Skip to content

Automate browser based workflows with AI

fix(sky-8861): close CodeQL py/incomplete-url-substring-sanitization (#5665)

Shuchang Zheng595955614e63
diff --git a/alembic/versions/2026_04_25_2140-70b5f11e3655_checksum_in_artifacts_workflow_system_.py b/alembic/versions/2026_04_25_2140-70b5f11e3655_checksum_in_artifacts_workflow_system_.py
new file mode 100644
index 000000000..3c4712e3e
--- /dev/null
+++ b/alembic/versions/2026_04_25_2140-70b5f11e3655_checksum_in_artifacts_workflow_system_.py
@@ -0,0 +1,42 @@
+"""checksum in artifacts; workflow_system_prompt
+
+Revision ID: 70b5f11e3655
+Revises: d1474f2d1581
+Create Date: 2026-04-25 21:40:15.008563+00:00
+
+"""
+
+from typing import Sequence, Union
+
+import sqlalchemy as sa
+
+from alembic import op
+
+# revision identifiers, used by Alembic.
+revision: str = "70b5f11e3655"
+down_revision: Union[str, None] = "d1474f2d1581"
+branch_labels: Union[str, Sequence[str], None] = None
+depends_on: Union[str, Sequence[str], None] = None
+
+
+def upgrade() -> None:
+ # ### commands auto generated by Alembic - please adjust! ###
+ op.add_column("artifacts", sa.Column("checksum", sa.String(), nullable=True))
+ op.add_column("observer_cruises", sa.Column("workflow_system_prompt", sa.UnicodeText(), nullable=True))
+ op.add_column("tasks", sa.Column("workflow_system_prompt", sa.UnicodeText(), nullable=True))
+ op.add_column(
+ "workflow_runs",
+ sa.Column(
+ "ignore_inherited_workflow_system_prompt", sa.Boolean(), server_default=sa.text("false"), nullable=False
+ ),
+ )
+ # ### end Alembic commands ###
+
+
+def downgrade() -> None:
+ # ### commands auto generated by Alembic - please adjust! ###
+ op.drop_column("workflow_runs", "ignore_inherited_workflow_system_prompt")
+ op.drop_column("tasks", "workflow_system_prompt")
+ op.drop_column("observer_cruises", "workflow_system_prompt")
+ op.drop_column("artifacts", "checksum")
+ # ### end Alembic commands ###
diff --git a/skyvern-frontend/src/routes/workflows/components/WorkflowVisualComparisonDrawer.tsx b/skyvern-frontend/src/routes/workflows/components/WorkflowVisualComparisonDrawer.tsx
index 604c88fbc..f7f35e9f7 100644
--- a/skyvern-frontend/src/routes/workflows/components/WorkflowVisualComparisonDrawer.tsx
+++ b/skyvern-frontend/src/routes/workflows/components/WorkflowVisualComparisonDrawer.tsx
@@ -131,6 +131,8 @@ function getWorkflowElements(version: WorkflowVersion) {
runSequentially: version.run_sequentially ?? false,
sequentialKey: version.sequential_key ?? null,
finallyBlockLabel: version.workflow_definition?.finally_block_label ?? null,
+ workflowSystemPrompt:
+ version.workflow_definition?.workflow_system_prompt ?? null,
};
return getElements(
diff --git a/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx b/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx
index febd3014a..203f68625 100644
--- a/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx
+++ b/skyvern-frontend/src/routes/workflows/debugger/Debugger.tsx
@@ -90,6 +90,8 @@ function Debugger() {
sequentialKey: workflow.sequential_key ?? null,
finallyBlockLabel:
workflow.workflow_definition?.finally_block_label ?? null,
+ workflowSystemPrompt:
+ workflow.workflow_definition?.workflow_system_prompt ?? null,
};
const elements = getElements(blocksToRender, settings, true);
diff --git a/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx b/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx
index 91f25ed0b..aa5f0bb23 100644
--- a/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/WorkflowEditor.tsx
@@ -82,6 +82,8 @@ function WorkflowEditor() {
sequentialKey: workflow.sequential_key ?? null,
finallyBlockLabel:
workflow.workflow_definition?.finally_block_label ?? null,
+ workflowSystemPrompt:
+ workflow.workflow_definition?.workflow_system_prompt ?? null,
};
const elements = getElements(blocksToRender, settings, !isGlobalWorkflow);
diff --git a/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx b/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx
index 259c63ed2..f884d32ed 100644
--- a/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/Workspace.tsx
@@ -1206,6 +1206,8 @@ function Workspace({
sequentialKey: workflowData.sequential_key ?? null,
finallyBlockLabel:
workflowData.workflow_definition?.finally_block_label ?? null,
+ workflowSystemPrompt:
+ workflowData.workflow_definition?.workflow_system_prompt ?? null,
};
const elements = getElements(
@@ -1254,6 +1256,8 @@ function Workspace({
sequentialKey: selectedVersion.sequential_key ?? null,
finallyBlockLabel:
selectedVersion.workflow_definition?.finally_block_label ?? null,
+ workflowSystemPrompt:
+ selectedVersion.workflow_definition?.workflow_system_prompt ?? null,
};
const elements = getElements(
@@ -2006,6 +2010,8 @@ function Workspace({
blocks: saveData.blocks,
finally_block_label:
saveData.settings.finallyBlockLabel ?? undefined,
+ workflow_system_prompt:
+ saveData.settings.workflowSystemPrompt ?? undefined,
});
// Convert current workflow definition YAML to blocks
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/ActionNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/ActionNode.tsx
index ce1b6aea9..752b8574d 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/ActionNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/ActionNode.tsx
@@ -38,6 +38,7 @@ import { useWorkflowRunQuery } from "@/routes/workflows/hooks/useWorkflowRunQuer
import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { DisableCache } from "../DisableCache";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
import { BlockExecutionOptions } from "../components/BlockExecutionOptions";
const urlTooltip =
@@ -270,6 +271,17 @@ function ActionNode({ id, data, type }: NodeProps<ActionNode>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
<Separator />
<div className="flex items-center justify-between">
<div className="flex gap-2">
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/types.ts
index b804ff6eb..e30b31e85 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/ActionNode/types.ts
@@ -36,6 +36,7 @@ export const actionNodeDefaultData: ActionNodeData = {
disableCache: false,
engine: RunEngine.SkyvernV1,
model: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isActionNode(node: Node): node is ActionNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/ExtractionNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/ExtractionNode.tsx
index f45b3df1c..11b877598 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/ExtractionNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/ExtractionNode.tsx
@@ -38,6 +38,7 @@ import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { useRerender } from "@/hooks/useRerender";
import { DisableCache } from "../DisableCache";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
import { BlockExecutionOptions } from "../components/BlockExecutionOptions";
import { AI_IMPROVE_CONFIGS } from "../../constants";
@@ -242,6 +243,17 @@ function ExtractionNode({ id, data, type }: NodeProps<ExtractionNode>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
</div>
</AccordionContent>
</AccordionItem>
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/types.ts
index 4b3c95147..71e619fa6 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/ExtractionNode/types.ts
@@ -30,6 +30,7 @@ export const extractionNodeDefaultData: ExtractionNodeData = {
disableCache: false,
engine: RunEngine.SkyvernV1,
model: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isExtractionNode(node: Node): node is ExtractionNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/FileDownloadNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/FileDownloadNode.tsx
index 9d4b75bbf..c96396784 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/FileDownloadNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/FileDownloadNode.tsx
@@ -38,6 +38,7 @@ import { useRerender } from "@/hooks/useRerender";
import { BROWSER_DOWNLOAD_TIMEOUT_SECONDS } from "@/api/types";
import { DisableCache } from "../DisableCache";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
import { BlockExecutionOptions } from "../components/BlockExecutionOptions";
import { AI_IMPROVE_CONFIGS } from "../../constants";
@@ -307,6 +308,17 @@ function FileDownloadNode({ id, data }: NodeProps<FileDownloadNode>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
<Separator />
<div className="space-y-2">
<div className="flex gap-2">
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/types.ts
index ce486f2d2..192a972a1 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/FileDownloadNode/types.ts
@@ -38,6 +38,7 @@ export const fileDownloadNodeDefaultData: FileDownloadNodeData = {
engine: RunEngine.SkyvernV1,
model: null,
downloadTimeout: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isFileDownloadNode(node: Node): node is FileDownloadNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/FileParserNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/FileParserNode.tsx
index 860bd99e7..87323718a 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/FileParserNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/FileParserNode.tsx
@@ -22,6 +22,8 @@ import { useWorkflowRunQuery } from "@/routes/workflows/hooks/useWorkflowRunQuer
import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { ModelSelector } from "@/components/ModelSelector";
import { useRecordingStore } from "@/store/useRecordingStore";
+import { Separator } from "@/components/ui/separator";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
const FILE_TYPE_OPTIONS: Array<{ value: FileParserFileType; label: string }> = [
{ value: "auto_detect", label: "Auto detect" },
@@ -188,6 +190,18 @@ function FileParserNode({ id, data }: NodeProps<FileParserNode>) {
update({ model: value });
}}
/>
+ <Separator />
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
</div>
</div>
</div>
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/types.ts
index 026f0d01c..ecbfeddd0 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/FileParserNode/types.ts
@@ -32,6 +32,7 @@ export const fileParserNodeDefaultData: FileParserNodeData = {
continueOnFailure: false,
jsonSchema: "null",
model: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isFileParserNode(node: AppNode): node is FileParserNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/IgnoreWorkflowSystemPrompt.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/IgnoreWorkflowSystemPrompt.tsx
new file mode 100644
index 000000000..844ea7b95
--- /dev/null
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/IgnoreWorkflowSystemPrompt.tsx
@@ -0,0 +1,37 @@
+import { HelpTooltip } from "@/components/HelpTooltip";
+import { Label } from "@/components/ui/label";
+import { Switch } from "@/components/ui/switch";
+
+function IgnoreWorkflowSystemPrompt({
+ ignoreWorkflowSystemPrompt,
+ editable,
+ onIgnoreWorkflowSystemPromptChange,
+}: {
+ ignoreWorkflowSystemPrompt: boolean;
+ editable: boolean;
+ onIgnoreWorkflowSystemPromptChange: (value: boolean) => void;
+}) {
+ return (
+ <div className="flex items-center justify-between gap-2">
+ <div className="flex min-w-0 items-center gap-2">
+ <Label className="truncate text-xs font-normal text-slate-300">
+ Ignore System Prompt
+ </Label>
+ <HelpTooltip content="When checked, this block ignores the workflow-level system prompt. Only relevant when the workflow has a workflow system prompt set." />
+ </div>
+ <div className="w-52 shrink-0">
+ <Switch
+ checked={ignoreWorkflowSystemPrompt}
+ onCheckedChange={(checked) => {
+ if (!editable) {
+ return;
+ }
+ onIgnoreWorkflowSystemPromptChange(checked);
+ }}
+ />
+ </div>
+ </div>
+ );
+}
+
+export { IgnoreWorkflowSystemPrompt };
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/LoginNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/LoginNode.tsx
index 654a80548..e05f04c0d 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/LoginNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/LoginNode.tsx
@@ -38,6 +38,7 @@ import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { useRerender } from "@/hooks/useRerender";
import { DisableCache } from "../DisableCache";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
import { BlockExecutionOptions } from "../components/BlockExecutionOptions";
import { AI_IMPROVE_CONFIGS } from "../../constants";
@@ -318,6 +319,17 @@ function LoginNode({ id, data, type }: NodeProps<LoginNode>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
<Separator />
<div className="space-y-2">
<div className="flex gap-2">
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/types.ts
index a6e874b1f..fd0a52c1c 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/LoginNode/types.ts
@@ -39,6 +39,7 @@ export const loginNodeDefaultData: LoginNodeData = {
terminateCriterion: "",
engine: RunEngine.SkyvernV1,
model: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isLoginNode(node: Node): node is LoginNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/NavigationNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/NavigationNode.tsx
index 3c6997168..e2f918e20 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/NavigationNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/NavigationNode.tsx
@@ -40,6 +40,7 @@ import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { RunEngine } from "@/api/types";
import { DisableCache } from "../DisableCache";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
import { BlockExecutionOptions } from "../components/BlockExecutionOptions";
import { AI_IMPROVE_CONFIGS } from "../../constants";
@@ -200,6 +201,17 @@ function NavigationNode({ id, data, type }: NodeProps<NavigationNode>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
<Separator />
<div className="space-y-2">
<div className="flex gap-2">
@@ -465,6 +477,17 @@ function NavigationNode({ id, data, type }: NodeProps<NavigationNode>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
<Separator />
<div className="flex items-center justify-between">
<div className="flex gap-2">
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/types.ts
index 9894be52f..11eb7afbb 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/NavigationNode/types.ts
@@ -49,6 +49,7 @@ export const navigationNodeDefaultData: NavigationNodeData = {
continueOnFailure: false,
disableCache: false,
includeActionHistoryInVerification: false,
+ ignoreWorkflowSystemPrompt: false,
// V2-specific fields
prompt: "",
maxSteps: MAX_STEPS_DEFAULT,
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/PDFParserNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/PDFParserNode.tsx
index 5530acd32..918acc25e 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/PDFParserNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/PDFParserNode.tsx
@@ -15,6 +15,8 @@ import { useWorkflowRunQuery } from "@/routes/workflows/hooks/useWorkflowRunQuer
import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { ModelSelector } from "@/components/ModelSelector";
import { useRecordingStore } from "@/store/useRecordingStore";
+import { Separator } from "@/components/ui/separator";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
function PDFParserNode({ id, data }: NodeProps<PDFParserNode>) {
const { editable, label } = data;
@@ -103,6 +105,18 @@ function PDFParserNode({ id, data }: NodeProps<PDFParserNode>) {
update({ model: value });
}}
/>
+ <Separator />
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
</div>
</div>
</div>
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/types.ts
index bcc4445eb..1ef39f323 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/PDFParserNode/types.ts
@@ -22,6 +22,7 @@ export const pdfParserNodeDefaultData: PDFParserNodeData = {
continueOnFailure: false,
jsonSchema: "null",
model: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isPdfParserNode(node: AppNode): node is PDFParserNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/StartNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/StartNode.tsx
index 2e497d454..492c756c5 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/StartNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/StartNode.tsx
@@ -74,6 +74,7 @@ interface StartSettings {
maxScreenshotScrollingTimes: number | null;
extraHttpHeaders: string | Record<string, unknown> | null;
finallyBlockLabel: string | null;
+ workflowSystemPrompt: string | null;
}
function StartNode({ id, data, parentId }: NodeProps<StartNode>) {
@@ -160,6 +161,9 @@ function StartNode({ id, data, parentId }: NodeProps<StartNode>) {
finallyBlockLabel: data.withWorkflowSettings
? data.finallyBlockLabel
: null,
+ workflowSystemPrompt: data.withWorkflowSettings
+ ? (data.workflowSystemPrompt ?? null)
+ : null,
};
};
@@ -539,6 +543,23 @@ function StartNode({ id, data, parentId }: NodeProps<StartNode>) {
</SelectContent>
</Select>
</div>
+ <div className="space-y-2">
+ <div className="flex items-center gap-2">
+ <Label>Workflow System Prompt</Label>
+ <HelpTooltip content="Applied to every LLM call in this workflow, including any sub-agents." />
+ </div>
+ <WorkflowBlockInputTextarea
+ nodeId={id}
+ onChange={(value) => {
+ update({
+ workflowSystemPrompt: value.length ? value : null,
+ });
+ }}
+ value={data.workflowSystemPrompt ?? ""}
+ placeholder="e.g. Format all dates as YYYY-MM-DD and all currency values as USD with two decimals."
+ className="nopan text-xs"
+ />
+ </div>
</div>
</AccordionContent>
</AccordionItem>
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 eb64838f9..61c95e2e9 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/StartNode/types.ts
@@ -19,6 +19,7 @@ export type WorkflowStartNodeData = {
runSequentially: boolean;
sequentialKey: string | null;
finallyBlockLabel: string | null;
+ workflowSystemPrompt: string | null;
label: "__start_block__";
showCode: boolean;
};
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/TaskNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/TaskNode.tsx
index 85ac49597..7c0eda766 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/TaskNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/TaskNode.tsx
@@ -40,6 +40,7 @@ import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { useRerender } from "@/hooks/useRerender";
import { DisableCache } from "../DisableCache";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
import { BlockExecutionOptions } from "../components/BlockExecutionOptions";
function TaskNode({ id, data, type }: NodeProps<TaskNode>) {
@@ -355,6 +356,17 @@ function TaskNode({ id, data, type }: NodeProps<TaskNode>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
<Separator />
<div className="flex items-center justify-between">
<div className="flex gap-2">
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/types.ts
index 9c1f47874..8390ce571 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/TaskNode/types.ts
@@ -47,6 +47,7 @@ export const taskNodeDefaultData: TaskNodeData = {
includeActionHistoryInVerification: false,
engine: RunEngine.SkyvernV1,
model: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isTaskNode(node: Node): node is TaskNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/Taskv2Node.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/Taskv2Node.tsx
index 5c106e9a8..d9a9625c7 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/Taskv2Node.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/Taskv2Node.tsx
@@ -29,6 +29,7 @@ import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { useBlockScriptStore } from "@/store/BlockScriptStore";
import { DisableCache } from "../DisableCache";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
function Taskv2Node({ id, data, type }: NodeProps<Taskv2Node>) {
const { editable, label } = data;
@@ -164,6 +165,17 @@ function Taskv2Node({ id, data, type }: NodeProps<Taskv2Node>) {
update({ disableCache });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
<Separator />
<div className="space-y-2">
<div className="flex gap-2">
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/types.ts
index b3f477c53..d3d1f16d9 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/Taskv2Node/types.ts
@@ -30,6 +30,7 @@ export const taskv2NodeDefaultData: Taskv2NodeData = {
disableCache: false,
model: null,
maxScreenshotScrolls: null,
+ ignoreWorkflowSystemPrompt: false,
};
export function isTaskV2Node(node: Node): node is Taskv2Node {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/TextPromptNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/TextPromptNode.tsx
index c92b69a83..04136a0ed 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/TextPromptNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/TextPromptNode.tsx
@@ -17,6 +17,7 @@ import { useWorkflowRunQuery } from "@/routes/workflows/hooks/useWorkflowRunQuer
import { useUpdate } from "@/routes/workflows/editor/useUpdate";
import { AI_IMPROVE_CONFIGS } from "../../constants";
import { useRecordingStore } from "@/store/useRecordingStore";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
function TextPromptNode({ id, data }: NodeProps<TextPromptNode>) {
const { editable, label } = data;
@@ -108,6 +109,14 @@ function TextPromptNode({ id, data }: NodeProps<TextPromptNode>) {
}}
suggestionContext={{ current_schema: data.jsonSchema }}
/>
+ <Separator />
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={data.ignoreWorkflowSystemPrompt ?? false}
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(ignoreWorkflowSystemPrompt) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
</div>
</div>
);
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/types.ts
index 8733e7f11..671cc8a74 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/TextPromptNode/types.ts
@@ -20,6 +20,7 @@ export const textPromptNodeDefaultData: TextPromptNodeData = {
continueOnFailure: false,
parameterKeys: [],
model: null,
+ ignoreWorkflowSystemPrompt: false,
} as const;
export function isTextPromptNode(node: AppNode): node is TextPromptNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/WorkflowTriggerNode.tsx b/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/WorkflowTriggerNode.tsx
index 10df329a6..33845fb57 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/WorkflowTriggerNode.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/WorkflowTriggerNode.tsx
@@ -34,6 +34,7 @@ import { useRecordingStore } from "@/store/useRecordingStore";
import { useWorkflowHasChangesStore } from "@/store/WorkflowHasChangesStore";
import { cn } from "@/util/utils";
import { BlockExecutionOptions } from "../components/BlockExecutionOptions";
+import { IgnoreWorkflowSystemPrompt } from "../IgnoreWorkflowSystemPrompt";
import { BrowserSessionSelector } from "./BrowserSessionSelector";
import {
PARENT_SESSION_VALUE,
@@ -337,6 +338,17 @@ function WorkflowTriggerNode({ id, data }: NodeProps<WorkflowTriggerNodeType>) {
update({ parameterKeys });
}}
/>
+ <IgnoreWorkflowSystemPrompt
+ ignoreWorkflowSystemPrompt={
+ data.ignoreWorkflowSystemPrompt ?? false
+ }
+ editable={editable}
+ onIgnoreWorkflowSystemPromptChange={(
+ ignoreWorkflowSystemPrompt,
+ ) => {
+ update({ ignoreWorkflowSystemPrompt });
+ }}
+ />
</AccordionContent>
</AccordionItem>
</Accordion>
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/types.ts
index c1160bdae..31595b5a7 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/WorkflowTriggerNode/types.ts
@@ -30,6 +30,7 @@ export const workflowTriggerNodeDefaultData: WorkflowTriggerNodeData = {
browserSessionId: "",
useParentBrowserSession: true,
parameterKeys: [],
+ ignoreWorkflowSystemPrompt: false,
};
export function isWorkflowTriggerNode(node: Node): node is WorkflowTriggerNode {
diff --git a/skyvern-frontend/src/routes/workflows/editor/nodes/types.ts b/skyvern-frontend/src/routes/workflows/editor/nodes/types.ts
index e2d38f4c7..39949e368 100644
--- a/skyvern-frontend/src/routes/workflows/editor/nodes/types.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/nodes/types.ts
@@ -12,6 +12,7 @@ export type NodeBaseData = {
model: WorkflowModel | null;
showCode?: boolean;
comparisonColor?: string;
+ ignoreWorkflowSystemPrompt?: boolean;
/**
* Optional metadata used for conditional branches.
* These values are only set on nodes that live within a conditional block.
diff --git a/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx b/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx
index 064914af2..642572c67 100644
--- a/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx
+++ b/skyvern-frontend/src/routes/workflows/editor/panels/WorkflowComparisonPanel.tsx
@@ -165,6 +165,8 @@ function getWorkflowElements(version: WorkflowVersion) {
runSequentially: version.run_sequentially ?? false,
sequentialKey: version.sequential_key ?? null,
finallyBlockLabel: version.workflow_definition?.finally_block_label ?? null,
+ workflowSystemPrompt:
+ version.workflow_definition?.workflow_system_prompt ?? null,
};
// Deep clone the blocks to ensure complete isolation from main editor
diff --git a/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts b/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts
index 732dcdc5a..b03f5bd7d 100644
--- a/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts
+++ b/skyvern-frontend/src/routes/workflows/editor/workflowEditorUtils.ts
@@ -548,6 +548,7 @@ function convertToNode(
nextLoopOnFailure: block.next_loop_on_failure,
editable,
model: block.model,
+ ignoreWorkflowSystemPrompt: block.ignore_workflow_system_prompt ?? false,
};
switch (block.block_type) {
case "conditional": {
@@ -1595,6 +1596,7 @@ function getElements(
runSequentially: settings.runSequentially,
sequentialKey: settings.sequentialKey,
finallyBlockLabel: settings.finallyBlockLabel ?? null,
+ workflowSystemPrompt: settings.workflowSystemPrompt ?? null,
}),
);
@@ -2310,6 +2312,8 @@ function getWorkflowBlock(
next_loop_on_failure: node.data.nextLoopOnFailure,
model: node.data.model,
next_block_label: nextBlockLabel,
+ ignore_workflow_system_prompt:
+ node.data.ignoreWorkflowSystemPrompt ?? false,
};
switch (node.type) {
case "task": {
@@ -2924,6 +2928,7 @@ function getWorkflowSettings(nodes: Array<AppNode>): WorkflowSettings {
runSequentially: false,
sequentialKey: null,
finallyBlockLabel: null,
+ workflowSystemPrompt: null,
};
const startNodes = nodes.filter(isStartNode);
const startNodeWithWorkflowSettings = startNodes.find(
@@ -2951,6 +2956,7 @@ function getWorkflowSettings(nodes: Array<AppNode>): WorkflowSettings {
runSequentially: data.runSequentially,
sequentialKey: data.sequentialKey,
finallyBlockLabel: data.finallyBlockLabel ?? null,
+ workflowSystemPrompt: data.workflowSystemPrompt ?? null,
};
}
return defaultSettings;
@@ -3804,6 +3810,8 @@ function convertBlocksToBlockYAML(
continue_on_failure: block.continue_on_failure,
next_loop_on_failure: block.next_loop_on_failure,
next_block_label: block.next_block_label,
+ ignore_workflow_system_prompt:
+ block.ignore_workflow_system_prompt ?? false,
};
switch (block.block_type) {
case "task": {
@@ -4198,6 +4206,8 @@ function convert(workflow: WorkflowApiResponse): WorkflowCreateYAMLRequest {
parameters: convertParametersToParameterYAML(userParameters),
blocks: convertBlocksToBlockYAML(workflow.workflow_definition.blocks),
finally_block_label: workflow.workflow_definition.finally_block_label,
+ workflow_system_prompt:
+ workflow.workflow_definition.workflow_system_prompt,
},
is_saved_task: workflow.is_saved_task,
status: workflow.status,
diff --git a/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts b/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts
index 16ab4709e..676aad32a 100644
--- a/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts
+++ b/skyvern-frontend/src/routes/workflows/types/workflowTypes.ts
@@ -292,6 +292,7 @@ export type WorkflowBlockBase = {
next_loop_on_failure?: boolean;
model: WorkflowModel | null;
next_block_label?: string | null;
+ ignore_workflow_system_prompt?: boolean;
};
export const BranchCriteriaTypes = {
@@ -603,6 +604,7 @@ export type WorkflowDefinition = {
parameters: Array<Parameter>;
blocks: Array<WorkflowBlock>;
finally_block_label?: string | null;
+ workflow_system_prompt?: string | null;
};
export type WorkflowApiResponse = {
@@ -652,6 +654,7 @@ export type WorkflowSettings = {
runSequentially: boolean;
sequentialKey: string | null;
finallyBlockLabel: string | null;
+ workflowSystemPrompt: string | null;
};
export type WorkflowModel = JsonObjectExtendable<{ model_name: string }>;
diff --git a/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts b/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts
index 72d91f417..54d3c1d27 100644
--- a/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts
+++ b/skyvern-frontend/src/routes/workflows/types/workflowYamlTypes.ts
@@ -29,6 +29,7 @@ export type WorkflowDefinitionYAML = {
parameters: Array<ParameterYAML>;
blocks: Array<BlockYAML>;
finally_block_label?: string | null;
+ workflow_system_prompt?: string | null;
};
export type ParameterYAML =
@@ -154,6 +155,7 @@ export type BlockYAMLBase = {
continue_on_failure?: boolean;
next_loop_on_failure?: boolean;
next_block_label?: string | null;
+ ignore_workflow_system_prompt?: boolean;
};
export type TaskBlockYAML = BlockYAMLBase & {
diff --git a/skyvern-frontend/src/store/WorkflowHasChangesStore.ts b/skyvern-frontend/src/store/WorkflowHasChangesStore.ts
index d344e7651..b13db6cf0 100644
--- a/skyvern-frontend/src/store/WorkflowHasChangesStore.ts
+++ b/skyvern-frontend/src/store/WorkflowHasChangesStore.ts
@@ -164,6 +164,8 @@ const useWorkflowSave = (opts?: WorkflowSaveOpts) => {
parameters: saveData.parameters,
blocks: saveData.blocks,
finally_block_label: saveData.settings.finallyBlockLabel ?? undefined,
+ workflow_system_prompt:
+ saveData.settings.workflowSystemPrompt ?? undefined,
},
is_saved_task: saveData.workflow.is_saved_task,
status: opts?.status ?? saveData.workflow.status,
diff --git a/skyvern/core/script_generations/real_skyvern_page_ai.py b/skyvern/core/script_generations/real_skyvern_page_ai.py
index 3c8b4d847..a65d56cc4 100644
--- a/skyvern/core/script_generations/real_skyvern_page_ai.py
+++ b/skyvern/core/script_generations/real_skyvern_page_ai.py
@@ -11,7 +11,8 @@ from playwright.async_api import Page
from skyvern.config import settings
from skyvern.constants import SKYVERN_PAGE_MAX_SCRAPING_RETRIES, SPECIAL_FIELD_VERIFICATION_CODE
-from skyvern.core.script_generations.skyvern_page_ai import SkyvernPageAi
+from skyvern.core.script_generations.skyvern_page_ai import SYSTEM_PROMPT_UNSET, SkyvernPageAi
+from skyvern.exceptions import WorkflowRunContextNotInitialized
from skyvern.forge import app
from skyvern.forge.prompts import prompt_engine
from skyvern.forge.sdk.api.files import validate_download_url
@@ -903,6 +904,7 @@ class RealSkyvernPageAi(SkyvernPageAi):
data: str | dict[str, Any] | None = None,
skip_refresh: bool = False,
include_extracted_text: bool = True,
+ system_prompt: str | None | Any = SYSTEM_PROMPT_UNSET,
) -> dict[str, Any] | list | str | None:
"""Extract information from the page using AI."""
@@ -915,6 +917,42 @@ class RealSkyvernPageAi(SkyvernPageAi):
prompt = _render_template_with_label(prompt, label=self.current_label)
local_datetime_str = datetime.now(tz_info).isoformat()
+ # Resolve the effective workflow_system_prompt for this run. Order:
+ # 1. Caller-passed value wins (including None — "block opted out,
+ # send no system prompt").
+ # 2. Block-recorded value from ``WorkflowRunContext``, populated by
+ # ``Block._apply_workflow_system_prompt`` in both the agent path
+ # (``format_potential_template_parameters``) and the script path
+ # (``_execute_single_block`` before ``exec``). Using the recorded
+ # value makes the block the single source of truth for the
+ # opt-out + resolved-string decision — script-path extractions
+ # hash to the same cache key and send the same LLM input the
+ # agent path would. A recorded ``None`` is a real opt-out, not a
+ # miss (SKY-9147).
+ # 3. Fall back to the run-wide effective prompt for non-block
+ # callers (standalone scripts, sdk routes, etc.) that never set
+ # ``current_label`` and never went through a Block.
+ workflow_system_prompt: str | None
+ if system_prompt is not SYSTEM_PROMPT_UNSET:
+ workflow_system_prompt = cast("str | None", system_prompt)
+ else:
+ workflow_system_prompt = None
+ workflow_run_context_for_prompt = None
+ if context and context.workflow_run_id:
+ try:
+ workflow_run_context_for_prompt = app.WORKFLOW_CONTEXT_MANAGER.get_workflow_run_context(
+ context.workflow_run_id
+ )
+ except WorkflowRunContextNotInitialized:
+ workflow_run_context_for_prompt = None
+
+ if workflow_run_context_for_prompt is not None:
+ recorded, value = workflow_run_context_for_prompt.get_block_workflow_system_prompt(self.current_label)
+ if recorded:
+ workflow_system_prompt = value
+ else:
+ workflow_system_prompt = workflow_run_context_for_prompt.resolve_effective_workflow_system_prompt()
+
# Render the prompt FIRST so the cache key hashes the exact string
# that will be sent to the LLM (captures economy-tree swaps and 2/3
# truncation inside load_prompt_with_elements).
@@ -981,6 +1019,7 @@ class RealSkyvernPageAi(SkyvernPageAi):
extracted_information_schema=post_ceiling_kwargs["extracted_information_schema"],
error_code_mapping=error_code_mapping_str,
llm_key=None,
+ workflow_system_prompt=workflow_system_prompt,
)
lookup_result = extraction_cache.lookup(workflow_run_id, cache_key)
except Exception:
@@ -1042,6 +1081,7 @@ class RealSkyvernPageAi(SkyvernPageAi):
screenshots=self.scraped_page.screenshots,
prompt_name="extract-information",
force_dict=False,
+ system_prompt=workflow_system_prompt,
)
# Validate and fill missing fields based on schema
diff --git a/skyvern/core/script_generations/skyvern_page.py b/skyvern/core/script_generations/skyvern_page.py
index fbf6dde99..af9947b20 100644
--- a/skyvern/core/script_generations/skyvern_page.py
+++ b/skyvern/core/script_generations/skyvern_page.py
@@ -1323,6 +1323,9 @@ class SkyvernPage(Page):
"""
data = kwargs.pop("data", None)
skip_refresh = kwargs.pop("skip_refresh", False)
+ extra_kwargs: dict[str, Any] = {}
+ if "system_prompt" in kwargs:
+ extra_kwargs["system_prompt"] = kwargs.pop("system_prompt")
return await self._ai.ai_extract(
prompt=prompt,
schema=schema,
@@ -1330,6 +1333,7 @@ class SkyvernPage(Page):
intention=intention,
data=data,
skip_refresh=skip_refresh,
+ **extra_kwargs,
)
async def validate(
diff --git a/skyvern/core/script_generations/skyvern_page_ai.py b/skyvern/core/script_generations/skyvern_page_ai.py
index 716b35eea..b839fe0bd 100644
--- a/skyvern/core/script_generations/skyvern_page_ai.py
+++ b/skyvern/core/script_generations/skyvern_page_ai.py
@@ -4,6 +4,12 @@ from typing import Any, Protocol
from skyvern.config import settings
+# Sentinel for the optional ``system_prompt`` parameter on ``ai_extract``.
+# Distinguishes "caller omitted the argument" (resolve from workflow context,
+# honoring the current block's ``ignore_workflow_system_prompt`` flag) from
+# "caller passed None" (opt out, send no system prompt).
+SYSTEM_PROMPT_UNSET: Any = object()
+
class SkyvernPageAi(Protocol):
"""Protocol defining the interface for AI-powered page interactions."""
@@ -67,6 +73,7 @@ class SkyvernPageAi(Protocol):
data: str | dict[str, Any] | None = None,
skip_refresh: bool = False,
include_extracted_text: bool = True,
+ system_prompt: str | None | Any = SYSTEM_PROMPT_UNSET,
) -> dict[str, Any] | list | str | None:
"""Extract information from the page using AI."""
...
diff --git a/skyvern/forge/agent.py b/skyvern/forge/agent.py
index 827089490..ba58a2cd3 100644
--- a/skyvern/forge/agent.py
+++ b/skyvern/forge/agent.py
@@ -393,6 +393,7 @@ class ForgeAgent:
retry=task_retry,
max_steps_per_run=task_block.max_steps_per_run,
error_code_mapping=task_block.error_code_mapping,
+ workflow_system_prompt=task_block.workflow_system_prompt,
include_action_history_in_verification=task_block.include_action_history_in_verification,
model=task_block.model,
max_screenshot_scrolling_times=workflow_run.max_screenshot_scrolls,
@@ -462,6 +463,7 @@ class ForgeAgent:
proxy_location=task_request.proxy_location,
extracted_information_schema=task_request.extracted_information_schema,
error_code_mapping=task_request.error_code_mapping,
+ workflow_system_prompt=task_request.workflow_system_prompt,
application=task_request.application,
include_action_history_in_verification=task_request.include_action_history_in_verification,
model=task_request.model,
@@ -1311,6 +1313,7 @@ class ForgeAgent:
prompt_name=prompt_name,
step=step,
screenshots=scraped_page.screenshots,
+ system_prompt=task.workflow_system_prompt,
)
else:
LOG.debug(
@@ -1917,6 +1920,7 @@ class ForgeAgent:
prompt_name="cua-answer-question",
step=step,
screenshots=scraped_page.screenshots,
+ system_prompt=task.workflow_system_prompt,
)
LOG.info("Skyvern response to CUA question", skyvern_response=skyvern_response)
resp_content = skyvern_response.get("answer")
@@ -2172,6 +2176,7 @@ class ForgeAgent:
prompt_name=prompt_name,
step=next_step,
screenshots=scraped_page.screenshots,
+ system_prompt=task.workflow_system_prompt,
)
LOG.info(
@@ -2499,6 +2504,7 @@ class ForgeAgent:
step=step,
screenshots=scraped_page_refreshed.screenshots,
prompt_name=prompt_name,
+ system_prompt=task.workflow_system_prompt,
)
result = CompleteVerifyResult.model_validate(verification_result)
if result.is_complete:
@@ -3286,7 +3292,9 @@ class ForgeAgent:
llm_api_handler = LLMAPIHandlerFactory.get_override_llm_api_handler(
task.llm_key, default=app.LLM_API_HANDLER
)
- json_response = await llm_api_handler(prompt=prompt, step=step, prompt_name="infer-action-type")
+ json_response = await llm_api_handler(
+ prompt=prompt, step=step, prompt_name="infer-action-type", system_prompt=task.workflow_system_prompt
+ )
if json_response.get("error"):
raise FailedToParseActionInstruction(
reason=json_response.get("thought"), error_type=json_response.get("error")
@@ -4736,7 +4744,11 @@ class ForgeAgent:
local_datetime=datetime.now(skyvern_context.ensure_context().tz_info).isoformat(),
)
json_response = await app.LLM_API_HANDLER(
- prompt=prompt, screenshots=screenshots, step=step, prompt_name="summarize-max-steps-reason"
+ prompt=prompt,
+ screenshots=screenshots,
+ step=step,
+ prompt_name="summarize-max-steps-reason",
+ system_prompt=task.workflow_system_prompt,
)
return MaxStepsReasonResponse.model_validate(json_response)
except Exception:
@@ -4879,6 +4891,7 @@ class ForgeAgent:
screenshots=screenshots,
step=step,
prompt_name="summarize-max-retries-reason",
+ system_prompt=task.workflow_system_prompt,
)
return MaxStepsReasonResponse.model_validate(json_response)
except Exception:
@@ -5256,6 +5269,7 @@ class ForgeAgent:
step=step,
screenshots=scraped_page.screenshots,
prompt_name=prompt_name,
+ system_prompt=task.workflow_system_prompt,
)
@staticmethod
@@ -5323,6 +5337,7 @@ class ForgeAgent:
data_extraction_goal=task.data_extraction_goal,
extracted_information_schema=post_ceiling_kwargs["data_extraction_schema"],
llm_key=None,
+ workflow_system_prompt=task.workflow_system_prompt,
)
lookup_result = extraction_cache.lookup(workflow_run_id, cache_key)
except Exception:
@@ -5375,7 +5390,10 @@ class ForgeAgent:
cache_path="agent",
)
data_extraction_summary_resp = await app.EXTRACTION_LLM_API_HANDLER(
- prompt=prompt, step=step, prompt_name="data-extraction-summary"
+ prompt=prompt,
+ step=step,
+ prompt_name="data-extraction-summary",
+ system_prompt=task.workflow_system_prompt,
)
if cache_key and isinstance(data_extraction_summary_resp, dict):
extraction_cache.store(workflow_run_id, cache_key, data_extraction_summary_resp)
diff --git a/skyvern/forge/sdk/artifact/manager.py b/skyvern/forge/sdk/artifact/manager.py
index 0af36fbdc..fd8c80144 100644
--- a/skyvern/forge/sdk/artifact/manager.py
+++ b/skyvern/forge/sdk/artifact/manager.py
@@ -272,6 +272,56 @@ class ArtifactManager:
path=path,
)
+ async def create_download_artifact(
+ self,
+ *,
+ organization_id: str,
+ run_id: str,
+ uri: str,
+ filename: str,
+ workflow_run_id: str | None = None,
+ checksum: str | None = None,
+ ) -> str:
+ """Register a downloaded file as an Artifact row without re-uploading.
+
+ The bytes already live at ``uri`` (the uploads bucket). We only record a
+ row so the file can be served through the signed ``/v1/artifacts/{id}/content``
+ endpoint.
+ """
+ # Idempotent on (run_id, uri): if a DOWNLOAD artifact already exists for the
+ # same physical file (e.g. a loop iteration re-uploads the same download dir),
+ # return the existing artifact_id so signed URLs stay stable across calls —
+ # otherwise ``loop_download_filter.to_downloaded_file_signature`` would treat
+ # every iteration's URL as new.
+ existing = await app.DATABASE.artifacts.find_download_artifact(
+ organization_id=organization_id,
+ run_id=run_id,
+ uri=uri,
+ )
+ if existing is not None:
+ return existing.artifact_id
+
+ artifact_id = generate_artifact_id()
+ context = skyvern_context.current()
+ if workflow_run_id is None and context is not None:
+ workflow_run_id = context.workflow_run_id
+ await app.DATABASE.artifacts.create_artifact(
+ artifact_id=artifact_id,
+ artifact_type=ArtifactType.DOWNLOAD,
+ uri=uri,
+ organization_id=organization_id,
+ run_id=run_id,
+ workflow_run_id=workflow_run_id,
+ checksum=checksum,
+ )
+ LOG.debug(
+ "Registered downloaded file as artifact",
+ artifact_id=artifact_id,
+ run_id=run_id,
+ filename=filename,
+ )
+ return artifact_id
+
async def create_thought_artifact(
self,
thought: Thought,
@@ -836,6 +886,25 @@ class ArtifactManager:
async def retrieve_artifact(self, artifact: Artifact) -> bytes | None:
return await app.STORAGE.retrieve_artifact(artifact)
+ def build_signed_content_url(
+ self,
+ artifact_id: str,
+ artifact_name: str | None = None,
+ artifact_type: str | None = None,
+ ) -> str:
+ """Return a signed ``/v1/artifacts/{id}/content`` URL for any artifact.
+
+ Non-bundled artifacts normally get a presigned S3 URL from
+ ``STORAGE.get_share_link``. This method always builds the Skyvern-origin
+ signed URL regardless of ``bundle_key`` — used for DOWNLOAD artifacts
+ so webhook payloads stay short and clients hit our origin.
+ """
+ return self._bundle_content_url(
+ artifact_id=artifact_id,
+ artifact_name=artifact_name,
+ artifact_type=artifact_type,
+ )
+
def _bundle_content_url(
self,
artifact_id: str,
diff --git a/skyvern/forge/sdk/artifact/models.py b/skyvern/forge/sdk/artifact/models.py
index 76c6d0ae4..e9161c3dc 100644
--- a/skyvern/forge/sdk/artifact/models.py
+++ b/skyvern/forge/sdk/artifact/models.py
@@ -61,6 +61,9 @@ class ArtifactType(StrEnum):
# Task archive: one ZIP per task containing task-level cleanup artifacts (HAR, console log, trace, final screenshot)
TASK_ARCHIVE = "task_archive"
+ # Files downloaded by the browser during a run (stored in the uploads bucket, not the artifacts bucket).
+ DOWNLOAD = "download"
+
class Artifact(BaseModel):
created_at: datetime = Field(
@@ -82,6 +85,7 @@ class Artifact(BaseModel):
artifact_type: ArtifactType
uri: str
bundle_key: str | None = None
+ checksum: str | None = None
task_id: str | None = None
step_id: str | None = None
workflow_run_id: str | None = None
diff --git a/skyvern/forge/sdk/artifact/storage/azure.py b/skyvern/forge/sdk/artifact/storage/azure.py
index 0923c4063..a869eb73a 100644
--- a/skyvern/forge/sdk/artifact/storage/azure.py
+++ b/skyvern/forge/sdk/artifact/storage/azure.py
@@ -8,6 +8,7 @@ import structlog
from skyvern.config import settings
from skyvern.constants import BROWSER_DOWNLOADING_SUFFIX, DOWNLOAD_FILE_PREFIX
+from skyvern.forge import app
from skyvern.forge.sdk.api.azure import AzureUri, StandardBlobTier
from skyvern.forge.sdk.api.files import (
calculate_sha256_for_file,
@@ -22,6 +23,7 @@ from skyvern.forge.sdk.artifact.models import Artifact, ArtifactType, LogEntityT
from skyvern.forge.sdk.artifact.storage.base import (
FILE_EXTENTSION_MAP,
BaseStorage,
+ _file_infos_from_download_artifacts,
)
from skyvern.forge.sdk.models import Step
from skyvern.forge.sdk.schemas.ai_suggestions import AISuggestion
@@ -405,16 +407,84 @@ class AzureStorage(BaseStorage):
organization_id=organization_id,
storage_tier=tier,
)
- # Upload file with checksum metadata
- await self.async_client.upload_file_from_path(
- uri=uri,
- file_path=fpath,
- metadata={"sha256_checksum": checksum, "original_filename": file},
- tier=tier,
- tags=tags,
- )
+ # Azure Blob metadata values must be ASCII; preserve the full
+ # filename via the blob path / Artifact URI instead.
+ metadata: dict[str, str] = {"sha256_checksum": checksum}
+ if file.isascii():
+ metadata["original_filename"] = file
+ # Catch upload failures so we never create an Artifact row for
+ # bytes that didn't actually land in storage.
+ try:
+ await self.async_client.upload_file_from_path(
+ uri=uri,
+ file_path=fpath,
+ metadata=metadata,
+ tier=tier,
+ tags=tags,
+ )
+ except Exception:
+ LOG.warning(
+ "Skipping downloaded file — Azure upload failed",
+ file=file,
+ organization_id=organization_id,
+ run_id=run_id,
+ exc_info=True,
+ )
+ continue
+
+ # Register the file as an Artifact so GET run output can serve it via
+ # the signed /v1/artifacts/{id}/content endpoint (SKY-8861). Persist
+ # the SHA-256 we already computed so retrieval doesn't need an
+ # extra blob HEAD per file.
+ if run_id is not None:
+ try:
+ await app.ARTIFACT_MANAGER.create_download_artifact(
+ organization_id=organization_id,
+ run_id=run_id,
+ uri=uri,
+ filename=file,
+ checksum=checksum,
+ )
+ except Exception:
+ LOG.warning(
+ "Failed to register downloaded file as artifact; falling back to SAS URLs for retrieval",
+ file=file,
+ organization_id=organization_id,
+ run_id=run_id,
+ exc_info=True,
+ )
async def get_downloaded_files(self, organization_id: str, run_id: str | None) -> list[FileInfo]:
+ # Artifact-first — see s3.py::get_downloaded_files for rationale. When
+ # the keyring isn't configured (OSS default) or no artifact rows exist
+ # (legacy run pre-SKY-8861) we fall back to the legacy listing path so
+ # downloaded files remain reachable.
+ if run_id is not None and settings.ARTIFACT_CONTENT_HMAC_KEYRING:
+ artifacts = await self._list_download_artifacts_safe(organization_id=organization_id, run_id=run_id)
+ if artifacts:
+ return _file_infos_from_download_artifacts(artifacts)
+
+ return await self._get_downloaded_files_via_blob_listing(organization_id=organization_id, run_id=run_id)
+
+ async def _list_download_artifacts_safe(self, *, organization_id: str, run_id: str) -> list[Artifact]:
+ try:
+ return await app.DATABASE.artifacts.list_artifacts_for_run_by_type(
+ run_id=run_id,
+ organization_id=organization_id,
+ artifact_type=ArtifactType.DOWNLOAD,
+ )
+ except Exception:
+ LOG.warning(
+ "Failed to look up download artifacts; falling back to SAS URLs",
+ organization_id=organization_id,
+ run_id=run_id,
+ exc_info=True,
+ )
+ return []
+
+ async def _get_downloaded_files_via_blob_listing(
+ self, *, organization_id: str, run_id: str | None
+ ) -> list[FileInfo]:
uri = f"azure://{settings.AZURE_STORAGE_CONTAINER_UPLOADS}/{DOWNLOAD_FILE_PREFIX}/{settings.ENV}/{organization_id}/{run_id}"
object_keys = await self.async_client.list_files(uri=uri)
if len(object_keys) == 0:
@@ -424,24 +494,22 @@ class AzureStorage(BaseStorage):
for key in object_keys:
object_uri = f"azure://{settings.AZURE_STORAGE_CONTAINER_UPLOADS}/{key}"
- # Get metadata (including checksum)
metadata = await self.async_client.get_file_metadata(object_uri, log_exception=False)
-
- # Create FileInfo object
filename = os.path.basename(key)
checksum = metadata.get("sha256_checksum") if metadata else None
+ display_name = metadata.get("original_filename", filename) if metadata else filename
- # Get SAS URL
sas_urls = await self.async_client.create_sas_urls([object_uri])
if not sas_urls:
continue
- file_info = FileInfo(
- url=sas_urls[0],
- checksum=checksum,
- filename=metadata.get("original_filename", filename) if metadata else filename,
+ file_infos.append(
+ FileInfo(
+ url=sas_urls[0],
+ checksum=checksum,
+ filename=display_name,
+ )
)
- file_infos.append(file_info)
return file_infos
diff --git a/skyvern/forge/sdk/artifact/storage/base.py b/skyvern/forge/sdk/artifact/storage/base.py
index 72f484402..4bbded38a 100644
--- a/skyvern/forge/sdk/artifact/storage/base.py
+++ b/skyvern/forge/sdk/artifact/storage/base.py
@@ -1,6 +1,7 @@
from abc import ABC, abstractmethod
from typing import BinaryIO
+from skyvern.forge import app
from skyvern.forge.sdk.artifact.models import Artifact, ArtifactType, LogEntityType
from skyvern.forge.sdk.models import Step
from skyvern.forge.sdk.schemas.ai_suggestions import AISuggestion
@@ -8,6 +9,33 @@ from skyvern.forge.sdk.schemas.files import FileInfo
from skyvern.forge.sdk.schemas.task_v2 import TaskV2, Thought
from skyvern.forge.sdk.schemas.workflow_runs import WorkflowRunBlock
+
+def _file_infos_from_download_artifacts(artifacts: list[Artifact]) -> list[FileInfo]:
+ """Build the API-shaped ``FileInfo`` list from DOWNLOAD artifact rows.
+
+ Filename is the URI basename (the save site writes ``{base_uri}/{file}``);
+ checksum and modified_at come straight from the row, so retrieval needs
+ zero S3 round-trips.
+ """
+ infos: list[FileInfo] = []
+ for artifact in artifacts:
+ filename = artifact.uri.rsplit("/", 1)[-1] if artifact.uri else ""
+ url = app.ARTIFACT_MANAGER.build_signed_content_url(
+ artifact_id=artifact.artifact_id,
+ artifact_name=filename,
+ artifact_type=ArtifactType.DOWNLOAD.value,
+ )
+ infos.append(
+ FileInfo(
+ url=url,
+ checksum=artifact.checksum,
+ filename=filename,
+ modified_at=artifact.created_at,
+ )
+ )
+ return infos
+
+
# TODO: This should be a part of the ArtifactType model
FILE_EXTENTSION_MAP: dict[ArtifactType, str] = {
ArtifactType.RECORDING: "webm",
diff --git a/skyvern/forge/sdk/artifact/storage/s3.py b/skyvern/forge/sdk/artifact/storage/s3.py
index 51b934ea5..a0f5f28e7 100644
--- a/skyvern/forge/sdk/artifact/storage/s3.py
+++ b/skyvern/forge/sdk/artifact/storage/s3.py
@@ -11,6 +11,7 @@ import zstandard as zstd
from skyvern.config import settings
from skyvern.constants import BROWSER_DOWNLOADING_SUFFIX, DOWNLOAD_FILE_PREFIX
+from skyvern.forge import app
from skyvern.forge.sdk.api.aws import AsyncAWSClient, S3StorageClass, S3Uri
from skyvern.forge.sdk.api.files import (
calculate_sha256_for_file,
@@ -24,6 +25,7 @@ from skyvern.forge.sdk.artifact.models import Artifact, ArtifactType, LogEntityT
from skyvern.forge.sdk.artifact.storage.base import (
FILE_EXTENTSION_MAP,
BaseStorage,
+ _file_infos_from_download_artifacts,
)
from skyvern.forge.sdk.models import Step
from skyvern.forge.sdk.schemas.ai_suggestions import AISuggestion
@@ -432,15 +434,91 @@ class S3Storage(BaseStorage):
organization_id=organization_id,
storage_class=storage_class,
)
- # Upload file with checksum metadata
- await self.async_client.upload_file_from_path(
- uri=uri,
- file_path=fpath,
- metadata={"sha256_checksum": checksum, "original_filename": file},
- storage_class=storage_class,
- )
+ # S3 object metadata only allows ASCII; non-ASCII filenames (CJK,
+ # emoji) would otherwise raise ParamValidationError at upload time.
+ # The full filename is still preserved in the S3 key and on the
+ # Artifact row's URI.
+ metadata: dict[str, str] = {"sha256_checksum": checksum}
+ if file.isascii():
+ metadata["original_filename"] = file
+ # Upload with raise_exception=True so a partial failure aborts
+ # this iteration and we never create an Artifact row for bytes
+ # that didn't actually land in S3.
+ try:
+ await self.async_client.upload_file_from_path(
+ uri=uri,
+ file_path=fpath,
+ metadata=metadata,
+ storage_class=storage_class,
+ raise_exception=True,
+ )
+ except Exception:
+ LOG.warning(
+ "Skipping downloaded file — S3 upload failed",
+ file=file,
+ organization_id=organization_id,
+ run_id=run_id,
+ exc_info=True,
+ )
+ continue
+
+ # Register the file as an Artifact so GET run output can serve it via
+ # the signed /v1/artifacts/{id}/content endpoint (SKY-8861). Persist
+ # the SHA-256 we already computed so retrieval doesn't need an
+ # extra S3 HEAD per file.
+ if run_id is not None:
+ try:
+ await app.ARTIFACT_MANAGER.create_download_artifact(
+ organization_id=organization_id,
+ run_id=run_id,
+ uri=uri,
+ filename=file,
+ checksum=checksum,
+ )
+ except Exception:
+ LOG.warning(
+ "Failed to register downloaded file as artifact; falling back to S3 listing for retrieval",
+ file=file,
+ organization_id=organization_id,
+ run_id=run_id,
+ exc_info=True,
+ )
async def get_downloaded_files(self, organization_id: str, run_id: str | None) -> list[FileInfo]:
+ # Artifact-first: when a run has DOWNLOAD artifact rows, return them as
+ # the source of truth — the row carries enough to build a short signed
+ # /v1/artifacts/{id}/content URL plus the SHA-256 we persisted at save
+ # time, so we skip the S3 LIST and per-file HEAD entirely (SKY-8861).
+ #
+ # If HMAC signing isn't configured (self-hosted OSS default), the signed
+ # endpoint requires an API key webhook consumers don't have, so we stay
+ # on the legacy S3-list+presign path even when rows exist.
+ if run_id is not None and settings.ARTIFACT_CONTENT_HMAC_KEYRING:
+ artifacts = await self._list_download_artifacts_safe(organization_id=organization_id, run_id=run_id)
+ if artifacts:
+ return _file_infos_from_download_artifacts(artifacts)
+
+ # Legacy fallback — runs predating SKY-8861 (no artifact rows) and
+ # OSS-default deployments without HMAC signing both arrive here.
+ return await self._get_downloaded_files_via_s3_listing(organization_id=organization_id, run_id=run_id)
+
+ async def _list_download_artifacts_safe(self, *, organization_id: str, run_id: str) -> list[Artifact]:
+ try:
+ return await app.DATABASE.artifacts.list_artifacts_for_run_by_type(
+ run_id=run_id,
+ organization_id=organization_id,
+ artifact_type=ArtifactType.DOWNLOAD,
+ )
+ except Exception:
+ LOG.warning(
+ "Failed to look up download artifacts; falling back to presigned S3 URLs",
+ organization_id=organization_id,
+ run_id=run_id,
+ exc_info=True,
+ )
+ return []
+
+ async def _get_downloaded_files_via_s3_listing(self, *, organization_id: str, run_id: str | None) -> list[FileInfo]:
bucket = settings.AWS_S3_BUCKET_UPLOADS
uri = f"s3://{bucket}/{DOWNLOAD_FILE_PREFIX}/{settings.ENV}/{organization_id}/{run_id}"
object_keys = await self.async_client.list_files(uri=uri)
@@ -451,25 +529,22 @@ class S3Storage(BaseStorage):
for key in object_keys:
object_uri = f"s3://{bucket}/{key}"
- # Get metadata (including checksum)
metadata = await self.async_client.get_file_metadata(object_uri, log_exception=False)
-
- # Create FileInfo object
filename = os.path.basename(key)
checksum = metadata.get("sha256_checksum") if metadata else None
+ display_name = metadata.get("original_filename", filename) if metadata else filename
- # Get presigned URL
presigned_urls = await self.async_client.create_presigned_urls([object_uri])
if not presigned_urls:
continue
- file_info = FileInfo(
- url=presigned_urls[0],
- checksum=checksum,
- filename=metadata.get("original_filename", filename) if metadata else filename,
+ file_infos.append(
+ FileInfo(
+ url=presigned_urls[0],
+ checksum=checksum,
+ filename=display_name,
+ )
)
- file_infos.append(file_info)
-
return file_infos
async def save_legacy_file(
diff --git a/skyvern/forge/sdk/cache/extraction_cache.py b/skyvern/forge/sdk/cache/extraction_cache.py
index 5333e5342..39c39cc36 100644
--- a/skyvern/forge/sdk/cache/extraction_cache.py
+++ b/skyvern/forge/sdk/cache/extraction_cache.py
@@ -40,6 +40,10 @@ Key derivation (shared with the cross-run tier):
correctness if an intra-task second-step extraction happens.
- llm_key — the caller's model override. Prevents stale hits when a user
changes models to retune quality.
+ - workflow_system_prompt — the workflow's workflow_system_prompt (or None).
+ The prompt is sent to the LLM as the `system` message; changing it
+ changes the output even if all user-prompt inputs are identical, so two
+ calls that differ only in workflow_system_prompt must not collide.
- Date is intentionally NOT in the key. Two calls on byte-identical page
content are semantically the same extraction regardless of wall-clock
date; relying on the content hash keeps hit rate up for scheduled
@@ -353,6 +357,7 @@ def compute_cache_key(
error_code_mapping: Any = None,
previous_extracted_information: Any = None,
llm_key: str | None = None,
+ workflow_system_prompt: str | None = None,
) -> str:
"""Return a stable sha256 hex digest for the inputs that affect extraction output.
@@ -391,6 +396,7 @@ def compute_cache_key(
_normalize(error_code_mapping),
_normalize(previous_extracted_information),
_s(llm_key),
+ _s(workflow_system_prompt),
]
joined = "\x1f".join(parts).encode("utf-8", errors="replace")
return hashlib.sha256(joined).hexdigest()
diff --git a/skyvern/forge/sdk/db/models.py b/skyvern/forge/sdk/db/models.py
index 5b4dbd1c8..87df33a87 100644
--- a/skyvern/forge/sdk/db/models.py
+++ b/skyvern/forge/sdk/db/models.py
@@ -103,6 +103,7 @@ class TaskModel(Base):
order = Column(Integer, nullable=True)
retry = Column(Integer, nullable=True)
error_code_mapping = Column(JSON, nullable=True)
+ workflow_system_prompt = Column(UnicodeText, nullable=True)
errors = Column(JSON, default=[], nullable=False)
max_steps_per_run = Column(Integer, nullable=True)
application = Column(String, nullable=True)
@@ -248,6 +249,7 @@ class ArtifactModel(Base):
uri = Column(String)
bundle_key = Column(String, nullable=True)
run_id = Column(String, nullable=True)
+ checksum = Column(String, nullable=True)
created_at = Column(DateTime, default=datetime.datetime.utcnow, nullable=False)
modified_at = Column(
DateTime,
@@ -426,6 +428,14 @@ class WorkflowRunModel(Base):
verification_code_identifier = Column(String, nullable=True)
verification_code_polling_started_at = Column(DateTime, nullable=True)
failure_category = Column(JSON, nullable=True)
+ # When True, this run was spawned by a WorkflowTriggerBlock whose
+ # ignore_workflow_system_prompt flag was set, and the child must not
+ # inherit the parent chain's workflow_system_prompt. Set at spawn time so
+ # async (Temporal-dispatched) child runs can honor the flag even though
+ # they start in a separate worker without in-process context.
+ ignore_inherited_workflow_system_prompt = Column(
+ Boolean, nullable=False, default=False, server_default=sqlalchemy.false()
+ )
queued_at = Column(DateTime, nullable=True)
started_at = Column(DateTime, nullable=True)
@@ -864,6 +874,7 @@ class TaskV2Model(Base):
proxy_location = Column(String, nullable=True)
extracted_information_schema = Column(JSON, nullable=True)
error_code_mapping = Column(JSON, nullable=True)
+ workflow_system_prompt = Column(UnicodeText, nullable=True)
max_steps = Column(Integer, nullable=True)
max_screenshot_scrolling_times = Column(Integer, nullable=True)
extra_http_headers = Column(JSON, nullable=True)
diff --git a/skyvern/forge/sdk/db/repositories/artifacts.py b/skyvern/forge/sdk/db/repositories/artifacts.py
index b2b0b1083..d01f118f3 100644
--- a/skyvern/forge/sdk/db/repositories/artifacts.py
+++ b/skyvern/forge/sdk/db/repositories/artifacts.py
@@ -47,6 +47,7 @@ class ArtifactsRepository(BaseRepository):
run_id: str | None = None,
thought_id: str | None = None,
ai_suggestion_id: str | None = None,
+ checksum: str | None = None,
) -> Artifact:
async with self.Session() as session:
new_artifact = ArtifactModel(
@@ -62,6 +63,7 @@ class ArtifactsRepository(BaseRepository):
run_id=run_id,
ai_suggestion_id=ai_suggestion_id,
organization_id=organization_id,
+ checksum=checksum,
)
session.add(new_artifact)
await session.commit()
@@ -357,6 +359,60 @@ class ArtifactsRepository(BaseRepository):
return convert_to_artifact(artifact, self.debug_enabled)
return None
+ @db_operation("find_download_artifact")
+ async def find_download_artifact(
+ self,
+ organization_id: str,
+ run_id: str,
+ uri: str,
+ ) -> Artifact | None:
+ """Return the existing DOWNLOAD artifact for ``(run_id, uri)`` if any.
+
+ Used by :meth:`ArtifactManager.create_download_artifact` to stay
+ idempotent: repeated saves of the same file in the same run (e.g.
+ within a loop block iteration) must reuse the existing artifact_id
+ so downstream URL-based dedup keeps seeing a stable URL.
+ """
+ async with self.Session() as session:
+ artifact = (
+ await session.scalars(
+ select(ArtifactModel)
+ .filter(ArtifactModel.run_id == run_id)
+ .filter(ArtifactModel.artifact_type == ArtifactType.DOWNLOAD)
+ .filter(ArtifactModel.organization_id == organization_id)
+ .filter(ArtifactModel.uri == uri)
+ .order_by(ArtifactModel.created_at.desc())
+ )
+ ).first()
+ if artifact:
+ return convert_to_artifact(artifact, self.debug_enabled)
+ return None
+
+ @db_operation("list_artifacts_for_run_by_type")
+ async def list_artifacts_for_run_by_type(
+ self,
+ run_id: str,
+ organization_id: str,
+ artifact_type: ArtifactType,
+ ) -> list[Artifact]:
+ """List all artifacts for a run filtered by type, using the dedicated ``run_id`` column.
+
+ Unlike :meth:`get_artifacts_for_run` this does not consult a ``RunReader`` —
+ it filters directly on the partial index ``ix_artifacts_run_id_partial`` and
+ returns the rows ordered by creation time.
+ """
+ async with self.Session() as session:
+ artifacts = (
+ await session.scalars(
+ select(ArtifactModel)
+ .filter(ArtifactModel.run_id == run_id)
+ .filter(ArtifactModel.artifact_type == artifact_type)
+ .filter(ArtifactModel.organization_id == organization_id)
+ .order_by(ArtifactModel.created_at)
+ )
+ ).all()
+ return [convert_to_artifact(a, self.debug_enabled) for a in artifacts]
+
@db_operation("get_artifact_for_run")
async def get_artifact_for_run(
self,
diff --git a/skyvern/forge/sdk/db/repositories/observer.py b/skyvern/forge/sdk/db/repositories/observer.py
index d0cbc86e5..a90028d8f 100644
--- a/skyvern/forge/sdk/db/repositories/observer.py
+++ b/skyvern/forge/sdk/db/repositories/observer.py
@@ -28,6 +28,7 @@ from skyvern.forge.sdk.db.utils import (
)
from skyvern.forge.sdk.schemas.task_v2 import TaskV2, TaskV2Status, Thought, ThoughtType
from skyvern.forge.sdk.schemas.workflow_runs import WorkflowRunBlock
+from skyvern.forge.sdk.utils.sanitization import sanitize_postgres_text
from skyvern.schemas.runs import ProxyLocationInput, RunEngine, ScriptRunResponse
from skyvern.schemas.workflows import BlockStatus, BlockType
@@ -196,12 +197,15 @@ class ObserverRepository(BaseRepository):
webhook_callback_url: str | None = None,
extracted_information_schema: dict | list | str | None = None,
error_code_mapping: dict | None = None,
+ workflow_system_prompt: str | None = None,
model: dict[str, Any] | None = None,
max_screenshot_scrolling_times: int | None = None,
extra_http_headers: dict[str, str] | None = None,
browser_address: str | None = None,
run_with: str | None = None,
) -> TaskV2:
+ if isinstance(workflow_system_prompt, str):
+ workflow_system_prompt = sanitize_postgres_text(workflow_system_prompt)
async with self.Session() as session:
new_task_v2 = TaskV2Model(
workflow_run_id=workflow_run_id,
@@ -215,6 +219,7 @@ class ObserverRepository(BaseRepository):
webhook_callback_url=webhook_callback_url,
extracted_information_schema=extracted_information_schema,
error_code_mapping=error_code_mapping,
+ workflow_system_prompt=workflow_system_prompt,
organization_id=organization_id,
model=model,
max_screenshot_scrolling_times=max_screenshot_scrolling_times,
diff --git a/skyvern/forge/sdk/db/repositories/tasks.py b/skyvern/forge/sdk/db/repositories/tasks.py
index eac12d56e..fc21c91c4 100644
--- a/skyvern/forge/sdk/db/repositories/tasks.py
+++ b/skyvern/forge/sdk/db/repositories/tasks.py
@@ -57,6 +57,7 @@ class TasksRepository(BaseRepository):
retry: int | None = None,
max_steps_per_run: int | None = None,
error_code_mapping: dict[str, str] | None = None,
+ workflow_system_prompt: str | None = None,
task_type: str = TaskType.general,
application: str | None = None,
include_action_history_in_verification: bool | None = None,
@@ -79,6 +80,7 @@ class TasksRepository(BaseRepository):
url = sanitize_postgres_text(url)
complete_criterion = _sanitize(complete_criterion)
terminate_criterion = _sanitize(terminate_criterion)
+ workflow_system_prompt = _sanitize(workflow_system_prompt)
async with self.Session() as session:
new_task = TaskModel(
@@ -102,6 +104,7 @@ class TasksRepository(BaseRepository):
retry=retry,
max_steps_per_run=max_steps_per_run,
error_code_mapping=error_code_mapping,
+ workflow_system_prompt=workflow_system_prompt,
application=application,
include_action_history_in_verification=include_action_history_in_verification,
model=model,
diff --git a/skyvern/forge/sdk/db/repositories/workflow_runs.py b/skyvern/forge/sdk/db/repositories/workflow_runs.py
index 480b4b88c..1864b27c8 100644
--- a/skyvern/forge/sdk/db/repositories/workflow_runs.py
+++ b/skyvern/forge/sdk/db/repositories/workflow_runs.py
@@ -163,6 +163,7 @@ class WorkflowRunsRepository(BaseRepository):
workflow_run_id: str | None = None,
trigger_type: WorkflowRunTriggerType | None = None,
workflow_schedule_id: str | None = None,
+ ignore_inherited_workflow_system_prompt: bool = False,
) -> WorkflowRun:
async with self.Session() as session:
kwargs: dict[str, Any] = {}
@@ -190,6 +191,7 @@ class WorkflowRunsRepository(BaseRepository):
code_gen=code_gen,
trigger_type=trigger_type.value if trigger_type else None,
workflow_schedule_id=workflow_schedule_id,
+ ignore_inherited_workflow_system_prompt=ignore_inherited_workflow_system_prompt,
**kwargs,
)
session.add(workflow_run)
diff --git a/skyvern/forge/sdk/db/utils.py b/skyvern/forge/sdk/db/utils.py
index eac4ed7ed..ae164c876 100644
--- a/skyvern/forge/sdk/db/utils.py
+++ b/skyvern/forge/sdk/db/utils.py
@@ -245,6 +245,7 @@ def convert_to_task(task_obj: TaskModel, debug_enabled: bool = False, workflow_p
retry=task_obj.retry,
max_steps_per_run=task_obj.max_steps_per_run,
error_code_mapping=task_obj.error_code_mapping,
+ workflow_system_prompt=task_obj.workflow_system_prompt,
errors=task_obj.errors,
application=task_obj.application,
model=task_obj.model,
@@ -372,6 +373,7 @@ def convert_to_artifact(artifact_model: ArtifactModel, debug_enabled: bool = Fal
artifact_type=ArtifactType[artifact_model.artifact_type.upper()],
uri=artifact_model.uri,
bundle_key=artifact_model.bundle_key,
+ checksum=artifact_model.checksum,
task_id=artifact_model.task_id,
step_id=artifact_model.step_id,
workflow_run_id=artifact_model.workflow_run_id,
@@ -488,6 +490,7 @@ def convert_to_workflow_run(
trigger_type=_safe_trigger_type(workflow_run_model.trigger_type),
workflow_schedule_id=workflow_run_model.workflow_schedule_id,
failure_category=workflow_run_model.failure_category,
+ ignore_inherited_workflow_system_prompt=workflow_run_model.ignore_inherited_workflow_system_prompt,
)
diff --git a/skyvern/forge/sdk/routes/agent_protocol.py b/skyvern/forge/sdk/routes/agent_protocol.py
index f5098a0dd..46c516491 100644
--- a/skyvern/forge/sdk/routes/agent_protocol.py
+++ b/skyvern/forge/sdk/routes/agent_protocol.py
@@ -1,7 +1,9 @@
import asyncio
import json
+import unicodedata
from enum import Enum
from typing import Annotated, Any
+from urllib.parse import quote
import structlog
import yaml
@@ -1459,10 +1461,106 @@ _ARTIFACT_CONTENT_TYPES: dict[ArtifactType, str] = {
ArtifactType.SCREENSHOT_ACTION: "image/png",
ArtifactType.SCREENSHOT_FINAL: "image/png",
ArtifactType.RECORDING: "video/webm",
+ ArtifactType.DOWNLOAD: "application/octet-stream",
}
_ARTIFACT_CONTENT_TYPE_DEFAULT = "application/json"
+def _sanitize_header_filename(name: str) -> str:
+ """Strip characters that would break or inject into a Content-Disposition header.
+
+ The artifact URI basename is derived from a user-controlled S3 key. Rejects:
+
+ - C0 (<0x20), DEL (0x7F), C1 (0x80-0x9F) — RFC 7230 violations.
+ - ``"`` and ``\\`` — would terminate or escape the quoted value.
+ - Unicode *format* (Cf) and *separator-line/paragraph* (Zl/Zp) chars —
+ includes bidi overrides (U+202E), ZWSP (U+200B), ZWNBSP (U+FEFF);
+ these enable filename spoofing in the browser download UI.
+ """
+ cleaned = []
+ for ch in name:
+ code = ord(ch)
+ if code < 0x20 or code == 0x7F or 0x80 <= code <= 0x9F:
+ continue
+ if ch in ('"', "\\"):
+ continue
+ if unicodedata.category(ch) in {"Cf", "Zl", "Zp"}:
+ continue
+ cleaned.append(ch)
+ return "".join(cleaned) or "download"
+
+
+def _ascii_fallback_filename(name: str) -> str:
+ """Best-effort ASCII form of ``name`` for the ``filename=`` parameter.
+
+ NFKD-normalizes and drops combining marks first so accented Latin
+ characters survive as their base letters (``fïlè.pdf`` → ``file.pdf``)
+ instead of being stripped entirely. The RFC 5987 ``filename*=UTF-8''...``
+ form still carries the full name for modern clients.
+
+ If the ASCII stem ends up empty (e.g. pure CJK or emoji names),
+ prepend ``download`` so legacy clients don't save a bare ``.pdf``
+ hidden file.
+ """
+ normalized = unicodedata.normalize("NFKD", name)
+ ascii_only = normalized.encode("ascii", "ignore").decode("ascii")
+ sanitized = _sanitize_header_filename(ascii_only)
+ stem, dot, ext = sanitized.rpartition(".")
+ if dot and not stem:
+ return f"download.{ext}"
+ return sanitized
+
+
+def _build_attachment_disposition(filename: str) -> str:
+ """Build a Content-Disposition header that survives non-ASCII filenames.
+
+ Emits both ``filename="<ascii>"`` (for legacy clients) and
+ ``filename*=UTF-8''<pct-encoded>`` (RFC 5987, for everything modern).
+ Ensures the header value is Latin-1 encodable so Starlette doesn't 500.
+ """
+ safe = _sanitize_header_filename(filename)
+ ascii_part = _ascii_fallback_filename(safe)
+ encoded = quote(safe, safe="")
+ return f"attachment; filename=\"{ascii_part}\"; filename*=UTF-8''{encoded}"
+
+
+def _artifact_filename_from_uri(uri: str | None) -> str:
+ """Extract the basename from an ``s3://``/``azure://`` URI without using
+ ``urlparse`` — that would split on ``?``/``#`` characters, which are legal
+ in S3 keys."""
+ if not uri:
+ return ""
+ return uri.rsplit("/", 1)[-1]
+
+
+def _artifact_response_config(artifact: Artifact) -> tuple[str, str]:
+ """Return (media_type, Content-Disposition) for the artifact content response.
+
+ DOWNLOAD artifacts use ``attachment`` disposition with the sanitized filename
+ so browsers never render user-supplied content inline (SKY-8862). All other
+ types keep the historical ``inline`` behaviour.
+ """
+ media_type = _ARTIFACT_CONTENT_TYPES.get(artifact.artifact_type, _ARTIFACT_CONTENT_TYPE_DEFAULT)
+ if artifact.artifact_type == ArtifactType.DOWNLOAD:
+ raw_name = _artifact_filename_from_uri(artifact.uri)
+ return media_type, _build_attachment_disposition(raw_name)
+ return media_type, "inline"
+
+
+def _artifact_content_response_headers(*, disposition: str, is_signed: bool) -> dict[str, str]:
+ """Response headers for the artifact content endpoint.
+
+ Includes ``X-Content-Type-Options: nosniff`` as defence-in-depth for
+ SKY-8862: even if something upstream strips the attachment disposition,
+ the browser will not sniff the octet-stream body back into HTML/PDF.
+ """
+ return {
+ "Content-Disposition": disposition,
+ "Cache-Control": f"private, max-age={ARTIFACT_URL_EXPIRY_SECONDS}" if is_signed else "private, no-cache",
+ "X-Content-Type-Options": "nosniff",
+ }
+
+
@base_router.get(
"/artifacts/{artifact_id}/content",
tags=["Artifacts"],
@@ -1529,16 +1627,15 @@ async def get_artifact_content(
status_code=http_status.HTTP_404_NOT_FOUND,
detail="Artifact content not available",
)
- media_type = _ARTIFACT_CONTENT_TYPES.get(artifact.artifact_type, _ARTIFACT_CONTENT_TYPE_DEFAULT)
+ media_type, content_disposition = _artifact_response_config(artifact)
is_signed = sig is not None and expiry is not None and kid is not None
- cache_control = f"private, max-age={ARTIFACT_URL_EXPIRY_SECONDS}" if is_signed else "private, no-cache"
return Response(
content=content,
media_type=media_type,
- headers={
- "Content-Disposition": "inline",
- "Cache-Control": cache_control,
- },
+ headers=_artifact_content_response_headers(
+ disposition=content_disposition,
+ is_signed=is_signed,
+ ),
)
diff --git a/skyvern/forge/sdk/schemas/task_v2.py b/skyvern/forge/sdk/schemas/task_v2.py
index 9e4de0449..c839c752b 100644
--- a/skyvern/forge/sdk/schemas/task_v2.py
+++ b/skyvern/forge/sdk/schemas/task_v2.py
@@ -46,6 +46,7 @@ class TaskV2(BaseModel):
webhook_failure_reason: str | None = None
extracted_information_schema: dict | list | str | None = None
error_code_mapping: dict | None = None
+ workflow_system_prompt: str | None = None
model: dict[str, Any] | None = None
queued_at: datetime | None = None
started_at: datetime | None = None
@@ -203,6 +204,7 @@ class TaskV2Request(BaseModel):
publish_workflow: bool = False
extracted_information_schema: dict | list | str | None = None
error_code_mapping: dict[str, str] | None = None
+ workflow_system_prompt: str | None = None
max_screenshot_scrolls: int | None = None
extra_http_headers: dict[str, str] | None = None
browser_address: str | None = None
diff --git a/skyvern/forge/sdk/schemas/tasks.py b/skyvern/forge/sdk/schemas/tasks.py
index 68670bd31..a3009bc95 100644
--- a/skyvern/forge/sdk/schemas/tasks.py
+++ b/skyvern/forge/sdk/schemas/tasks.py
@@ -69,6 +69,11 @@ class TaskBase(BaseModel):
}
],
)
+ workflow_system_prompt: str | None = Field(
+ default=None,
+ description="System prompt applied to every LLM call for this task. Inherited from the workflow-level workflow_system_prompt when unset.",
+ examples=["Never guess at an answer. If unsure, respond with UNKNOWN."],
+ )
proxy_location: ProxyLocationInput = Field(
default=None,
description=PROXY_LOCATION_DOC_STRING,
diff --git a/skyvern/forge/sdk/workflow/context_manager.py b/skyvern/forge/sdk/workflow/context_manager.py
index b1d277379..e7d735944 100644
--- a/skyvern/forge/sdk/workflow/context_manager.py
+++ b/skyvern/forge/sdk/workflow/context_manager.py
@@ -1,4 +1,5 @@
import copy
+from datetime import datetime, timezone
from typing import TYPE_CHECKING, Any, Awaitable, Callable, Self
import structlog
@@ -26,7 +27,7 @@ from skyvern.forge.sdk.schemas.organizations import Organization
from skyvern.forge.sdk.schemas.tasks import TaskStatus
from skyvern.forge.sdk.services.bitwarden import BitwardenConstants, BitwardenService
from skyvern.forge.sdk.services.credentials import AzureVaultConstants, OnePasswordConstants, parse_totp_secret
-from skyvern.forge.sdk.workflow.exceptions import OutputParameterKeyCollisionError
+from skyvern.forge.sdk.workflow.exceptions import MissingJinjaVariables, OutputParameterKeyCollisionError
from skyvern.forge.sdk.workflow.models.parameter import (
PARAMETER_TYPE,
AWSSecretParameter,
@@ -45,6 +46,7 @@ from skyvern.forge.sdk.workflow.models.parameter import (
WorkflowParameterType,
)
from skyvern.utils.strings import generate_random_string
+from skyvern.utils.templating import get_missing_variables
if TYPE_CHECKING:
from skyvern.forge.sdk.workflow.models.workflow import Workflow, WorkflowRunParameter
@@ -83,6 +85,7 @@ class WorkflowRunContext:
],
block_outputs: dict[str, Any] | None = None,
workflow: "Workflow | None" = None,
+ inherited_workflow_system_prompt: str | None = None,
) -> Self:
# key is label name
workflow_run_context = cls(
@@ -92,6 +95,7 @@ class WorkflowRunContext:
workflow_run_id=workflow_run_id,
aws_client=aws_client,
workflow=workflow,
+ inherited_workflow_system_prompt=inherited_workflow_system_prompt,
)
workflow_run_context.organization_id = organization.organization_id
@@ -170,12 +174,34 @@ class WorkflowRunContext:
workflow_run_id: str,
aws_client: AsyncAWSClient,
workflow: "Workflow | None" = None,
+ inherited_workflow_system_prompt: str | None = None,
) -> None:
self.workflow_title = workflow_title
self.workflow_id = workflow_id
self.workflow_permanent_id = workflow_permanent_id
self.workflow_run_id = workflow_run_id
self.workflow = workflow
+ # Joined raw workflow_system_prompt(s) from ancestor workflows (outermost
+ # first) collected by walking workflow_run.parent_workflow_run_id at
+ # execute_workflow time. Jinja-rendered on demand and concatenated with
+ # this workflow's own workflow_system_prompt inside
+ # resolve_effective_workflow_system_prompt so parent-workflow rules flow
+ # into every child block and agent (SKY-9147).
+ self.inherited_workflow_system_prompt = inherited_workflow_system_prompt
+ # Sentinel for the lazy-resolved effective workflow_system_prompt cache.
+ # Using a sentinel (not None) so "resolved to None" is distinguishable
+ # from "not yet resolved". Invalidated by set_workflow() because late
+ # hydration can change the workflow's own workflow_system_prompt.
+ self._effective_workflow_system_prompt_cache: str | None = None
+ self._effective_workflow_system_prompt_resolved: bool = False
+ # Per-block record of the effective workflow_system_prompt once a block
+ # has run through ``Block._apply_workflow_system_prompt``. Keyed by
+ # block label. ``None`` is a valid recorded value (block opted out).
+ # Read by the script path (``RealSkyvernPageAi.ai_extract``) so a
+ # cached-script extraction uses the same string the agent path would
+ # — single source of truth, no re-resolving the opt-out from the
+ # workflow definition in two places (SKY-9147).
+ self._block_workflow_system_prompts: dict[str, str | None] = {}
self.blocks_metadata: dict[str, BlockMetadata] = {}
self.parameters: dict[str, PARAMETER_TYPE] = {}
self.values: dict[str, Any] = {}
@@ -193,6 +219,12 @@ class WorkflowRunContext:
This is used when the workflow is fetched from the database as a fallback.
"""
self.workflow = workflow
+ # Late-hydrated workflow may carry a different workflow_system_prompt than
+ # what was visible at construction time. Drop the cache so the next
+ # resolve_effective_workflow_system_prompt() re-renders against the new
+ # definition.
+ self._effective_workflow_system_prompt_resolved = False
+ self._effective_workflow_system_prompt_cache = None
def get_parameter(self, key: str) -> Parameter:
return self.parameters[key]
@@ -225,6 +257,108 @@ class WorkflowRunContext:
label = ""
return self.blocks_metadata.get(label, BlockMetadata())
+ def record_block_workflow_system_prompt(self, label: str, value: str | None) -> None:
+ """Record the effective ``workflow_system_prompt`` a block resolved to.
+
+ Called by ``Block._apply_workflow_system_prompt`` (agent path) and by
+ the script-path dispatch before handing execution to cached code. Both
+ paths use the same recorded value in ``ai_extract`` so agent and
+ script extractions for the same block hash to the same cache key and
+ the same LLM input.
+ """
+ if label:
+ self._block_workflow_system_prompts[label] = value
+
+ def get_block_workflow_system_prompt(self, label: str | None) -> tuple[bool, str | None]:
+ """Return ``(recorded, value)`` for a block label.
+
+ ``recorded`` is True only when the block has actually run through
+ ``_apply_workflow_system_prompt`` — a recorded ``None`` (opt-out) is
+ distinguished from "never recorded" so callers can fall back safely
+ for non-block invocations (e.g. standalone scripts).
+ """
+ if label and label in self._block_workflow_system_prompts:
+ return True, self._block_workflow_system_prompts[label]
+ return False, None
+
+ def resolve_effective_workflow_system_prompt(self) -> str | None:
+ """Return the effective workflow-level system prompt for this run.
+
+ Concatenates any prompt inherited from ancestor workflows (propagated via
+ ``WorkflowTriggerBlock`` — outermost first) with this workflow's own
+ ``workflow_system_prompt``. Jinja substitutions are rendered against this
+ run's values for both portions so ancestor templates can still reference
+ common variables like ``workflow_title``; placeholders that only exist in
+ the parent's context render empty under non-strict mode. Parts join with
+ a blank line so distinct rule sets stay readable to the LLM. Returns
+ ``None`` when nothing is configured at any level so callers can short-
+ circuit on a simple falsy check.
+
+ The resolved string is cached on first call and reused for the life of
+ the run so every block sees the same effective prompt — and the LLM
+ cache keys that derive from it stay stable across blocks. The cache is
+ invalidated in ``set_workflow`` for the late-hydration path.
+ """
+ if self._effective_workflow_system_prompt_resolved:
+ return self._effective_workflow_system_prompt_cache
+ own_raw: str | None = None
+ if self.workflow is not None and self.workflow.workflow_definition is not None:
+ candidate = self.workflow.workflow_definition.workflow_system_prompt
+ # ``isinstance`` guard: a malformed workflow definition (or a test
+ # MagicMock whose attribute access returns another mock) could
+ # hand us a non-string here. Jinja's ``from_string`` would then
+ # raise ``Can't compile non template nodes`` deep inside the
+ # render path. Narrowing to ``str`` keeps the fallback silent.
+ if isinstance(candidate, str):
+ own_raw = candidate
+ inherited = (
+ self.inherited_workflow_system_prompt if isinstance(self.inherited_workflow_system_prompt, str) else None
+ )
+ inherited_resolved = self.render_workflow_level_template(inherited) if inherited else None
+ own_resolved = self.render_workflow_level_template(own_raw) if own_raw else None
+ parts = [p for p in (inherited_resolved, own_resolved) if p]
+ resolved = "\n\n".join(parts) if parts else None
+ self._effective_workflow_system_prompt_cache = resolved
+ self._effective_workflow_system_prompt_resolved = True
+ return resolved
+
+ def render_workflow_level_template(self, raw_template: str) -> str:
+ """Render a Jinja template against workflow-scoped variables only.
+
+ Shared by every path that resolves the workflow-level workflow_system_prompt
+ (block execution, script-path ai_extract) so both produce the same string —
+ same cache key, same LLM output. Deliberately omits block-scoped context:
+ a workflow-wide prompt has no single "current block" to bind against.
+ """
+ if not raw_template:
+ return raw_template
+
+ template_data: dict[str, Any] = self.values.copy()
+ template_data.setdefault("workflow_title", self.workflow_title)
+ template_data.setdefault("workflow_id", self.workflow_id)
+ template_data.setdefault("workflow_permanent_id", self.workflow_permanent_id)
+ template_data.setdefault("workflow_run_id", self.workflow_run_id)
+ template_data.setdefault("browser_session_id", self.browser_session_id or "")
+ template_data.setdefault("current_date", datetime.now(timezone.utc).strftime("%Y-%m-%d"))
+ template_data["workflow_run_outputs"] = self.workflow_run_outputs
+ template_data["workflow_run_summary"] = self.build_workflow_run_summary()
+
+ if missing_variables := get_missing_variables(raw_template, template_data):
+ if settings.WORKFLOW_TEMPLATING_STRICTNESS == "strict":
+ raise MissingJinjaVariables(template=raw_template, variables=missing_variables)
+ # Non-strict mode silently renders undefined variables as empty strings,
+ # which makes typos like {{ persona }} invisible to the user. Emit a
+ # warning so the operator has a breadcrumb when a workflow_system_prompt
+ # isn't picking up the value they expected.
+ LOG.warning(
+ "Undefined Jinja variables in workflow-level template; rendering them as empty strings",
+ missing_variables=missing_variables,
+ workflow_run_id=self.workflow_run_id,
+ workflow_permanent_id=self.workflow_permanent_id,
+ )
+
+ return jinja_sandbox_env.from_string(raw_template).render(template_data)
+
async def _should_include_secrets_in_templates(self) -> bool:
"""
Check if secrets should be included in template formatting based on experimentation provider.
@@ -1266,6 +1400,7 @@ class WorkflowContextManager:
],
block_outputs: dict[str, Any] | None = None,
workflow: "Workflow | None" = None,
+ inherited_workflow_system_prompt: str | None = None,
) -> WorkflowRunContext:
workflow_run_context = await WorkflowRunContext.init(
self.aws_client,
@@ -1280,6 +1415,7 @@ class WorkflowContextManager:
secret_parameters,
block_outputs,
workflow,
+ inherited_workflow_system_prompt=inherited_workflow_system_prompt,
)
self.workflow_run_contexts[workflow_run_id] = workflow_run_context
return workflow_run_context
diff --git a/skyvern/forge/sdk/workflow/models/block.py b/skyvern/forge/sdk/workflow/models/block.py
index 9a1eb9282..9687c9e90 100644
--- a/skyvern/forge/sdk/workflow/models/block.py
+++ b/skyvern/forge/sdk/workflow/models/block.py
@@ -261,6 +261,19 @@ class Block(BaseModel, abc.ABC):
continue_on_failure: bool = False
model: dict[str, Any] | None = None
disable_cache: bool = False
+ # Opt-out from workflow-level workflow_system_prompt inheritance (and, on a
+ # WorkflowTriggerBlock, from propagating the parent chain's prompt into the
+ # spawned child run). A no-op for deterministic blocks that don't call an LLM.
+ ignore_workflow_system_prompt: bool = False
+ # Runtime cache populated by ``Block._apply_workflow_system_prompt`` — not
+ # user-settable. Excluded from serialization (``model_dump`` / JSON / API
+ # responses) so the resolved prompt doesn't leak into logs, workflow
+ # definition round-trips, or responses that weren't meant to carry it.
+ # Deliberately absent from the BlockYAML schema so it can never be set
+ # through YAML or the API. The user-facing opt-out is
+ # ``ignore_workflow_system_prompt``. Only consumed by block types that call
+ # an LLM; deterministic blocks ignore it.
+ workflow_system_prompt: str | None = Field(default=None, exclude=True)
# Only valid for blocks inside a for loop block
# Whether to continue to the next iteration when the block fails
@@ -517,6 +530,42 @@ class Block(BaseModel, abc.ABC):
return template.render(template_data)
+ def _apply_workflow_system_prompt(
+ self,
+ workflow_run_context: WorkflowRunContext,
+ ) -> None:
+ """Resolve the workflow-level ``workflow_system_prompt`` for this block and
+ materialize it onto ``self.workflow_system_prompt``.
+
+ Concatenates any prompt inherited from ancestor workflows (propagated through
+ ``WorkflowTriggerBlock``) with this workflow's own ``workflow_system_prompt``.
+ Jinja substitutions on this workflow's own prompt are resolved against
+ ``workflow_run_context``; the inherited portion is already resolved at the
+ trigger boundary.
+
+ Shared by every block type that needs to inherit the workflow system prompt
+ into its own ``workflow_system_prompt`` runtime cache before dispatching an
+ LLM call. Callers invoke this inside ``format_potential_template_parameters``
+ so the value is available at execute time. ``workflow_system_prompt`` on each
+ block is a runtime cache — it's deliberately absent from the BlockYAML schema
+ and not user-settable.
+
+ When a block opts out via ``ignore_workflow_system_prompt``, this leaves
+ the block's own ``workflow_system_prompt`` untouched (falling back to the
+ system default if none is set). The opt-out covers both this workflow's
+ prompt and any inherited prompt from parent workflows.
+ """
+ if self.ignore_workflow_system_prompt:
+ # Record the opt-out so the script path (``ai_extract``) reads the
+ # same decision instead of re-resolving the flag from the
+ # definition. See ``WorkflowRunContext.record_block_workflow_system_prompt``.
+ workflow_run_context.record_block_workflow_system_prompt(self.label, None)
+ return
+ resolved = workflow_run_context.resolve_effective_workflow_system_prompt()
+ if resolved is not None:
+ self.workflow_system_prompt = resolved
+ workflow_run_context.record_block_workflow_system_prompt(self.label, resolved)
+
@classmethod
def get_subclasses(cls) -> tuple[type[Block], ...]:
return tuple(cls.__subclasses__())
@@ -809,6 +858,10 @@ class BaseTaskBlock(Block):
for error_code, error_description in merged_mapping.items()
}
+ # Materialize the workflow-level workflow_system_prompt onto this block so
+ # ForgeAgent.create_task can hand it off to the Task row verbatim.
+ self._apply_workflow_system_prompt(workflow_run_context)
+
@staticmethod
async def get_task_order(workflow_run_id: str, current_retry: int) -> tuple[int, int]:
"""
@@ -2787,6 +2840,8 @@ class TextPromptBlock(Block):
if self.json_schema:
self.json_schema = self._render_schema_templates(self.json_schema, workflow_run_context)
+ self._apply_workflow_system_prompt(workflow_run_context)
+
async def send_prompt(
self,
prompt: str,
@@ -2839,6 +2894,7 @@ class TextPromptBlock(Block):
response = await llm_api_handler(
prompt=prompt,
prompt_name="text-prompt",
+ system_prompt=self.workflow_system_prompt,
workflow_run_block_id=workflow_run_block_id,
organization_id=organization_id,
)
@@ -3786,6 +3842,8 @@ class FileParserBlock(Block):
self.file_url, workflow_run_context
)
+ self._apply_workflow_system_prompt(workflow_run_context)
+
def _detect_file_type_from_url(self, file_url: str, file_path: str | None = None) -> FileType:
"""Detect file type based on file extension in the URL, with magic-byte fallback."""
url_parsed = urlparse(file_url)
@@ -4024,6 +4082,8 @@ class FileParserBlock(Block):
llm_api_handler = LLMAPIHandlerFactory.get_override_llm_api_handler(
self.override_llm_key, default=app.LLM_API_HANDLER
)
+ # OCR transcription intentionally skips system_prompt
+ # It still applies to the downstream extract-information-from-file-text call.
llm_response = await llm_api_handler(
prompt=llm_prompt,
prompt_name="extract-text-from-image",
@@ -4055,6 +4115,8 @@ class FileParserBlock(Block):
llm_api_handler = LLMAPIHandlerFactory.get_override_llm_api_handler(
self.override_llm_key, default=app.LLM_API_HANDLER
)
+ # OCR transcription intentionally skips system_prompt — see
+ # _parse_pdf_file_with_vision_ocr for rationale.
llm_response = await llm_api_handler(
prompt=llm_prompt,
prompt_name="extract-text-from-image",
@@ -4175,6 +4237,7 @@ class FileParserBlock(Block):
prompt=llm_prompt,
prompt_name="extract-information-from-file-text",
force_dict=False,
+ system_prompt=self.workflow_system_prompt,
workflow_run_block_id=workflow_run_block_id,
organization_id=organization_id,
)
@@ -4348,6 +4411,8 @@ class PDFParserBlock(Block):
self.file_url, workflow_run_context
)
+ self._apply_workflow_system_prompt(workflow_run_context)
+
async def execute(
self,
workflow_run_id: str,
@@ -4416,6 +4481,7 @@ class PDFParserBlock(Block):
prompt=llm_prompt,
prompt_name="extract-information-from-file-text",
force_dict=False,
+ system_prompt=self.workflow_system_prompt,
workflow_run_block_id=workflow_run_block_id,
organization_id=organization_id,
)
@@ -4837,6 +4903,10 @@ class TaskV2Block(Block):
)
self.totp_verification_url = prepend_scheme_and_validate_url(self.totp_verification_url)
+ # Materialize the workflow-level workflow_system_prompt onto this block so
+ # execute() can hand it off to the TaskV2 row verbatim.
+ self._apply_workflow_system_prompt(workflow_run_context)
+
async def execute(
self,
workflow_run_id: str,
@@ -4912,6 +4982,7 @@ class TaskV2Block(Block):
totp_identifier=resolved_totp_identifier,
totp_verification_url=resolved_totp_verification_url,
max_screenshot_scrolling_times=workflow_run.max_screenshot_scrolls,
+ workflow_system_prompt=self.workflow_system_prompt,
)
await app.DATABASE.observer.update_task_v2(
task_v2.observer_cruise_id, status=TaskV2Status.queued, organization_id=organization_id
@@ -7091,6 +7162,7 @@ class WorkflowTriggerBlock(Block):
workflow_permanent_id=resolved_workflow_permanent_id,
organization=organization,
parent_workflow_run_id=workflow_run_id,
+ ignore_inherited_workflow_system_prompt=self.ignore_workflow_system_prompt,
)
except Exception as e:
error_msg = get_user_facing_exception_message(e)
@@ -7106,6 +7178,9 @@ class WorkflowTriggerBlock(Block):
)
try:
+ # The opt-out flag is persisted on the child's workflow_run row at
+ # spawn time (setup_workflow_run above), so execute_workflow reads
+ # it from the DB. This works identically for sync and async triggers.
final_run = await app.WORKFLOW_SERVICE.execute_workflow(
workflow_run_id=triggered_run_id,
api_key=None,
@@ -7173,6 +7248,12 @@ class WorkflowTriggerBlock(Block):
browser_session_id=resolved_browser_session_id,
)
try:
+ # ``run_workflow`` persists this flag to the child's
+ # workflow_run row via its internal setup_workflow_run call,
+ # then dispatches to Temporal without passing the flag
+ # separately; the worker reads it back from the DB inside
+ # ``execute_workflow``. Symmetric with the sync branch above
+ # — the flag is written once, at spawn time, for both paths.
triggered_workflow_run = await run_workflow(
workflow_id=resolved_workflow_permanent_id,
organization=organization,
@@ -7180,6 +7261,7 @@ class WorkflowTriggerBlock(Block):
request=None,
background_tasks=None,
parent_workflow_run_id=workflow_run_id,
+ ignore_inherited_workflow_system_prompt=self.ignore_workflow_system_prompt,
)
except Exception as e:
error_msg = get_user_facing_exception_message(e)
diff --git a/skyvern/forge/sdk/workflow/models/workflow.py b/skyvern/forge/sdk/workflow/models/workflow.py
index 3d80d677b..361e59a0b 100644
--- a/skyvern/forge/sdk/workflow/models/workflow.py
+++ b/skyvern/forge/sdk/workflow/models/workflow.py
@@ -56,6 +56,7 @@ class WorkflowDefinition(BaseModel):
blocks: List[BlockTypeVar]
finally_block_label: str | None = None
error_code_mapping: dict[str, str] | None = None
+ workflow_system_prompt: str | None = None
def validate(self) -> None:
all_labels: set[str] = set()
@@ -199,6 +200,7 @@ class WorkflowRun(BaseModel):
code_gen: bool | None = None
trigger_type: WorkflowRunTriggerType | None = None
workflow_schedule_id: str | None = None
+ ignore_inherited_workflow_system_prompt: bool = False
@field_validator("run_with", mode="before")
@classmethod
diff --git a/skyvern/forge/sdk/workflow/service.py b/skyvern/forge/sdk/workflow/service.py
index ec55caff7..439563597 100644
--- a/skyvern/forge/sdk/workflow/service.py
+++ b/skyvern/forge/sdk/workflow/service.py
@@ -80,6 +80,7 @@ from skyvern.forge.sdk.workflow.models.block import (
ForLoopBlock,
NavigationBlock,
TaskV2Block,
+ WorkflowTriggerBlock,
compute_conditional_scopes,
get_all_blocks,
)
@@ -649,6 +650,7 @@ class WorkflowService:
workflow_run_id: str | None = None,
trigger_type: WorkflowRunTriggerType | None = None,
workflow_schedule_id: str | None = None,
+ ignore_inherited_workflow_system_prompt: bool = False,
) -> WorkflowRun:
"""
Create a workflow run and its parameters. Validate the workflow and the organization. If there are missing
@@ -708,6 +710,7 @@ class WorkflowService:
workflow_run_id=workflow_run_id,
trigger_type=trigger_type,
workflow_schedule_id=workflow_schedule_id,
+ ignore_inherited_workflow_system_prompt=ignore_inherited_workflow_system_prompt,
)
LOG.info(
f"Created workflow run {workflow_run.workflow_run_id} for workflow {workflow.workflow_id}",
@@ -940,6 +943,79 @@ class WorkflowService:
return None
+ async def _collect_inherited_workflow_system_prompt(
+ self,
+ parent_workflow_run_id: str | None,
+ ) -> str | None:
+ """Walk up the parent workflow-run chain and join each ancestor's raw
+ ``workflow_system_prompt`` (outermost first). Returns None when no ancestor
+ has one set. A depth cap matches ``WorkflowTriggerBlock.MAX_TRIGGER_DEPTH``
+ to keep the traversal bounded against malformed chains.
+
+ This reads raw prompt strings from each ancestor's ``workflow_definition``
+ without Jinja rendering — the child's context will render them later via
+ ``WorkflowRunContext.resolve_effective_workflow_system_prompt``. Using raw
+ strings here avoids depending on the parent's live ``WorkflowRunContext``,
+ which isn't available for async/fire-and-forget child runs.
+
+ Chain-break on opt-out: when an ancestor has ``ignore_inherited_workflow_system_prompt``
+ set, its own prompt is still included (it ran without its own ancestors'
+ prompts, but its own rules remain its statement to descendants), but the
+ traversal stops there. A workflow explicitly opting out of its parents'
+ rules creates a clean boundary for itself and everything it triggers —
+ otherwise descendants would silently reintroduce prompts the opted-out
+ workflow rejected.
+ """
+ # Two-phase walk to keep DB round trips bounded. Phase 1 is an
+ # inherently sequential chain walk (each ``parent_workflow_run_id`` is
+ # only known after fetching the previous run), capped at
+ # ``MAX_TRIGGER_DEPTH``. Phase 2 batches the independent workflow-
+ # definition fetches with ``asyncio.gather`` so all N definition
+ # lookups happen in one concurrent burst instead of N sequential
+ # awaits — brings the worst case from 2N round trips down to
+ # N + 1 (depth-bounded at 10). A deeper optimization (single
+ # recursive CTE across workflow_runs + workflows) is possible if
+ # trigger chains ever get deep enough to matter.
+ chain: list[tuple[str, bool]] = [] # [(workflow_id, ignore_inherited), ...] outermost child first
+ current_parent_id: str | None = parent_workflow_run_id
+ visited: set[str] = set()
+ depth = 0
+ while current_parent_id and depth < WorkflowTriggerBlock.MAX_TRIGGER_DEPTH:
+ if current_parent_id in visited:
+ break
+ visited.add(current_parent_id)
+ parent_run = await app.DATABASE.workflow_runs.get_workflow_run(current_parent_id)
+ if parent_run is None:
+ break
+ chain.append((parent_run.workflow_id, parent_run.ignore_inherited_workflow_system_prompt))
+ if parent_run.ignore_inherited_workflow_system_prompt:
+ break
+ current_parent_id = parent_run.parent_workflow_run_id
+ depth += 1
+
+ if not chain:
+ return None
+
+ # Fetch all ancestor workflow definitions concurrently.
+ ancestor_workflows = await asyncio.gather(
+ *(self.get_workflow(workflow_id=workflow_id) for workflow_id, _ in chain),
+ return_exceptions=False,
+ )
+
+ prompts: list[str] = []
+ for workflow in ancestor_workflows:
+ if workflow is None or workflow.workflow_definition is None:
+ continue
+ raw = workflow.workflow_definition.workflow_system_prompt
+ if raw:
+ prompts.append(raw)
+
+ if not prompts:
+ return None
+ # Outermost ancestor first so child-local rules appear after broader rules.
+ prompts.reverse()
+ return "\n\n".join(prompts)
+
@traced(name="skyvern.workflow.execute", role="wrapper")
async def execute_workflow(
self,
@@ -951,7 +1027,15 @@ class WorkflowService:
browser_session_id: str | None = None,
need_call_webhook: bool = True,
) -> WorkflowRun:
- """Execute a workflow."""
+ """Execute a workflow.
+
+ When the workflow_run row has ``ignore_inherited_workflow_system_prompt``
+ set (populated at spawn time by a ``WorkflowTriggerBlock`` whose
+ ``ignore_workflow_system_prompt`` flag is True), the child workflow
+ starts with a clean slate — no inherited prompt from the ancestor
+ chain. Persisting the intent on the row means the flag is honored for
+ both sync and async (Temporal-dispatched) trigger modes.
+ """
organization_id = organization.organization_id
LOG.info(
@@ -1008,6 +1092,19 @@ class WorkflowService:
# Get all <workflow parameter, workflow run parameter> tuples
wp_wps_tuples = await self.get_workflow_run_parameter_tuples(workflow_run_id=workflow_run_id)
workflow_output_parameters = await self.get_workflow_output_parameters(workflow_id=workflow.workflow_id)
+ # Collect resolved workflow_system_prompt from every ancestor workflow so child
+ # blocks inherit them (SKY-9147). We read each parent's workflow_definition from
+ # the DB because the parent's in-memory WorkflowRunContext may be gone by the
+ # time a fire-and-forget child runs on its own worker. Jinja placeholders in
+ # ancestor prompts are rendered against this run's values; parent-only
+ # parameters will simply render empty in non-strict mode.
+ inherited_workflow_system_prompt = (
+ None
+ if workflow_run.ignore_inherited_workflow_system_prompt
+ else await self._collect_inherited_workflow_system_prompt(
+ parent_workflow_run_id=workflow_run.parent_workflow_run_id,
+ )
+ )
try:
await app.WORKFLOW_CONTEXT_MANAGER.initialize_workflow_run_context(
organization,
@@ -1021,6 +1118,7 @@ class WorkflowService:
secret_parameters,
block_outputs,
workflow,
+ inherited_workflow_system_prompt=inherited_workflow_system_prompt,
)
except Exception as e:
LOG.exception(
@@ -2121,6 +2219,22 @@ class WorkflowService:
block_label=block.label,
run_signature=script_block.run_signature,
)
+ # Script path skips the block's own execute() (which is where
+ # format_potential_template_parameters runs in the agent path),
+ # so we apply the workflow_system_prompt here to thread the
+ # block-resolved value into the ``WorkflowRunContext`` cache.
+ # ``ai_extract`` reads from that cache so the script-generated
+ # extraction honors ``ignore_workflow_system_prompt`` the same
+ # way the agent path does — same string, same cache key.
+ try:
+ workflow_run_context = app.WORKFLOW_CONTEXT_MANAGER.get_workflow_run_context(workflow_run_id)
+ block._apply_workflow_system_prompt(workflow_run_context)
+ except Exception:
+ LOG.warning(
+ "Failed to apply workflow_system_prompt for script-path block; continuing",
+ block_label=block.label,
+ exc_info=True,
+ )
block_exec_start = time.monotonic()
try:
vars_dict = vars(loaded_script_module) if loaded_script_module else {}
@@ -3750,6 +3864,7 @@ class WorkflowService:
workflow_run_id: str | None = None,
trigger_type: WorkflowRunTriggerType | None = None,
workflow_schedule_id: str | None = None,
+ ignore_inherited_workflow_system_prompt: bool = False,
) -> WorkflowRun:
# validate the browser session or profile id
browser_profile_id = workflow_request.browser_profile_id
@@ -3840,6 +3955,7 @@ class WorkflowService:
workflow_run_id=workflow_run_id,
trigger_type=trigger_type,
workflow_schedule_id=workflow_schedule_id,
+ ignore_inherited_workflow_system_prompt=ignore_inherited_workflow_system_prompt,
)
async def _update_workflow_run_status(
diff --git a/skyvern/forge/sdk/workflow/workflow_definition_converter.py b/skyvern/forge/sdk/workflow/workflow_definition_converter.py
index 63718c7fe..1df8c0a5b 100644
--- a/skyvern/forge/sdk/workflow/workflow_definition_converter.py
+++ b/skyvern/forge/sdk/workflow/workflow_definition_converter.py
@@ -327,6 +327,7 @@ def convert_workflow_definition(
version=dag_version,
finally_block_label=workflow_definition_yaml.finally_block_label,
error_code_mapping=workflow_definition_yaml.error_code_mapping,
+ workflow_system_prompt=workflow_definition_yaml.workflow_system_prompt,
)
LOG.info(
@@ -383,6 +384,7 @@ def _build_block_kwargs(
"continue_on_failure": block_yaml.continue_on_failure,
"next_loop_on_failure": block_yaml.next_loop_on_failure,
"model": block_yaml.model,
+ "ignore_workflow_system_prompt": block_yaml.ignore_workflow_system_prompt,
}
diff --git a/skyvern/library/skyvern_browser_page_ai.py b/skyvern/library/skyvern_browser_page_ai.py
index f9693df12..ac26533e1 100644
--- a/skyvern/library/skyvern_browser_page_ai.py
+++ b/skyvern/library/skyvern_browser_page_ai.py
@@ -15,7 +15,7 @@ from skyvern.client import (
RunSdkActionRequestAction_Validate,
)
from skyvern.config import settings
-from skyvern.core.script_generations.skyvern_page_ai import SkyvernPageAi
+from skyvern.core.script_generations.skyvern_page_ai import SYSTEM_PROMPT_UNSET, SkyvernPageAi
if TYPE_CHECKING:
from skyvern.library.skyvern_browser import SkyvernBrowser
@@ -168,13 +168,15 @@ class SdkSkyvernPageAi(SkyvernPageAi):
data: str | dict[str, Any] | None = None,
skip_refresh: bool = False,
include_extracted_text: bool = True,
+ system_prompt: str | None | Any = SYSTEM_PROMPT_UNSET,
) -> dict[str, Any] | list | str | None:
"""Extract information from the page using AI via API call.
- Note: skip_refresh and include_extracted_text are accepted for Protocol
- compatibility but not forwarded to the API. The server-side controls
- both via the Task record on the SDK HTTP path. The optimizations only
- take effect on the direct RealSkyvernPageAI path (MCP local browser).
+ Note: skip_refresh, include_extracted_text, and system_prompt are
+ accepted for Protocol compatibility but not forwarded to the API. The
+ server-side controls them via the Task record on the SDK HTTP path.
+ The optimizations only take effect on the direct RealSkyvernPageAI
+ path (MCP local browser).
"""
LOG.info("AI extract", prompt=prompt, workflow_run_id=self._browser.workflow_run_id)
diff --git a/skyvern/schemas/workflows.py b/skyvern/schemas/workflows.py
index 9d903071f..47cd945ce 100644
--- a/skyvern/schemas/workflows.py
+++ b/skyvern/schemas/workflows.py
@@ -321,6 +321,16 @@ def sanitize_workflow_yaml_with_references(workflow_yaml: dict[str, Any]) -> dic
workflow_definition["parameters"], old_output_key, new_output_key
)
+ # workflow_system_prompt is rendered through Jinja at execution time, so
+ # references inside it need the same rename treatment as block fields.
+ if isinstance(workflow_definition.get("workflow_system_prompt"), str):
+ workflow_definition["workflow_system_prompt"] = _replace_references_in_value(
+ workflow_definition["workflow_system_prompt"], old_output_key, new_output_key
+ )
+ workflow_definition["workflow_system_prompt"] = _replace_references_in_value(
+ workflow_definition["workflow_system_prompt"], old_label, new_label
+ )
+
# Step 4: Update all parameter key references
for old_key, new_key in param_key_mapping.items():
# Update Jinja references in blocks (e.g., {{ old_key }})
@@ -339,6 +349,11 @@ def sanitize_workflow_yaml_with_references(workflow_yaml: dict[str, Any]) -> dic
workflow_definition["parameters"], old_key, new_key
)
+ if isinstance(workflow_definition.get("workflow_system_prompt"), str):
+ workflow_definition["workflow_system_prompt"] = _replace_references_in_value(
+ workflow_definition["workflow_system_prompt"], old_key, new_key
+ )
+
# Rewrite workflow-level error_code_mapping atomically so substitutions don't chain
# (e.g. foo-bar -> foo_bar and foo_bar -> foo_bar_2 must not combine into foo_bar_2).
if "error_code_mapping" in workflow_definition:
@@ -614,6 +629,10 @@ class BlockYAML(BaseModel, abc.ABC):
)
continue_on_failure: bool = False
model: dict[str, Any] | None = None
+ # Opt-out from workflow-level workflow_system_prompt inheritance (and, on a
+ # WorkflowTriggerBlock, from propagating the parent chain's prompt into the
+ # spawned child run). A no-op for deterministic blocks that don't call an LLM.
+ ignore_workflow_system_prompt: bool = False
# Only valid for blocks inside a for loop block
# Whether to continue to the next iteration when the block fails
next_loop_on_failure: bool = False
@@ -1105,6 +1124,7 @@ class WorkflowDefinitionYAML(BaseModel):
blocks: list[BLOCK_YAML_TYPES]
finally_block_label: str | None = None
error_code_mapping: dict[str, str] | None = None
+ workflow_system_prompt: str | None = None
@model_validator(mode="after")
def validate_unique_block_labels(self) -> "WorkflowDefinitionYAML":
diff --git a/skyvern/services/task_v2_service.py b/skyvern/services/task_v2_service.py
index 01d0c3120..f1aa5a8ed 100644
--- a/skyvern/services/task_v2_service.py
+++ b/skyvern/services/task_v2_service.py
@@ -145,6 +145,7 @@ async def _summarize_max_steps_failure_reason(
screenshots=screenshots,
prompt_name="task_v2_summarize-max-steps-reason",
thought=thought,
+ system_prompt=task_v2.workflow_system_prompt,
)
return json_response.get("reasoning", ""), json_response.get("failure_categories")
except Exception:
@@ -252,6 +253,7 @@ async def initialize_task_v2(
parent_workflow_run_id: str | None = None,
extracted_information_schema: dict | list | str | None = None,
error_code_mapping: dict | None = None,
+ workflow_system_prompt: str | None = None,
create_task_run: bool = False,
model: dict[str, Any] | None = None,
max_screenshot_scrolling_times: int | None = None,
@@ -270,6 +272,7 @@ async def initialize_task_v2(
proxy_location=proxy_location,
extracted_information_schema=extracted_information_schema,
error_code_mapping=error_code_mapping,
+ workflow_system_prompt=workflow_system_prompt,
model=model,
max_screenshot_scrolling_times=max_screenshot_scrolling_times,
extra_http_headers=extra_http_headers,
@@ -381,6 +384,7 @@ async def initialize_task_v2_metadata(
prompt=metadata_prompt,
thought=thought,
prompt_name="task_v2_generate_metadata",
+ system_prompt=task_v2.workflow_system_prompt,
)
# validate
@@ -829,6 +833,7 @@ async def run_task_v2_helper(
screenshots=scraped_page.screenshots,
thought=thought,
prompt_name="task_v2",
+ system_prompt=task_v2.workflow_system_prompt,
)
LOG.info(
"Task v2 response",
@@ -1079,6 +1084,7 @@ async def run_task_v2_helper(
screenshots=completion_screenshots,
thought=thought,
prompt_name="task_v2_check_completion",
+ system_prompt=task_v2.workflow_system_prompt,
)
LOG.info(
"Task v2 completion check response",
@@ -1465,6 +1471,7 @@ async def _generate_loop_task(
screenshots=scraped_page.screenshots,
thought=thought_task_in_loop,
prompt_name="task_v2_generate_task_block",
+ system_prompt=task_v2.workflow_system_prompt,
)
LOG.info("Task in loop metadata response", task_in_loop_metadata_response=task_in_loop_metadata_response)
navigation_goal = task_in_loop_metadata_response.get("navigation_goal")
@@ -1571,6 +1578,7 @@ async def _generate_extraction_task(
task_v2=task_v2,
prompt_name="task_v2_generate_extraction_task",
organization_id=task_v2.organization_id,
+ system_prompt=task_v2.workflow_system_prompt,
)
break
except (InvalidLLMResponseFormat, EmptyLLMResponseError) as e:
@@ -1971,6 +1979,7 @@ async def _summarize_task_v2(
screenshots=screenshots,
thought=thought,
prompt_name="task_v2_summary",
+ system_prompt=task_v2.workflow_system_prompt,
)
LOG.info("Task v2 summary response", task_v2_summary_resp=task_v2_summary_resp)
diff --git a/skyvern/services/workflow_service.py b/skyvern/services/workflow_service.py
index 033b11d53..6c1e793df 100644
--- a/skyvern/services/workflow_service.py
+++ b/skyvern/services/workflow_service.py
@@ -27,6 +27,7 @@ async def prepare_workflow(
code_gen: bool | None = None,
parent_workflow_run_id: str | None = None,
trigger_type: WorkflowRunTriggerType | None = None,
+ ignore_inherited_workflow_system_prompt: bool = False,
) -> WorkflowRun:
"""
Prepare a workflow to be run.
@@ -47,6 +48,7 @@ async def prepare_workflow(
code_gen=code_gen,
parent_workflow_run_id=parent_workflow_run_id,
trigger_type=trigger_type,
+ ignore_inherited_workflow_system_prompt=ignore_inherited_workflow_system_prompt,
)
workflow = await app.WORKFLOW_SERVICE.get_workflow_by_permanent_id(
@@ -87,6 +89,7 @@ async def run_workflow(
block_outputs: dict[str, t.Any] | None = None,
parent_workflow_run_id: str | None = None,
trigger_type: WorkflowRunTriggerType | None = None,
+ ignore_inherited_workflow_system_prompt: bool = False,
) -> WorkflowRun:
workflow_run = await prepare_workflow(
workflow_id=workflow_id,
@@ -98,6 +101,7 @@ async def run_workflow(
request_id=request_id,
parent_workflow_run_id=parent_workflow_run_id,
trigger_type=trigger_type,
+ ignore_inherited_workflow_system_prompt=ignore_inherited_workflow_system_prompt,
)
await AsyncExecutorFactory.get_executor().execute_workflow(
diff --git a/skyvern/webeye/actions/handler.py b/skyvern/webeye/actions/handler.py
index 1ba7ed098..9dd9314a3 100644
--- a/skyvern/webeye/actions/handler.py
+++ b/skyvern/webeye/actions/handler.py
@@ -4463,6 +4463,7 @@ async def extract_information_for_navigation_goal(
error_code_mapping=error_code_mapping_str,
previous_extracted_information=post_ceiling_kwargs["previous_extracted_information"],
llm_key=llm_key_override,
+ workflow_system_prompt=task.workflow_system_prompt,
)
if is_retry_step:
# Proactively evict the in-run entry. The cross-run tier will be
@@ -4548,6 +4549,7 @@ async def extract_information_for_navigation_goal(
# reasons unrelated to cache correctness.
prompt_name="extract-information",
force_dict=False,
+ system_prompt=task.workflow_system_prompt,
)
# Apply the same post-processing the miss path applies so the
# comparison is apples-to-apples against the cached value.
@@ -4696,6 +4698,7 @@ async def extract_information_for_navigation_goal(
screenshots=scraped_page.screenshots,
prompt_name="extract-information",
force_dict=False,
+ system_prompt=task.workflow_system_prompt,
)
# Validate and fill missing fields based on schema
diff --git a/tests/unit/test_ai_extract_system_prompt_optout.py b/tests/unit/test_ai_extract_system_prompt_optout.py
new file mode 100644
index 000000000..70ec83973
--- /dev/null
+++ b/tests/unit/test_ai_extract_system_prompt_optout.py
@@ -0,0 +1,263 @@
+"""Regression tests: script-path ``ai_extract`` must honor the current block's
+``ignore_workflow_system_prompt`` opt-out, matching agent-path behavior.
+
+Without this, a cached script-path extraction still injects the workflow prompt
+even when the block has opted out — the two execution modes diverge for the
+same block (SKY-9147).
+"""
+
+from __future__ import annotations
+
+import asyncio
+from datetime import datetime, timezone
+from typing import Any
+from unittest.mock import MagicMock
+
+import pytest
+
+from skyvern.core.script_generations import real_skyvern_page_ai as module
+from skyvern.core.script_generations.skyvern_page_ai import SYSTEM_PROMPT_UNSET
+from skyvern.forge.sdk.workflow.models.block import ExtractionBlock
+from skyvern.forge.sdk.workflow.models.parameter import OutputParameter, ParameterType
+from skyvern.forge.sdk.workflow.models.workflow import Workflow, WorkflowDefinition
+
+
+def _make_output_parameter() -> OutputParameter:
+ now = datetime.now(timezone.utc)
+ return OutputParameter(
+ parameter_type=ParameterType.OUTPUT,
+ key="extract1_output",
+ description="test output",
+ output_parameter_id="op_extract1",
+ workflow_id="w_test",
+ created_at=now,
+ modified_at=now,
+ )
+
+
+def _make_workflow_with_block(*, ignore_workflow_system_prompt: bool, workflow_system_prompt: str | None) -> Workflow:
+ block = ExtractionBlock(
+ label="extract1",
+ output_parameter=_make_output_parameter(),
+ data_extraction_goal="Extract things",
+ ignore_workflow_system_prompt=ignore_workflow_system_prompt,
+ )
+ now = datetime.now(timezone.utc)
+ return Workflow(
+ workflow_id="w_test",
+ organization_id="o_test",
+ title="test",
+ workflow_permanent_id="wpid_test",
+ version=1,
+ is_saved_task=False,
+ workflow_definition=WorkflowDefinition(
+ parameters=[],
+ blocks=[block],
+ workflow_system_prompt=workflow_system_prompt,
+ ),
+ created_at=now,
+ modified_at=now,
+ )
+
+
+def _run_ai_extract(
+ monkeypatch: pytest.MonkeyPatch,
+ *,
+ workflow: Workflow | None,
+ current_label: str | None,
+ extra_kwargs: dict[str, Any] | None = None,
+) -> dict[str, Any]:
+ """Drive ``RealSkyvernPageAi.ai_extract`` with mocks that capture both the
+ cache-key ``system_prompt`` and the LLM-handler ``system_prompt``. Returns
+ a dict with keys ``cache_system_prompt`` and ``llm_system_prompt``."""
+
+ captured: dict[str, Any] = {}
+
+ def fake_compute_cache_key(**kwargs: Any) -> str:
+ captured["cache_system_prompt"] = kwargs.get("workflow_system_prompt")
+ return "fake-cache-key"
+
+ def fake_lookup(*_args: Any, **_kwargs: Any) -> None:
+ # Cache miss — let the flow fall through to the handler.
+ return None
+
+ def fake_load_prompt(**_kwargs: Any) -> tuple[str, dict[str, Any]]:
+ return "rendered-prompt", {
+ "extracted_text": None,
+ "extracted_information_schema": None,
+ }
+
+ async def fake_handler(*, system_prompt: Any = None, **_ignored: Any) -> dict[str, Any]:
+ captured["llm_system_prompt"] = system_prompt
+ return {}
+
+ # Build a minimal fake ``WorkflowRunContext``. The single-block workflows
+ # this helper builds go through ``_apply_workflow_system_prompt`` in the
+ # real script-path dispatch, so simulate that by pre-recording the block's
+ # effective value — ``None`` when the block opts out, the rendered prompt
+ # otherwise. ``ai_extract`` reads this recorded value verbatim, which is
+ # exactly the contract Andrew's fix enforces.
+ if workflow is not None:
+ block = workflow.workflow_definition.blocks[0] if workflow.workflow_definition.blocks else None
+ workflow_prompt = workflow.workflow_definition.workflow_system_prompt
+ recorded_value: str | None = (
+ None if (block is not None and getattr(block, "ignore_workflow_system_prompt", False)) else workflow_prompt
+ )
+ has_block_record = block is not None and current_label == block.label
+ ctx = MagicMock()
+ ctx.workflow = workflow
+ ctx.resolve_effective_workflow_system_prompt = MagicMock(return_value=workflow_prompt)
+ ctx.get_block_workflow_system_prompt = MagicMock(
+ return_value=(has_block_record, recorded_value),
+ )
+
+ def get_run_ctx(_run_id: str) -> Any:
+ return ctx
+
+ monkeypatch.setattr(
+ module.app.WORKFLOW_CONTEXT_MANAGER,
+ "get_workflow_run_context",
+ get_run_ctx,
+ )
+
+ skyvern_ctx = MagicMock()
+ skyvern_ctx.workflow_run_id = "wr_test"
+ skyvern_ctx.tz_info = None
+ skyvern_ctx.organization_id = None
+ skyvern_ctx.task_id = None
+ skyvern_ctx.step_id = None
+ skyvern_ctx.script_mode = False
+
+ monkeypatch.setattr(module.skyvern_context, "current", lambda: skyvern_ctx)
+ monkeypatch.setattr(module, "load_prompt_with_elements_tracked", fake_load_prompt)
+ monkeypatch.setattr(module.extraction_cache, "compute_cache_key", fake_compute_cache_key)
+ monkeypatch.setattr(module.extraction_cache, "lookup", fake_lookup)
+ monkeypatch.setattr(module.app, "EXTRACTION_LLM_API_HANDLER", fake_handler)
+
+ scraped_page = MagicMock()
+ scraped_page.url = "https://example.test"
+ scraped_page.extracted_text = "page text"
+ scraped_page.screenshots = []
+ scraped_page.build_element_tree = MagicMock(return_value="<a>link</a>")
+ scraped_page.support_economy_elements_tree = MagicMock(return_value=False)
+ scraped_page.last_used_element_tree_html = None
+
+ page = module.RealSkyvernPageAi.__new__(module.RealSkyvernPageAi)
+ page.scraped_page = scraped_page
+ page.current_label = current_label
+
+ async def fake_refresh(*_args: Any, **_kwargs: Any) -> None:
+ return None
+
+ monkeypatch.setattr(page, "_refresh_scraped_page", fake_refresh)
+
+ asyncio.run(
+ page.ai_extract(
+ prompt="Extract things",
+ schema={"type": "object"},
+ **(extra_kwargs or {}),
+ )
+ )
+
+ return captured
+
+
+# ---------------------------------------------------------------------------
+# The bug Andrew flagged: block opts out, but script-path still injects prompt.
+# ---------------------------------------------------------------------------
+
+
+def test_opted_out_block_sends_no_system_prompt_to_llm(monkeypatch: pytest.MonkeyPatch) -> None:
+ workflow = _make_workflow_with_block(
+ ignore_workflow_system_prompt=True,
+ workflow_system_prompt="WORKFLOW RULES.",
+ )
+ captured = _run_ai_extract(monkeypatch, workflow=workflow, current_label="extract1")
+
+ assert captured["llm_system_prompt"] is None
+
+
+def test_opted_out_block_cache_key_omits_system_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
+ """Cache-key parity with agent path: an opted-out block's extractions must
+ hash the same way whether the workflow has a prompt set or not. Otherwise a
+ user toggling ``workflow_system_prompt`` at the workflow level silently
+ invalidates caches for blocks that explicitly opted out."""
+ workflow = _make_workflow_with_block(
+ ignore_workflow_system_prompt=True,
+ workflow_system_prompt="WORKFLOW RULES.",
+ )
+ captured = _run_ai_extract(monkeypatch, workflow=workflow, current_label="extract1")
+
+ assert captured["cache_system_prompt"] is None
+
+
+# ---------------------------------------------------------------------------
+# Non-opted-out block still inherits the workflow prompt (no regression of the
+# base feature).
+# ---------------------------------------------------------------------------
+
+
+def test_non_opted_out_block_receives_workflow_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
+ workflow = _make_workflow_with_block(
+ ignore_workflow_system_prompt=False,
+ workflow_system_prompt="WORKFLOW RULES.",
+ )
+ captured = _run_ai_extract(monkeypatch, workflow=workflow, current_label="extract1")
+
+ assert captured["llm_system_prompt"] == "WORKFLOW RULES."
+ assert captured["cache_system_prompt"] == "WORKFLOW RULES."
+
+
+# ---------------------------------------------------------------------------
+# Explicit parameter overrides: caller can pass ``system_prompt`` directly,
+# bypassing the block-flag lookup. Includes the "explicit None" escape hatch.
+# ---------------------------------------------------------------------------
+
+
+def test_explicit_system_prompt_overrides_workflow(monkeypatch: pytest.MonkeyPatch) -> None:
+ workflow = _make_workflow_with_block(
+ ignore_workflow_system_prompt=False,
+ workflow_system_prompt="WORKFLOW RULES.",
+ )
+ captured = _run_ai_extract(
+ monkeypatch,
+ workflow=workflow,
+ current_label="extract1",
+ extra_kwargs={"system_prompt": "EXPLICIT RULES."},
+ )
+
+ assert captured["llm_system_prompt"] == "EXPLICIT RULES."
+ assert captured["cache_system_prompt"] == "EXPLICIT RULES."
+
+
+def test_explicit_none_opts_out_even_when_workflow_has_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
+ """Explicit ``None`` is the escape hatch: caller says "send no system prompt"
+ regardless of the workflow or block state."""
+ workflow = _make_workflow_with_block(
+ ignore_workflow_system_prompt=False,
+ workflow_system_prompt="WORKFLOW RULES.",
+ )
+ captured = _run_ai_extract(
+ monkeypatch,
+ workflow=workflow,
+ current_label="extract1",
+ extra_kwargs={"system_prompt": None},
+ )
+
+ assert captured["llm_system_prompt"] is None
+ assert captured["cache_system_prompt"] is None
+
+
+def test_sentinel_default_is_not_leaked_to_handler(monkeypatch: pytest.MonkeyPatch) -> None:
+ """The sentinel must never reach the LLM handler — if the fallback path
+ silently forwarded it, the handler would see a non-None, non-string object
+ and fail or pass a bogus ``system_prompt`` to the model."""
+ workflow = _make_workflow_with_block(
+ ignore_workflow_system_prompt=False,
+ workflow_system_prompt=None, # no prompt at all
+ )
+ captured = _run_ai_extract(monkeypatch, workflow=workflow, current_label="extract1")
+
+ assert captured["llm_system_prompt"] is None
+ assert captured["llm_system_prompt"] is not SYSTEM_PROMPT_UNSET
+ assert captured["cache_system_prompt"] is None
diff --git a/tests/unit/test_block_workflow_system_prompt_inheritance.py b/tests/unit/test_block_workflow_system_prompt_inheritance.py
new file mode 100644
index 000000000..ae6e12302
--- /dev/null
+++ b/tests/unit/test_block_workflow_system_prompt_inheritance.py
@@ -0,0 +1,489 @@
+"""Tests for workflow-level workflow_system_prompt inheritance into blocks at execution time."""
+
+from datetime import datetime, timezone
+from unittest.mock import MagicMock
+
+from skyvern.forge.sdk.workflow.context_manager import WorkflowRunContext
+from skyvern.forge.sdk.workflow.models.block import (
+ FileParserBlock,
+ PDFParserBlock,
+ TaskBlock,
+ TaskV2Block,
+ TextPromptBlock,
+)
+from skyvern.forge.sdk.workflow.models.parameter import OutputParameter, ParameterType
+from skyvern.forge.sdk.workflow.models.workflow import Workflow, WorkflowDefinition
+from skyvern.schemas.workflows import FileType
+
+
+def _make_output_parameter() -> OutputParameter:
+ now = datetime.now(timezone.utc)
+ return OutputParameter(
+ parameter_type=ParameterType.OUTPUT,
+ key="task1_output",
+ description="test output",
+ output_parameter_id="op_task1",
+ workflow_id="w_test",
+ created_at=now,
+ modified_at=now,
+ )
+
+
+def _make_task_block(workflow_system_prompt: str | None = None) -> TaskBlock:
+ return TaskBlock(
+ label="task1",
+ output_parameter=_make_output_parameter(),
+ title="task title",
+ workflow_system_prompt=workflow_system_prompt,
+ )
+
+
+def _make_task_v2_block(workflow_system_prompt: str | None = None) -> TaskV2Block:
+ return TaskV2Block(
+ label="task1",
+ output_parameter=_make_output_parameter(),
+ prompt="user goal",
+ workflow_system_prompt=workflow_system_prompt,
+ )
+
+
+def _make_workflow(workflow_system_prompt: str | None) -> Workflow:
+ workflow_definition = WorkflowDefinition(
+ parameters=[],
+ blocks=[],
+ workflow_system_prompt=workflow_system_prompt,
+ )
+ now = datetime.now(timezone.utc)
+ return Workflow(
+ workflow_id="w_test",
+ organization_id="o_test",
+ title="test",
+ workflow_permanent_id="wpid_test",
+ version=1,
+ is_saved_task=False,
+ workflow_definition=workflow_definition,
+ created_at=now,
+ modified_at=now,
+ )
+
+
+def _make_workflow_run_context(
+ workflow_system_prompt: str | None,
+ inherited_workflow_system_prompt: str | None = None,
+) -> WorkflowRunContext:
+ ctx = WorkflowRunContext(
+ workflow_title="test",
+ workflow_id="w_test",
+ workflow_permanent_id="wpid_test",
+ workflow_run_id="wr_test",
+ aws_client=MagicMock(),
+ workflow=_make_workflow(workflow_system_prompt),
+ inherited_workflow_system_prompt=inherited_workflow_system_prompt,
+ )
+ return ctx
+
+
+class TestTaskBlockSystemPromptInheritance:
+ def test_block_inherits_workflow_prompt_when_none(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Never guess. If unsure, say UNKNOWN.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Never guess. If unsure, say UNKNOWN."
+
+ def test_both_none_stays_none(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context(workflow_system_prompt=None)
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_jinja_substitution_resolves_workflow_parameters(self) -> None:
+ """Global system prompt should support Jinja substitution against workflow parameters."""
+ block = _make_task_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Respond in the style of {{ style }}.")
+ ctx.values["style"] = "a formal English butler"
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Respond in the style of a formal English butler."
+
+
+class TestTaskV2BlockSystemPromptInheritance:
+ def test_block_inherits_workflow_prompt_when_none(self) -> None:
+ block = _make_task_v2_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Never guess.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Never guess."
+
+ def test_both_none_stays_none(self) -> None:
+ block = _make_task_v2_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context(workflow_system_prompt=None)
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+
+def _make_text_prompt_block(workflow_system_prompt: str | None = None) -> TextPromptBlock:
+ return TextPromptBlock(
+ label="prompt1",
+ output_parameter=_make_output_parameter(),
+ prompt="what is 2 + 2?",
+ workflow_system_prompt=workflow_system_prompt,
+ )
+
+
+def _make_file_parser_block(workflow_system_prompt: str | None = None) -> FileParserBlock:
+ return FileParserBlock(
+ label="fileparser1",
+ output_parameter=_make_output_parameter(),
+ file_url="https://example.com/file.csv",
+ file_type=FileType.CSV,
+ workflow_system_prompt=workflow_system_prompt,
+ )
+
+
+def _make_pdf_parser_block(workflow_system_prompt: str | None = None) -> PDFParserBlock:
+ return PDFParserBlock(
+ label="pdfparser1",
+ output_parameter=_make_output_parameter(),
+ file_url="https://example.com/file.pdf",
+ workflow_system_prompt=workflow_system_prompt,
+ )
+
+
+class TestTextPromptBlockSystemPromptInheritance:
+ def test_block_inherits_workflow_prompt_when_none(self) -> None:
+ block = _make_text_prompt_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Answer only in Spanish.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Answer only in Spanish."
+
+ def test_both_none_stays_none(self) -> None:
+ block = _make_text_prompt_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context(workflow_system_prompt=None)
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_jinja_substitution_resolves_workflow_parameters(self) -> None:
+ block = _make_text_prompt_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Respond in the style of {{ style }}.")
+ ctx.values["style"] = "a pirate"
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Respond in the style of a pirate."
+
+
+class TestFileParserBlockSystemPromptInheritance:
+ def test_block_inherits_workflow_prompt_when_none(self) -> None:
+ block = _make_file_parser_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Only respond with structured data.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Only respond with structured data."
+
+ def test_both_none_stays_none(self) -> None:
+ block = _make_file_parser_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context(workflow_system_prompt=None)
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+
+class TestPDFParserBlockSystemPromptInheritance:
+ def test_block_inherits_workflow_prompt_when_none(self) -> None:
+ block = _make_pdf_parser_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Summarize in English.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Summarize in English."
+
+ def test_both_none_stays_none(self) -> None:
+ block = _make_pdf_parser_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context(workflow_system_prompt=None)
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+
+class TestChildWorkflowInheritsParentWorkflowSystemPrompt:
+ """SKY-9147: parent workflow_trigger workflow_system_prompt must flow into child blocks."""
+
+ def test_child_inherits_parent_when_child_unset(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context(
+ workflow_system_prompt=None,
+ inherited_workflow_system_prompt="Omit the word 'not'.",
+ )
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Omit the word 'not'."
+
+ def test_child_concatenates_parent_and_own_prompt(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context(
+ workflow_system_prompt="Respond in French.",
+ inherited_workflow_system_prompt="Omit the word 'not'.",
+ )
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Omit the word 'not'.\n\nRespond in French."
+
+ def test_ignore_flag_drops_both_inherited_and_own(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ block.ignore_workflow_system_prompt = True
+ ctx = _make_workflow_run_context(
+ workflow_system_prompt="Respond in French.",
+ inherited_workflow_system_prompt="Omit the word 'not'.",
+ )
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_inherited_only_with_no_workflow_attached(self) -> None:
+ """Child context may carry inherited rules even when workflow is not hydrated."""
+ block = _make_text_prompt_block(workflow_system_prompt=None)
+ ctx = WorkflowRunContext(
+ workflow_title="child",
+ workflow_id="w_child",
+ workflow_permanent_id="wpid_child",
+ workflow_run_id="wr_child",
+ aws_client=MagicMock(),
+ workflow=None,
+ inherited_workflow_system_prompt="Be concise.",
+ )
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Be concise."
+
+
+class TestIgnoreWorkflowSystemPromptPerBlock:
+ """SKY-9147: per-block opt-out short-circuits inheritance across every LLM-consuming block."""
+
+ def test_task_block_opt_out_skips_workflow_prompt(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ block.ignore_workflow_system_prompt = True
+ ctx = _make_workflow_run_context("Be concise.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_task_v2_block_opt_out_skips_workflow_prompt(self) -> None:
+ block = _make_task_v2_block(workflow_system_prompt=None)
+ block.ignore_workflow_system_prompt = True
+ ctx = _make_workflow_run_context("Be concise.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_text_prompt_block_opt_out_skips_workflow_prompt(self) -> None:
+ block = _make_text_prompt_block(workflow_system_prompt=None)
+ block.ignore_workflow_system_prompt = True
+ ctx = _make_workflow_run_context("Answer only in Spanish.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_file_parser_block_opt_out_skips_workflow_prompt(self) -> None:
+ block = _make_file_parser_block(workflow_system_prompt=None)
+ block.ignore_workflow_system_prompt = True
+ ctx = _make_workflow_run_context("Only respond with structured data.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_pdf_parser_block_opt_out_skips_workflow_prompt(self) -> None:
+ block = _make_pdf_parser_block(workflow_system_prompt=None)
+ block.ignore_workflow_system_prompt = True
+ ctx = _make_workflow_run_context("Summarize in English.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt is None
+
+ def test_opt_out_default_false_preserves_inheritance(self) -> None:
+ """Regression guard: omitting the field leaves default inheritance behavior intact."""
+ block = _make_task_block(workflow_system_prompt=None)
+ assert block.ignore_workflow_system_prompt is False
+ ctx = _make_workflow_run_context("Be concise.")
+
+ block.format_potential_template_parameters(ctx)
+
+ assert block.workflow_system_prompt == "Be concise."
+
+
+class TestWorkflowTriggerPersistsOptOutOnChildRun:
+ """SKY-9147: the trigger-block flag is persisted on the spawned child's
+ workflow_run row so both sync and async (Temporal-dispatched) child
+ executions honor it uniformly when they read their own row.
+ """
+
+ def test_workflow_run_has_skip_inherited_field(self) -> None:
+ """WorkflowRun Pydantic model carries the persisted flag."""
+ now = datetime.now(timezone.utc)
+ from skyvern.forge.sdk.workflow.models.workflow import WorkflowRun, WorkflowRunStatus
+
+ run = WorkflowRun(
+ workflow_run_id="wr_child",
+ workflow_id="w_child",
+ workflow_permanent_id="wpid_child",
+ organization_id="o_test",
+ status=WorkflowRunStatus.created,
+ created_at=now,
+ modified_at=now,
+ ignore_inherited_workflow_system_prompt=True,
+ )
+
+ assert run.ignore_inherited_workflow_system_prompt is True
+
+ def test_workflow_run_defaults_false(self) -> None:
+ """Existing code paths that omit the field continue to inherit."""
+ now = datetime.now(timezone.utc)
+ from skyvern.forge.sdk.workflow.models.workflow import WorkflowRun, WorkflowRunStatus
+
+ run = WorkflowRun(
+ workflow_run_id="wr_child",
+ workflow_id="w_child",
+ workflow_permanent_id="wpid_child",
+ organization_id="o_test",
+ status=WorkflowRunStatus.created,
+ created_at=now,
+ modified_at=now,
+ )
+
+ assert run.ignore_inherited_workflow_system_prompt is False
+
+
+class TestWorkflowDefinitionYAMLRoundTrip:
+ def test_workflow_system_prompt_survives_yaml_roundtrip(self) -> None:
+ """Regression guard: workflow_system_prompt must roundtrip through WorkflowCreateYAMLRequest."""
+ from skyvern.schemas.workflows import WorkflowCreateYAMLRequest
+
+ payload = {
+ "title": "test",
+ "workflow_definition": {
+ "version": 1,
+ "parameters": [],
+ "blocks": [],
+ "workflow_system_prompt": "Never guess.",
+ },
+ }
+ request = WorkflowCreateYAMLRequest.model_validate(payload)
+ assert request.workflow_definition.workflow_system_prompt == "Never guess."
+
+
+class TestBlockWorkflowSystemPromptNotSerialized:
+ """The runtime cache must not leak through ``model_dump`` or JSON
+ serialization — it's a per-run transient, not part of the block's
+ authored shape."""
+
+ def test_unset_field_absent_from_model_dump(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ assert "workflow_system_prompt" not in block.model_dump()
+
+ def test_resolved_runtime_value_absent_from_model_dump(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Never guess.")
+ block.format_potential_template_parameters(ctx)
+
+ # Invariant confirmed at runtime: inheritance populated the cache…
+ assert block.workflow_system_prompt == "Never guess."
+ # …but it still doesn't escape via serialization.
+ dumped = block.model_dump()
+ assert "workflow_system_prompt" not in dumped
+ assert "Never guess." not in block.model_dump_json()
+
+
+class TestBlockWorkflowSystemPromptRecordedOnContext:
+ """``Block._apply_workflow_system_prompt`` records its decision on the
+ ``WorkflowRunContext`` so both the agent path (which uses the block's
+ own ``workflow_system_prompt`` field) and the script path (which reads
+ from the context cache via ``ai_extract``) see the same value — single
+ source of truth for the opt-out (SKY-9147)."""
+
+ def test_non_opted_out_block_records_resolved_value(self) -> None:
+ block = _make_task_block(workflow_system_prompt=None)
+ ctx = _make_workflow_run_context("Answer only in Spanish.")
+
+ block.format_potential_template_parameters(ctx)
+
+ recorded, value = ctx.get_block_workflow_system_prompt(block.label)
+ assert recorded is True
+ assert value == "Answer only in Spanish."
+
+ def test_opted_out_block_records_none(self) -> None:
+ block = TaskBlock(
+ label="task1",
+ output_parameter=_make_output_parameter(),
+ title="task title",
+ ignore_workflow_system_prompt=True,
+ )
+ ctx = _make_workflow_run_context("WORKFLOW RULES.")
+
+ block.format_potential_template_parameters(ctx)
+
+ # Recorded explicitly so ``ai_extract`` reads ``None`` (opt-out) rather
+ # than falling through to ``resolve_effective_workflow_system_prompt``.
+ recorded, value = ctx.get_block_workflow_system_prompt(block.label)
+ assert recorded is True
+ assert value is None
+
+ def test_unknown_label_returns_not_recorded(self) -> None:
+ ctx = _make_workflow_run_context("whatever")
+ recorded, value = ctx.get_block_workflow_system_prompt("never-seen")
+ assert recorded is False
+ assert value is None
+
+
+class TestResolveEffectiveWorkflowSystemPromptRejectsNonString:
+ """``resolve_effective_workflow_system_prompt`` must treat a non-string
+ ``workflow_system_prompt`` as absent rather than passing it to Jinja.
+
+ Regression: a malformed workflow definition (or a test fixture whose
+ attribute access returns a ``MagicMock``) previously flowed a non-str
+ into ``env.from_string`` and exploded with ``Can't compile non template
+ nodes`` from deep inside the template compiler."""
+
+ def test_magicmock_workflow_definition_yields_none(self) -> None:
+ ctx = _make_workflow_run_context(workflow_system_prompt=None)
+ mock_workflow = MagicMock()
+ # Attribute access on a MagicMock returns another truthy MagicMock,
+ # mirroring what ``get_workflow_by_permanent_id`` returns in tests
+ # that don't set up a real Workflow.
+ ctx.set_workflow(mock_workflow)
+
+ assert ctx.resolve_effective_workflow_system_prompt() is None
+
+ def test_non_string_inherited_prompt_yields_none(self) -> None:
+ ctx = WorkflowRunContext(
+ workflow_title="test",
+ workflow_id="w_test",
+ workflow_permanent_id="wpid_test",
+ workflow_run_id="wr_test",
+ aws_client=MagicMock(),
+ inherited_workflow_system_prompt=MagicMock(), # non-str
+ )
+
+ assert ctx.resolve_effective_workflow_system_prompt() is None
diff --git a/tests/unit/test_collect_inherited_workflow_system_prompt.py b/tests/unit/test_collect_inherited_workflow_system_prompt.py
new file mode 100644
index 000000000..7d02a60ba
--- /dev/null
+++ b/tests/unit/test_collect_inherited_workflow_system_prompt.py
@@ -0,0 +1,270 @@
+"""Unit tests for ``WorkflowService._collect_inherited_workflow_system_prompt``.
+
+The helper walks the parent workflow-run chain via
+``app.DATABASE.workflow_runs.get_workflow_run`` + ``WorkflowService.get_workflow``
+and joins each ancestor's raw ``workflow_system_prompt`` outermost-first. These
+tests pin the chain-break-on-opt-out, cycle detection, outermost-first ordering,
+and depth-cap semantics described in the helper's docstring (SKY-9147).
+"""
+
+from __future__ import annotations
+
+from datetime import datetime, timezone
+from unittest.mock import AsyncMock
+
+import pytest
+
+from skyvern.forge import app
+from skyvern.forge.sdk.workflow.models.block import WorkflowTriggerBlock
+from skyvern.forge.sdk.workflow.models.workflow import (
+ Workflow,
+ WorkflowDefinition,
+ WorkflowRun,
+ WorkflowRunStatus,
+)
+from skyvern.forge.sdk.workflow.service import WorkflowService
+
+
+def _make_workflow(workflow_id: str, workflow_system_prompt: str | None) -> Workflow:
+ now = datetime.now(timezone.utc)
+ return Workflow(
+ workflow_id=workflow_id,
+ organization_id="o_test",
+ title=f"workflow {workflow_id}",
+ workflow_permanent_id=f"wpid_{workflow_id}",
+ version=1,
+ is_saved_task=False,
+ workflow_definition=WorkflowDefinition(
+ parameters=[],
+ blocks=[],
+ workflow_system_prompt=workflow_system_prompt,
+ ),
+ created_at=now,
+ modified_at=now,
+ )
+
+
+def _make_run(
+ *,
+ run_id: str,
+ workflow_id: str,
+ parent_run_id: str | None,
+ skip_inherited: bool = False,
+) -> WorkflowRun:
+ now = datetime.now(timezone.utc)
+ return WorkflowRun(
+ workflow_run_id=run_id,
+ workflow_id=workflow_id,
+ workflow_permanent_id=f"wpid_{workflow_id}",
+ organization_id="o_test",
+ status=WorkflowRunStatus.running,
+ parent_workflow_run_id=parent_run_id,
+ ignore_inherited_workflow_system_prompt=skip_inherited,
+ created_at=now,
+ modified_at=now,
+ )
+
+
+def _install_chain(
+ monkeypatch: pytest.MonkeyPatch,
+ runs: dict[str, WorkflowRun],
+ workflows: dict[str, Workflow],
+) -> tuple[AsyncMock, AsyncMock]:
+ """Wire mocked ``get_workflow_run`` + ``get_workflow`` against the provided
+ dicts. Returns both mocks so tests can assert call counts."""
+
+ get_run = AsyncMock(side_effect=lambda run_id: runs.get(run_id))
+ monkeypatch.setattr(app.DATABASE.workflow_runs, "get_workflow_run", get_run)
+
+ get_workflow = AsyncMock(side_effect=lambda workflow_id: workflows.get(workflow_id))
+ monkeypatch.setattr(WorkflowService, "get_workflow", get_workflow)
+
+ return get_run, get_workflow
+
+
+async def test_returns_none_when_parent_run_id_is_none(monkeypatch: pytest.MonkeyPatch) -> None:
+ _install_chain(monkeypatch, runs={}, workflows={})
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id=None)
+
+ assert result is None
+
+
+async def test_returns_none_when_no_ancestor_has_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
+ """Chain exists but every ancestor's ``workflow_system_prompt`` is None/empty."""
+ runs = {
+ "wr_parent": _make_run(run_id="wr_parent", workflow_id="w_parent", parent_run_id="wr_grandparent"),
+ "wr_grandparent": _make_run(run_id="wr_grandparent", workflow_id="w_grandparent", parent_run_id=None),
+ }
+ workflows = {
+ "w_parent": _make_workflow("w_parent", workflow_system_prompt=None),
+ "w_grandparent": _make_workflow("w_grandparent", workflow_system_prompt=""),
+ }
+ _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_parent")
+
+ assert result is None
+
+
+async def test_single_ancestor_with_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
+ runs = {"wr_parent": _make_run(run_id="wr_parent", workflow_id="w_parent", parent_run_id=None)}
+ workflows = {"w_parent": _make_workflow("w_parent", workflow_system_prompt="Respond in English.")}
+ _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_parent")
+
+ assert result == "Respond in English."
+
+
+async def test_outermost_first_ordering(monkeypatch: pytest.MonkeyPatch) -> None:
+ """Chain: great-grandparent -> grandparent -> parent -> (child). The helper
+ walks up bottom-up starting from the parent; the returned string must join
+ outermost-first (great-grandparent before grandparent before parent)."""
+ runs = {
+ "wr_parent": _make_run(run_id="wr_parent", workflow_id="w_parent", parent_run_id="wr_grand"),
+ "wr_grand": _make_run(run_id="wr_grand", workflow_id="w_grand", parent_run_id="wr_great"),
+ "wr_great": _make_run(run_id="wr_great", workflow_id="w_great", parent_run_id=None),
+ }
+ workflows = {
+ "w_parent": _make_workflow("w_parent", workflow_system_prompt="PARENT rule."),
+ "w_grand": _make_workflow("w_grand", workflow_system_prompt="GRAND rule."),
+ "w_great": _make_workflow("w_great", workflow_system_prompt="GREAT rule."),
+ }
+ _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_parent")
+
+ assert result == "GREAT rule.\n\nGRAND rule.\n\nPARENT rule."
+
+
+async def test_chain_break_includes_opted_out_ancestors_own_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
+ """When an ancestor has ``ignore_inherited_workflow_system_prompt=True``, its own
+ prompt is still collected but traversal stops — the grandparent's rules must
+ NOT appear in the result (SKY-9147 design: an opted-out workflow rejects its
+ parents' rules but still propagates its own to descendants)."""
+ runs = {
+ "wr_parent": _make_run(
+ run_id="wr_parent",
+ workflow_id="w_parent",
+ parent_run_id="wr_grand",
+ skip_inherited=True,
+ ),
+ "wr_grand": _make_run(run_id="wr_grand", workflow_id="w_grand", parent_run_id=None),
+ }
+ workflows = {
+ "w_parent": _make_workflow("w_parent", workflow_system_prompt="PARENT rule."),
+ "w_grand": _make_workflow("w_grand", workflow_system_prompt="GRAND rule."),
+ }
+ _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_parent")
+
+ assert result == "PARENT rule."
+
+
+async def test_chain_break_opted_out_ancestor_with_no_own_prompt(monkeypatch: pytest.MonkeyPatch) -> None:
+ """When the opted-out ancestor has no prompt of its own and its ancestors do,
+ the result is None — traversal still stops at the opted-out boundary."""
+ runs = {
+ "wr_parent": _make_run(
+ run_id="wr_parent",
+ workflow_id="w_parent",
+ parent_run_id="wr_grand",
+ skip_inherited=True,
+ ),
+ "wr_grand": _make_run(run_id="wr_grand", workflow_id="w_grand", parent_run_id=None),
+ }
+ workflows = {
+ "w_parent": _make_workflow("w_parent", workflow_system_prompt=None),
+ "w_grand": _make_workflow("w_grand", workflow_system_prompt="GRAND rule."),
+ }
+ _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_parent")
+
+ assert result is None
+
+
+async def test_cycle_detection_breaks_loop(monkeypatch: pytest.MonkeyPatch) -> None:
+ """A parent chain that points back to an already-visited run must not infinite-
+ loop. The helper should collect each unique ancestor's prompt exactly once."""
+ runs = {
+ "wr_a": _make_run(run_id="wr_a", workflow_id="w_a", parent_run_id="wr_b"),
+ "wr_b": _make_run(run_id="wr_b", workflow_id="w_b", parent_run_id="wr_a"),
+ }
+ workflows = {
+ "w_a": _make_workflow("w_a", workflow_system_prompt="A rule."),
+ "w_b": _make_workflow("w_b", workflow_system_prompt="B rule."),
+ }
+ get_run, _ = _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_a")
+
+ assert result == "B rule.\n\nA rule."
+ # Each unique run fetched exactly once — no infinite loop.
+ assert get_run.await_count == 2
+
+
+async def test_missing_parent_run_breaks_chain(monkeypatch: pytest.MonkeyPatch) -> None:
+ """If ``get_workflow_run`` returns None mid-walk (soft-deleted / race),
+ traversal stops cleanly and returns whatever was collected so far."""
+ runs = {
+ "wr_parent": _make_run(run_id="wr_parent", workflow_id="w_parent", parent_run_id="wr_missing"),
+ }
+ workflows = {"w_parent": _make_workflow("w_parent", workflow_system_prompt="PARENT rule.")}
+ _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_parent")
+
+ assert result == "PARENT rule."
+
+
+async def test_missing_parent_workflow_skips_that_ancestor(monkeypatch: pytest.MonkeyPatch) -> None:
+ """If ``get_workflow`` returns None (deleted definition) the traversal still
+ continues up the chain, silently skipping that level rather than aborting."""
+ runs = {
+ "wr_parent": _make_run(run_id="wr_parent", workflow_id="w_parent_missing", parent_run_id="wr_grand"),
+ "wr_grand": _make_run(run_id="wr_grand", workflow_id="w_grand", parent_run_id=None),
+ }
+ workflows = {"w_grand": _make_workflow("w_grand", workflow_system_prompt="GRAND rule.")}
+ _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_parent")
+
+ assert result == "GRAND rule."
+
+
+async def test_depth_cap_bounds_traversal(monkeypatch: pytest.MonkeyPatch) -> None:
+ """A malformed deep chain must stop at ``MAX_TRIGGER_DEPTH`` — the helper
+ collects ancestors up to the cap and returns those without hanging."""
+ depth = WorkflowTriggerBlock.MAX_TRIGGER_DEPTH + 5
+ runs: dict[str, WorkflowRun] = {}
+ workflows: dict[str, Workflow] = {}
+ for i in range(depth):
+ run_id = f"wr_{i}"
+ workflow_id = f"w_{i}"
+ parent_id = f"wr_{i + 1}" if i + 1 < depth else None
+ runs[run_id] = _make_run(run_id=run_id, workflow_id=workflow_id, parent_run_id=parent_id)
+ workflows[workflow_id] = _make_workflow(workflow_id, workflow_system_prompt=f"rule {i}.")
+
+ get_run, _ = _install_chain(monkeypatch, runs, workflows)
+
+ result = await WorkflowService()._collect_inherited_workflow_system_prompt(parent_workflow_run_id="wr_0")
+
+ # Exactly MAX_TRIGGER_DEPTH ancestors fetched; deeper ones dropped.
+ assert get_run.await_count == WorkflowTriggerBlock.MAX_TRIGGER_DEPTH
+ assert result is not None
+ # Parts are "rule 0." … "rule (cap-1)." joined outermost-first.
+ expected_parts = [f"rule {i}." for i in reversed(range(WorkflowTriggerBlock.MAX_TRIGGER_DEPTH))]
+ assert result == "\n\n".join(expected_parts)
diff --git a/tests/unit/test_data_extraction_summary_schema_cap.py b/tests/unit/test_data_extraction_summary_schema_cap.py
index 8da4cba8c..312d721d0 100644
--- a/tests/unit/test_data_extraction_summary_schema_cap.py
+++ b/tests/unit/test_data_extraction_summary_schema_cap.py
@@ -18,7 +18,7 @@ def _run_create_extract_action(monkeypatch, extracted_information_schema):
captured.update(kwargs)
return original_load_prompt(template_name, **kwargs)
- async def fake_handler(*, prompt, step, prompt_name):
+ async def fake_handler(*, prompt, step, prompt_name, **_ignored):
captured["prompt"] = prompt
return {"summary": "ok"}
diff --git a/tests/unit/test_downloaded_files_artifact_urls.py b/tests/unit/test_downloaded_files_artifact_urls.py
new file mode 100644
index 000000000..b169d63db
--- /dev/null
+++ b/tests/unit/test_downloaded_files_artifact_urls.py
@@ -0,0 +1,507 @@
+"""Tests for downloaded_files migration to short artifact URLs (SKY-8861)."""
+
+from __future__ import annotations
+
+from unittest.mock import AsyncMock, MagicMock, patch
+from urllib.parse import urlparse
+
+import pytest
+
+from skyvern.forge.sdk.artifact.manager import ArtifactManager
+from skyvern.forge.sdk.artifact.models import Artifact, ArtifactType
+from skyvern.forge.sdk.artifact.storage.s3 import S3Storage
+
+
+def _is_amazonaws_s3_url(url: str) -> bool:
+ """Strict check that ``url`` is a real ``*.s3.amazonaws.com`` URL.
+
+ Avoids the substring trap CodeQL flags as ``py/incomplete-url-substring-sanitization``
+ — ``"s3.amazonaws.com" in url`` matches ``http://evil.com/?x=s3.amazonaws.com``
+ and similar bypasses. Parse the URL and check the hostname suffix instead.
+ """
+ host = urlparse(url).hostname
+ if host is None:
+ return False
+ return host == "s3.amazonaws.com" or host.endswith(".s3.amazonaws.com")
+
+
+async def test_create_download_artifact_is_idempotent_per_run_and_uri():
+ """A repeat save (e.g. inside a loop) must return the existing artifact_id so
+ downstream URL-based dedup (``loop_download_filter``) keeps seeing a stable URL.
+ """
+ manager = ArtifactManager()
+
+ existing = Artifact(
+ artifact_id="a_existing",
+ artifact_type=ArtifactType.DOWNLOAD,
+ uri="s3://skyvern-uploads/downloads/local/o_1/wr_1/file.pdf",
+ organization_id="o_1",
+ run_id="wr_1",
+ workflow_run_id="wr_1",
+ created_at="2026-04-23T00:00:00Z",
+ modified_at="2026-04-23T00:00:00Z",
+ )
+ find_existing = AsyncMock(return_value=existing)
+ mock_db_create = AsyncMock()
+
+ with (
+ patch(
+ "skyvern.forge.sdk.artifact.manager.app.DATABASE.artifacts.find_download_artifact",
+ find_existing,
+ ),
+ patch(
+ "skyvern.forge.sdk.artifact.manager.app.DATABASE.artifacts.create_artifact",
+ mock_db_create,
+ ),
+ ):
+ artifact_id = await manager.create_download_artifact(
+ organization_id="o_1",
+ run_id="wr_1",
+ workflow_run_id="wr_1",
+ uri=existing.uri,
+ filename="file.pdf",
+ )
+
+ assert artifact_id == "a_existing"
+ mock_db_create.assert_not_awaited()
+
+
+async def test_create_download_artifact_inserts_row_without_uploading():
+ """create_download_artifact only writes a DB row; bytes are already in S3."""
+ manager = ArtifactManager()
+
+ mock_db_create = AsyncMock(
+ return_value=Artifact(
+ artifact_id="a_abc123",
+ artifact_type=ArtifactType.DOWNLOAD,
+ uri="s3://skyvern-uploads/download/prod/o_1/wr_1/file.pdf",
+ organization_id="o_1",
+ run_id="wr_1",
+ workflow_run_id="wr_1",
+ created_at="2026-04-23T00:00:00Z",
+ modified_at="2026-04-23T00:00:00Z",
+ )
+ )
+ mock_store = AsyncMock()
+ find_existing = AsyncMock(return_value=None)
+
+ with (
+ patch(
+ "skyvern.forge.sdk.artifact.manager.app.DATABASE.artifacts.find_download_artifact",
+ find_existing,
+ ),
+ patch("skyvern.forge.sdk.artifact.manager.app.DATABASE.artifacts.create_artifact", mock_db_create),
+ patch("skyvern.forge.sdk.artifact.manager.app.STORAGE.store_artifact", mock_store),
+ patch("skyvern.forge.sdk.artifact.manager.app.STORAGE.store_artifact_from_path", mock_store),
+ ):
+ artifact_id = await manager.create_download_artifact(
+ organization_id="o_1",
+ run_id="wr_1",
+ workflow_run_id="wr_1",
+ uri="s3://skyvern-uploads/download/prod/o_1/wr_1/file.pdf",
+ filename="file.pdf",
+ )
+
+ assert artifact_id.startswith("a_")
+ mock_db_create.assert_awaited_once()
+ _, kwargs = mock_db_create.call_args
+ assert kwargs["artifact_type"] == ArtifactType.DOWNLOAD
+ assert kwargs["uri"] == "s3://skyvern-uploads/download/prod/o_1/wr_1/file.pdf"
+ assert kwargs["organization_id"] == "o_1"
+ assert kwargs["run_id"] == "wr_1"
+ assert kwargs["workflow_run_id"] == "wr_1"
+ mock_store.assert_not_awaited()
+
+
+async def test_save_downloaded_files_registers_artifact_per_file(tmp_path):
+ """After uploading each file to S3, save_downloaded_files should create an
+ Artifact row so later retrieval can build short /v1/artifacts URLs."""
+ download_dir = tmp_path / "downloads"
+ download_dir.mkdir()
+ (download_dir / "invoice.pdf").write_bytes(b"%PDF-1.4 ...")
+ (download_dir / "report.csv").write_bytes(b"a,b,c\n1,2,3\n")
+
+ storage = S3Storage()
+ storage.async_client = MagicMock()
+ storage.async_client.upload_file_from_path = AsyncMock()
+
+ mock_create_download = AsyncMock(return_value="a_new")
+ mock_artifact_manager = MagicMock()
+ mock_artifact_manager.create_download_artifact = mock_create_download
+
+ with (
+ patch("skyvern.forge.sdk.artifact.storage.s3.get_download_dir", return_value=str(download_dir)),
+ patch.object(storage, "_get_storage_class_for_org", new=AsyncMock(return_value=MagicMock())),
+ patch("skyvern.forge.sdk.artifact.storage.s3.calculate_sha256_for_file", return_value="sha-xyz"),
+ patch("skyvern.forge.sdk.artifact.storage.s3.app") as app_module,
+ ):
+ app_module.ARTIFACT_MANAGER = mock_artifact_manager
+ await storage.save_downloaded_files(organization_id="o_1", run_id="wr_1")
+
+ assert mock_create_download.await_count == 2
+ uris = {call.kwargs["uri"] for call in mock_create_download.await_args_list}
+ filenames = {call.kwargs["filename"] for call in mock_create_download.await_args_list}
+ assert filenames == {"invoice.pdf", "report.csv"}
+ assert all(u.startswith("s3://") and "/downloads/" in u and "/o_1/wr_1/" in u for u in uris)
+ for call in mock_create_download.await_args_list:
+ assert call.kwargs["organization_id"] == "o_1"
+ assert call.kwargs["run_id"] == "wr_1"
+
+
+def _make_artifact(
+ artifact_id: str,
+ uri: str,
+ run_id: str = "wr_1",
+ *,
+ checksum: str | None = None,
+ created_at: str = "2026-04-23T00:00:00Z",
+) -> Artifact:
+ return Artifact(
+ artifact_id=artifact_id,
+ artifact_type=ArtifactType.DOWNLOAD,
+ uri=uri,
+ organization_id="o_1",
+ run_id=run_id,
+ workflow_run_id=run_id if run_id.startswith("wr_") else None,
+ checksum=checksum,
+ created_at=created_at,
+ modified_at=created_at,
+ )
+
+
+_DUMMY_KEYRING_JSON = '{"current_kid": "k1", "keys": {"k1": {"secret": "0000000000000000000000000000000000000000000000000000000000000000"}}}'
+
+
+def keyring_configured():
+ """Simulate cloud-style config: HMAC keyring is set so the artifact URL branch is active.
+ Unit tests default to no keyring to match the OSS default, so tests that exercise the
+ short-URL path must opt in."""
+ from skyvern.config import settings
+
+ with patch.object(settings, "ARTIFACT_CONTENT_HMAC_KEYRING", _DUMMY_KEYRING_JSON):
+ yield
+
+
+async def test_get_downloaded_files_uses_artifact_urls_when_rows_exist(keyring_configured):
+ """When DOWNLOAD artifact rows exist, retrieval skips S3 entirely:
+ URL, checksum, filename, modified_at all come straight from the row."""
+ storage = S3Storage()
+ storage.async_client = MagicMock()
+ storage.async_client.list_files = AsyncMock() # must NOT be called
+ storage.async_client.get_file_metadata = AsyncMock() # must NOT be called
+ storage.async_client.create_presigned_urls = AsyncMock() # must NOT be called
+
+ artifact = _make_artifact(
+ "a_42",
+ "s3://skyvern-uploads/downloads/local/o_1/wr_1/invoice.pdf",
+ checksum="sha-from-db",
+ )
+ mock_list = AsyncMock(return_value=[artifact])
+ build_url = MagicMock(return_value="https://api.skyvern.com/v1/artifacts/a_42/content?expiry=x&kid=y&sig=z")
+
+ with patch("skyvern.forge.sdk.artifact.storage.base.app") as base_app:
+ with patch("skyvern.forge.sdk.artifact.storage.s3.app") as s3_app:
+ s3_app.DATABASE.artifacts.list_artifacts_for_run_by_type = mock_list
+ base_app.ARTIFACT_MANAGER.build_signed_content_url = build_url
+ result = await storage.get_downloaded_files(organization_id="o_1", run_id="wr_1")
+
+ assert len(result) == 1
+ assert result[0].url.startswith("https://api.skyvern.com/v1/artifacts/a_42/content")
+ assert result[0].filename == "invoice.pdf"
+ assert result[0].checksum == "sha-from-db"
+ assert result[0].modified_at is not None
+ storage.async_client.list_files.assert_not_awaited()
+ storage.async_client.get_file_metadata.assert_not_awaited()
+ storage.async_client.create_presigned_urls.assert_not_awaited()
+ mock_list.assert_awaited_once_with(run_id="wr_1", organization_id="o_1", artifact_type=ArtifactType.DOWNLOAD)
+
+
+async def test_get_downloaded_files_preserves_artifact_row_order(keyring_configured):
+ """Artifact rows are returned ASC by created_at; FileInfo list must follow the
+ same order (matches save order, drives loop_download_filter signatures)."""
+ storage = S3Storage()
+ storage.async_client = MagicMock()
+
+ first = _make_artifact(
+ "a_1",
+ "s3://skyvern-uploads/downloads/local/o_1/wr_1/first.pdf",
+ created_at="2026-04-23T00:00:00Z",
+ )
+ second = _make_artifact(
+ "a_2",
+ "s3://skyvern-uploads/downloads/local/o_1/wr_1/second.pdf",
+ created_at="2026-04-23T00:01:00Z",
+ )
+ mock_list = AsyncMock(return_value=[first, second])
+ build_url = MagicMock(
+ side_effect=lambda artifact_id, **_: f"https://api.skyvern.com/v1/artifacts/{artifact_id}/content"
+ )
+
+ with patch("skyvern.forge.sdk.artifact.storage.base.app") as base_app:
+ with patch("skyvern.forge.sdk.artifact.storage.s3.app") as s3_app:
+ s3_app.DATABASE.artifacts.list_artifacts_for_run_by_type = mock_list
+ base_app.ARTIFACT_MANAGER.build_signed_content_url = build_url
+ result = await storage.get_downloaded_files(organization_id="o_1", run_id="wr_1")
+
+ assert [fi.filename for fi in result] == ["first.pdf", "second.pdf"]
+
+
+async def test_get_downloaded_files_falls_back_to_presigned_for_legacy_runs(keyring_configured):
+ """Production-cloud legacy run: keyring IS configured, but the run pre-dates SKY-8861
+ so no artifact rows exist. Files in S3 must still surface as presigned URLs — the
+ whole point of keeping the fallback path."""
+ storage = S3Storage()
+ storage.async_client = MagicMock()
+ s3_key = "downloads/local/o_1/wr_old/legacy.pdf"
+ storage.async_client.list_files = AsyncMock(return_value=[s3_key])
+ storage.async_client.get_file_metadata = AsyncMock(
+ return_value={"sha256_checksum": "sha-old", "original_filename": "legacy.pdf"}
+ )
+ storage.async_client.create_presigned_urls = AsyncMock(
+ return_value=["https://skyvern-uploads.s3.amazonaws.com/...?sig=old"]
+ )
+
+ mock_list = AsyncMock(return_value=[]) # no artifact rows for this legacy run
+ build_url = MagicMock() # must NOT be called
+
+ with patch("skyvern.forge.sdk.artifact.storage.base.app") as base_app:
+ with patch("skyvern.forge.sdk.artifact.storage.s3.app") as s3_app:
+ s3_app.DATABASE.artifacts.list_artifacts_for_run_by_type = mock_list
+ base_app.ARTIFACT_MANAGER.build_signed_content_url = build_url
+ result = await storage.get_downloaded_files(organization_id="o_1", run_id="wr_old")
+
+ assert len(result) == 1
+ assert result[0].filename == "legacy.pdf"
+ assert result[0].checksum == "sha-old"
+ assert _is_amazonaws_s3_url(result[0].url)
+ build_url.assert_not_called()
+ storage.async_client.list_files.assert_awaited_once()
+ storage.async_client.create_presigned_urls.assert_awaited_once()
+
+
+async def test_get_downloaded_files_falls_back_to_presigned_when_keyring_unset(tmp_path):
+ """Self-hosted OSS deployments without ARTIFACT_CONTENT_HMAC_KEYRING must keep
+ serving presigned S3 URLs — the short Skyvern URL would be unsigned and the
+ content endpoint would 401 without an API key."""
+ from skyvern.config import settings
+
+ storage = S3Storage()
+ storage.async_client = MagicMock()
+ s3_key = "downloads/local/o_1/wr_1/invoice.pdf"
+ storage.async_client.list_files = AsyncMock(return_value=[s3_key])
+ storage.async_client.get_file_metadata = AsyncMock(
+ return_value={"sha256_checksum": "sha-abc", "original_filename": "invoice.pdf"}
+ )
+ storage.async_client.create_presigned_urls = AsyncMock(
+ return_value=["https://skyvern-uploads.s3.amazonaws.com/...?sig=fallback"]
+ )
+
+ artifact = _make_artifact("a_42", f"s3://skyvern-uploads/{s3_key}")
+ mock_list = AsyncMock(return_value=[artifact])
+ build_url = MagicMock()
+
+ with (
+ patch("skyvern.forge.sdk.artifact.storage.s3.app") as app_module,
+ patch.object(settings, "ARTIFACT_CONTENT_HMAC_KEYRING", None),
+ ):
+ app_module.DATABASE.artifacts.list_artifacts_for_run_by_type = mock_list
+ app_module.ARTIFACT_MANAGER.build_signed_content_url = build_url
+ result = await storage.get_downloaded_files(organization_id="o_1", run_id="wr_1")
+
+ assert len(result) == 1
+ assert _is_amazonaws_s3_url(result[0].url)
+ build_url.assert_not_called()
+
+
+async def test_get_downloaded_files_artifact_lookup_failure_falls_back_to_listing(keyring_configured):
+ """If the DB lookup raises (transient outage), retrieval must not 500 the
+ run-output API — fall through to the legacy S3-listing path so files
+ still surface as presigned URLs."""
+ storage = S3Storage()
+ storage.async_client = MagicMock()
+ s3_key = "downloads/local/o_1/wr_1/recoverable.pdf"
+ storage.async_client.list_files = AsyncMock(return_value=[s3_key])
+ storage.async_client.get_file_metadata = AsyncMock(
+ return_value={"sha256_checksum": "sha-recover", "original_filename": "recoverable.pdf"}
+ )
+ storage.async_client.create_presigned_urls = AsyncMock(
+ return_value=["https://skyvern-uploads.s3.amazonaws.com/...?sig=fallback"]
+ )
+
+ mock_list = AsyncMock(side_effect=RuntimeError("DB unreachable"))
+
+ with patch("skyvern.forge.sdk.artifact.storage.s3.app") as app_module:
+ app_module.DATABASE.artifacts.list_artifacts_for_run_by_type = mock_list
+ result = await storage.get_downloaded_files(organization_id="o_1", run_id="wr_1")
+
+ assert len(result) == 1
+ assert _is_amazonaws_s3_url(result[0].url)
+ storage.async_client.list_files.assert_awaited_once()
+
+
+async def test_content_endpoint_download_returns_attachment_with_filename():
+ """DOWNLOAD artifacts must serve with attachment disposition so browsers don't render
+ PDFs inline (defeats the SKY-8862 XSS-via-PDF mitigation)."""
+ from skyvern.forge.sdk.routes.agent_protocol import _artifact_response_config
+
+ artifact = _make_artifact("a_dl", "s3://skyvern-uploads/downloads/local/o_1/wr_1/invoice.pdf")
+ media_type, disposition = _artifact_response_config(artifact)
+ assert media_type == "application/octet-stream"
+ assert disposition.startswith("attachment;")
+ assert 'filename="invoice.pdf"' in disposition
+
+
+def test_content_endpoint_non_download_stays_inline():
+ """Existing artifact types keep the inline disposition we had before."""
+ from skyvern.forge.sdk.routes.agent_protocol import _artifact_response_config
+
+ screenshot = Artifact(
+ artifact_id="a_ss",
+ artifact_type=ArtifactType.SCREENSHOT_FINAL,
+ uri="s3://skyvern-artifacts/.../final.png",
+ organization_id="o_1",
+ created_at="2026-04-23T00:00:00Z",
+ modified_at="2026-04-23T00:00:00Z",
+ )
+ media_type, disposition = _artifact_response_config(screenshot)
+ assert media_type == "image/png"
+ assert disposition == "inline"
+
+
+def test_content_endpoint_download_non_ascii_filename_does_not_crash_header_encoding():
+ """Starlette encodes response headers as Latin-1. Unicode filenames must use RFC 5987
+ (filename*=UTF-8''...) with an ASCII fallback so the endpoint does not 500."""
+ from starlette.responses import Response
+
+ from skyvern.forge.sdk.routes.agent_protocol import _artifact_response_config
+
+ artifact = _make_artifact("a_unicode", "s3://skyvern-uploads/downloads/local/o_1/wr_1/文档.pdf")
+ media_type, disposition = _artifact_response_config(artifact)
+ # Must not raise — Starlette's header encoding rejects non-Latin-1 bytes.
+ Response(content=b"x", media_type=media_type, headers={"Content-Disposition": disposition})
+ assert "filename*=UTF-8''" in disposition
+ assert "%E6%96%87%E6%A1%A3.pdf" in disposition or "%e6%96%87%e6%a1%a3.pdf" in disposition
+
+
+def test_sanitize_header_filename_strips_crlf_and_quotes_directly():
+ """Direct unit test on _sanitize_header_filename so we know the function works
+ even when urlparse isn't part of the chain (defense in depth)."""
+ from skyvern.forge.sdk.routes.agent_protocol import _sanitize_header_filename
+
+ assert _sanitize_header_filename('evil"pdf') == "evilpdf"
+ assert _sanitize_header_filename("hello\r\nworld.pdf") == "helloworld.pdf"
+ assert _sanitize_header_filename("back\\slash.pdf") == "backslash.pdf"
+ assert _sanitize_header_filename("") == "download"
+
+
+def test_content_endpoint_download_filename_preserves_question_and_hash():
+ """S3 keys may legitimately contain '?' or '#'; urlparse would otherwise strip them."""
+ from skyvern.forge.sdk.routes.agent_protocol import _artifact_response_config
+
+ artifact = _make_artifact("a_q", "s3://skyvern-uploads/downloads/local/o_1/wr_1/report?v=2#a.pdf")
+ _, disposition = _artifact_response_config(artifact)
+ assert "report%3Fv%3D2%23a.pdf" in disposition or "report?v=2#a.pdf" in disposition
+
+
+def test_sanitize_header_filename_strips_bidi_and_format_characters():
+ """Unicode bidi overrides and format chars (ZWSP, RLO, ZWNBSP) enable filename
+ spoofing in the browser's download UI (``invoice\\u202efdp.exe`` -> ``invoice.exe.pdf``)."""
+ from skyvern.forge.sdk.routes.agent_protocol import _sanitize_header_filename
+
+ assert "\u202e" not in _sanitize_header_filename("invoice\u202efdp.exe")
+ assert "\u200b" not in _sanitize_header_filename("stealth\u200b.pdf")
+ assert "\ufeff" not in _sanitize_header_filename("bom\ufeff.pdf")
+
+
+def test_ascii_fallback_filename_preserves_stem_for_pure_unicode_names():
+ """Pure non-ASCII names (e.g. CJK, emoji) must not reduce to a bare ``.pdf`` hidden
+ dotfile after the NFKD strip; fall back to a ``download`` stem instead."""
+ from skyvern.forge.sdk.routes.agent_protocol import _ascii_fallback_filename
+
+ assert _ascii_fallback_filename("文档.pdf") == "download.pdf"
+ assert _ascii_fallback_filename("🎉.pdf") == "download.pdf"
+ # Accented Latin still transliterates to keep the stem.
+ assert _ascii_fallback_filename("fïlè.pdf") == "file.pdf"
+
+
+def test_sanitize_header_filename_strips_control_characters():
+ """NUL/DEL/C1 control chars are valid Latin-1 bytes but violate RFC 7230 header syntax."""
+ from skyvern.forge.sdk.routes.agent_protocol import _sanitize_header_filename
+
+ assert "\x00" not in _sanitize_header_filename("evil\x00.pdf")
+ assert "\x7f" not in _sanitize_header_filename("evil\x7f.pdf")
+ assert "\x1b" not in _sanitize_header_filename("evil\x1b.pdf")
+ assert "\x80" not in _sanitize_header_filename("evil\x80.pdf")
+
+
+async def test_content_endpoint_sets_nosniff_header_end_to_end():
+ """Hit the real route through a FastAPI TestClient and verify the response
+ actually carries X-Content-Type-Options: nosniff on the DOWNLOAD path.
+
+ Defence-in-depth for SKY-8862: prevents a refactor from silently dropping
+ the header without this suite noticing."""
+ import json
+
+ from fastapi import FastAPI
+ from fastapi.testclient import TestClient
+
+ from skyvern.config import settings
+ from skyvern.forge.sdk.artifact.signing import sign_artifact_url
+ from skyvern.forge.sdk.routes.routers import base_router
+
+ artifact = _make_artifact("a_e2e", "s3://skyvern-uploads/downloads/local/o_1/wr_1/report.pdf")
+ keyring_json = json.dumps({"current_kid": "k1", "keys": {"k1": {"secret": "0" * 64, "created_at": "2026-04-23"}}})
+
+ with (
+ patch.object(settings, "ARTIFACT_CONTENT_HMAC_KEYRING", keyring_json),
+ patch.object(settings, "SKYVERN_BASE_URL", "http://testserver"),
+ patch("skyvern.forge.sdk.routes.agent_protocol.app") as app_module,
+ ):
+ app_module.DATABASE.artifacts.get_artifact_by_id_no_org = AsyncMock(return_value=artifact)
+ app_module.ARTIFACT_MANAGER.retrieve_artifact = AsyncMock(return_value=b"%PDF-1.4 fake body")
+
+ from skyvern.forge.sdk.artifact.signing import parse_keyring
+
+ signed_url = sign_artifact_url(
+ base_url="http://testserver",
+ artifact_id=artifact.artifact_id,
+ keyring=parse_keyring(keyring_json),
+ artifact_name="report.pdf",
+ artifact_type="download",
+ )
+
+ test_app = FastAPI()
+ test_app.include_router(base_router, prefix="/v1")
+ client = TestClient(test_app)
+ resp = client.get(signed_url.replace("http://testserver", ""))
+
+ assert resp.status_code == 200, resp.text
+ assert resp.headers.get("X-Content-Type-Options") == "nosniff"
+ assert resp.headers.get("Content-Disposition", "").startswith("attachment;")
+ assert resp.content == b"%PDF-1.4 fake body"
+
+
+def test_content_endpoint_download_filename_strips_header_injection():
+ """URI-derived filenames go straight into a Content-Disposition header;
+ CR/LF and raw quotes must be stripped to prevent header injection."""
+ from skyvern.forge.sdk.routes.agent_protocol import _artifact_response_config
+
+ artifact = _make_artifact(
+ "a_bad",
+ 's3://skyvern-uploads/downloads/local/o_1/wr_1/evil"\r\nSet-Cookie: x=y.pdf',
+ )
+ _, disposition = _artifact_response_config(artifact)
+ assert "\r" not in disposition
+ assert "\n" not in disposition
+ assert disposition.count('"') == 2 # only the pair around filename
diff --git a/tests/unit/test_extract_information_text_optout.py b/tests/unit/test_extract_information_text_optout.py
index de5abf2b6..a4f49fb78 100644
--- a/tests/unit/test_extract_information_text_optout.py
+++ b/tests/unit/test_extract_information_text_optout.py
@@ -196,7 +196,7 @@ def _capture_ai_extract_kwargs(monkeypatch, include_extracted_text: bool):
async def fake_refresh(*_args, **_kwargs):
return None
- async def fake_handler(*, prompt, step, screenshots, prompt_name, force_dict):
+ async def fake_handler(*, prompt, step, screenshots, prompt_name, force_dict, **_ignored):
return {}
monkeypatch.setattr(module, "load_prompt_with_elements_tracked", fake_load_prompt_with_elements_tracked)
@@ -258,7 +258,7 @@ def _capture_ai_extract_kwargs_with_schema(monkeypatch, schema):
async def fake_refresh(*_args, **_kwargs):
return None
- async def fake_handler(*, prompt, step, screenshots, prompt_name, force_dict):
+ async def fake_handler(*, prompt, step, screenshots, prompt_name, force_dict, **_ignored):
return {}
monkeypatch.setattr(module, "load_prompt_with_elements_tracked", fake_load_prompt_with_elements_tracked)
diff --git a/tests/unit/test_workflow_parameter_validation.py b/tests/unit/test_workflow_parameter_validation.py
index 1a68b4eb8..ce0df96d2 100644
--- a/tests/unit/test_workflow_parameter_validation.py
+++ b/tests/unit/test_workflow_parameter_validation.py
@@ -625,3 +625,40 @@ class TestSanitizeWorkflowYamlWithReferences:
goal = result["workflow_definition"]["blocks"][0]["navigation_goal"]
assert "{{ block_1 }}" in goal
assert "{{ block_1_output }}" in goal
+
+ def test_sanitize_updates_output_references_in_workflow_system_prompt(self) -> None:
+ """Output references inside the workflow-level workflow_system_prompt must
+ be rewritten when the referenced block label is sanitized. The global
+ prompt is resolved through Jinja at execution time, so its references
+ need the same renaming treatment as block-level fields."""
+ workflow_yaml = {
+ "title": "Test Workflow",
+ "workflow_definition": {
+ "parameters": [],
+ "blocks": [{"label": "my-block", "block_type": "task"}],
+ "workflow_system_prompt": "Honor {{ my-block_output }} for every downstream block.",
+ },
+ }
+ result = sanitize_workflow_yaml_with_references(workflow_yaml)
+ assert result["workflow_definition"]["blocks"][0]["label"] == "my_block"
+ assert "{{ my_block_output }}" in result["workflow_definition"]["workflow_system_prompt"]
+
+ def test_sanitize_parameter_key_updates_jinja_references_in_workflow_system_prompt(self) -> None:
+ """Parameter-key references inside the workflow-level workflow_system_prompt
+ must be rewritten when the parameter key is sanitized."""
+ workflow_yaml = {
+ "title": "Test Workflow",
+ "workflow_definition": {
+ "parameters": [
+ {
+ "key": "user-input",
+ "parameter_type": "workflow",
+ }
+ ],
+ "blocks": [],
+ "workflow_system_prompt": "Always respond in the style of {{ user-input }}.",
+ },
+ }
+ result = sanitize_workflow_yaml_with_references(workflow_yaml)
+ assert result["workflow_definition"]["parameters"][0]["key"] == "user_input"
+ assert "{{ user_input }}" in result["workflow_definition"]["workflow_system_prompt"]
diff --git a/tests/unit/test_workflow_system_prompt_llm_forwarding.py b/tests/unit/test_workflow_system_prompt_llm_forwarding.py
new file mode 100644
index 000000000..08e9bdb3a
--- /dev/null
+++ b/tests/unit/test_workflow_system_prompt_llm_forwarding.py
@@ -0,0 +1,387 @@
+"""Regression tests: workflow-level workflow_system_prompt must reach the LLM handler.
+
+Complements ``test_block_workflow_system_prompt_inheritance.py`` (which asserts
+that the workflow prompt flows onto the block model). These tests assert the
+next hop — that each block/function that makes an LLM call actually forwards
+``system_prompt`` as a kwarg to the handler. Without this, inheritance passes
+quietly but the prompt has no effect on output (the bug the user hit on the
+extraction path).
+"""
+
+from __future__ import annotations
+
+import asyncio
+from datetime import datetime, timezone
+from unittest.mock import AsyncMock, MagicMock
+
+from skyvern.forge.sdk.cache import extraction_cache
+from skyvern.forge.sdk.workflow.models.block import FileParserBlock, PDFParserBlock, TextPromptBlock
+from skyvern.forge.sdk.workflow.models.parameter import OutputParameter, ParameterType
+from skyvern.schemas.workflows import FileType
+
+# ---------------------------------------------------------------------------
+# Shared helpers
+# ---------------------------------------------------------------------------
+
+
+def _make_output_parameter() -> OutputParameter:
+ now = datetime.now(timezone.utc)
+ return OutputParameter(
+ parameter_type=ParameterType.OUTPUT,
+ key="task1_output",
+ description="test output",
+ output_parameter_id="op_task1",
+ workflow_id="w_test",
+ created_at=now,
+ modified_at=now,
+ )
+
+
+def _make_scraped_page():
+ refreshed = MagicMock()
+ refreshed.extracted_text = "page text"
+ refreshed.url = "https://example.test"
+ refreshed.screenshots = []
+ refreshed.build_element_tree = MagicMock(return_value="<a>link</a>")
+ refreshed.support_economy_elements_tree = MagicMock(return_value=False)
+ refreshed.last_used_element_tree_html = None
+
+ scraped_page = MagicMock()
+ scraped_page.refresh = AsyncMock(return_value=refreshed)
+ scraped_page.screenshots = []
+ return scraped_page
+
+
+def _make_task(*, system_prompt: str | None) -> MagicMock:
+ task = MagicMock()
+ task.navigation_goal = None
+ task.navigation_payload = None
+ task.extracted_information = None
+ task.extracted_information_schema = {"type": "object"}
+ task.data_extraction_goal = "Extract documents"
+ task.error_code_mapping = None
+ task.llm_key = None
+ task.workflow_run_id = "wfr_sysprompt"
+ task.task_id = "tsk_sysprompt"
+ task.workflow_permanent_id = "wpid_sysprompt"
+ task.organization_id = "o_sysprompt"
+ task.include_extracted_text = True
+ task.workflow_system_prompt = system_prompt
+ return task
+
+
+# ---------------------------------------------------------------------------
+# extract-information handler (the OP's exact failure path)
+# ---------------------------------------------------------------------------
+
+
+def test_extract_information_forwards_system_prompt(monkeypatch) -> None:
+ """Regression for the OP's bug: an ExtractionBlock whose Task carries a
+ ``system_prompt`` (inherited from workflow.workflow_system_prompt) must
+ forward it to the extract-information LLM call."""
+ from skyvern.webeye.actions import handler
+
+ extraction_cache._reset_for_tests()
+
+ captured: dict = {}
+
+ async def fake_llm(**kwargs):
+ captured.update(kwargs)
+ return {"quotes": "SHOUTED QUOTE"}
+
+ monkeypatch.setattr(
+ handler,
+ "load_prompt_with_elements_tracked",
+ lambda **kwargs: ("rendered-prompt", dict(kwargs)),
+ )
+ monkeypatch.setattr(handler, "ensure_context", lambda: MagicMock(tz_info=None))
+ monkeypatch.setattr(handler.service_utils, "is_cua_task", AsyncMock(return_value=False))
+ monkeypatch.setattr(
+ handler.LLMAPIHandlerFactory,
+ "get_override_llm_api_handler",
+ lambda llm_key, default: fake_llm,
+ )
+ monkeypatch.setattr(
+ handler.app.AGENT_FUNCTION,
+ "should_shadow_extraction_cache_hit",
+ AsyncMock(return_value=False),
+ )
+ monkeypatch.setattr(
+ handler.app.AGENT_FUNCTION,
+ "lookup_cross_run_extraction_cache",
+ AsyncMock(return_value=None),
+ )
+ monkeypatch.setattr(
+ handler.app.AGENT_FUNCTION,
+ "store_cross_run_extraction_cache",
+ AsyncMock(return_value=None),
+ )
+
+ task = _make_task(system_prompt="Respond only in uppercase.")
+ step = MagicMock(step_id="stp_sp", retry_index=0)
+ scraped_page = _make_scraped_page()
+
+ asyncio.run(handler.extract_information_for_navigation_goal(task=task, step=step, scraped_page=scraped_page))
+
+ assert captured.get("system_prompt") == "Respond only in uppercase."
+ extraction_cache._reset_for_tests()
+
+
+def test_extract_information_passes_none_system_prompt_when_task_has_none(monkeypatch) -> None:
+ """No workflow_system_prompt → task.workflow_system_prompt is None → handler receives None.
+ Locks in that we don't accidentally invent a default or drop the kwarg."""
+ from skyvern.webeye.actions import handler
+
+ extraction_cache._reset_for_tests()
+
+ captured: dict = {}
+
+ async def fake_llm(**kwargs):
+ captured.update(kwargs)
+ return {"quotes": "x"}
+
+ monkeypatch.setattr(
+ handler,
+ "load_prompt_with_elements_tracked",
+ lambda **kwargs: ("rendered-prompt", dict(kwargs)),
+ )
+ monkeypatch.setattr(handler, "ensure_context", lambda: MagicMock(tz_info=None))
+ monkeypatch.setattr(handler.service_utils, "is_cua_task", AsyncMock(return_value=False))
+ monkeypatch.setattr(
+ handler.LLMAPIHandlerFactory,
+ "get_override_llm_api_handler",
+ lambda llm_key, default: fake_llm,
+ )
+ monkeypatch.setattr(
+ handler.app.AGENT_FUNCTION,
+ "should_shadow_extraction_cache_hit",
+ AsyncMock(return_value=False),
+ )
+ monkeypatch.setattr(
+ handler.app.AGENT_FUNCTION,
+ "lookup_cross_run_extraction_cache",
+ AsyncMock(return_value=None),
+ )
+ monkeypatch.setattr(
+ handler.app.AGENT_FUNCTION,
+ "store_cross_run_extraction_cache",
+ AsyncMock(return_value=None),
+ )
+
+ task = _make_task(system_prompt=None)
+ step = MagicMock(step_id="stp_none", retry_index=0)
+ scraped_page = _make_scraped_page()
+
+ asyncio.run(handler.extract_information_for_navigation_goal(task=task, step=step, scraped_page=scraped_page))
+
+ assert "system_prompt" in captured
+ assert captured["system_prompt"] is None
+ extraction_cache._reset_for_tests()
+
+
+# ---------------------------------------------------------------------------
+# data-extraction-summary (agent.py path)
+# ---------------------------------------------------------------------------
+
+
+def test_data_extraction_summary_forwards_system_prompt(monkeypatch) -> None:
+ from skyvern.forge import agent as agent_module
+
+ captured: dict = {}
+
+ async def fake_handler(**kwargs):
+ captured.update(kwargs)
+ return {"summary": "ok"}
+
+ monkeypatch.setattr(agent_module.app, "EXTRACTION_LLM_API_HANDLER", fake_handler)
+ monkeypatch.setattr(
+ agent_module.skyvern_context,
+ "ensure_context",
+ lambda: MagicMock(tz_info=None, workflow_run_id="wr_sp"),
+ )
+ monkeypatch.setattr(agent_module.extraction_cache, "compute_cache_key", lambda **_: None)
+ monkeypatch.setattr(agent_module.extraction_cache, "lookup", lambda *a, **k: None)
+
+ task = _make_task(system_prompt="Answer in French.")
+ step = MagicMock(step_id="stp_sp", order=0)
+ scraped_page = MagicMock(url="https://example.test")
+
+ asyncio.run(agent_module.ForgeAgent.create_extract_action(task=task, step=step, scraped_page=scraped_page))
+
+ assert captured.get("system_prompt") == "Answer in French."
+
+
+# ---------------------------------------------------------------------------
+# Extraction cache key — system_prompt must be part of the key
+# ---------------------------------------------------------------------------
+
+
+def test_extraction_cache_key_changes_with_system_prompt() -> None:
+ """Two calls that differ only in system_prompt must produce different
+ digests — otherwise a user switching workflow_system_prompt mid-run would
+ get stale output from a prior key's cached value."""
+ base_kwargs: dict = dict(
+ call_path="handler",
+ element_tree="<a>link</a>",
+ extracted_text="text",
+ current_url="https://example.test",
+ data_extraction_goal="Extract docs",
+ extracted_information_schema={"type": "object"},
+ navigation_payload=None,
+ error_code_mapping=None,
+ previous_extracted_information=None,
+ llm_key=None,
+ )
+
+ key_no_sp = extraction_cache.compute_cache_key(**base_kwargs, workflow_system_prompt=None)
+ key_a = extraction_cache.compute_cache_key(**base_kwargs, workflow_system_prompt="Answer in Spanish.")
+ key_b = extraction_cache.compute_cache_key(**base_kwargs, workflow_system_prompt="Answer in French.")
+
+ assert key_no_sp != key_a
+ assert key_a != key_b
+ # Same prompt → same key (determinism).
+ key_a2 = extraction_cache.compute_cache_key(**base_kwargs, workflow_system_prompt="Answer in Spanish.")
+ assert key_a == key_a2
+
+
+# ---------------------------------------------------------------------------
+# TextPromptBlock.send_prompt
+# ---------------------------------------------------------------------------
+
+
+def test_text_prompt_block_forwards_system_prompt(monkeypatch) -> None:
+ from skyvern.forge.sdk.workflow.models import block as block_module
+
+ captured: dict = {}
+
+ async def fake_llm(**kwargs):
+ captured.update(kwargs)
+ return {"llm_response": "hi"}
+
+ async def fake_default_handler(*args, **kwargs):
+ return None
+
+ async def fake_resolve(self, workflow_run_id, organization_id):
+ return fake_default_handler
+
+ monkeypatch.setattr(TextPromptBlock, "_resolve_default_llm_handler", fake_resolve)
+ monkeypatch.setattr(
+ block_module.LLMAPIHandlerFactory,
+ "get_override_llm_api_handler",
+ lambda llm_key, default: fake_llm,
+ )
+
+ block = TextPromptBlock(
+ label="p",
+ output_parameter=_make_output_parameter(),
+ prompt="What is the meaning of life?",
+ workflow_system_prompt="Respond as Shakespeare.",
+ )
+
+ asyncio.run(
+ block.send_prompt(
+ prompt="What is the meaning of life?",
+ parameter_values={},
+ workflow_run_id="wfr_sp",
+ organization_id="o_sp",
+ workflow_run_block_id=None,
+ )
+ )
+
+ assert captured.get("system_prompt") == "Respond as Shakespeare."
+
+
+# ---------------------------------------------------------------------------
+# FileParserBlock._extract_with_ai
+# ---------------------------------------------------------------------------
+
+
+def test_file_parser_block_forwards_system_prompt(monkeypatch) -> None:
+ from skyvern.forge.sdk.workflow.models import block as block_module
+
+ captured: dict = {}
+
+ async def fake_llm(**kwargs):
+ captured.update(kwargs)
+ return {"output": {}}
+
+ monkeypatch.setattr(
+ block_module.LLMAPIHandlerFactory,
+ "get_override_llm_api_handler",
+ lambda llm_key, default: fake_llm,
+ )
+
+ block = FileParserBlock(
+ label="fp",
+ output_parameter=_make_output_parameter(),
+ file_url="https://example.com/file.csv",
+ file_type=FileType.CSV,
+ workflow_system_prompt="Parse as JSON only.",
+ )
+
+ asyncio.run(block._extract_with_ai("hello,world\n1,2", workflow_run_context=MagicMock()))
+
+ assert captured.get("system_prompt") == "Parse as JSON only."
+
+
+# ---------------------------------------------------------------------------
+# PDFParserBlock.execute — system_prompt forwarded to LLM call
+# ---------------------------------------------------------------------------
+
+
+def test_pdf_parser_block_forwards_system_prompt(monkeypatch) -> None:
+ from skyvern.forge.sdk.workflow.models import block as block_module
+
+ captured: dict = {}
+
+ async def fake_llm(**kwargs):
+ captured.update(kwargs)
+ return {"output": {}}
+
+ # Patch the app-level handler directly since PDFParserBlock uses app.LLM_API_HANDLER.
+ monkeypatch.setattr(block_module.app, "LLM_API_HANDLER", fake_llm)
+ monkeypatch.setattr(
+ block_module,
+ "download_file",
+ AsyncMock(return_value="/tmp/file.pdf"),
+ )
+ monkeypatch.setattr(block_module, "extract_pdf_file", lambda *a, **k: "extracted text")
+
+ workflow_run_context = MagicMock()
+ workflow_run_context.has_parameter = MagicMock(return_value=False)
+ workflow_run_context.has_value = MagicMock(return_value=False)
+ workflow_run_context.workflow = None
+ workflow_run_context.resolve_effective_workflow_system_prompt = MagicMock(return_value=None)
+
+ monkeypatch.setattr(
+ PDFParserBlock,
+ "get_workflow_run_context",
+ staticmethod(lambda workflow_run_id: workflow_run_context),
+ )
+ monkeypatch.setattr(
+ PDFParserBlock,
+ "record_output_parameter_value",
+ AsyncMock(return_value=None),
+ )
+ monkeypatch.setattr(
+ PDFParserBlock,
+ "build_block_result",
+ AsyncMock(return_value=MagicMock()),
+ )
+
+ block = PDFParserBlock(
+ label="pdf",
+ output_parameter=_make_output_parameter(),
+ file_url="https://example.com/file.pdf",
+ workflow_system_prompt="Summarize in one sentence.",
+ )
+
+ asyncio.run(
+ block.execute(
+ workflow_run_id="wfr_sp_pdf",
+ workflow_run_block_id="wrb_sp_pdf",
+ organization_id="o_sp_pdf",
+ )
+ )
+
+ assert captured.get("system_prompt") == "Summarize in one sentence."