pi-subagents 0.57.0 → 0.58.0

package/src/runs/foreground/subagent-executor.ts

@@ -75,7 +75,7 @@ import { usageBudgetExceededMessage, usageBudgetState, validateUsageBudgetConfig
 import { intersectSubagentCapabilityCeilings, resolveCurrentSubagentCapabilityCeiling, type ResolvedSubagentCapabilityCeiling } from "../shared/capability-ceiling.ts";
 import { isAgentContractV1 } from "../shared/agent-contract.ts";
 import { normalizeExtensionBindings, type ExtensionBindings } from "../shared/extension-bindings.ts";
-import { finalizeSingleOutput, injectSingleOutputInstruction, normalizeSingleOutputOverride, resolveSingleOutputPath, validateFileOnlyOutputMode } from "../shared/single-output.ts";
+import { finalizeSingleOutput, injectSingleOutputInstruction, normalizeSingleOutputOverride, outputPathMappingFromTask, resolveSingleOutputPath, validateFileOnlyOutputMode } from "../shared/single-output.ts";
 import { assertJsonSchemaObject, cleanupStructuredOutputRuntime, createStructuredOutputRuntime } from "../shared/structured-output.ts";
 import { compactForegroundDetails, getSingleResultOutput, mapConcurrent, readStatus, resolveChildCwd, sumResultsCost, sumResultsUsage } from "../../shared/utils.ts";
 import { createTaskMutationArbiter } from "../shared/llm-intent-arbiter.ts";
@@ -169,6 +169,7 @@ import {
 	type ToolBudgetConfig,
 	type TurnBudgetConfig,
 	type UsageBudgetConfig,
+	type WorkflowTerminalOutcome,
 	type SubagentRunMode,
 	type SubagentState,
 	DIRS,
@@ -418,6 +419,7 @@ type ForkThinkingOverrideForTask = (agentName: string, idx?: number, modelOverri
 interface ExecutionContextData {
 	params: SubagentParamsLike;
 	effectiveCwd: string;
+	requestedCwd?: string;
 	ctx: ExtensionContext;
 	signal: AbortSignal;
 	onUpdate?: (r: AgentToolResult<Details>) => void;
@@ -1793,6 +1795,7 @@ async function resumeAsyncRun(input: {
 		systemPrompt: "",
 		systemPromptMode: recoveryDescriptor.systemPromptMode,
 		inheritProjectContext: recoveryDescriptor.inheritProjectContext,
+		inheritGlobalContext: recoveryDescriptor.inheritGlobalContext,
 		inheritSkills: recoveryDescriptor.inheritSkills,
 		source: "project",
 		filePath: recoveryDescriptor.agentFilePath ?? path.join(getProjectSubagentsDir(recoveryDescriptor.cwd), "recovery-agent"),
@@ -3237,6 +3240,7 @@ async function runAsyncPath(data: ExecutionContextData, deps: ExecutorDeps): Pro
 			ctx: asyncCtx,
 			availableModels,
 			cwd: effectiveCwd,
+			requestedCwd: data.requestedCwd,
 			maxOutput: params.maxOutput,
 			artifactsDir: artifactConfig.enabled ? artifactsDir : undefined,
 			artifactConfig,
@@ -3701,6 +3705,7 @@ async function runSinglePath(data: ExecutionContextData, deps: ExecutorDeps): Pr
 	try {
 		r = await runSync(ctx.cwd, agents, params.agent!, task, compactOptional<Parameters<typeof runSync>[4]>({
 			permissions: deps.config.permissions,
+			runtimeSnapshotHost: deps.pi,
 			parentSessionId: ctx.sessionManager.getSessionId() ?? undefined,
 			llmIntentArbiter: createTaskMutationArbiter(ctx),
 			...workflowForegroundSteeringLaunchOptions(foregroundControl, 0),
@@ -3708,6 +3713,7 @@ async function runSinglePath(data: ExecutionContextData, deps: ExecutorDeps): Pr
 			unknownAgentDiagnosticContext: data.unknownAgentDiagnosticContext,
 			runFanoutBudget: params.runFanoutAdmitted ? data.runFanoutBudget : { ...data.runFanoutBudget, parentPath: `${data.runFanoutBudget.parentPath ? `${data.runFanoutBudget.parentPath}/` : ""}single` },
 			cwd: singleCwd,
+			requestedCwd: data.requestedCwd,
 			signal,
 			interruptSignal: interruptController.signal,
 			allowIntercomDetach: agentConfig.systemPrompt?.includes(INTERCOM_BRIDGE_MARKER) === true,
@@ -3751,6 +3757,9 @@ async function runSinglePath(data: ExecutionContextData, deps: ExecutorDeps): Pr
 			},
 			onDetachedExit: (result) => {
 				try {
+					if (worktreeSetup) {
+						finalizeSingleWorktreeHandoff({ worktreeSetup, artifactsDir, runId, cwd: sourceCwd, agent: params.agent!, result });
+					}
 					try {
 						updateRememberedForegroundChild(deps.state, { runId, mode: "single", cwd: singleCwd, sessionId: data.parentSessionId, index: 0, result, events: deps.pi.events, notify: true });
 					} catch {
@@ -3998,6 +4007,7 @@ function workflowChildResult(
 	result: AgentToolResult<Details>,
 	childParams: Record<string, unknown> = {},
 	resumeState?: SubagentState,
+	forcedTerminalOutcome?: WorkflowTerminalOutcome,
 ): WorkflowScriptChildResult {
 	const receiptOutput = result.content.map((part) => part.type === "text" ? part.text : "").filter(Boolean).join("\n");
 	const output = result.details.results.length === 1 && result.details.results[0]?.finalOutput !== undefined
@@ -4010,6 +4020,12 @@ function workflowChildResult(
 	const detached = result.details.results.some((child) => child.detached);
 	const interrupted = result.details.results.some((child) => child.interrupted);
 	const stopped = result.details.results.some((child) => child.stopped);
+	const terminalOutcome = forcedTerminalOutcome
+		?? (result.details.results.some((child) => child.timedOut)
+			? { state: "partial" as const, reason: "timeout" as const }
+			: result.details.usageBudget?.exhausted || result.details.results.some((child) => child.turnBudgetExceeded || child.toolBudgetBlocked)
+				? { state: "partial" as const, reason: "budget_exhausted" as const }
+				: undefined);
 	const ok = result.isError !== true && !detached && !interrupted && !stopped;
 	const artifactPaths = new Set<string>();
 	if (result.details.asyncDir) artifactPaths.add(result.details.asyncDir);
@@ -4042,6 +4058,7 @@ function workflowChildResult(
 	const continuationRunIds = [...new Set([resumeSourceRunId, runId].filter((value): value is string => Boolean(value)))];
 	const outputReference = result.details.results.find((child) => child.savedOutputPath)?.savedOutputPath
 		?? result.details.results.find((child) => child.outputReference?.path)?.outputReference?.path;
+	const outputPathMapping = typeof childParams.task === "string" ? outputPathMappingFromTask(childParams.task, outputReference) : undefined;
 	const externalResult = result.details.results.length === 1 && result.details.results[0]?.runner?.type === "external-cli" ? result.details.results[0] : undefined;
 	const externalStatus = result.details.asyncDir ? readStatus(result.details.asyncDir) : undefined;
 	const externalStatusStep = externalStatus?.steps?.length === 1 && externalStatus.steps[0]?.runner?.type === "external-cli" ? externalStatus.steps[0] : undefined;
@@ -4052,6 +4069,7 @@ function workflowChildResult(
 	return {
 		key,
 		ok,
+		...(terminalOutcome ? { terminalOutcome } : {}),
 		...(resolvedAgents.length === 1 ? { agent: resolvedAgents[0] } : {}),
 		...(runId ? { runId } : {}),
 		output,
@@ -4063,6 +4081,7 @@ function workflowChildResult(
 		...(requestedContext ? { requestedContext } : {}),
 		...(resolvedContext ? { resolvedContext } : {}),
 		...(outputReference ? { outputReference } : {}),
+		...(outputPathMapping ? { outputPathMapping } : {}),
 		...(externalAdapter ? { externalAdapter } : {}),
 		resumability,
 		continuation: { runIds: continuationRunIds },
@@ -4071,6 +4090,13 @@ function workflowChildResult(
 	};
 }
 
+function workflowOutputPathMappingSummary(children: WorkflowScriptChildResult[]): string {
+	const mappings = children.flatMap((child) => child.outputPathMapping
+		? [`'${child.key}': requested ${child.outputPathMapping.requestedPath} -> saved ${child.outputPathMapping.savedPath}`]
+		: []);
+	return mappings.length > 0 ? ` Output path mappings: ${mappings.join("; ")}.` : "";
+}
+
 function workflowSteerReceipt(key: string, result: AgentToolResult<Details>): WorkflowSteerResult {
 	const steering = result.details.steering;
 	const error = result.content.map((part) => part.type === "text" ? part.text : "").filter(Boolean).join("\n") || undefined;
@@ -4110,8 +4136,14 @@ function terminalWorkflowReceipt(
 	state: WorkflowReceiptState,
 	children: WorkflowScriptChildResult[],
 	workflowChildren?: WorkflowReceipt["workflowChildren"],
+	terminalOutcome?: WorkflowTerminalOutcome,
 ): WorkflowReceipt {
-	return buildWorkflowReceipt({ workflowRunId, state, children, workflowChildren });
+	return buildWorkflowReceipt({ workflowRunId, state, children, workflowChildren, terminalOutcome });
+}
+
+function workflowFailureTerminalOutcome(error: unknown, _children: WorkflowScriptChildResult[], usageBudget: ReturnType<typeof usageBudgetState>): WorkflowTerminalOutcome | undefined {
+	if (usageBudget?.exhausted) return { state: "partial", reason: "budget_exhausted" };
+	return error instanceof WorkflowScriptError && error.errorKind === "timeout" ? { state: "partial", reason: "timeout" } : undefined;
 }
 
 function workflowFailureMessage(error: unknown, workflowRunId: string, children: WorkflowScriptChildResult[]): string {
@@ -4791,7 +4823,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 									version: 1,
 									childId: entry.key,
 									status: "stopped",
-									reason: "user",
+									reason: "subagent-action",
 									source: "async",
 									stepIndex: status.steps?.indexOf(projectedStep),
 									agent: projectedStep.agent,
@@ -4836,7 +4868,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 							launch: async (key, childParams, workflowSignal, admission) => {
 								if (workflowUsageBudget.budget && childParams.async === true) return workflowChildResult(key, buildRequestedModeError(childParams as SubagentParamsLike, "workflow usageBudget does not support async runs.run launches."), childParams, deps.state);
 								const budgetState = usageBudgetState(workflowUsageBudget.budget, sumResultsCost(workflowResults));
-								if (budgetState?.exhausted) return workflowChildResult(key, buildRequestedModeError(childParams as SubagentParamsLike, usageBudgetExceededMessage(budgetState)), childParams, deps.state);
+								if (budgetState?.exhausted) return workflowChildResult(key, buildRequestedModeError(childParams as SubagentParamsLike, usageBudgetExceededMessage(budgetState)), childParams, deps.state, { state: "partial", reason: "budget_exhausted" });
 								const childPhase = typeof childParams.phase === "string" && childParams.phase.trim() ? childParams.phase.trim() : undefined;
 								const childLabel = typeof childParams.label === "string" && childParams.label.trim() ? childParams.label.trim() : undefined;
 								recordMissionWorkflowChild(missionBinding, workflowRunId, key, {
@@ -4923,7 +4955,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 						});
 						const returnPreview = formatWorkflowValue(workflow.value).slice(0, 1_000);
 						const emitPreview = workflow.emits.length > 0 ? ` Emitted: ${workflow.emits.map(formatWorkflowValue).join(", ").slice(0, 1_000)}` : "";
-						const summary = `Workflow completed with ${workflow.children.length} child run(s). Return: ${returnPreview}${emitPreview} Trace: ${workflow.trace.length} event(s).`;
+						const summary = `Workflow completed with ${workflow.children.length} child run(s). Return: ${returnPreview}${emitPreview} Trace: ${workflow.trace.length} event(s).${workflowOutputPathMappingSummary(workflow.children)}`;
 						const outputWarning = writeWorkflowAggregateOutput(workflowAggregateOutputPath, summary, producedChildOutputPaths);
 						const resultSummary = appendWorkflowOutputWarning(summary, outputWarning);
 						const workflowUsage = sumResultsUsage(workflowResults);
@@ -4936,7 +4968,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 						} catch (receiptError) {
 							appendWorkflowEvent({ type: "subagent.workflow.receipt_write_failed", error: `Failed to persist async workflow receipt: ${receiptError instanceof Error ? receiptError.message : String(receiptError)}` });
 						}
-						if (!writeWorkflowResult({ id: workflowRunId, runId: workflowRunId, toolCallId, agent: "workflow", mode: "workflow", success: true, state: "complete", summary: resultSummary, output: resultSummary, workflowChildren, results: workflow.children.map((child) => ({ workflowKey: child.key, ...(child.agent ? { agent: child.agent } : {}), ...(child.runId ? { runId: child.runId } : {}), output: child.output, outputState: child.output.trim() || child.structuredOutput !== undefined ? "present" : "absent", structuredOutput: child.structuredOutput, success: child.ok, ...(child.stopped ? { stopped: true } : {}), ...(child.interrupted ? { interrupted: true } : {}), ...(child.artifactPaths[0] ? { artifactPaths: { outputPath: child.artifactPaths[0] } } : {}) })), workflow: status.workflow, ...(workflowReceipt ? { workflowReceipt } : {}), asyncDir, cwd: workflowCwd, sessionId: currentSessionId, completionOwnerId, ...(requestParams.scheduleOrigin ? { scheduleOrigin: requestParams.scheduleOrigin } : {}), timestamp: Date.now(), durationMs: Date.now() - startedAt })) return;
+						if (!writeWorkflowResult({ id: workflowRunId, runId: workflowRunId, toolCallId, agent: "workflow", mode: "workflow", success: true, state: "complete", summary: resultSummary, output: resultSummary, workflowChildren, results: workflow.children.map((child) => ({ workflowKey: child.key, ...(child.agent ? { agent: child.agent } : {}), ...(child.runId ? { runId: child.runId } : {}), output: child.output, outputState: child.output.trim() || child.structuredOutput !== undefined ? "present" : "absent", structuredOutput: child.structuredOutput, success: child.ok, ...(child.outputPathMapping ? { outputPathMapping: child.outputPathMapping } : {}), ...(child.stopped ? { stopped: true } : {}), ...(child.interrupted ? { interrupted: true } : {}), ...(child.artifactPaths[0] ? { artifactPaths: { outputPath: child.artifactPaths[0] } } : {}) })), workflow: status.workflow, ...(workflowReceipt ? { workflowReceipt } : {}), asyncDir, cwd: workflowCwd, sessionId: currentSessionId, completionOwnerId, ...(requestParams.scheduleOrigin ? { scheduleOrigin: requestParams.scheduleOrigin } : {}), timestamp: Date.now(), durationMs: Date.now() - startedAt })) return;
 						persist();
 						persistClosed = true;
 						appendWorkflowEvent({ type: "subagent.workflow.completed", state: status.state, ...(status.error ? { error: status.error } : {}) });
@@ -4965,23 +4997,24 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 							const promoted = promotePausedWorkflowIfSettled(status);
 							if (promoted) status = promoted;
 						}
-						const terminalSummary = status.state === "complete"
+						const terminalSummary = `${status.state === "complete"
 							? "Workflow completed after detached child finished."
-							: status.error ?? (pauseForDetached ? "Workflow paused." : "Workflow failed.");
+							: status.error ?? (pauseForDetached ? "Workflow paused." : "Workflow failed.")}${workflowOutputPathMappingSummary(partial.children)}`;
 						const outputWarning = writeWorkflowAggregateOutput(workflowAggregateOutputPath, terminalSummary, producedChildOutputPaths);
 						const resultSummary = appendWorkflowOutputWarning(terminalSummary, outputWarning);
 						const receiptState: WorkflowReceiptState = status.state === "complete" ? "complete" : status.state === "paused" ? "paused" : status.state === "stopped" ? "stopped" : "failed";
-						const receipt = terminalWorkflowReceipt(workflowRunId, receiptState, partial.children, workflowChildren);
+						const terminalOutcome = workflowFailureTerminalOutcome(error, partial.children, usageBudgetState(workflowUsageBudget.budget, sumResultsCost(workflowResults)));
+						const receipt = terminalWorkflowReceipt(workflowRunId, receiptState, partial.children, workflowChildren, terminalOutcome);
 						let workflowReceipt: { path: string; receipt: WorkflowReceipt } | undefined;
 						try {
 							workflowReceipt = { path: writeWorkflowReceipt(asyncDir, receipt), receipt };
 						} catch (receiptError) {
 							appendWorkflowEvent({ type: "subagent.workflow.receipt_write_failed", error: `Failed to persist async workflow receipt: ${receiptError instanceof Error ? receiptError.message : String(receiptError)}` });
 						}
-						if (!writeWorkflowResult({ id: workflowRunId, runId: workflowRunId, toolCallId, agent: "workflow", mode: "workflow", success: status.state === "complete", state: status.state, summary: resultSummary, error: status.state === "complete" ? undefined : status.error, stopped: status.stopped, activityState: status.activityState, workflowChildren, results: partial.children.map((child) => ({ workflowKey: child.key, ...(child.agent ? { agent: child.agent } : {}), ...(child.runId ? { runId: child.runId } : {}), output: child.output, outputState: child.output.trim() || child.structuredOutput !== undefined ? "present" : "absent", structuredOutput: child.structuredOutput, success: child.ok, ...(child.stopped ? { stopped: true } : {}), ...(child.interrupted ? { interrupted: true } : {}), ...(child.detached && status.state !== "complete" ? { detached: true } : {}), ...(child.artifactPaths[0] ? { artifactPaths: { outputPath: child.artifactPaths[0] } } : {}) })), workflow: status.workflow, ...(workflowReceipt ? { workflowReceipt } : {}), asyncDir, cwd: workflowCwd, sessionId: currentSessionId, completionOwnerId, ...(requestParams.scheduleOrigin ? { scheduleOrigin: requestParams.scheduleOrigin } : {}), timestamp: Date.now(), durationMs: Date.now() - startedAt })) return;
+						if (!writeWorkflowResult({ id: workflowRunId, runId: workflowRunId, toolCallId, agent: "workflow", mode: "workflow", success: status.state === "complete", state: status.state, summary: resultSummary, error: status.state === "complete" ? undefined : status.error, stopped: status.stopped, activityState: status.activityState, workflowChildren, ...(terminalOutcome ? { terminalOutcome } : {}), results: partial.children.map((child) => ({ workflowKey: child.key, ...(child.agent ? { agent: child.agent } : {}), ...(child.runId ? { runId: child.runId } : {}), output: child.output, outputState: child.output.trim() || child.structuredOutput !== undefined ? "present" : "absent", structuredOutput: child.structuredOutput, success: child.ok, ...(child.terminalOutcome ? { terminalOutcome: child.terminalOutcome } : {}), ...(child.outputPathMapping ? { outputPathMapping: child.outputPathMapping } : {}), ...(child.stopped ? { stopped: true } : {}), ...(child.interrupted ? { interrupted: true } : {}), ...(child.detached && status.state !== "complete" ? { detached: true } : {}), ...(child.artifactPaths[0] ? { artifactPaths: { outputPath: child.artifactPaths[0] } } : {}) })), workflow: status.workflow, ...(workflowReceipt ? { workflowReceipt } : {}), asyncDir, cwd: workflowCwd, sessionId: currentSessionId, completionOwnerId, ...(requestParams.scheduleOrigin ? { scheduleOrigin: requestParams.scheduleOrigin } : {}), timestamp: Date.now(), durationMs: Date.now() - startedAt })) return;
 						persist();
 						persistClosed = true;
-						appendWorkflowEvent({ type: "subagent.workflow.completed", state: status.state, ...(status.error ? { error: status.error } : {}), ...(status.activityState ? { activityState: status.activityState } : {}) });
+						appendWorkflowEvent({ type: "subagent.workflow.completed", state: status.state, ...(terminalOutcome ? { terminalOutcome } : {}), ...(status.error ? { error: status.error } : {}), ...(status.activityState ? { activityState: status.activityState } : {}) });
 					} finally {
 						persistClosed = true;
 						deps.state.workflowControllers?.delete(workflowRunId);
@@ -5041,7 +5074,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 						if (delegatedWorkflowPermit && childParams.async === true) throw new Error("Workflow child permit supports one foreground child only.");
 						if (workflowUsageBudget.budget && childParams.async === true) return workflowChildResult(key, buildRequestedModeError(childParams as SubagentParamsLike, "workflow usageBudget does not support async runs.run launches."), childParams, deps.state);
 						const budgetState = usageBudgetState(workflowUsageBudget.budget, sumResultsCost(workflowResults));
-						if (budgetState?.exhausted) return workflowChildResult(key, buildRequestedModeError(childParams as SubagentParamsLike, usageBudgetExceededMessage(budgetState)), childParams, deps.state);
+						if (budgetState?.exhausted) return workflowChildResult(key, buildRequestedModeError(childParams as SubagentParamsLike, usageBudgetExceededMessage(budgetState)), childParams, deps.state, { state: "partial", reason: "budget_exhausted" });
 						const childPhase = typeof childParams.phase === "string" && childParams.phase.trim() ? childParams.phase.trim() : undefined;
 						const childLabel = typeof childParams.label === "string" && childParams.label.trim() ? childParams.label.trim() : undefined;
 						recordMissionWorkflowChild(missionBinding, _id, key, {
@@ -5111,6 +5144,8 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 				if (workflow.emits.length > 0) sections.push(`Emitted:\n${workflow.emits.map(formatWorkflowValue).join("\n")}`);
 				if (workflow.console.length > 0) sections.push(`Console:\n${workflow.console.map((entry) => `[${entry.level}] ${entry.text}`).join("\n")}`);
 				if (traceLines.length > 0) sections.push(`Call trace:\n${traceLines.join("\n")}`);
+				const outputMappings = workflowOutputPathMappingSummary(workflow.children).trim();
+				if (outputMappings) sections.push(outputMappings);
 				const workflowText = sections.join("\n\n");
 				const outputWarning = writeWorkflowAggregateOutput(workflowAggregateOutputPath, workflowText, producedChildOutputPaths);
 				const displayText = appendWorkflowOutputWarning(workflowText, outputWarning);
@@ -5126,11 +5161,14 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 				if (partial.emits.length > 0) sections.push(`Emitted:\n${partial.emits.map(formatWorkflowValue).join("\n")}`);
 				if (partial.console.length > 0) sections.push(`Console:\n${partial.console.map((entry) => `[${entry.level}] ${entry.text}`).join("\n")}`);
 				if (traceLines.length > 0) sections.push(`Call trace:\n${traceLines.join("\n")}`);
+				const outputMappings = workflowOutputPathMappingSummary(partial.children).trim();
+				if (outputMappings) sections.push(outputMappings);
 				const workflowText = sections.join("\n\n");
 				const outputWarning = writeWorkflowAggregateOutput(workflowAggregateOutputPath, workflowText, producedChildOutputPaths);
 				const displayText = appendWorkflowOutputWarning(workflowText, outputWarning);
 				const workflowChildren = workflowChildSummary({ parentToolCallId: _id, workflowRunId: _id, workflowState: "failed", inventoryComplete: true, trace: partial.trace, children: partial.children });
-				const receipt = terminalWorkflowReceipt(_id, "failed", partial.children, workflowChildren);
+				const terminalOutcome = workflowFailureTerminalOutcome(error, partial.children, usageBudgetState(workflowUsageBudget.budget, sumResultsCost(workflowResults)));
+				const receipt = terminalWorkflowReceipt(_id, "failed", partial.children, workflowChildren, terminalOutcome);
 				return attachWorkflowMission(withRunFanoutBudget({
 					content: [{ type: "text", text: displayText }],
 					isError: true,
@@ -5139,6 +5177,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 			}
 		}
 		const directParams = requestParams;
+		const requestedCwd = directParams.cwd;
 		const requestCwd = resolveRequestedCwd(ctx.cwd, directParams.cwd);
 		const paramsWithResolvedCwd = directParams.cwd === undefined ? directParams : { ...directParams, cwd: requestCwd };
 		const action = paramsWithResolvedCwd.action;
@@ -5612,7 +5651,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 						if (!isStoppableAsyncStatusStep(resolution.child.step)) return { content: [{ type: "text", text: `Child '${paramsWithResolvedCwd.childId}' in async run '${targetRunId}' is ${resolution.child.step.status}; stop only supports pending or running children.` }], isError: true, details: { mode: "management", results: [] } };
 						const stopChild = deps.state.workflowChildStops?.get(targetRunId!);
 						if (!stopChild) return { content: [{ type: "text", text: `Workflow ${targetRunId} child stop is unavailable in this extension runtime.` }], isError: true, details: { mode: "management", results: [] } };
-						if (!stopChild(resolution.child.id)) return { content: [{ type: "text", text: `Child '${paramsWithResolvedCwd.childId}' in workflow ${workflowRunId} is not available to stop.` }], isError: true, details: { mode: "management", results: [] } };
+						if (!stopChild(resolution.child.id, `Workflow child '${resolution.child.id}' stopped.`)) return { content: [{ type: "text", text: `Child '${paramsWithResolvedCwd.childId}' in workflow ${workflowRunId} is not available to stop.` }], isError: true, details: { mode: "management", results: [] } };
 						try {
 							fs.appendFileSync(path.join(asyncJob.asyncDir, "events.jsonl"), `${JSON.stringify({
 								type: "subagent.child-status",
@@ -5621,7 +5660,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 								childId: resolution.child.id,
 								status: "stopping",
 								ts: Date.now(),
-								reason: "user",
+								reason: "subagent-action",
 								source: "async",
 								stepIndex: resolution.child.index,
 								agent: resolution.child.step.agent,
@@ -5635,17 +5674,13 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 						}
 						return { content: [{ type: "text", text: `Stop requested for child ${resolution.child.id} in async workflow ${workflowRunId}.` }], details: { mode: "management", results: [] } };
 					}
-					workflowController.abort(new Error("Workflow stopped by user."));
+					workflowController.abort(new Error("Workflow stopped."));
 					return { content: [{ type: "text", text: `Stop requested for async workflow ${targetRunId}.` }], details: { mode: "management", results: [] } };
 				}
 				let resolved: ResolvedSubagentRunId | undefined;
 				if (paramsWithResolvedCwd.dir) {
 					try {
 						const location = resolveAsyncRunLocation(paramsWithResolvedCwd, DIRS.async, DIRS.results);
-						const existingStatus = readStatus(location.asyncDir ?? "");
-						if (existingStatus?.mode === "workflow" && existingStatus.state === "running") {
-							return { content: [{ type: "text", text: `Workflow ${existingStatus.runId} is not controlled by this extension runtime; reload recovery cannot stop it safely.` }], isError: true, details: { mode: "management", results: [] } };
-						}
 						const stopResult = stopAsyncRun(deps.state, location.resolvedId ?? targetRunId ?? path.basename(location.asyncDir ?? paramsWithResolvedCwd.dir), deps.kill, location, paramsWithResolvedCwd.childId);
 						return stopResult ?? { content: [{ type: "text", text: `No running or queued async run was found for '${targetRunId ?? paramsWithResolvedCwd.dir}'.` }], isError: true, details: { mode: "management", results: [] } };
 					} catch (error) {
@@ -5662,12 +5697,6 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 				}
 				if (resolved?.kind === "nested") return { content: [{ type: "text", text: "action='stop' supports current-session top-level async runs only." }], isError: true, details: { mode: "management", results: [] } };
 				if (resolved?.kind === "foreground") return { content: [{ type: "text", text: "action='stop' supports async runs only. Use action='interrupt' for foreground runs." }], isError: true, details: { mode: "management", results: [] } };
-				if (resolved?.kind === "async") {
-					const existingStatus = readStatus(resolved.location.asyncDir ?? "");
-					if (existingStatus?.mode === "workflow" && existingStatus.state === "running") {
-						return { content: [{ type: "text", text: `Workflow ${resolved.id} is not controlled by this extension runtime; reload recovery cannot stop it safely.` }], isError: true, details: { mode: "management", results: [] } };
-					}
-				}
 				const stopResult = stopAsyncRun(
 					deps.state,
 					resolved?.kind === "async" ? resolved.id : targetRunId,
@@ -5869,6 +5898,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 						parentModel,
 						forkAvailableModels,
 						agentConfig?.modelProvider ?? parentModel?.provider,
+						{ source: "inherited" },
 					);
 				const candidates = buildModelCandidates(
 					primaryModel,
@@ -6085,6 +6115,7 @@ export function createSubagentExecutor(deps: ExecutorDeps): {
 		const execData: ExecutionContextData = omitUndefinedProperties({
 			params: effectiveParams,
 			effectiveCwd,
+			requestedCwd,
 			ctx,
 			signal,
 			onUpdate: onUpdateWithContext,

package/src/runs/foreground/workflow-detach-reconcile.ts

@@ -9,11 +9,15 @@ import {
 	type IntercomEventBus,
 	type SingleResult,
 	type SubagentState,
+	type WorkflowRecoveryAction,
+	type WorkflowTerminalOutcome,
+	type WorkflowTerminalResolution,
 } from "../../shared/types.ts";
 import { updateActiveRunIndex } from "../background/active-run-index.ts";
 import { resultFilePath, writeAsyncResultFile } from "../background/result-files.ts";
 import { resolveAsyncResumeTarget } from "../background/async-resume.ts";
 import { externalCliReceiptMetadata, normalizeExternalCliRunnerStatus } from "../shared/external-cli-contract.ts";
+import { outputPathMappingFromTask } from "../shared/single-output.ts";
 import { readWorkflowReceipt, workflowReceiptPath, writeWorkflowReceipt, type WorkflowReceipt } from "../../workflows/workflow-receipt.ts";
 import { workflowChildSummary } from "../../workflows/workflow-child-summary.ts";
 
@@ -31,18 +35,33 @@ function childSucceeded(result: Pick<SingleResult, "exitCode" | "error" | "inter
 
 const UNSUPPORTED_DETACHED_WORKFLOW_CONTINUATION = "unsupported-continuation: detached workflow child settled, but JavaScript workflow continuation was not persisted. Resume the workflow explicitly instead of treating the completed child as top-level workflow completion.";
 const INTERRUPTED_DETACHED_CHILD = "Interrupted. Waiting for explicit next action.";
+type WorkflowStatusStep = NonNullable<AsyncStatus["steps"]>[number] & {
+	outputPathMapping?: { requestedPath: string; savedPath: string };
+	interrupted?: boolean;
+};
+
+function findWorkflowStep(status: AsyncStatus, childRunId: string, workflowKey?: string): WorkflowStatusStep | undefined {
+	return status.steps?.find((candidate) => candidate.runId === childRunId)
+		?? (workflowKey ? status.steps?.find((candidate) => candidate.workflowKey === workflowKey) : undefined) as WorkflowStatusStep | undefined;
+}
 
 export function applyDetachedChildToPausedWorkflow(
 	status: AsyncStatus,
-	input: { childRunId: string; result: Pick<SingleResult, "exitCode" | "error" | "interrupted" | "sessionFile">; workflowKey?: string },
+	input: { childRunId: string; result: Pick<SingleResult, "exitCode" | "error" | "interrupted" | "sessionFile" | "stopped">; workflowKey?: string },
 ): AsyncStatus | undefined {
 	if (status.mode !== "workflow" || status.state !== "paused") return undefined;
 	const next = cloneWorkflowStatus(status);
-	const step = next.steps?.find((candidate) => candidate.runId === input.childRunId)
-		?? (input.workflowKey ? next.steps?.find((candidate) => candidate.workflowKey === input.workflowKey) : undefined);
+	const step = findWorkflowStep(next, input.childRunId, input.workflowKey);
 	if (!step) return undefined;
 	const succeeded = childSucceeded(input.result);
-	const failedSiblingError = next.steps?.find((candidate) => candidate !== step && candidate.status === "failed" && candidate.error)?.error;
+	const failedSiblingError = next.steps?.find((candidate) => {
+		const candidateStep = candidate as WorkflowStatusStep;
+		return candidateStep !== step && candidateStep.status === "failed" && !candidateStep.interrupted && candidateStep.error;
+	})?.error;
+	const interruptedChildError = next.steps?.find((candidate) => {
+		const candidateStep = candidate as WorkflowStatusStep;
+		return candidateStep.status === "failed" && candidateStep.interrupted && candidateStep.error;
+	})?.error;
 	const updatedAt = Date.now();
 	step.status = succeeded ? "completed" : "failed";
 	step.endedAt = updatedAt;
@@ -50,13 +69,24 @@ export function applyDetachedChildToPausedWorkflow(
 	delete step.currentTool;
 	delete step.currentToolStartedAt;
 	if (input.result.sessionFile) step.sessionFile = input.result.sessionFile;
-	if (succeeded) delete step.error;
-	else if (input.result.interrupted) step.error = input.result.error ?? INTERRUPTED_DETACHED_CHILD;
-	else if (input.result.error) step.error = input.result.error;
+	if (succeeded) {
+		delete step.error;
+		delete step.interrupted;
+	} else if (input.result.interrupted || input.result.stopped) {
+		step.error = input.result.error ?? INTERRUPTED_DETACHED_CHILD;
+		step.interrupted = true;
+		if (input.result.stopped) step.stopped = true;
+	} else {
+		delete step.interrupted;
+		delete step.stopped;
+		if (input.result.error) step.error = input.result.error;
+	}
 	next.lastUpdate = updatedAt;
 	const promoted = promotePausedWorkflowIfSettled(next);
-	if (promoted?.state === "failed" && input.result.interrupted && !failedSiblingError) promoted.error = input.result.error ?? INTERRUPTED_DETACHED_CHILD;
+	if (promoted?.state === "failed" && failedSiblingError) promoted.error = failedSiblingError;
+	else if (promoted?.state === "failed" && (input.result.interrupted || input.result.stopped)) promoted.error = input.result.error ?? INTERRUPTED_DETACHED_CHILD;
 	else if (promoted?.state === "failed" && input.result.error) promoted.error = input.result.error;
+	else if (promoted?.state === "failed" && interruptedChildError) promoted.error = interruptedChildError;
 	const resolved = promoted ?? next;
 	const workflowState = resolved.state === "complete" ? "completed" : resolved.state === "paused" ? "paused" : resolved.state === "stopped" ? "stopped" : "failed";
 	resolved.workflowChildren = workflowChildSummary({ parentToolCallId: resolved.toolCallId ?? resolved.runId, workflowRunId: resolved.runId, workflowState, inventoryComplete: true, trace: resolved.workflow?.trace, steps: resolved.steps });
@@ -82,8 +112,37 @@ export function promotePausedWorkflowIfSettled(status: AsyncStatus): AsyncStatus
 	return next;
 }
 
-function workflowResultChildren(status: AsyncStatus, childRunId: string, result: SingleResult, existingResults: unknown): unknown {
+function workflowResolution(status: AsyncStatus, result: Pick<SingleResult, "interrupted">): WorkflowTerminalResolution | undefined {
+	if (status.state !== "complete" && status.state !== "failed") return undefined;
+	if (status.steps?.some((step) => {
+		const candidate = step as WorkflowStatusStep;
+		return candidate.status === "failed" && !candidate.interrupted;
+	})) return "failed-child";
+	if (result.interrupted || status.steps?.some((step) => {
+		const candidate = step as WorkflowStatusStep;
+		return candidate.status === "stopped" || candidate.stopped || candidate.interrupted || candidate.error === INTERRUPTED_DETACHED_CHILD;
+	})) return "interrupted-child";
+	return "settled-awaiting-resume";
+}
+
+function workflowRecovery(receipt: WorkflowReceipt | undefined): WorkflowRecoveryAction[] {
+	if (!receipt) return [];
+	return Object.values(receipt.entries).flatMap((entry) => entry.resumability.state === "resumable"
+		? [{ key: entry.key, call: "runs.run" as const, resume: { workflowRunId: receipt.workflowRunId, key: entry.key, latest: true as const }, taskRequired: true as const }]
+		: []);
+}
+
+function resultTerminalOutcome(result: Pick<SingleResult, "timedOut" | "turnBudgetExceeded" | "toolBudgetBlocked">): WorkflowTerminalOutcome | undefined {
+	if (result.timedOut) return { state: "partial", reason: "timeout" };
+	if (result.turnBudgetExceeded || result.toolBudgetBlocked) return { state: "partial", reason: "budget_exhausted" };
+	return undefined;
+}
+
+function workflowResultChildren(status: AsyncStatus, childRunId: string, result: SingleResult, existingResults: unknown, receipt?: WorkflowReceipt): unknown {
 	const output = getSingleResultOutput(result);
+	const outputReference = result.savedOutputPath ?? result.outputReference?.path;
+	const outputPathMapping = outputPathMappingFromTask(result.task, outputReference);
+	const terminalOutcome = resultTerminalOutcome(result);
 	if (Array.isArray(existingResults)) {
 		return existingResults.map((entry) => {
 			if (!entry || typeof entry !== "object" || Array.isArray(entry)) return entry;
@@ -95,28 +154,56 @@ function workflowResultChildren(status: AsyncStatus, childRunId: string, result:
 				output,
 				outputState: output.trim() ? "present" : "absent",
 				detached: undefined,
-				...(result.interrupted ? { interrupted: true } : {}),
+				...(outputReference ? { outputReference } : {}),
+				...(outputPathMapping ? { outputPathMapping } : {}),
+				...(result.interrupted || result.stopped ? { interrupted: true } : {}),
+				...(result.stopped ? { stopped: true } : {}),
+				...(terminalOutcome ? { terminalOutcome } : {}),
 				...(result.error ? { error: result.error } : {}),
 			};
 		});
 	}
-	return status.steps?.map((step) => ({
+	return status.steps?.map((step: WorkflowStatusStep) => ({
 		workflowKey: step.workflowKey,
 		agent: step.agent,
 		runId: step.runId,
 		success: step.status === "completed" || step.status === "complete",
 		output: step.runId === childRunId ? output : "",
 		outputState: step.runId === childRunId && output.trim() ? "present" : "absent",
-		...(step.runId === childRunId && result.interrupted ? { interrupted: true } : {}),
+		...(step.runId === childRunId && outputReference ? { outputReference } : step.workflowKey && receipt?.entries[step.workflowKey]?.outputReference ? { outputReference: receipt.entries[step.workflowKey]!.outputReference } : {}),
+		...(step.runId === childRunId && outputPathMapping ? { outputPathMapping } : step.outputPathMapping ? { outputPathMapping: step.outputPathMapping } : {}),
+		...(step.interrupted ? { interrupted: true } : {}),
+		...(step.runId === childRunId && terminalOutcome ? { terminalOutcome } : step.workflowKey && receipt?.entries[step.workflowKey]?.terminalOutcome ? { terminalOutcome: receipt.entries[step.workflowKey]!.terminalOutcome } : {}),
+		...(step.stopped ? { stopped: true } : {}),
 		...(step.error ? { error: step.error } : {}),
 	}));
 }
 
+function workflowResultOutputPathMappingSummary(results: unknown): string {
+	if (!Array.isArray(results)) return "";
+	const mappings = results.flatMap((entry) => {
+		if (!entry || typeof entry !== "object" || Array.isArray(entry)) return [];
+		const child = entry as Record<string, unknown>;
+		const mapping = child.outputPathMapping;
+		if (!mapping || typeof mapping !== "object" || Array.isArray(mapping)) return [];
+		const { requestedPath, savedPath } = mapping as Record<string, unknown>;
+		if (typeof requestedPath !== "string" || typeof savedPath !== "string") return [];
+		const key = typeof child.workflowKey === "string" ? child.workflowKey : "child";
+		return [`'${key}': requested ${requestedPath} -> saved ${savedPath}`];
+	});
+	return mappings.length > 0 ? ` Output path mappings: ${mappings.join("; ")}.` : "";
+}
+
 function publishedWorkflowResult(status: AsyncStatus, childRunId: string, result: SingleResult, asyncDir: string, existing?: Record<string, unknown>, receipt?: WorkflowReceipt): Record<string, unknown> {
 	const sessionId = status.sessionId ?? (typeof existing?.sessionId === "string" ? existing.sessionId : undefined);
-	const summary = status.state === "complete"
+	const resolution = workflowResolution(status, result);
+	const recovery = workflowRecovery(receipt);
+	const results = workflowResultChildren(status, childRunId, result, existing?.results, receipt);
+	const summary = `${resolution === "settled-awaiting-resume"
+		? `Workflow lanes settled after detached child ${childRunId} finished. JavaScript workflow continuation was not persisted.${recovery.length ? " Use the listed keyed recovery action to continue a child." : " No retained child is resumable."}`
+		: status.state === "complete"
 		? `Workflow completed after detached child ${childRunId} finished.`
-		: status.error ?? (typeof existing?.summary === "string" ? existing.summary : undefined);
+		: status.error ?? (typeof existing?.summary === "string" ? existing.summary : "Workflow failed.")}${workflowResultOutputPathMappingSummary(results)}`;
 	return {
 		...(existing ?? {}),
 		id: status.runId,
@@ -131,9 +218,11 @@ function publishedWorkflowResult(status: AsyncStatus, childRunId: string, result
 		activityState: status.activityState,
 		endedAt: status.endedAt,
 		timestamp: Date.now(),
-		results: workflowResultChildren(status, childRunId, result, existing?.results),
+		results,
 		workflow: status.workflow,
 		workflowChildren: status.workflowChildren,
+		...(receipt?.terminalOutcome ? { terminalOutcome: receipt.terminalOutcome } : {}),
+		...(resolution ? { workflowResolution: resolution, recovery } : {}),
 		reconciledFromDetachedChild: childRunId,
 		...(receipt ? { workflowReceipt: { path: path.join(asyncDir, "workflow-receipt.json"), receipt } } : {}),
 		asyncDir,
@@ -143,7 +232,7 @@ function publishedWorkflowResult(status: AsyncStatus, childRunId: string, result
 	};
 }
 
-function reconcileWorkflowReceipt(status: AsyncStatus, childRunId: string, result: SingleResult, asyncDir: string): WorkflowReceipt | undefined {
+function reconcileWorkflowReceipt(status: AsyncStatus, childRunId: string, result: SingleResult, asyncDir: string, resolution: WorkflowTerminalResolution | undefined): WorkflowReceipt | undefined {
 	const receiptPath = workflowReceiptPath(DIRS.async, status.runId);
 	if (!fs.existsSync(receiptPath)) return undefined;
 	const receipt = readWorkflowReceipt(DIRS.async, status.runId);
@@ -165,6 +254,7 @@ function reconcileWorkflowReceipt(status: AsyncStatus, childRunId: string, resul
 	const externalRunner = normalizeExternalCliRunnerStatus(result.runner?.type === "external-cli" ? result.runner : externalStep?.runner);
 	const externalProcess = result.externalProcess ?? externalStep?.externalProcess;
 	const externalAdapter = externalRunner ? externalCliReceiptMetadata({ runner: externalRunner, externalProcess, outputReference }) : entry.externalAdapter;
+	const childTerminalOutcome = resultTerminalOutcome(result) ?? entry.terminalOutcome;
 	if (externalAdapter) resumability = { state: "not-resumable", reason: externalAdapter.nonResumableReason };
 	const updatedEntry: WorkflowReceipt["entries"][string] = resumability.state === "resumable"
 		? {
@@ -174,6 +264,7 @@ function reconcileWorkflowReceipt(status: AsyncStatus, childRunId: string, resul
 			latestRunId: entry.latestRunId ?? childRunId,
 			resumability,
 			...(outputReference ? { outputReference } : {}),
+			...(childTerminalOutcome ? { terminalOutcome: childTerminalOutcome } : {}),
 			...(externalAdapter ? { externalAdapter } : {}),
 		}
 		: {
@@ -182,6 +273,7 @@ function reconcileWorkflowReceipt(status: AsyncStatus, childRunId: string, resul
 			...(step.context ? { resolvedContext: step.context } : {}),
 			resumability,
 			...(outputReference ? { outputReference } : {}),
+			...(childTerminalOutcome ? { terminalOutcome: childTerminalOutcome } : {}),
 			...(externalAdapter ? { externalAdapter } : {}),
 		};
 	const next: WorkflowReceipt = {
@@ -192,7 +284,14 @@ function reconcileWorkflowReceipt(status: AsyncStatus, childRunId: string, resul
 			[key]: updatedEntry,
 		},
 		workflowChildren: status.workflowChildren,
+		...(receipt.terminalOutcome ? { terminalOutcome: receipt.terminalOutcome } : {}),
+		...(resolution ? { workflowResolution: resolution } : {}),
 	};
+	if (resolution) next.recovery = workflowRecovery(next);
+	else {
+		delete next.workflowResolution;
+		delete next.recovery;
+	}
 	writeWorkflowReceipt(asyncDir, next);
 	return next;
 }
@@ -224,6 +323,10 @@ export function reconcileDetachedWorkflowChildCompletion(input: {
 		workflowKey: input.workflowKey,
 	});
 	if (!next) return false;
+	const outputReference = input.result.savedOutputPath ?? input.result.outputReference?.path;
+	const outputPathMapping = outputPathMappingFromTask(input.result.task, outputReference);
+	const settledStep = findWorkflowStep(next, input.childRunId, input.workflowKey);
+	if (settledStep && outputPathMapping) settledStep.outputPathMapping = outputPathMapping;
 	writeAtomicJson(path.join(asyncDir, "status.json"), next);
 	updateActiveRunIndex(asyncDir, next.state, next.toolCallId);
 	if (job) {
@@ -243,8 +346,9 @@ export function reconcileDetachedWorkflowChildCompletion(input: {
 	}
 	let receipt: WorkflowReceipt | undefined;
 	let receiptError: string | undefined;
+	const resolution = workflowResolution(next, input.result);
 	try {
-		receipt = reconcileWorkflowReceipt(next, input.childRunId, input.result, asyncDir);
+		receipt = reconcileWorkflowReceipt(next, input.childRunId, input.result, asyncDir, resolution);
 	} catch (error) {
 		receiptError = `Failed to reconcile async workflow receipt: ${error instanceof Error ? error.message : String(error)}`;
 	}
@@ -260,11 +364,15 @@ export function reconcileDetachedWorkflowChildCompletion(input: {
 		});
 	}
 	if (next.state === "complete" || next.state === "failed") {
+		if (!resolution) throw new Error(`Terminal detached workflow '${input.workflowRunId}' has no resolution classification.`);
 		appendDetachedWorkflowEvent(asyncDir, {
 			ts: Date.now(),
 			runId: input.workflowRunId,
 			type: "subagent.workflow.completed",
 			state: next.state,
+			workflowResolution: resolution,
+			...(receipt?.terminalOutcome ? { terminalOutcome: receipt.terminalOutcome } : {}),
+			recovery: workflowRecovery(receipt),
 			...(next.error ? { error: next.error } : {}),
 			reconciledFromDetachedChild: input.childRunId,
 		});
@@ -276,6 +384,9 @@ export function reconcileDetachedWorkflowChildCompletion(input: {
 			agent: "workflow",
 			success: next.state === "complete",
 			state: next.state,
+			workflowResolution: resolution,
+			...(receipt?.terminalOutcome ? { terminalOutcome: receipt.terminalOutcome } : {}),
+			recovery: workflowRecovery(receipt),
 			summary: typeof published.summary === "string" ? published.summary : (next.state === "complete" ? "Workflow completed." : next.error),
 			reconciledFromDetachedChild: input.childRunId,
 			...(Array.isArray(published.results) ? { results: published.results } : {}),

package/src/runs/shared/completion-guard.ts

@@ -47,6 +47,7 @@ interface CompletionMutationGuardInput {
 	messages: Message[];
 	tools?: string[];
 	mcpDirectTools?: string[];
+	mutationTools?: string[];
 	toolAvailabilityError?: string;
 	mutationEvidence?: TrackedMutationEvidence;
 }
@@ -128,7 +129,7 @@ function hasCheckpointMutationEvidence(message: Message): boolean {
 		&& data.beforeCommit !== data.afterCommit;
 }
 
-export function hasMutationToolCall(messages: Message[]): boolean {
+export function hasMutationToolCall(messages: Message[], mutationTools?: readonly string[]): boolean {
 	for (const message of messages) {
 		if (hasCheckpointMutationEvidence(message)) return true;
 		if (message.role !== "assistant") continue;
@@ -138,7 +139,7 @@ export function hasMutationToolCall(messages: Message[]): boolean {
 			const args = typeof part.arguments === "object" && part.arguments !== null && !Array.isArray(part.arguments)
 				? part.arguments as Record<string, unknown>
 				: {};
-			if (isMutatingTool(part.name, args)) return true;
+			if (isMutatingTool(part.name, args, mutationTools)) return true;
 		}
 	}
 	return false;
@@ -239,7 +240,7 @@ export function evaluateCompletionMutationGuard(input: CompletionMutationGuardIn
 	const expectedMutation = hasMutationToolCapability(input.tools, input.mcpDirectTools)
 		? expectsImplementationMutation(input.agent, input.task)
 		: false;
-	const attemptedMutation = hasMutationToolCall(input.messages) || input.mutationEvidence?.attemptedMutation === true;
+	const attemptedMutation = hasMutationToolCall(input.messages, input.mutationTools) || input.mutationEvidence?.attemptedMutation === true;
 	const noEditChallengeComplete = isImplementationChallengeTask(input.task)
 		&& reportsNoBetterChallengeChange(input.messages);
 	return {