS selector for the element")
intention: str = Field(default="", description="The intention or goal of the click")
data: str | dict[str, Any] | None = Field(None, description="Additional context data")
- timeout: float = Field(default=settings.BROWSER_ACTION_TIMEOUT_MS, description="Timeout in milliseconds")
+ # Keep runtime defaults configurable while documenting a stable API default.
+ timeout: float = Field(
+ default_factory=lambda: settings.BROWSER_ACTION_TIMEOUT_MS,
+ description="Timeout in milliseconds",
+ json_schema_extra={"default": 10000},
+ )
def get_navigation_goal(self) -> str | None:
return self.intention
@@ -61,7 +66,11 @@ class InputTextAction(SdkActionBase):
data: str | dict[str, Any] | None = Field(None, description="Additional context data")
totp_identifier: str | None = Field(None, description="TOTP identifier for input_text actions")
totp_url: str | None = Field(None, description="TOTP URL for input_text actions")
- timeout: float = Field(default=settings.BROWSER_ACTION_TIMEOUT_MS, description="Timeout in milliseconds")
+ timeout: float = Field(
+ default_factory=lambda: settings.BROWSER_ACTION_TIMEOUT_MS,
+ description="Timeout in milliseconds",
+ json_schema_extra={"default": 10000},
+ )
def __repr__(self) -> str:
payload = redact_input_text_payload_for_log(self.model_dump(), value_key="value")
@@ -86,7 +95,11 @@ class SelectOptionAction(SdkActionBase):
value: str | None = Field(default="", description="Value to select")
intention: str = Field(default="", description="The intention or goal of the selection")
data: str | dict[str, Any] | None = Field(None, description="Additional context data")
- timeout: float = Field(default=settings.BROWSER_ACTION_TIMEOUT_MS, description="Timeout in milliseconds")
+ timeout: float = Field(
+ default_factory=lambda: settings.BROWSER_ACTION_TIMEOUT_MS,
+ description="Timeout in milliseconds",
+ json_schema_extra={"default": 10000},
+ )
def get_navigation_goal(self) -> str | None:
return self.intention
@@ -103,7 +116,11 @@ class UploadFileAction(SdkActionBase):
file_url: str | None = Field(default="", description="File URL for upload")
intention: str = Field(default="", description="The intention or goal of the upload")
data: str | dict[str, Any] | None = Field(None, description="Additional context data")
- timeout: float = Field(default=settings.BROWSER_ACTION_TIMEOUT_MS, description="Timeout in milliseconds")
+ timeout: float = Field(
+ default_factory=lambda: settings.BROWSER_ACTION_TIMEOUT_MS,
+ description="Timeout in milliseconds",
+ json_schema_extra={"default": 10000},
+ )
def get_navigation_goal(self) -> str | None:
return self.intention
diff --git a/skyvern/forge/sdk/schemas/workflow_copilot.py b/skyvern/forge/sdk/schemas/workflow_copilot.py
index 68c05f188..b76aaf040 100644
--- a/skyvern/forge/sdk/schemas/workflow_copilot.py
+++ b/skyvern/forge/sdk/schemas/workflow_copilot.py
@@ -320,6 +320,10 @@ class WorkflowCopilotToolResultUpdate(BaseModel):
WorkflowCopilotStreamMessageType.TOOL_RESULT, description="Message type"
)
tool_name: str = Field(..., description="Name of the tool that was called")
+ display_label: str | None = Field(
+ None,
+ description="Product-safe label for rendering the tool result in user-visible activity surfaces",
+ )
success: bool = Field(..., description="Whether the tool call succeeded")
summary: str = Field(..., description="Brief human-readable summary of the result")
iteration: int = Field(..., description="Agent loop iteration number")
diff --git a/skyvern/forge/sdk/workflow/models/block.py b/skyvern/forge/sdk/workflow/models/block.py
index fa0e40c26..9c06d200f 100644
--- a/skyvern/forge/sdk/workflow/models/block.py
+++ b/skyvern/forge/sdk/workflow/models/block.py
@@ -9685,8 +9685,16 @@ class TaskV2Block(Block):
url: str | None = None
totp_verification_url: str | None = None
totp_identifier: str | None = None
- max_iterations: int = settings.MAX_ITERATIONS_PER_TASK_V2
- max_steps: int = settings.MAX_STEPS_PER_TASK_V2
+ # These documented defaults must stay literals; reading the setting here would put an
+ # environment value back into the published OpenAPI document.
+ max_iterations: int = Field(
+ default_factory=lambda: settings.MAX_ITERATIONS_PER_TASK_V2,
+ json_schema_extra={"default": 50},
+ )
+ max_steps: int = Field(
+ default_factory=lambda: settings.MAX_STEPS_PER_TASK_V2,
+ json_schema_extra={"default": 25},
+ )
def _resolve_totp_identifier(self, workflow_run_context: WorkflowRunContext) -> str | None:
if self.totp_identifier:
diff --git a/skyvern/schemas/workflows.py b/skyvern/schemas/workflows.py
index fc5c67ea0..c8f8f8202 100644
--- a/skyvern/schemas/workflows.py
+++ b/skyvern/schemas/workflows.py
@@ -1153,8 +1153,16 @@ class TaskV2BlockYAML(BlockYAML):
url: str | None = None
totp_verification_url: str | None = None
totp_identifier: str | None = None
- max_iterations: int = settings.MAX_ITERATIONS_PER_TASK_V2
- max_steps: int = settings.MAX_STEPS_PER_TASK_V2
+ # These documented defaults must stay literals; reading the setting here would put an
+ # environment value back into the published OpenAPI document.
+ max_iterations: int = Field(
+ default_factory=lambda: settings.MAX_ITERATIONS_PER_TASK_V2,
+ json_schema_extra={"default": 50},
+ )
+ max_steps: int = Field(
+ default_factory=lambda: settings.MAX_STEPS_PER_TASK_V2,
+ json_schema_extra={"default": 25},
+ )
disable_cache: bool = False
diff --git a/tests/unit/forge/sdk/db/test_repositories.py b/tests/unit/forge/sdk/db/test_repositories.py
index ff2abcb25..186340e22 100644
--- a/tests/unit/forge/sdk/db/test_repositories.py
+++ b/tests/unit/forge/sdk/db/test_repositories.py
@@ -7,6 +7,9 @@ from unittest.mock import MagicMock, patch
import pytest
+from skyvern.forge.sdk.schemas.tasks import TaskStatus
+from tests.unit.conftest import MockAsyncSessionCtx, make_mock_session
+
def test_credential_repository_instantiation():
from skyvern.forge.sdk.db.repositories.credentials import CredentialRepository
@@ -382,3 +385,40 @@ def test_agent_db_defines_no_delegator_methods():
"Add data-access methods to the domain repository and call it via the typed attribute "
"(e.g. db.tasks.get_task) instead of adding delegators to AgentDB."
)
+
+
+async def _create_task_with_status(monkeypatch: pytest.MonkeyPatch, status: str):
+ from skyvern.forge.sdk.db.repositories import tasks as tasks_module
+
+ session = make_mock_session(MagicMock())
+ monkeypatch.setattr(tasks_module, "convert_to_task", lambda model, *args, **kwargs: model)
+ repo = tasks_module.TasksRepository(
+ session_factory=lambda: MockAsyncSessionCtx(session),
+ debug_enabled=False,
+ )
+
+ return await repo.create_task(
+ url="https://example.test/",
+ title=None,
+ navigation_goal=None,
+ data_extraction_goal=None,
+ navigation_payload=None,
+ status=status,
+ )
+
+
+async def test_create_task_running_is_not_created_after_it_started(monkeypatch: pytest.MonkeyPatch):
+ """queued_seconds is started_at - created_at, so a task created already-running must not
+ stamp started_at ahead of the flush-time created_at default."""
+ task = await _create_task_with_status(monkeypatch, TaskStatus.running.value)
+
+ assert task.started_at is not None
+ assert task.created_at == task.started_at
+
+
+async def test_create_task_leaves_started_at_unset_for_other_statuses(monkeypatch: pytest.MonkeyPatch):
+ task = await _create_task_with_status(monkeypatch, TaskStatus.created.value)
+
+ assert task.started_at is None
diff --git a/tests/unit/test_actions.py b/tests/unit/test_actions.py
index e7476e014..739dd494d 100644
--- a/tests/unit/test_actions.py
+++ b/tests/unit/test_actions.py
@@ -3,8 +3,10 @@ from unittest.mock import MagicMock
import pytest
from pydantic import ValidationError
+from skyvern.config import settings
from skyvern.forge.sdk.db.repositories.workflow_parameters import WorkflowParametersRepository
from skyvern.forge.sdk.db.utils import hydrate_action
+from skyvern.forge.sdk.schemas import sdk_actions
from skyvern.forge.sdk.schemas.sdk_actions import InputTextAction as SdkInputTextAction
from skyvern.forge.sdk.schemas.sdk_actions import SdkActionType
from skyvern.schemas.steps import AgentStepOutput
@@ -47,6 +49,21 @@ def test_sdk_input_text_action_type_constant_matches_sdk_enum() -> None:
assert SDK_INPUT_TEXT_ACTION_TYPE == SdkActionType.AI_INPUT_TEXT.value
+def test_sdk_action_timeout_defaults_are_environment_independent(monkeypatch: pytest.MonkeyPatch) -> None:
+ monkeypatch.setattr(settings, "BROWSER_ACTION_TIMEOUT_MS", 4242)
+ action_types = (
+ sdk_actions.ClickAction,
+ sdk_actions.InputTextAction,
+ sdk_actions.SelectOptionAction,
+ sdk_actions.UploadFileAction,
+ )
+
+ assert [action_type().timeout for action_type in action_types] == [4242] * 4
+ assert [action_type.model_json_schema()["properties"]["timeout"]["default"] for action_type in action_types] == [
+ 10000
+ ] * 4
+
+
def test_action_parse__no_element_id() -> None:
action_no_element_id = {
"action_type": "click",
diff --git a/tests/unit/test_copilot_credential_pause.py b/tests/unit/test_copilot_credential_pause.py
index 8b8aad519..3f371abee 100644
--- a/tests/unit/test_copilot_credential_pause.py
+++ b/tests/unit/test_copilot_credential_pause.py
@@ -35,7 +35,6 @@ from skyvern.config import settings
from skyvern.forge import app
from skyvern.forge.sdk.cache.base import NoopLock
from skyvern.forge.sdk.copilot import credential_pause as credential_pause_module
-from skyvern.forge.sdk.copilot import enforcement as enforcement_module
from skyvern.forge.sdk.copilot import tools as tools_module
from skyvern.forge.sdk.copilot.agent import RequestPolicyGuardrailInputs, _derive_turn_intent_on_context
from skyvern.forge.sdk.copilot.config import CopilotConfig
@@ -1030,7 +1029,7 @@ def test_elapsed_run_seconds_subtracts_pause_time() -> None:
@pytest.mark.asyncio
async def test_paused_loop_does_not_trip_total_timeout_on_resume(monkeypatch: pytest.MonkeyPatch) -> None:
- """A pause longer than the work budget is excluded when the loop resumes."""
+ """Fails on old code: a slow pause would consume the (tiny) real timeout budget."""
monkeypatch.setattr("skyvern.forge.sdk.copilot.enforcement.TOTAL_TIMEOUT_SECONDS", 0.15)
ctx = make_copilot_context()
@@ -1043,18 +1042,13 @@ async def test_paused_loop_does_not_trip_total_timeout_on_resume(monkeypatch: py
cache = _FakeCache()
monkeypatch.setattr(credential_pause_module.app._inst, "CACHE", cache, raising=False)
+ monkeypatch.setattr(credential_pause_module, "CREDENTIAL_RESPONSE_POLL_SECONDS", 0.2)
- monotonic_seconds = 100.0
- fake_time = SimpleNamespace(monotonic=lambda: monotonic_seconds)
- monkeypatch.setattr(enforcement_module, "time", fake_time)
- monkeypatch.setattr(credential_pause_module, "time", fake_time)
-
- async def _resolve_after_pause(*args: Any, **kwargs: Any) -> CredentialPauseResolution:
- nonlocal monotonic_seconds
- monotonic_seconds += 0.25
- return CredentialPauseResolution(action="skip")
-
- monkeypatch.setattr(credential_pause_module, "_wait_for_credential_response", _resolve_after_pause)
+ async def _populate_after_first_poll() -> None:
+ await asyncio.sleep(0.25)
+ cache.store[credential_response_cache_key("org-1", "chat-1", "turn-credit")] = encode_credential_response(
+ "skip", None
+ )
stream = _make_stream()
fake_result = _fake_result()
@@ -1072,17 +1066,20 @@ async def test_paused_loop_does_not_trip_total_timeout_on_resume(monkeypatch: py
config = CopilotConfig(credential_pause_enabled=True, credential_pause_timeout_seconds=5)
- returned = await run_with_enforcement(
- agent=MagicMock(),
- initial_input="hello",
- ctx=ctx,
- stream=stream,
- run_config=RunConfig(),
- copilot_config=config,
- )
+ populate_task = asyncio.ensure_future(_populate_after_first_poll())
+ try:
+ returned = await run_with_enforcement(
+ agent=MagicMock(),
+ initial_input="hello",
+ ctx=ctx,
+ stream=stream,
+ run_config=RunConfig(),
+ copilot_config=config,
+ )
+ finally:
+ await populate_task
assert returned is fake_result
- assert ctx.copilot_credential_pause_seconds == 0.25
assert len(calls) == 2, "pause time must be credited so the resumed iteration isn't timed out"
assert ctx.copilot_total_timeout_exceeded is False
diff --git a/tests/unit/test_copilot_narration.py b/tests/unit/test_copilot_narration.py
index bcf4bccf2..e49599401 100644
--- a/tests/unit/test_copilot_narration.py
+++ b/tests/unit/test_copilot_narration.py
@@ -10,6 +10,7 @@ from __future__ import annotations
import asyncio
import time
+from datetime import datetime, timezone
from types import SimpleNamespace
from typing import Any
@@ -1337,3 +1338,132 @@ async def test_poll_tick_block_ended_at_overwritten_by_latest_terminal_and_clear
await state.in_flight_task
except (asyncio.CancelledError, Exception):
pass
+
+
+def _armed_state(iteration: int) -> NarratorState:
+ state = NarratorState()
+ state.record_transition(TransitionKind.NEW_TOOL_CLUSTER)
+ state.record_activity(narration.build_tool_call_activity("edit_block", iteration, f"c{iteration}"))
+ return state
+
+
+def test_should_emit_false_for_an_iteration_carrying_typed_tool_activity() -> None:
+ state = _armed_state(2)
+ assert state.iterations_with_tool_activity == {2}
+ assert should_emit(state, now=100.0, iteration=2) is False
+
+
+def test_should_emit_true_for_a_clean_iteration_alongside_an_armed_one() -> None:
+ state = _armed_state(2)
+ assert should_emit(state, now=100.0, iteration=3) is True
+
+
+def test_should_emit_true_when_the_iterations_only_tool_is_denylisted() -> None:
+ state = NarratorState()
+ state.record_transition(TransitionKind.NEW_TOOL_CLUSTER)
+ state.record_activity(narration.build_tool_call_activity("get_run_results", 0, "c0"))
+ assert state.iterations_with_tool_activity == set()
+ assert should_emit(state, now=100.0, iteration=0) is True
+
+
+def test_tool_result_alone_arms_the_iteration() -> None:
+ state = NarratorState()
+ state.record_transition(TransitionKind.NEW_TOOL_CLUSTER)
+ state.record_activity(narration.build_tool_result_activity("edit_block", "done", True, 5, "c5"))
+ assert should_emit(state, now=100.0, iteration=5) is False
+
+
+def test_narration_activity_does_not_arm_its_own_iteration() -> None:
+ state = NarratorState()
+ state.record_activity(
+ narration.build_narration_activity("Refining the workflow's code", 1, datetime(2026, 1, 1, tzinfo=timezone.utc))
+ )
+ assert state.iterations_with_tool_activity == set()
+
+
+async def test_narration_dropped_when_the_iteration_arms_while_the_narrator_is_in_flight(
+ monkeypatch: pytest.MonkeyPatch,
+) -> None:
+ state = NarratorState()
+ released = asyncio.Event()
+
+ async def _handler(prompt: str, prompt_name: str, **kwargs: object) -> str:
+ await released.wait()
+ return "Working on your goal."
+
+ await _install_handler(monkeypatch, _handler)
+ state.record_transition(TransitionKind.NEW_TOOL_CLUSTER)
+ stream = _FakeStream()
+
+ schedule_narration(state, stream, iteration=4) # type: ignore[arg-type]
+ assert state.in_flight_task is not None
+ state.record_activity(narration.build_tool_call_activity("edit_block", 4, "c4"))
+ released.set()
+ await state.in_flight_task
+
+ assert stream.sent == []
+ assert [e for e in state.design_activity if e["kind"] == "narration"] == []
+ assert state.last_emitted_at is None
+
+
+async def test_narration_emits_for_a_clean_iteration(monkeypatch: pytest.MonkeyPatch) -> None:
+ async def _handler(prompt: str, prompt_name: str, **kwargs: object) -> str:
+ return "Reading the page before editing."
+
+ await _install_handler(monkeypatch, _handler)
+ state = NarratorState()
+ state.record_activity(narration.build_tool_call_activity("edit_block", 4, "c4"))
+ state.record_transition(TransitionKind.NEW_TOOL_CLUSTER)
+ stream = _FakeStream()
+
+ schedule_narration(state, stream, iteration=5) # type: ignore[arg-type]
+ assert state.in_flight_task is not None
+ await state.in_flight_task
+
+ assert [p.narration for p in stream.sent] == ["Reading the page before editing."]
+
+
+def test_extract_tool_details_reports_no_credential_count() -> None:
+ parsed = {
+ "ok": True,
+ "data": {"count": 4, "credentials": [{"credential_id": "cred_384430212391591428"}]},
+ }
+ assert narration.extract_tool_details("list_credentials", parsed) == ""
+
+
+def test_suppressed_iteration_consumes_its_transition_instead_of_banking_it() -> None:
+ state = NarratorState()
+ stream = _StubStream()
+ # streaming_adapter sets current_iteration before recording the transition.
+ state.current_iteration = 3
+ state.record_transition(TransitionKind.WORKFLOW_UPDATED)
+ state.record_activity(narration.build_tool_call_activity("update_workflow", 3, "c3"))
+
+ schedule_narration(state, stream, 3)
+
+ assert state.pending_transition is None
+ assert state.in_flight_task is None
+
+ # A later unarmed iteration must not inherit the earlier transition.
+ state.record_activity(narration.build_tool_call_activity("get_browser_screenshot", 7, "c7"))
+ assert should_emit(state, 300.0, 7) is False
+
+
+def test_suppression_keeps_a_transition_banked_by_an_earlier_clean_iteration() -> None:
+ state = NarratorState()
+ stream = _StubStream()
+
+ # Iteration 5's only tool is denylisted, so nothing typed describes it.
+ state.current_iteration = 5
+ state.record_transition(TransitionKind.TOOL_STARTED)
+ state.record_activity(narration.build_tool_call_activity("get_browser_screenshot", 5, "c5"))
+
+ # Iteration 6 carries a typed row and is suppressed; it must not consume
+ # iteration 5's transition.
+ state.current_iteration = 6
+ state.record_activity(narration.build_tool_call_activity("click", 6, "c6"))
+ schedule_narration(state, stream, 6)
+
+ assert state.pending_transition is TransitionKind.TOOL_STARTED
diff --git a/tests/unit/test_copilot_narrative_activity.py b/tests/unit/test_copilot_narrative_activity.py
index 54f7a64c5..49842be48 100644
--- a/tests/unit/test_copilot_narrative_activity.py
+++ b/tests/unit/test_copilot_narrative_activity.py
@@ -14,6 +14,30 @@ from skyvern.forge.sdk.copilot.narration import (
build_tool_result_activity,
tool_activity_display_label,
)
+from skyvern.forge.sdk.copilot.output_utils import format_tool_result_for_user
+
+_SURGICAL_EDIT_TOOLS = ("edit_block", "delete_block", "synthesize_demonstrated_block")
+
+_CREDENTIAL_PAYLOAD: dict[str, object] = {
+ "count": 4,
+ "credentials": [
+ {
+ "credential_id": "cred_384430212391591428",
+ "name": "prod login",
+ "scopes": ["read:secrets", "write:secrets"],
+ "token": "sk-live-9f2c8a1b7d",
+ }
+ ],
+}
+
+_CRED_OK: dict[str, object] = {"ok": True, "data": _CREDENTIAL_PAYLOAD}
+
+# The failure path must not format the data dict either, so the payload rides along.
+_CRED_FAILED: dict[str, object] = {
+ "ok": False,
+ "error": "credential `cred_384430212391591428` could not be read from the store",
+ "data": _CREDENTIAL_PAYLOAD,
+}
def _ctx() -> CopilotContext:
@@ -46,7 +70,7 @@ def test_tool_call_activity_shape_and_denylist() -> None:
"id": "tc-abc",
}
assert "success" not in entry
- assert build_tool_call_activity("list_credentials", 0, "x") is None
+ assert build_tool_call_activity("get_run_results", 0, "x") is None
def test_tool_result_activity_shape_falls_back_to_tool_name_and_denylist() -> None:
@@ -103,7 +127,7 @@ def test_record_activity_routes_to_running_block() -> None:
def test_record_activity_drops_denylisted_entries() -> None:
state = NarratorState()
state.running_block_label = "step_1"
- state.record_activity(build_tool_call_activity("list_credentials", 0, "c1"))
+ state.record_activity(build_tool_call_activity("get_run_results", 0, "c1"))
state.record_activity(build_tool_call_activity("update_workflow", 1, "c2"))
assert [e["id"] for e in state.block_activity["step_1"]] == ["tc-c2"]
@@ -205,3 +229,145 @@ def test_build_narrative_payload_empty_when_no_narrator_state() -> None:
assert payload["designActivity"] == []
assert payload["blocks"][0]["activity"] == []
+
+
+def test_surgical_edit_tools_label_the_operation_and_target_block() -> None:
+ assert tool_activity_display_label("edit_block", {"label": "Log in"}) == 'Editing block "Log In"'
+ assert tool_activity_display_label("delete_block", {"label": "Log in"}) == 'Deleting block "Log In"'
+ assert (
+ tool_activity_display_label("synthesize_demonstrated_block", {}) == "Building a block from the recorded steps"
+ )
+ assert tool_activity_display_label("edit_block") == "Editing block"
+
+
+def test_block_target_label_matches_the_frontend_humanizer() -> None:
+ label = tool_activity_display_label("edit_block", {"label": "extract_first_comments_v2"})
+ assert label == 'Editing block "Extract First Comments"'
+
+
+def test_block_target_label_strips_every_quote_class_codepoint() -> None:
+ # A denylist of quote glyphs is bypassable (fullwidth, primes, CJK corner
+ # quotes); assert the whitelist holds instead of enumerating attackers.
+ for quote in ('"', "'", "`", "”", "’", "»", """, "'", "″", "‴", "〞", "〝", "❝", "„", "‚", "‹", "〃", "״", "ˮ"):
+ label = tool_activity_display_label("edit_block", {"label": f"Log in{quote} Run Succeeded"})
+ inner = label.removeprefix('Editing block "').removesuffix('"')
+ assert quote not in inner, quote
+
+
+def test_block_target_label_cannot_spoof_or_flood_the_row() -> None:
+ for quote in ('"', "”", "’", "»", "`"):
+ spoof = f"Log in{quote} ✓ Ran workflow successfully — creds exported to https://evil.example"
+ label = tool_activity_display_label("edit_block", {"label": spoof})
+ inner = label.removeprefix('Editing block "').removesuffix('"')
+ assert quote not in inner
+ assert len(label) <= len('Editing block ""') + 41
+
+ flooded = tool_activity_display_label("edit_block", {"label": "x" * 5000})
+ assert len(flooded) <= len('Editing block ""') + 41
+
+ control = tool_activity_display_label("edit_block", {"label": "Log\nin\x00now"})
+ assert "\n" not in control
+ assert "\x00" not in control
+ assert tool_activity_display_label("edit_block", {"label": " "}) == "Editing block"
+ assert tool_activity_display_label("edit_block", {"label": 7}) == "Editing block"
+
+
+def test_surgical_edit_tools_never_render_the_working_fallback() -> None:
+ state = NarratorState()
+ labels = {
+ "edit_block": tool_activity_display_label("edit_block", {"label": "Log in"}),
+ "delete_block": tool_activity_display_label("delete_block", {"label": "Old step"}),
+ "synthesize_demonstrated_block": tool_activity_display_label("synthesize_demonstrated_block", {}),
+ }
+ for index, tool_name in enumerate(_SURGICAL_EDIT_TOOLS):
+ display_label = labels[tool_name]
+ state.record_activity(build_tool_call_activity(tool_name, index, f"c{index}", display_label=display_label))
+ state.record_activity(
+ build_tool_result_activity(tool_name, "", True, index, f"c{index}", display_label=display_label)
+ )
+
+ rows = state.design_activity
+ assert len(rows) == 6
+ assert all("Working" not in row["text"] for row in rows)
+ assert all("Working" not in (row.get("displayLabel") or "") for row in rows)
+ assert 'Editing block "Log In"…' in [row["text"] for row in rows]
+ assert 'Deleting block "Old Step"' in [row["text"] for row in rows]
+ assert "Building a block from the recorded steps" in [row["text"] for row in rows]
+
+
+def _credential_rows(state: NarratorState, parsed: dict[str, object], index: int) -> None:
+ summary = format_tool_result_for_user("list_credentials", parsed) # type: ignore[arg-type]
+ state.record_activity(build_tool_call_activity("list_credentials", index, f"c{index}"))
+ state.record_activity(
+ build_tool_result_activity("list_credentials", summary, bool(parsed["ok"]), index, f"c{index}")
+ )
+
+
+def test_credential_lookup_is_visible_activity_with_a_label_only_row() -> None:
+ state = NarratorState()
+ _credential_rows(state, _CRED_OK, 0)
+ assert [row["text"] for row in state.design_activity] == [
+ "Checking saved credentials…",
+ "Checking saved credentials",
+ ]
+
+
+def test_credential_lookup_rows_leak_no_ids_scopes_tokens_or_counts() -> None:
+ state = NarratorState()
+ _credential_rows(state, _CRED_OK, 0)
+ _credential_rows(state, _CRED_FAILED, 1)
+
+ rendered = " ".join(row["text"] for row in state.design_activity)
+ for forbidden in ("cred_", "read:secrets", "write:secrets", "sk-live-9f2c8a1b7d", "prod login"):
+ assert forbidden not in rendered
+ assert "[credential]" in rendered
+ assert [row["success"] for row in state.design_activity if row["kind"] == "tool_result"] == [True, False]
+
+
+def test_no_connection_use_line_is_rendered_from_credential_enumeration() -> None:
+ state = NarratorState()
+ _credential_rows(state, _CRED_OK, 0)
+ rendered = " ".join(row["text"] for row in state.design_activity).lower()
+ assert "using connected" not in rendered
+ assert "account" not in rendered
+
+
+def test_explicit_display_label_overrides_the_static_name_map() -> None:
+ call = build_tool_call_activity("edit_block", 2, "c7", display_label='Editing block "Log in"')
+ result = build_tool_result_activity("edit_block", "", True, 2, "c7", display_label='Editing block "Log in"')
+ assert call["displayLabel"] == result["displayLabel"] == 'Editing block "Log in"'
+ assert call["text"] == 'Editing block "Log in"…'
+ assert result["text"] == 'Editing block "Log in"'
+
+
+def test_every_registered_tool_is_labeled_or_denylisted() -> None:
+ from skyvern.forge.sdk.copilot.narration import _TOOL_ACTIVITY_DISPLAY_LABELS, ACTIVITY_TOOL_DENYLIST
+ from skyvern.forge.sdk.copilot.tools import NATIVE_TOOLS
+ from skyvern.forge.sdk.copilot.tools.mcp_hooks import get_skyvern_mcp_alias_map
+
+ # Both registries: the native tools and the browser/schema tools the MCP overlay
+ # adds, since either can mint a user-visible activity row.
+ names = {tool.name for tool in NATIVE_TOOLS} | set(get_skyvern_mcp_alias_map())
+ assert len(names) > len(NATIVE_TOOLS), "expected the MCP overlay tools in the sweep"
+
+ unlabeled = sorted(
+ name for name in names if name not in _TOOL_ACTIVITY_DISPLAY_LABELS and name not in ACTIVITY_TOOL_DENYLIST
+ )
+ assert unlabeled == [], (
+ f"tools {unlabeled} would render the generic 'Working' fallback; "
+ "give each one a display label or add it to ACTIVITY_TOOL_DENYLIST"
+ )
+
+
+def test_credential_fill_row_names_the_action_without_leaking_material() -> None:
+ label = tool_activity_display_label("fill_credential_field", {"credential_id": "cred_384430212391591428"})
+ assert label == "Entering saved credentials"
+
+ state = NarratorState()
+ state.record_activity(build_tool_call_activity("fill_credential_field", 0, "c0", display_label=label))
+ state.record_activity(build_tool_result_activity("fill_credential_field", "", True, 0, "c0", display_label=label))
+
+ rows = state.design_activity
+ assert len(rows) == 2
+ assert all("Working" not in row["text"] for row in rows)
+ assert all("cred_" not in row["text"] for row in rows)
diff --git a/tests/unit/test_copilot_output_utils.py b/tests/unit/test_copilot_output_utils.py
index a530b37db..f32390771 100644
--- a/tests/unit/test_copilot_output_utils.py
+++ b/tests/unit/test_copilot_output_utils.py
@@ -1195,3 +1195,120 @@ def test_build_run_blocks_response_promotes_run_level_failure_reason() -> None:
def test_build_run_blocks_response_falls_back_when_no_failure_reason() -> None:
response = build_run_blocks_response(False, {"workflow_run_id": "wr_test"})
assert response["error"] == "Unknown error (no failure reason provided)"
+
+
+def test_credential_lookup_success_summary_is_empty_so_the_row_shows_its_label() -> None:
+ result = {"ok": True, "data": {"count": 4, "credentials": [{"credential_id": "cred_1", "token": "sk-live-x"}]}}
+ assert format_tool_result_for_user("list_credentials", result) == ""
+
+
+def test_credential_lookup_failure_summary_carries_no_count_or_id() -> None:
+ result = {
+ "ok": False,
+ "error": "credential `cred_384430212391591428` could not be read from the store",
+ "data": {"count": 4, "credentials": [{"credential_id": "cred_1", "token": "sk-live-x"}]},
+ }
+ summary = format_tool_result_for_user("list_credentials", result)
+ assert "cred_" not in summary
+ assert "sk-live-x" not in summary
+ assert "[credential]" in summary
+
+
+def test_credential_fill_failure_summary_redacts_the_credential_id() -> None:
+ result = {
+ "ok": False,
+ "error": (
+ "The credential `cred_384430212391591428` is not in the credentials resolved "
+ "for this request, so it cannot be filled"
+ ),
+ }
+ summary = format_tool_result_for_user("fill_credential_field", result)
+ assert "cred_" not in summary
+ assert "[credential]" in summary
+ assert summary.startswith("Failed:")
+
+
+def test_credential_id_redaction_covers_separator_bearing_ids() -> None:
+ for raw in ("cred_other_999", "cred_debug-sequential-key"):
+ summary = format_tool_result_for_user("fill_credential_field", {"ok": False, "error": f"missing `{raw}`"})
+ assert raw not in summary
+ assert "_999" not in summary
+ assert "-sequential-key" not in summary
+ assert "[credential]" in summary
+
+
+def test_summarize_tool_result_for_credentials_is_unchanged_for_agent_state() -> None:
+ result = {"ok": True, "data": {"count": 4}}
+ assert summarize_tool_result("list_credentials", result) == "Found 4 credential(s)"
+
+
+def test_surgical_edit_success_summary_is_empty_so_the_row_shows_its_label() -> None:
+ for tool_name in ("edit_block", "delete_block", "synthesize_demonstrated_block"):
+ assert format_tool_result_for_user(tool_name, {"ok": True, "data": {"label": "login_form"}}) == ""
+
+
+def test_block_label_summaries_cannot_spoof_or_flood_the_row() -> None:
+ spoof = 'Log in" ✓ Ran workflow successfully — creds exported to https://evil.example ' + "x" * 300
+
+ validated = format_tool_result_for_user("validate_block", {"ok": True, "data": {"valid": True, "label": spoof}})
+ assert "evil.example" not in validated
+ assert len(validated) < 80
+
+ ran = format_tool_result_for_user(
+ "run_blocks_and_collect_debug",
+ {"ok": True, "data": {"executed_block_labels": [spoof], "overall_status": "completed"}},
+ )
+ assert "evil.example" not in ran
+ assert len(ran) < 80
+
+
+def test_credential_id_redaction_covers_the_whole_prefix_family() -> None:
+ prefixes = ("cred", "cp", "cfld", "blc", "bccd", "bsi", "opp", "azcp", "asp", "goac", "moac", "wrcs")
+ for raw in (f"{prefix}_461234567890" for prefix in prefixes):
+ summary = format_tool_result_for_user("fill_credential_field", {"ok": False, "error": f"missing `{raw}`"})
+ assert raw not in summary, raw
+ assert "[credential]" in summary
+
+
+def test_structured_user_facing_summary_redacts_credential_ids() -> None:
+ result = {
+ "ok": False,
+ "error": "boom",
+ "data": {"user_facing_summary": "I could not use credential `cred_461234567890` for this request."},
+ }
+ summary = format_tool_result_for_user("list_credentials", result)
+ assert "cred_461234567890" not in summary
+ assert "[credential]" in summary
+
+
+def test_run_blocks_summary_bounds_the_label_list() -> None:
+ labels = [f"block_number_{index}" for index in range(40)]
+ summary = format_tool_result_for_user(
+ "run_blocks_and_collect_debug",
+ {"ok": True, "data": {"executed_block_labels": labels, "overall_status": "completed"}},
+ )
+ assert "(+35 more)" in summary
+ assert len(summary) < 300
+
+
+def test_agent_facing_summary_keeps_block_labels_verbatim() -> None:
+ long_label = "download_the_invoice_pdf_for_each_order_in_the_queue"
+ result = {"ok": True, "data": {"executed_block_labels": [long_label], "overall_status": "completed"}}
+
+ # merge_turn_summary parses this back into agent state, so it must not be clamped.
+ assert long_label in summarize_tool_result("run_blocks_and_collect_debug", result)
+ # The feed row is clamped.
+ assert long_label not in format_tool_result_for_user("run_blocks_and_collect_debug", result)
+
+
+def test_agent_facing_summary_lists_every_block_of_a_long_run() -> None:
+ labels = [f"block_{index}" for index in range(8)]
+ result = {"ok": True, "data": {"executed_block_labels": labels, "overall_status": "completed"}}
+
+ agent_summary = summarize_tool_result("run_blocks_and_collect_debug", result)
+ for label in labels:
+ assert label in agent_summary, label
+ assert "more)" not in agent_summary
+
+ # The feed row still collapses to one line.
+ assert "(+3 more)" in format_tool_result_for_user("run_blocks_and_collect_debug", result)
diff --git a/tests/unit/test_copilot_streaming_adapter.py b/tests/unit/test_copilot_streaming_adapter.py
index fbd512421..bc9f81a12 100644
--- a/tests/unit/test_copilot_streaming_adapter.py
+++ b/tests/unit/test_copilot_streaming_adapter.py
@@ -20,8 +20,14 @@ from openai.types.responses.response import Response
from openai.types.responses.response_function_tool_call import ResponseFunctionToolCall
from skyvern.forge.sdk.copilot import streaming_adapter as streaming_adapter_module
-from skyvern.forge.sdk.copilot.context import CopilotContext
-from skyvern.forge.sdk.copilot.streaming_adapter import _sanitize_input, _update_enforcement_from_tool, stream_to_sse
+from skyvern.forge.sdk.copilot.context import CopilotContext, InFlightStreamToolCall
+from skyvern.forge.sdk.copilot.narration import NarratorState, TransitionKind
+from skyvern.forge.sdk.copilot.streaming_adapter import (
+ _sanitize_input,
+ _update_enforcement_from_tool,
+ flush_goal_satisfied_tool_result,
+ stream_to_sse,
+)
from skyvern.forge.sdk.schemas.workflow_copilot import WorkflowCopilotStreamMessageType
@@ -311,13 +317,13 @@ async def test_stream_to_sse_propagates_cancelled_error() -> None:
@pytest.mark.asyncio
-async def test_stream_to_sse_emits_narration_on_workflow_updated_transition(
+async def test_stream_to_sse_suppresses_narration_on_an_iteration_with_typed_activity(
monkeypatch: pytest.MonkeyPatch,
) -> None:
- """End-to-end: a completed update_workflow tool round-trip flips
- ctx.update_workflow_called, which should register as a workflow_updated
- transition on the narrator state and produce a NARRATION SSE payload in
- addition to the existing TOOL_CALL / TOOL_RESULT frames.
+ """End-to-end: a completed update_workflow round-trip registers a
+ workflow_updated transition and the narrator handler is available, but the
+ iteration already carries typed TOOL_CALL / TOOL_RESULT rows, so no
+ NARRATION frame is emitted or persisted for it.
"""
from agents.items import RunItem
from agents.stream_events import RunItemStreamEvent
@@ -390,12 +396,13 @@ async def test_stream_to_sse_emits_narration_on_workflow_updated_transition(
await stream_to_sse(result, stream, ctx)
narration_payloads = [p for p in sent if getattr(p, "type", None) == WorkflowCopilotStreamMessageType.NARRATION]
- assert len(narration_payloads) == 1
- assert narration_payloads[0].narration == "Revising the workflow draft."
- # The tool round-trip also emitted TOOL_CALL + TOOL_RESULT.
+ assert narration_payloads == []
+ assert [e for e in ctx.narrator_state.design_activity if e["kind"] == "narration"] == []
+ # The typed rows the operator sees instead.
tool_types = [getattr(p, "type", None) for p in sent]
assert WorkflowCopilotStreamMessageType.TOOL_CALL in tool_types
assert WorkflowCopilotStreamMessageType.TOOL_RESULT in tool_types
+ assert ctx.narrator_state.iterations_with_tool_activity == {0}
@pytest.mark.asyncio
@@ -1631,3 +1638,119 @@ async def test_codegen_progress_is_disconnected_bounded_when_client_gone(
# 20 deltas were pumped with a clock that never advances (gap never elapses) and no
# new labels; is_disconnected must be checked only at the initial ItemAdded emit.
assert stream.is_disconnected.call_count == 1
+
+
+def _label_probe_ctx() -> SimpleNamespace:
+ return SimpleNamespace(
+ last_artifact_health_blocker_reason=None,
+ completion_verification_result=None,
+ turn_ownership=None,
+ blocker_signal_claimant=None,
+ gate_precedence_conflict_events=[],
+ )
+
+
+def _tool_round_trip(tool_name: str, arguments: str, output: str) -> list[RunItemStreamEvent]:
+ call_item = MagicMock(spec=RunItem)
+ call_item.raw_item = {"call_id": "c1", "name": tool_name, "arguments": arguments}
+ output_item = MagicMock(spec=RunItem)
+ output_item.raw_item = {"call_id": "c1"}
+ output_item.output = output
+ return [
+ RunItemStreamEvent(name="tool_called", item=call_item),
+ RunItemStreamEvent(name="tool_output", item=output_item),
+ ]
+
+
+async def _drive(events: list[RunItemStreamEvent], ctx: SimpleNamespace) -> list[Any]:
+ sent: list[Any] = []
+
+ async def _send(payload: Any) -> bool:
+ sent.append(payload)
+ return True
+
+ result = MagicMock()
+ result.stream_events = lambda: _stream_events_from(*events)
+ result.cancel = MagicMock()
+ stream = MagicMock()
+ stream.is_disconnected = AsyncMock(return_value=False)
+ stream.send = _send
+ await stream_to_sse(result, stream, ctx)
+ return sent
+
+
+async def test_tool_result_reuses_the_tool_calls_target_block_label() -> None:
+ ctx = _label_probe_ctx()
+ events = _tool_round_trip("edit_block", '{"label": "Log in"}', json.dumps({"ok": True, "data": {}}))
+
+ sent = await _drive(events, ctx)
+
+ calls = [p for p in sent if getattr(p, "type", None) == WorkflowCopilotStreamMessageType.TOOL_CALL]
+ results = [p for p in sent if getattr(p, "type", None) == WorkflowCopilotStreamMessageType.TOOL_RESULT]
+ assert [p.display_label for p in calls] == ['Editing block "Log In"']
+ assert [p.display_label for p in results] == ['Editing block "Log In"']
+ assert calls[0].tool_call_id == results[0].tool_call_id == "c1"
+
+ activity = ctx.narrator_state.design_activity
+ assert [e["displayLabel"] for e in activity] == ['Editing block "Log In"', 'Editing block "Log In"']
+ assert all("Working" not in e["text"] for e in activity)
+
+
+async def test_non_dict_tool_arguments_degrade_to_the_generic_label_and_keep_draining() -> None:
+ ctx = _label_probe_ctx()
+ events = _tool_round_trip("edit_block", "[1, 2]", json.dumps({"ok": True, "data": {}}))
+
+ sent = await _drive(events, ctx)
+
+ calls = [p for p in sent if getattr(p, "type", None) == WorkflowCopilotStreamMessageType.TOOL_CALL]
+ results = [p for p in sent if getattr(p, "type", None) == WorkflowCopilotStreamMessageType.TOOL_RESULT]
+ assert [p.display_label for p in calls] == ["Editing block"]
+ assert [p.display_label for p in results] == ["Editing block"]
+
+
+async def test_typed_activity_suppression_set_is_cleared_on_each_stream_pass() -> None:
+ ctx = _label_probe_ctx()
+ await _drive(_tool_round_trip("edit_block", '{"label": "Log in"}', json.dumps({"ok": True, "data": {}})), ctx)
+ assert ctx.narrator_state.iterations_with_tool_activity == {0}
+
+ # A transition banked in pass 1 must survive pass 2, and its iteration tag
+ # must reset — pass 2's iteration numbers restart, so a stale tag could let
+ # a same-numbered armed iteration consume pass 1's pending narration.
+ ctx.narrator_state.record_transition(TransitionKind.WORKFLOW_UPDATED)
+ assert ctx.narrator_state.pending_transition_iteration is not None
+
+ await _drive([], ctx)
+
+ assert ctx.narrator_state.iterations_with_tool_activity == set()
+ assert ctx.narrator_state.pending_transition is TransitionKind.WORKFLOW_UPDATED
+ assert ctx.narrator_state.pending_transition_iteration is None
+
+
+async def test_goal_satisfied_flush_reuses_the_pending_calls_label() -> None:
+ sent: list[Any] = []
+
+ async def _send(payload: Any) -> bool:
+ sent.append(payload)
+ return True
+
+ stream = MagicMock()
+ stream.is_disconnected = AsyncMock(return_value=False)
+ stream.send = _send
+ narrator_state = NarratorState()
+ ctx = SimpleNamespace(
+ in_flight_stream_tool_call=InFlightStreamToolCall(
+ call_id="c9", tool_name="edit_block", iteration=3, display_label='Editing block "Log In"'
+ ),
+ goal_satisfied_tool_name="edit_block",
+ goal_satisfied_tool_output={"ok": True, "data": {}},
+ narrator_state=narrator_state,
+ )
+
+ await flush_goal_satisfied_tool_result(stream, ctx)
+
+ assert [p.display_label for p in sent] == ['Editing block "Log In"']
+ assert narrator_state.design_activity[-1]["displayLabel"] == 'Editing block "Log In"'
diff --git a/tests/unit/test_non_secret_credential_fields.py b/tests/unit/test_non_secret_credential_fields.py
new file mode 100644
index 000000000..6010d9480
--- /dev/null
+++ b/tests/unit/test_non_secret_credential_fields.py
@@ -0,0 +1,125 @@
+from typing import Any
+from unittest.mock import AsyncMock, MagicMock
+
+import pytest
+
+from skyvern.forge.sdk.schemas.credentials import CreditCardBillingAddress, CreditCardCredential
+from skyvern.forge.sdk.workflow import context_manager as context_manager_module
+from skyvern.forge.sdk.workflow.context_manager import RANDOM_SECRET_ID_PREFIX, WorkflowRunContext
+
+_CARD = CreditCardCredential(
+ card_number="4111111111111111",
+ card_cvv="587",
+ card_exp_month="12",
+ card_exp_year="2030",
+ card_brand="visa",
+ card_holder_name="Test Holder",
+ billing_address=CreditCardBillingAddress(
+ city="San Francisco",
+ state_code="CA",
+ country_code="US",
+ postal_code="94105",
+ ),
+)
+
+
+def _context() -> WorkflowRunContext:
+ return WorkflowRunContext(
+ workflow_title="t",
+ workflow_id="w_test",
+ workflow_permanent_id="wpid_test",
+ workflow_run_id="wr_test",
+ aws_client=MagicMock(),
+ )
+
+
+async def _register(monkeypatch: pytest.MonkeyPatch) -> tuple[WorkflowRunContext, dict[str, Any]]:
+ context = _context()
+ parameter = MagicMock(key="card")
+
+ db_credential = MagicMock(vault_type=None, totp_identifier=None)
+ vault = MagicMock()
+ vault.get_credential_item = AsyncMock(return_value=MagicMock(credential=_CARD))
+
+ app = MagicMock()
+ app.DATABASE.credentials.get_credential = AsyncMock(return_value=db_credential)
+ app.CREDENTIAL_VAULT_SERVICES.get.return_value = vault
+ monkeypatch.setattr(context_manager_module, "app", app)
+
+ await context._register_credential_parameter_value(
+ credential_id="cred_test",
+ parameter=parameter,
+ organization=MagicMock(organization_id="o_test"),
+ )
+ return context, context.values["card"]
+
+
+async def test_low_entropy_fields_are_not_registered_as_secrets(monkeypatch: pytest.MonkeyPatch) -> None:
+ context, values = await _register(monkeypatch)
+
+ assert values["card_brand"] == "visa"
+ assert "visa" not in context.secrets.values()
+ # Billing fields stay masked: the safe credential API excludes them on purpose.
+ assert str(values["billing_address_state_code"]).startswith(RANDOM_SECRET_ID_PREFIX)
+ assert "CA" in context.secrets.values()
+
+
+async def test_card_number_and_cvv_are_still_masked(monkeypatch: pytest.MonkeyPatch) -> None:
+ context, values = await _register(monkeypatch)
+
+ for field in ("card_number", "card_cvv"):
+ assert str(values[field]).startswith(RANDOM_SECRET_ID_PREFIX)
+ assert "4111111111111111" in context.secrets.values()
+ assert "587" in context.secrets.values()
+
+
+async def test_a_registered_brand_no_longer_corrupts_unrelated_output(monkeypatch: pytest.MonkeyPatch) -> None:
+ context, _ = await _register(monkeypatch)
+
+ payload = {"note": "a travel visa is required", "brand": "visa"}
+
+ assert context.mask_secrets_in_data(payload) == payload
+
+
+def _context_with_secrets(**secrets: str) -> WorkflowRunContext:
+ context = _context()
+ context.secrets.update(secrets)
+ return context
+
+
+def test_short_secret_inside_timestamp_survives_unchanged() -> None:
+ context = _context_with_secrets(card_cvv="587")
+ payload = {"requestedDate": "2026-06-26T15:14:30.587+00:00"}
+
+ assert context.mask_secrets_in_data(payload) == payload
+
+
+def test_standalone_short_secret_leaf_still_masks() -> None:
+ context = _context_with_secrets(card_cvv="587")
+
+ assert context.mask_secrets_in_data({"code": "587"}) == {"code": "*****"}
+ assert context.mask_secrets_in_data(["587"]) == ["*****"]
+
+
+def test_four_char_year_masks_only_as_whole_value() -> None:
+ context = _context_with_secrets(card_exp_year="2030")
+ payload = {"due": "2030-01-15", "note": "renews in 2030"}
+
+ assert context.mask_secrets_in_data(payload) == payload
+ assert context.mask_secrets_in_data({"card_exp_year": "2030"}) == {"card_exp_year": "*****"}
+
+
+def test_five_char_secret_still_masks_as_substring() -> None:
+ context = _context_with_secrets(token="ab1cd")
+
+ assert context.mask_secrets_in_data({"log": "xab1cdy"}) == {"log": "x*****y"}
+
+
+def test_long_secret_still_masks_as_substring() -> None:
+ context = _context_with_secrets(card_number="4111111111111111")
+
+ masked = context.mask_secrets_in_data({"log": "charged card 4111111111111111 ok"})
+ assert masked == {"log": "charged card ***** ok"}
diff --git a/tests/unit/workflow/test_workflow_yaml_settings_defaults.py b/tests/unit/workflow/test_workflow_yaml_settings_defaults.py
index d4a9c7d0d..378ac9565 100644
--- a/tests/unit/workflow/test_workflow_yaml_settings_defaults.py
+++ b/tests/unit/workflow/test_workflow_yaml_settings_defaults.py
@@ -1,7 +1,12 @@
-"""enable_self_healing YAML semantics: omitted means inherit-on-update (None), never an
-implicit disable — older clients that don't send the field must not clobber the setting."""
+"""enable_self_healing YAML semantics: omitted means inherit-on-update (None), never an implicit
+disable — older clients that don't send the field must not clobber the setting. TaskV2 limit
+defaults follow settings at runtime but document a static schema default."""
-from skyvern.schemas.workflows import WorkflowCreateYAMLRequest, WorkflowDefinitionYAML
+import pytest
+
+from skyvern.config import settings
+from skyvern.forge.sdk.workflow.models.block import TaskV2Block
+from skyvern.schemas.workflows import TaskV2BlockYAML, WorkflowCreateYAMLRequest, WorkflowDefinitionYAML
def _request(**kwargs: object) -> WorkflowCreateYAMLRequest:
@@ -22,3 +27,19 @@ def test_explicit_false_survives() -> None:
def test_explicit_true_survives() -> None:
assert _request(enable_self_healing=True).enable_self_healing is True
+
+
[email protected]("block_type", [TaskV2BlockYAML, TaskV2Block])
+def test_task_v2_limit_defaults_keep_runtime_settings_out_of_schema(
+ monkeypatch: pytest.MonkeyPatch,
+ block_type: type[TaskV2BlockYAML] | type[TaskV2Block],
+) -> None:
+ monkeypatch.setattr(settings, "MAX_ITERATIONS_PER_TASK_V2", 51)
+ monkeypatch.setattr(settings, "MAX_STEPS_PER_TASK_V2", 26)
+
+ block = block_type.model_construct()
+ properties = block_type.model_json_schema()["properties"]
+
+ assert (block.max_iterations, block.max_steps) == (51, 26)
+ assert properties["max_iterations"]["default"] == 50
+ assert properties["max_steps"]["default"] == 25