pi-subagents 0.57.0 → 0.58.0

package/src/runs/background/subagent-runner.ts

@@ -14,6 +14,7 @@ import { createChildTranscriptWriter, type ChildTranscriptWriter } from "../../s
 import { closeSteerInbox, consumeInterruptRequest, consumeSteerRequests, deliverInterruptRequest, deliverStopRequest, deliverTimeoutRequest, enqueueStepSteer, steerAcksDir, steerCapabilityPath, stepSteerInboxDir, watchAsyncControlInbox, type SteerAck, type SteerCapability, type SteerRequest, type StopRequest } from "./control-channel.ts";
 import { appendJsonl as appendRawJsonl, formatOutputArtifactContent, getArtifactPaths, writeArtifact, writeMetadata } from "../../shared/artifacts.ts";
 import { PI_CODING_AGENT_PACKAGE, getPiSpawnCommand, resolveInstalledPiPackageRoot } from "../shared/pi-spawn.ts";
+import { preflightLaunchCwd } from "../shared/launch-cwd.ts";
 import { captureSingleOutputSnapshot, extractChildWrittenOutput, finalizeSingleOutput, formatSavedOutputReference, injectOutputPathSystemPrompt, injectSingleOutputInstruction, resolveSingleOutput, type SingleOutputSnapshot } from "../shared/single-output.ts";
 import {
 	type ActivityState,
@@ -49,6 +50,7 @@ import {
 	type SteeringTargetState,
 	type SteeringTargetStatus,
 	type SubagentChildStatusEvent,
+	type SettlementDiagnostic,
 	DEFAULT_MAX_OUTPUT,
 	type MaxOutputConfig,
 	SUBAGENT_LIFECYCLE_ARTIFACT_VERSION,
@@ -101,7 +103,7 @@ import { createOwnedProcessTreeController, type OwnedProcessTreeController } fro
 import { createSteeringStatus, recordSteeringRequest, steeringStatus, terminalSteeringNoticeState, updateSteeringTarget } from "./steering.ts";
 import { attachPostExitStdioGuard, trySignalChild } from "../../shared/post-exit-stdio-guard.ts";
 import { PROMPT_REDACTED, detectSubagentError, extractTextFromContent, extractToolArgsPreview, getFinalOutput, hasEmptyTerminalAssistantResponse, readStatus } from "../../shared/utils.ts";
-import { evaluateCompletionMutationGuard, validateImplementationToolContract } from "../shared/completion-guard.ts";
+import { evaluateCompletionMutationGuard, expectsImplementationMutation, hasMutationToolCapability, validateImplementationToolContract } from "../shared/completion-guard.ts";
 import {
 	createMutatingFailureState,
 	didMutatingToolFail,
@@ -124,7 +126,7 @@ import {
 	formatWorktreeTaskCwdConflict,
 	type WorktreeSetup,
 } from "../shared/worktree.ts";
-import { resolveEffectiveThinking } from "../../shared/model-info.ts";
+import { findModelInfo, resolveEffectiveThinking } from "../../shared/model-info.ts";
 import { assertThinkingWithinCeiling, decodeThinkingCeiling, SUBAGENT_THINKING_CEILING_ENV } from "../../shared/thinking-ceiling.ts";
 import { launchBindingDigest } from "../../shared/launch-contract.ts";
 import { writeInitialProgressFile } from "../../shared/settings.ts";
@@ -543,6 +545,7 @@ interface RunPiStreamingResult {
 	observedMutationAttempt?: boolean;
 	structuredOutputToolInvoked?: boolean;
 	structuredOutputMessageStartIndex?: number;
+	structuredOutput?: unknown;
 	watchdog?: ChildWatchdogStateSnapshot;
 	runtimeAcknowledgedExtensions?: RuntimeAcknowledgedChildExtensionsV1;
 	processInstanceId: string;
@@ -553,6 +556,7 @@ interface RunPiStreamingResult {
 	currentToolArgs?: string;
 	currentPath?: string;
 	afterCompactionSettlement?: boolean;
+	effects?: import("../../shared/types.ts").EffectsProjection;
 }
 
 const MAX_CHILD_FAILURE_DIAGNOSTIC_CHARS = 8_192;
@@ -561,16 +565,18 @@ function formatChildFailureDiagnostic(input: {
 	error: string | undefined;
 	afterCompactionSettlement?: boolean;
 	missingFileOnlyOutput?: string;
+	missingStructuredOutput?: string;
 }): string | undefined {
-	if (!input.error) return undefined;
 	const notes = [
 		input.afterCompactionSettlement ? "Child failure followed session compaction and agent settlement." : undefined,
 		input.missingFileOnlyOutput ? `Required file-only output was not produced: ${input.missingFileOnlyOutput.slice(0, 2_048)}` : undefined,
+		input.missingStructuredOutput ? `Required structured output was not produced: ${input.missingStructuredOutput.slice(0, 2_048)}` : undefined,
 	].filter((note): note is string => Boolean(note));
 	if (notes.length === 0) return input.error;
 	const context = notes.join("\n");
 	const errorLimit = MAX_CHILD_FAILURE_DIAGNOSTIC_CHARS - (context ? context.length + 1 : 0);
-	return `${input.error.slice(0, Math.max(0, errorLimit))}${context ? `\n${context}` : ""}`;
+	const baseError = input.error || "Subagent failed.";
+	return `${baseError.slice(0, Math.max(0, errorLimit))}${context ? `\n${context}` : ""}`;
 }
 
 function runPiStreaming(
@@ -596,6 +602,7 @@ function runPiStreaming(
 	orcaProgressTab?: OrcaProgressTab,
 	expectedModelForVerification?: string,
 	modelVerificationRegistry?: Array<{ provider: string; id: string; fullId: string }>,
+	mutationTools?: readonly string[],
 ): Promise<RunPiStreamingResult> {
 	return new Promise((resolve) => {
 		const startedAt = Date.now();
@@ -791,7 +798,7 @@ function runPiStreaming(
 					structuredOutputToolInvoked = true;
 					structuredOutputMessageStartIndex = messages.length;
 				}
-				observedMutationAttempt = observedMutationAttempt || isMutatingTool(event.toolName, event.args);
+				observedMutationAttempt = observedMutationAttempt || isMutatingTool(event.toolName, event.args, mutationTools);
 				const toolArgs = extractToolArgsPreview(event.args ?? {});
 				writeOutputLine(toolArgs ? `${event.toolName}: ${toolArgs}` : event.toolName);
 				return;
@@ -1297,7 +1304,7 @@ interface SingleStepContext {
 	nestedRoute?: NestedRouteInfo;
 	capabilityCeiling?: ResolvedSubagentCapabilityCeiling;
 	runFanoutBudget?: RunFanoutBudgetDescriptor;
-	onAttemptStart?: (attempt: { model?: string; thinking?: string }) => void;
+	onAttemptStart?: (attempt: { model?: string; thinking?: string; contextLimit?: number }) => void;
 	onChildEvent?: (event: ChildEvent) => void;
 	onWriterProcess?: (writer: { state: "none" | "spawning" } | { state: "running"; pid: number }) => void;
 	onExternalProcess?: (process: ExternalProcessStatus) => void;
@@ -1399,6 +1406,8 @@ async function runSingleStepInner(
 			model: step.model,
 			modelCandidates: step.modelCandidates,
 			mcpDirectTools: step.mcpDirectTools,
+			mcpConfig: step.mcpConfig,
+			runtimeServerNames: step.runtimeServerNames,
 			cwd: step.cwd ?? ctx.cwd,
 			requireReadTool: Boolean(step.skills?.length),
 			structuredOutput: Boolean(effectiveStructuredOutput),
@@ -1594,6 +1603,10 @@ async function runSingleStepInner(
 		});
 	}
 
+	const effectiveCwd = step.cwd ?? ctx.cwd;
+	const cwdError = preflightLaunchCwd(step.requestedCwd ?? effectiveCwd, effectiveCwd);
+	if (cwdError) return { agent: step.agent, output: cwdError, error: cwdError, exitCode: 1, context: step.context };
+
 	const candidates = step.modelCandidates !== undefined
 		? step.modelCandidates.length > 0 ? step.modelCandidates : [undefined]
 		: step.model
@@ -1636,7 +1649,11 @@ async function runSingleStepInner(
 			const message = error instanceof Error ? error.message : String(error);
 			return omitUndefinedProperties({ agent: step.agent, output: message, error: message, exitCode: 1, context: step.context, thinkingCeiling: step.thinkingCeiling });
 		}
-		ctx.onAttemptStart?.(omitUndefinedProperties({ model: candidate, thinking: resolveEffectiveThinking(candidate, step.thinking) }));
+		ctx.onAttemptStart?.(omitUndefinedProperties({
+			model: candidate,
+			thinking: resolveEffectiveThinking(candidate, step.thinking),
+			contextLimit: findModelInfo(candidate, step.modelVerificationRegistry)?.contextWindow,
+		}));
 		const outputSnapshot = captureSingleOutputSnapshot(step.outputPath);
 		if (effectiveStructuredOutput) {
 			try {
@@ -1665,6 +1682,7 @@ async function runSingleStepInner(
 			sessionFile: step.sessionFile,
 			model: candidate,
 			inheritProjectContext: step.inheritProjectContext,
+			inheritGlobalContext: step.inheritGlobalContext,
 			inheritSkills: step.inheritSkills,
 			requireReadTool: Boolean(step.skills?.length),
 			tools: step.tools,
@@ -1675,6 +1693,8 @@ async function runSingleStepInner(
 			systemPrompt: appendTurnBudgetSystemPrompt(step.systemPrompt ?? "", ctx.turnBudget),
 			systemPromptMode: step.systemPromptMode,
 			mcpDirectTools: step.mcpDirectTools,
+			mcpConfig: step.mcpConfig,
+			runtimeServerNames: step.runtimeServerNames,
 			capabilityCeiling: step.capabilityCeiling ?? ctx.capabilityCeiling,
 			cwd: step.cwd ?? ctx.cwd,
 			promptFileStem: step.agent,
@@ -1718,6 +1738,8 @@ async function runSingleStepInner(
 				model: step.model,
 				modelCandidates: step.modelCandidates,
 				mcpDirectTools: step.mcpDirectTools,
+				mcpConfig: step.mcpConfig,
+				runtimeServerNames: step.runtimeServerNames,
 				cwd: step.cwd ?? ctx.cwd,
 				requireReadTool: Boolean(step.skills?.length),
 				structuredOutput: Boolean(effectiveStructuredOutput),
@@ -1737,6 +1759,7 @@ async function runSingleStepInner(
 				systemPrompt: appendTurnBudgetSystemPrompt(step.systemPrompt ?? "", ctx.turnBudget),
 				systemPromptMode: step.systemPromptMode,
 				inheritProjectContext: step.inheritProjectContext,
+				inheritGlobalContext: step.inheritGlobalContext,
 				inheritSkills: step.inheritSkills,
 				skills: step.skills,
 				tools: toolPlan.effectiveToolAllowlist,
@@ -1773,6 +1796,7 @@ async function runSingleStepInner(
 			ctx.orcaProgressTab,
 			expectedModelForVerification,
 			step.modelVerificationRegistry,
+			step.mutationTools,
 		);
 		if (run.processCloseObservedAt !== undefined) {
 			writerProcesses.push({
@@ -1863,6 +1887,7 @@ async function runSingleStepInner(
 			: undefined;
 		const completionGuardEnabled = isAgentContractV1(step.agentContract) ? step.completionGuard === true : step.completionGuard !== false;
 		const completionToolPlan = resolvedTaskToolPlan;
+		const completionTools = completionToolPlan ? (completionToolPlan.explicitToolAllowlist ? completionToolPlan.effectiveToolAllowlist : undefined) : step.tools;
 		const mutationEvidence = collectTrackedMutationEvidence(mutationSnapshot, step.cwd ?? ctx.cwd);
 		finalMutationEvidence = mutationEvidence;
 		const completionMutationEvidence = ctx.trackedMutationEvidenceForCompletionGuard === false ? undefined : mutationEvidence;
@@ -1871,8 +1896,9 @@ async function runSingleStepInner(
 				agent: step.agent,
 				task: taskForCompletionGuard,
 				messages: run.messages,
-				tools: completionToolPlan ? (completionToolPlan.explicitToolAllowlist ? completionToolPlan.effectiveToolAllowlist : undefined) : step.tools,
+				tools: completionTools,
 				mcpDirectTools: completionToolPlan?.effectiveMcpTools ?? step.mcpDirectTools,
+				mutationTools: step.mutationTools,
 				toolAvailabilityError,
 				mutationEvidence: completionMutationEvidence,
 			}))
@@ -1880,6 +1906,9 @@ async function runSingleStepInner(
 		const mutationAttemptObserved = run.observedMutationAttempt === true || completionMutationEvidence?.attemptedMutation === true;
 		const completionGuardTriggered = completionGuard?.triggered === true && !mutationAttemptObserved;
 		const completionGuardBlocked = completionGuard?.blocked === true;
+		const mutationExpected = completionGuard?.expectedMutation ?? (completionGuardEnabled
+			&& hasMutationToolCapability(completionTools, completionToolPlan?.effectiveMcpTools ?? step.mcpDirectTools)
+			&& expectsImplementationMutation(step.agent, taskForCompletionGuard));
 		const fileMutationEffect = completionGuard
 			? {
 				status: completionGuardBlocked ? "blocked" as const : completionGuard.expectedMutation ? completionGuardTriggered ? "missing" as const : "observed" as const : "not-applicable" as const,
@@ -1940,7 +1969,24 @@ async function runSingleStepInner(
 			toolBudgetBlocked = Boolean(blockedMessage);
 			toolBudget = toolBudgetState(step.toolBudget, toolMessages.length, blockedMessage ? (blockedMessage as { toolName?: string }).toolName : undefined);
 		}
-		finalResult = { ...run, exitCode: effectiveExitCode, model: candidate ?? run.model, error, structuredOutput, runtimeAcknowledgedExtensions, ...(step.agentContract ? { agentContract: step.agentContract } : {}), ...(fileMutationEffect ? { effects: { fileMutation: fileMutationEffect } } : {}) } as RunPiStreamingResult & { structuredOutput?: unknown; agentContract?: import("../../shared/types.ts").AgentContract; effects?: import("../../shared/types.ts").EffectsProjection };
+		const requiredOutput = step.outputMode === "file-only" && step.outputPath
+			? { kind: "file-only" as const, path: step.outputPath, missing: !fs.existsSync(step.outputPath) }
+			: effectiveStructuredOutput
+				? { kind: "structured" as const, path: effectiveStructuredOutput.outputPath, missing: !fs.existsSync(effectiveStructuredOutput.outputPath) }
+			: undefined;
+		const settlementDiagnostic: SettlementDiagnostic | undefined = effectiveExitCode !== 0 || completionGuardTriggered || completionGuardBlocked
+			? {
+				finalTextPresent: Boolean(stripAcceptanceReport(run.finalOutput).trim()),
+				mutation: {
+					expected: mutationExpected,
+					attempted: completionGuard?.attemptedMutation === true || mutationAttemptObserved,
+					observed: mutationEvidence.attemptedMutation,
+				},
+				...(requiredOutput ? { requiredOutput } : {}),
+				afterCompactionSettlement: run.afterCompactionSettlement === true,
+			}
+			: undefined;
+		finalResult = { ...run, exitCode: effectiveExitCode, model: candidate ?? run.model, error, structuredOutput, runtimeAcknowledgedExtensions, ...(step.agentContract ? { agentContract: step.agentContract } : {}), ...(fileMutationEffect || settlementDiagnostic ? { effects: { ...(fileMutationEffect ? { fileMutation: fileMutationEffect } : {}), ...(settlementDiagnostic ? { settlementDiagnostic } : {}) } } : {}) } as RunPiStreamingResult;
 		if (run.turnBudgetExceeded) break modelAttemptsLoop;
 		if (run.stopped || run.timedOut || ctx.timeoutSignal?.aborted || ctx.stopSignal?.aborted || ctx.skipAcceptance?.()) break modelAttemptsLoop;
 		if (attempt.success || completionGuardTriggered) break modelAttemptsLoop;
@@ -2105,13 +2151,12 @@ async function runSingleStepInner(
 					: finalResult?.error ?? (intercomDetachReceipt ? INTERCOM_DETACH_RECEIPT : undefined);
 	const effectiveFinalError = formatChildFailureDiagnostic({
 		error: baseFinalError,
-		afterCompactionSettlement: finalResult?.afterCompactionSettlement,
-		missingFileOnlyOutput: effectiveFinalExitCode !== 0
-			&& finalResult?.afterCompactionSettlement
-			&& step.outputMode === "file-only"
-			&& step.outputPath
-			&& !fs.existsSync(step.outputPath)
-			? step.outputPath
+		afterCompactionSettlement: effectiveFinalExitCode !== 0 ? finalResult?.afterCompactionSettlement : undefined,
+		missingFileOnlyOutput: effectiveFinalExitCode !== 0 && finalResult?.effects?.settlementDiagnostic?.requiredOutput?.kind === "file-only" && finalResult.effects.settlementDiagnostic.requiredOutput.missing
+			? finalResult.effects.settlementDiagnostic.requiredOutput.path
+			: undefined,
+		missingStructuredOutput: effectiveFinalExitCode !== 0 && finalResult?.effects?.settlementDiagnostic?.requiredOutput?.kind === "structured" && finalResult.effects.settlementDiagnostic.requiredOutput.missing
+			? finalResult.effects.settlementDiagnostic.requiredOutput.path
 			: undefined,
 	});
 
@@ -2514,6 +2559,7 @@ async function runSubagent(
 					...(transcriptPath ? { transcriptPath } : {}),
 					skills: task.skills,
 					model: task.model,
+					...(task.contextLimit !== undefined ? { contextLimit: task.contextLimit } : {}),
 					thinking: task.thinking,
 					attemptedModels: task.modelCandidates && task.modelCandidates.length > 0 ? task.modelCandidates : task.model ? [task.model] : undefined,
 					recentTools: [],
@@ -2531,6 +2577,7 @@ async function runSubagent(
 				label: step.label ?? step.parallel.label ?? `Dynamic fanout (${step.collect.as})`,
 				outputName: step.collect.as,
 				structured: Boolean(step.collect.outputSchema),
+				...(step.parallel.contextLimit !== undefined ? { contextLimit: step.parallel.contextLimit } : {}),
 				...(step.agentContract ? { agentContract: step.agentContract } : {}),
 				...(step.capabilityCeiling ? { capabilityCeiling: step.capabilityCeiling } : {}),
 				...(step.thinkingCeiling ? { thinkingCeiling: step.thinkingCeiling } : {}),
@@ -2563,6 +2610,7 @@ async function runSubagent(
 				...(transcriptPath ? { transcriptPath } : {}),
 				skills: step.skills,
 				model: step.model,
+				...(step.contextLimit !== undefined ? { contextLimit: step.contextLimit } : {}),
 				thinking: step.thinking,
 				attemptedModels: step.modelCandidates && step.modelCandidates.length > 0 ? step.modelCandidates : step.model ? [step.model] : undefined,
 				recentTools: [],
@@ -3424,11 +3472,12 @@ async function runSubagent(
 		}
 		pendingStepSteers.push(...remaining);
 	};
-	const updateStepModel = (flatIndex: number, model: string | undefined, thinking: string | undefined, now = Date.now()): void => {
+	const updateStepModel = (flatIndex: number, model: string | undefined, thinking: string | undefined, contextLimit?: number, now = Date.now()): void => {
 		const step = statusPayload.steps[flatIndex];
 		if (!step) return;
 		setOptionalProperty(step, "model", model);
 		setOptionalProperty(step, "thinking", thinking);
+		setOptionalProperty(step, "contextLimit", contextLimit);
 		statusPayload.lastUpdate = now;
 		writeStatusPayload();
 	};
@@ -3514,7 +3563,7 @@ async function runSubagent(
 			return;
 		}
 		if (event.type === "tool_execution_start" && event.toolName) {
-			const mutates = isMutatingTool(event.toolName, event.args);
+			const mutates = isMutatingTool(event.toolName, event.args, flatSteps[flatIndex]?.mutationTools);
 			const currentPath = resolveCurrentPath(event.toolName, event.args);
 			const argsPreview = extractToolArgsPreview(event.args ?? {});
 			const blocksSupervisor = isBlockingSupervisorTool(event.toolName, event.args);
@@ -4059,6 +4108,7 @@ async function runSubagent(
 					...(transcriptPath ? { transcriptPath } : {}),
 					...(task.skills ? { skills: task.skills } : {}),
 					...(task.model ? { model: task.model } : {}),
+					...(task.contextLimit !== undefined ? { contextLimit: task.contextLimit } : {}),
 					...(task.thinking ? { thinking: task.thinking } : {}),
 					...(task.thinkingCeiling ? { thinkingCeiling: task.thinkingCeiling } : {}),
 					...(task.modelCandidates && task.modelCandidates.length > 0 ? { attemptedModels: task.modelCandidates } : task.model ? { attemptedModels: [task.model] } : {}),
@@ -4188,7 +4238,7 @@ async function runSubagent(
 					stopMessage,
 					turnBudget: config.turnBudget,
 					toolTimeoutMs: task.toolTimeoutMs ?? config.toolTimeoutMs,
-					onAttemptStart: (attempt) => updateStepModel(fi, attempt.model, attempt.thinking),
+					onAttemptStart: (attempt) => updateStepModel(fi, attempt.model, attempt.thinking, attempt.contextLimit),
 					onChildEvent: (event) => updateStepFromChildEvent(fi, event),
 					onWriterProcess,
 					onExternalProcess: (process) => updateExternalProcess(fi, process),
@@ -4584,7 +4634,7 @@ async function runSubagent(
 							stopMessage,
 							turnBudget: config.turnBudget,
 							toolTimeoutMs: taskForRun.toolTimeoutMs ?? config.toolTimeoutMs,
-							onAttemptStart: (attempt) => updateStepModel(fi, attempt.model, attempt.thinking),
+							onAttemptStart: (attempt) => updateStepModel(fi, attempt.model, attempt.thinking, attempt.contextLimit),
 							onChildEvent: (event) => updateStepFromChildEvent(fi, event),
 							onWriterProcess,
 							onExternalProcess: (process) => updateExternalProcess(fi, process),
@@ -4939,7 +4989,7 @@ async function runSubagent(
 				stopMessage,
 				turnBudget: config.turnBudget,
 				toolTimeoutMs: seqStep.toolTimeoutMs ?? config.toolTimeoutMs,
-				onAttemptStart: (attempt) => updateStepModel(flatIndex, attempt.model, attempt.thinking),
+				onAttemptStart: (attempt) => updateStepModel(flatIndex, attempt.model, attempt.thinking, attempt.contextLimit),
 				onChildEvent: (event) => updateStepFromChildEvent(flatIndex, event),
 				onWriterProcess,
 				onExternalProcess: (process) => updateExternalProcess(flatIndex, process),

package/src/runs/foreground/execution.ts

@@ -58,6 +58,7 @@ import { effectiveToolTimeoutMs, formatToolTimeoutMessage, resolveToolTimeoutMs,
 import { evaluateCompletionMutationGuard, validateImplementationToolContract } from "../shared/completion-guard.ts";
 import { arbitrateCompletionGuardRescue } from "../shared/llm-intent-arbiter.ts";
 import { getPiSpawnCommand } from "../shared/pi-spawn.ts";
+import { preflightLaunchCwd } from "../shared/launch-cwd.ts";
 import { createJsonlWriter } from "../../shared/jsonl-writer.ts";
 import { createOrcaProgressTab, type OrcaProgressTab } from "../shared/orca-progress-tabs.ts";
 import { attachPostExitStdioGuard, trySignalChild } from "../../shared/post-exit-stdio-guard.ts";
@@ -346,6 +347,7 @@ async function runSingleAttempt(
 		thinking: effectiveThinking,
 		systemPromptMode: agent.systemPromptMode,
 		inheritProjectContext: agent.inheritProjectContext,
+		inheritGlobalContext: agent.inheritGlobalContext,
 		inheritSkills: agent.inheritSkills,
 		requireReadTool: Boolean(shared.resolvedSkillNames?.length),
 		tools: agent.tools,
@@ -381,6 +383,7 @@ async function runSingleAttempt(
 		capabilityCeiling: options.capabilityCeiling,
 		thinkingCeiling: options.thinkingCeiling,
 		extensionBindings: options.extensionBindings,
+		runtimeSnapshotHost: options.runtimeSnapshotHost,
 	});
 	if (!shared.launchWarnings.emitted && warnings.length > 0) {
 		for (const warning of warnings) console.warn(`[pi-subagents] ${warning}`);
@@ -403,6 +406,7 @@ async function runSingleAttempt(
 		inheritedCapabilityCeiling: decodeSubagentCapabilityCeiling(process.env[SUBAGENT_CAPABILITY_CEILING_ENV]),
 		agentName: agent.name,
 		permissionRules,
+		runtimeSnapshotHost: options.runtimeSnapshotHost,
 	});
 	const contractTools = toolPlan.explicitToolAllowlist ? toolPlan.effectiveToolAllowlist : undefined;
 	const contractError = validateImplementationToolContract({
@@ -446,6 +450,7 @@ async function runSingleAttempt(
 		systemPrompt: effectiveSystemPrompt,
 		systemPromptMode: agent.systemPromptMode,
 		inheritProjectContext: agent.inheritProjectContext,
+		inheritGlobalContext: agent.inheritGlobalContext,
 		inheritSkills: agent.inheritSkills,
 		skills: shared.resolvedSkillNames ?? [],
 		tools: toolPlan.effectiveToolAllowlist,
@@ -1064,7 +1069,7 @@ async function runSingleAttempt(
 				if (options.toolBudget) {
 					result.toolBudget = toolBudgetState(options.toolBudget, progress.toolCount);
 				}
-				const mutates = isMutatingTool(evt.toolName, toolArgs);
+				const mutates = isMutatingTool(evt.toolName, toolArgs, agent.mutationTools);
 				observedMutationAttempt = observedMutationAttempt || mutates;
 				pendingToolResult = { tool: evt.toolName ?? "tool", path: activeTool?.path, mutates, startedAt: now };
 				fireUpdate();
@@ -1583,6 +1588,7 @@ async function runSingleAttempt(
 			messages: result.messages ?? [],
 			tools: contractTools,
 			mcpDirectTools: toolPlan.effectiveMcpTools,
+			mutationTools: agent.mutationTools,
 			toolAvailabilityError,
 			mutationEvidence,
 		})
@@ -1688,6 +1694,19 @@ async function runSyncCompletionInner(
 	task: string,
 	options: RunSyncOptions,
 ): Promise<SingleResult> {
+	const effectiveCwd = options.cwd ?? runtimeCwd;
+	const cwdError = preflightLaunchCwd(options.requestedCwd ?? effectiveCwd, effectiveCwd);
+	if (cwdError) {
+		return redactResultPrompt(withRunContext({
+			index: options.index ?? 0,
+			agent: agentName,
+			task,
+			exitCode: 1,
+			messages: [],
+			usage: emptyUsage(),
+			error: cwdError,
+		}, options.context));
+	}
 	options = {
 		...options,
 		capabilityCeiling: intersectSubagentCapabilityCeilings(options.capabilityCeiling ?? resolveCurrentSubagentCapabilityCeiling(options.parentSessionId), decodeSubagentCapabilityCeiling(process.env[SUBAGENT_CAPABILITY_CEILING_ENV])),