pi-subagents 0.52.0 → 0.53.0

package/src/extension/index.ts

@@ -19,7 +19,8 @@ import * as path from "node:path";
 import type { AgentToolResult } from "@earendil-works/pi-agent-core";
 import { keyText, type ExtensionAPI, type ExtensionContext, type ToolDefinition } from "@earendil-works/pi-coding-agent";
 import { Box, Container, Spacer, Text, truncateToWidth, visibleWidth, wrapTextWithAnsi, type Component } from "@earendil-works/pi-tui";
-import { discoverAgents } from "../agents/agents.ts";
+import { discoverAgents, discoverAgentsAll, type AgentConfig, type AgentScope } from "../agents/agents.ts";
+import { clearRuntimeAgentsForPi, listRuntimeAgentConfigs, mergeRuntimeAgents } from "../agents/runtime-agent-registry.ts";
 import { ensureAccessibleDir } from "../shared/accessible-dir.ts";
 import { cleanupAllArtifactDirs, cleanupOldArtifacts, getArtifactsDir } from "../shared/artifacts.ts";
 import { resolveCurrentSessionId } from "../shared/session-identity.ts";
@@ -489,6 +490,18 @@ export default function registerSubagentExtension(pi: ExtensionAPI): void {
 		if (scheduledRunManager.observedCompletionRunIds().size > 0) return true;
 		return missionObserverResultCandidateFiles(DIRS.results).length > 0;
 	};
+	const discoverAgentsForRuntime = (cwd: string, scope: AgentScope) => {
+		const discovered = discoverAgents(cwd, scope);
+		if (listRuntimeAgentConfigs(pi).length === 0) return discovered;
+		const all = discoverAgentsAll(cwd);
+		const configuredAgents: AgentConfig[] = [
+			...all.builtin,
+			...all.package,
+			...all.user,
+			...all.project,
+		];
+		return mergeRuntimeAgents(pi, discovered, configuredAgents);
+	};
 	const { ensurePoller, refreshWidget, handleStarted, handleComplete, resetJobs, restoreActiveJobs, dispose: disposeAsyncJobTracker } = createAsyncJobTracker(pi, state, DIRS.async, {
 		widgetEnabled: asyncWidgetEnabled,
 		onJobTerminal: () => refreshResultDelivery(),
@@ -504,6 +517,7 @@ export default function registerSubagentExtension(pi: ExtensionAPI): void {
 			observedCompletionRunIds: () => scheduledRunManager.observedCompletionRunIds(),
 			hasDeliveryDemand: hasResultDeliveryDemand,
 			deliverIntercomResults: config.intercomBridge?.resultDelivery === true,
+			resultScanLogging: config.resultScanLogging ?? "all",
 		},
 	);
 	const { startResultWatcher, primeExistingResults, stopResultWatcher } = resultWatcher;
@@ -538,7 +552,7 @@ export default function registerSubagentExtension(pi: ExtensionAPI): void {
 		tempArtifactsDir,
 		getSubagentSessionRoot,
 		expandTilde,
-		discoverAgents,
+		discoverAgents: discoverAgentsForRuntime,
 		activateSupervisorTransport: () => supervisorChannel.activateTransport(),
 		refreshResultDelivery: () => refreshResultDelivery(),
 	});
@@ -881,6 +895,7 @@ export default function registerSubagentExtension(pi: ExtensionAPI): void {
 			if (runtimeCleaned) return;
 			runtimeCleaned = true;
 			const shuttingDownParentSession = parentSessionEnvValue;
+			clearRuntimeAgentsForPi(pi);
 			clearTimeout(resultIndexCleanupTimer);
 			clearTimeout(asyncRetentionTimer);
 			asyncRetentionAbort.abort();

package/src/extension/rpc.ts

@@ -27,7 +27,7 @@ export const SUBAGENT_RPC_REQUEST_EVENT = "subagents:rpc:v1:request";
 export const SUBAGENT_RPC_READY_EVENT = "subagents:rpc:v1:ready";
 export const SUBAGENT_RPC_REPLY_EVENT_PREFIX = "subagents:rpc:v1:reply:";
 
-export const SUBAGENT_RPC_METHODS = ["ping", "status", "spawn", "steer", "interrupt", "stop", "resume"] as const;
+export const SUBAGENT_RPC_METHODS = ["ping", "status", "manage", "spawn", "steer", "interrupt", "stop", "resume"] as const;
 export type SubagentRpcMethod = typeof SUBAGENT_RPC_METHODS[number];
 
 export interface SubagentRpcRequestEnvelope {
@@ -58,6 +58,18 @@ export type SubagentRpcReplyEnvelope<T = unknown> = {
 	};
 };
 
+export const SUBAGENT_RPC_MANAGEMENT_ACTIONS = [
+	"schedule.list",
+	"schedule.show",
+	"schedule.history",
+	"schedule.pause",
+	"schedule.resume",
+	"schedule.run",
+	"schedule.delete",
+] as const;
+
+type SubagentRpcManagementAction = typeof SUBAGENT_RPC_MANAGEMENT_ACTIONS[number];
+
 type SubagentRpcErrorCode =
 	| "invalid_request"
 	| "invalid_params"
@@ -375,6 +387,7 @@ function pingData(ctx: ExtensionContext | null) {
 		methods: [...SUBAGENT_RPC_METHODS],
 		capabilities: {
 			status: true,
+			managementActions: [...SUBAGENT_RPC_MANAGEMENT_ACTIONS],
 			fleetStatus: { version: 1 },
 			asyncStatusSnapshot: { kind: ASYNC_STATUS_SNAPSHOT_KIND, version: ASYNC_STATUS_SNAPSHOT_VERSION },
 			asyncSpawn: true,
@@ -412,6 +425,30 @@ async function executeChecked(
 	return dataFromToolResult(result);
 }
 
+function manageParams(params: unknown): SubagentParamsLike {
+	const input = assertRecordParams(params, "manage");
+	if (typeof input.action !== "string" || !(SUBAGENT_RPC_MANAGEMENT_ACTIONS as readonly string[]).includes(input.action)) {
+		throw new SubagentRpcError(
+			"invalid_params",
+			`RPC manage action must be one of: ${SUBAGENT_RPC_MANAGEMENT_ACTIONS.join(", ")}.`,
+		);
+	}
+	if (input.id !== undefined && (typeof input.id !== "string" || !input.id.trim())) {
+		throw new SubagentRpcError("invalid_params", "RPC manage id must be a non-empty string.");
+	}
+	const action = input.action as SubagentRpcManagementAction;
+	const requiresId = action !== "schedule.list";
+	if (requiresId && typeof input.id !== "string") {
+		throw new SubagentRpcError("invalid_params", `RPC manage ${action} requires id.`);
+	}
+	const output: SubagentParamsLike = {
+		action,
+		...(typeof input.id === "string" ? { id: input.id.trim() } : {}),
+	};
+	assertSubagentParams(output, "RPC manage params");
+	return output;
+}
+
 function spawnParams(params: unknown): SubagentParamsLike {
 	const input = assertRecordParams(params, "spawn");
 	const normalized = normalizePublicSubagentExecution(input);
@@ -535,6 +572,9 @@ async function handleRequest(
 	if (request.method === "ping") return pingData(ctx);
 	if (!ctx) throw new SubagentRpcError("no_active_session", "No active extension context for subagent RPC.");
 
+	if (request.method === "manage") {
+		return executeChecked(options, ctx, request.requestId, request.method, manageParams(request.params));
+	}
 	if (request.method === "spawn") {
 		return executeChecked(options, ctx, request.requestId, request.method, spawnParams(request.params));
 	}

package/src/extension/schemas.ts

@@ -307,13 +307,13 @@ const SubagentParamProperties = {
 		],
 		description: "Agent config for create/update. Object or JSON string."
 	})),
-	workflowScript: Type.Optional(Type.String({ minLength: 1, description: "Trusted inline JavaScript statement body. Normally async unless asyncByDefault:false; set async:true when async matters. Use async:false only when the parent must block until completion, never for reviews or gates. Use explicit return for output. Use top-level await, plain helper functions, or explicit Promise chains; nested async function, arrow, and method helpers are rejected. Use await runs.run(key, {agent, task, worktree?, gate?}) or runs.run(key, {resume, task}), runs.all([...]), await runs.steer(key, message, {mode?, index?, ackTimeoutMs?}), runs.status(id), runs.ref(s), emit(value), console, and return. For ordinary parallel fanout, use await runs.all([{key, agent, task}, ...]); do not read .output from unawaited runs.run launches. Stored runs.run promises are only for advanced rolling fanout, and each must later be observed with direct await, Promise.race, or Promise.all. runs.steer targets a prior stable child key, never a raw run id, and must be awaited or returned. Mission workflows also have async state.get(key) and state.set(key, JSONValue). Compose sequential and parallel phases dynamically. Set worktree:true at workflow or child level for a separate managed worktree; child fields override workflow defaults. gate is one host-run command and cannot be combined with acceptance. runs.run accepts one child only. No filesystem, shell, Pi tools, or host globals." })),
+	workflowScript: Type.Optional(Type.String({ minLength: 1, description: "Trusted inline JavaScript statement body. Normally async unless asyncByDefault:false; set async:true when async matters. Use async:false only when the parent must block until completion, never for reviews or gates. Use explicit return for output. Use top-level await, plain helper functions, or explicit Promise chains; nested async function, arrow, and method helpers are rejected. Use await runs.run(key, {agent, task, worktree?, gate?}) or runs.run(key, {resume, task}), where resume is a retained run id or {workflowRunId,key,latest:true} from a durable async workflow receipt. Use runs.all([...]), await runs.steer(key, message, {mode?, index?, ackTimeoutMs?}), runs.status(id), runs.ref(s), emit(value), console, and return. For ordinary parallel fanout, use await runs.all([{key, agent, task}, ...]); do not read .output from unawaited runs.run launches. Stored runs.run promises are only for advanced rolling fanout, and each must later be observed with direct await, Promise.race, or Promise.all. runs.steer targets a prior stable child key, never a raw run id, and must be awaited or returned. Mission workflows also have async state.get(key) and state.set(key, JSONValue). Compose sequential and parallel phases dynamically. Set worktree:true at workflow or child level for a separate managed worktree; child fields override workflow defaults. gate is one host-run command and cannot be combined with acceptance. runs.run accepts one child only. No filesystem, shell, Pi tools, or host globals." })),
 	chatProgress: Type.Optional(Type.String({ enum: ["auto", "off", "live-card"], description: "WorkflowScript chat progress projection. auto shows a live in-chat card only for watched foreground workflows in the same Git repository; it is off otherwise. Explicit live-card requires same-repository async:false; async workflows should omit chatProgress or use auto/off." })),
 	isolation: Type.Optional(Type.String({ enum: ["none", "worktree"], description: "Workflow child isolation. none runs in the shared cwd; worktree requires managed git worktree isolation." })),
 	worktree: Type.Optional(Type.Boolean({ description: "Managed child isolation. true gives each workflow child a separate git worktree; an individual runs.run/runs.all item can override a workflow default with worktree:false." })),
 	context: Type.Optional(Type.String({
-		enum: ["fresh", "fork"],
-		description: "'fresh' or 'fork' to branch from parent session. Explicit context overrides every child. If omitted, config defaultSubagentContext wins over each agent defaultContext; implicit fork needs a persisted parent session and leaf, else fresh.",
+		enum: ["fresh", "fork", "profile"],
+		description: "'fresh' or 'fork' to branch from parent session, or 'profile' to require the selected agent's declared defaultContext. Explicit fresh/fork overrides every child; profile ignores config defaultSubagentContext and fails when an agent has no defaultContext. If omitted, config defaultSubagentContext wins over each agent defaultContext; implicit fork needs a persisted parent session and leaf, else fresh.",
 	})),
 	async: Type.Optional(Type.Boolean({ description: "Run in background unless asyncByDefault:false. Set false only when the parent must block until completion." })),
 	timeoutMs: Type.Optional(Type.Integer({ minimum: 1, description: "Timeout. Foreground and single async runs use config timeoutMs, else 30m; async composites have no default parent deadline. Alias maxRuntimeMs." })),
@@ -341,7 +341,7 @@ const SubagentParamProperties = {
 	})),
 	outputMode: Type.Optional(OutputModeOverride),
 	skill: Type.Optional(SkillOverride),
-	model: Type.Optional(Type.String({ description: "Default child model override (e.g. 'anthropic/claude-sonnet-4')" })),
+	model: Type.Optional(Type.String({ description: "Default child model override. Full provider/id values are accepted; bare ids resolve from the active registry." })),
 	outputSchema: Type.Optional(JsonSchemaObject),
 	agentContract: Type.Optional(AgentContractOverride),
 	acceptance: Type.Optional(AcceptanceOverride),
@@ -370,6 +370,9 @@ const SubagentWaitParamsSchema = Type.Object({
 		minimum: 1,
 		description: "Give up waiting after this many milliseconds (the runs keep going regardless). Defaults to 1800000 (30 minutes).",
 	})),
+	stopOnAttention: Type.Optional(Type.Boolean({
+		description: "Blocking waits stop when a run needs attention by default. Set false to keep waiting through idle or long-thinking attention; supervisor/contact requests still stop the wait.",
+	})),
 });
 
 export const SubagentWaitParams = keepTopLevelParameterDescriptions(SubagentWaitParamsSchema);

package/src/extension/tool-description.ts

@@ -36,7 +36,7 @@ EXECUTION:
 • WORKFLOW SCRIPT: { workflowScript: "return runs.run('main', {agent:'worker', task:'...'})" }. Use stable-key runs.run for one child and await runs.all([{key,agent,task}, ...]) for ordinary parallel children; do not read .output from unawaited runs.run launches. Stored runs.run promises are only for advanced rolling fanout and each must later be observed with direct await, Promise.race, or Promise.all. Ordinary JavaScript provides sequence, branching, filtering, retries, and aggregation. workflowScript is an ordinary JavaScript statement body, so use an explicit return for a useful result. Use top-level await, plain helper functions, or explicit Promise chains; nested async function, arrow, and method helpers are rejected. For task text with Markdown fences or shell blocks, build quoted lines instead of nesting raw template literals: \`const task=["Run:","\`\`\`bash","npm test","\`\`\`"].join("\\n")\`. Scripts normally start async unless config sets asyncByDefault:false; set async:true explicitly when async behavior matters. Pass async:false only when the parent must block until completion, never for final reviews or gates. Same-repo blocking workflows default to a live in-chat card; explicit live-card requires same-repository async:false, so async workflows should omit chatProgress or use auto/off. Workflow-level child controls default onto each runs.run launch, and explicit child fields override them. Use {action:"children.list"} to list recent retained workflow children with resumable/not-resumable reasons. Resume only rows reported resumable. For a simple follow-up or implementation challenge, use {action:"resume", id:"run-id", message:"..."}. Resume keeps the stored agent/model/tool contract. If no resumable child is listed, launch a same-role fallback challenge and label it as fallback. Inside workflowScript, continue one with runs.run(key, {resume:"run-id", task:"follow-up"}); workflow resumes wait for completed output, and loops must continue from each latest returned runId. Await runs.steer(key, message, {mode?, index?, ackTimeoutMs?}) to guide a prior keyed child without exposing its run id; receipts are queued, delivered, missed, or failed. Always await or return runs.steer. For repository mutation lanes, set worktree:true on the workflow or individual runs.run/runs.all item for managed isolation; each parallel child gets a separate worktree and handoff artifact. A workflow usageBudget is enforced once across the workflow. Available globals are runs.run, runs.all, runs.steer, runs.status, runs.ref/refs, emit, console, and standard JavaScript only. Workflows get async state.get(key) and state.set(key, JSONValue) through their automatic or explicit mission; mission:false workflows do not have a state global. Scripts cannot access filesystem, shell, arbitrary Pi tools, or host globals.
 • Sequential example: { workflowScript: "const a = await runs.run('analyze', {agent:'agent-a', task:'Analyze the request'}); return (await runs.run('plan', {agent:'agent-b', task:'Plan from: '+a.output})).output" }
 • Parallel example: { workflowScript: "const [a,b] = await runs.all([{key:'correctness',agent:'agent-a',task:'Review correctness'},{key:'tests',agent:'agent-b',task:'Review tests'}]); return {correctness:a.output,tests:b.output}" }
-• Optional context is "fresh" or "fork". Explicit context wins. When omitted, config defaultSubagentContext wins over agent defaultContext. timeoutMs/maxRuntimeMs apply to foreground and async workflows; foreground workflows default to 30 minutes and async workflows have no default timeout. Omit acceptance for reviewer/read-only calls; evidence levels end at verified, and acceptance.review.required requests independent writer review.
+• Optional context is "fresh", "fork", or "profile". profile requires the selected agent's declared defaultContext and ignores config defaultSubagentContext. Explicit fresh/fork wins. When omitted, config defaultSubagentContext wins over agent defaultContext. timeoutMs/maxRuntimeMs apply to foreground and async workflows; foreground workflows default to 30 minutes and async workflows have no default timeout. Omit acceptance for reviewer/read-only calls; evidence levels end at verified, and acceptance.review.required requests independent writer review.
 • Durable mission attachment is automatic by default. Use missionId to attach an existing mission, mission:{...} to override auto-create, or mission:false for ephemeral work. A mission object needs exactly one non-empty title or summary; objective and labels are optional. goal may only be true and requires budget:{tokens}.
 
 MANAGEMENT / CONTROL (use action; omit execution fields):
@@ -53,7 +53,7 @@ EXECUTE:
 • SINGLE {agent:"worker",task:"..."} starts exactly one child through the workflow runtime. Workflow-level fields remain child defaults. Do not combine agent/task with action or workflowScript.
 • SCRIPT {workflowScript:"return runs.run('main', {agent:'worker', task:'...'})"}. Use stable-key runs.run for one child and await runs.all([{key,agent,task}, ...]) for ordinary parallel work; do not read .output from unawaited runs.run launches. Stored runs.run promises are only for advanced rolling fanout and each must later be observed with direct await, Promise.race, or Promise.all. Await runs.steer(key,message,options?) to guide a prior keyed child; it returns queued, delivered, missed, or failed and never accepts a raw run id. Always await or return steering calls. Use {action:"children.list"} for recent retained workflow children and resume only rows reported resumable. Use {action:"resume",id:"run-id",message:"..."} for a simple follow-up or challenge; resume keeps the stored agent/model/tool contract. If none is resumable, launch a same-role fallback challenge and label it as fallback. Inside workflowScript use runs.run(key,{resume:"run-id",task:"follow-up"}) when the script must wait for completion and continue from the latest returned runId. Workflows get async state.get/state.set through their automatic or explicit mission; mission:false does not. Scripts are ordinary JavaScript statement bodies; use explicit return for a useful result. Use top-level await, plain helper functions, or explicit Promise chains; nested async function, arrow, and method helpers are rejected. For task text with Markdown fences or shell blocks, build quoted lines instead of nesting raw template literals: \`const task=["Run:","\`\`\`bash","npm test","\`\`\`"].join("\\n")\`. Use JavaScript for sequence, branching, retries, and aggregation. For repository mutation lanes, use worktree:true on the workflow or runs.run/runs.all item for managed isolation. Scripts normally start async unless config sets asyncByDefault:false; set async:true explicitly when async behavior matters. async:false blocks the parent until completion and auto-enables a same-repo live chat card unless chatProgress is off; explicit live-card requires same-repository async:false, so async workflows should omit chatProgress or use auto/off.
 • Example: {workflowScript:"const [a,b]=await runs.all([{key:'a',agent:'agent-a',task:'Implement A',worktree:true},{key:'b',agent:'agent-b',task:'Implement B',worktree:true}]); return [a.output,b.output]"}
-• context can be fresh or fork. Explicit context wins; omitted context follows defaultSubagentContext before agent defaultContext. timeoutMs/maxRuntimeMs apply to foreground and async workflows; foreground workflows default to 30 minutes and async workflows have no default timeout. Omit acceptance for reviewer/read-only calls.
+• context can be fresh, fork, or profile. profile requires the selected agent's declared defaultContext and ignores defaultSubagentContext. Explicit fresh/fork wins; omitted context follows defaultSubagentContext before agent defaultContext. timeoutMs/maxRuntimeMs apply to foreground and async workflows; foreground workflows default to 30 minutes and async workflows have no default timeout. Omit acceptance for reviewer/read-only calls.
 
 MANAGE / CONTROL:
 • Use action without execution fields for list/get/models/guide/authoring, refine/refine.show/refine.rollback, mission, watchdog, status, interrupt, stop, resume, steer, script-only scheduling, diagnostics, and other management actions. guide reads shipped current-version docs by topic.

package/src/runs/background/async-execution.ts

@@ -1406,7 +1406,7 @@ export function executeAsyncSingle(
 	const effectiveOutput = normalizeSingleOutputOverride(params.output, agentConfig.output);
 	const outputPath = resolveSingleOutputPath(effectiveOutput, ctx.cwd, instructionCwd, params.outputBaseDir ?? (artifactsDir ? path.join(artifactsDir, "outputs", id) : undefined));
 	systemPrompt = injectOutputPathSystemPrompt(systemPrompt, outputPath, agentConfig);
-	const outputMode = params.outputMode ?? "inline";
+	const outputMode = params.outputMode ?? agentConfig.outputMode ?? "inline";
 	const validationError = validateFileOnlyOutputMode(outputMode, outputPath, `Async single run (${agent})`);
 	if (validationError) return formatAsyncStartError("single", validationError);
 	const taskWithOutputInstruction = injectSingleOutputInstruction(task, outputPath, agentConfig);
@@ -1534,6 +1534,7 @@ export function executeAsyncSingle(
 		...(params.structuredOutputSchema ? { structuredOutputSchema: params.structuredOutputSchema } : {}),
 		...(params.acceptance !== undefined ? { acceptance: params.acceptance } : {}),
 		...(controlConfig ? { controlConfig } : {}),
+		...(params.context ? { context: params.context } : {}),
 		...(params.intercomBridge !== undefined ? { intercomBridge: params.intercomBridge } : {}),
 		...(deadlineAt !== undefined ? { absoluteDeadlineAt: deadlineAt } : {}),
 		...(initialTurnBudget ? { initialTurnBudget: { maxTurns: initialTurnBudget.maxTurns, graceTurns: initialTurnBudget.graceTurns } } : {}),

package/src/runs/background/async-job-tracker.ts

@@ -324,7 +324,8 @@ export function createAsyncJobTracker(pi: Pick<ExtensionAPI, "events">, state: S
 	};
 
 	const refreshJob = (job: AsyncJobState): boolean => {
-		const widgetStateBefore = widgetRenderKey(job);
+		const widgetExpanded = state.lastUiContext?.hasUI ? state.lastUiContext.ui.getToolsExpanded?.() ?? false : false;
+		const widgetStateBefore = widgetRenderKey(job, widgetExpanded);
 		let nestedRefreshFailed = false;
 		const refreshNestedProjection = () => {
 			try {
@@ -422,7 +423,7 @@ export function createAsyncJobTracker(pi: Pick<ExtensionAPI, "events">, state: S
 						scheduleCleanup(job.asyncId);
 					}
 				}
-				return widgetRenderKey(job) !== widgetStateBefore;
+				return widgetRenderKey(job, widgetExpanded) !== widgetStateBefore;
 			}
 			if (job.status === "queued") {
 				job.status = "running";
@@ -439,7 +440,7 @@ export function createAsyncJobTracker(pi: Pick<ExtensionAPI, "events">, state: S
 			rememberFleetJob(state, job);
 			if (!hasLiveNestedDescendants(job.nestedChildren) && !state.cleanupTimers.has(job.asyncId)) scheduleCleanup(job.asyncId);
 		}
-		return widgetRenderKey(job) !== widgetStateBefore;
+		return widgetRenderKey(job, widgetExpanded) !== widgetStateBefore;
 	};
 
 	const scheduleJobRefresh = (asyncId: string, delayMs = EVENT_REFRESH_DEBOUNCE_MS) => {

package/src/runs/background/async-resume.ts

@@ -321,7 +321,7 @@ export function readAsyncRecoveryDescriptor(asyncDir: string | undefined): Steer
 		"version", "launchContractDigest", "sourceRunId", "agentContract", "agent", "sessionFile", "cwd", "model", "modelOverrideFromParent", "fallbackModels", "thinking", "tools", "extensions",
 		"subagentOnlyExtensions", "mcpDirectTools", "systemPrompt", "systemPromptMode", "inheritProjectContext", "inheritSkills", "skills",
 		"skillPath", "agentFilePath", "completionGuard", "memory", "outputPath", "outputMode", "structuredOutputSchema", "acceptance", "sessionDir", "artifactConfig",
-		"artifactsDir", "maxOutput", "controlConfig", "intercomBridge", "absoluteDeadlineAt", "initialTurnBudget", "initialToolBudget", "maxSubagentDepth", "share", "capabilityCeiling",
+		"artifactsDir", "maxOutput", "controlConfig", "context", "intercomBridge", "absoluteDeadlineAt", "initialTurnBudget", "initialToolBudget", "maxSubagentDepth", "share", "capabilityCeiling",
 		"launchResolvedExtensions", "runFanoutBudget",
 	]);
 	for (const field of Object.keys(parsed)) {
@@ -345,6 +345,7 @@ export function readAsyncRecoveryDescriptor(asyncDir: string | undefined): Steer
 	}
 	if (parsed.systemPromptMode !== "append" && parsed.systemPromptMode !== "replace") throw new Error(`Invalid async recovery descriptor '${descriptorPath}': systemPromptMode is invalid.`);
 	if (parsed.outputMode !== "inline" && parsed.outputMode !== "file-only") throw new Error(`Invalid async recovery descriptor '${descriptorPath}': outputMode is invalid.`);
+	if (parsed.context !== undefined && parsed.context !== "fresh" && parsed.context !== "fork") throw new Error(`Invalid async recovery descriptor '${descriptorPath}': context is invalid.`);
 	if (parsed.modelOverrideFromParent !== undefined && typeof parsed.modelOverrideFromParent !== "boolean") throw new Error(`Invalid async recovery descriptor '${descriptorPath}': modelOverrideFromParent must be a boolean.`);
 	for (const field of ["inheritProjectContext", "inheritSkills", "share"] as const) {
 		if (typeof parsed[field] !== "boolean") throw new Error(`Invalid async recovery descriptor '${descriptorPath}': ${field} must be a boolean.`);

package/src/runs/background/async-status-snapshot.ts

@@ -227,12 +227,21 @@ function snapshotBytes(snapshot: AsyncStatusSnapshotV1): number {
 }
 
 function enforceByteLimit(snapshot: AsyncStatusSnapshotV1): void {
-	while (snapshot.runs.length > 0 && snapshotBytes(snapshot) > snapshot.caps.maxSerializedBytes) {
-		snapshot.runs.pop();
-		snapshot.omitted.runs += 1;
-		snapshot.omitted.byteLimitExceeded = true;
+	if (snapshotBytes(snapshot) <= snapshot.caps.maxSerializedBytes) return;
+	snapshot.omitted.byteLimitExceeded = true;
+	const runs = snapshot.runs;
+	const initialOmittedRuns = snapshot.omitted.runs;
+	let lower = 0;
+	let upper = Math.max(0, runs.length - 1);
+	while (lower < upper) {
+		const retained = Math.ceil((lower + upper) / 2);
+		snapshot.runs = runs.slice(0, retained);
+		snapshot.omitted.runs = initialOmittedRuns + runs.length - retained;
+		if (snapshotBytes(snapshot) <= snapshot.caps.maxSerializedBytes) lower = retained;
+		else upper = retained - 1;
 	}
-	if (snapshotBytes(snapshot) > snapshot.caps.maxSerializedBytes) snapshot.omitted.byteLimitExceeded = true;
+	snapshot.runs = runs.slice(0, lower);
+	snapshot.omitted.runs = initialOmittedRuns + runs.length - lower;
 }
 
 export function buildAsyncStatusSnapshot(jobs: Iterable<AsyncJobState>, options: AsyncStatusSnapshotOptions = {}): AsyncStatusSnapshotV1 {

package/src/runs/background/auto-drain.ts

@@ -58,6 +58,7 @@ export async function drainOutstandingWork(deps: AutoDrainDeps): Promise<void> {
 				now,
 				stopOnAttention: false,
 				failOnFailedRuns: true,
+				failOnAttention: true,
 			},
 		);
 		if (waitResult.isError) {

package/src/runs/background/result-watcher.ts

@@ -57,6 +57,8 @@ type ResultWatcherDeps = {
 	coalesceDelayMs?: number;
 	/** Returns true while a durable completion source needs periodic delivery checks. */
 	hasDeliveryDemand?: () => boolean;
+	/** Control how slow result-index scans are logged. Defaults to \"all\". */
+	resultScanLogging?: "all" | "activity" | "off";
 	platform?: NodeJS.Platform;
 };
 
@@ -604,6 +606,12 @@ export function createResultWatcher(
 	const logScanStats = (stats: ResultScanStats) => {
 		const elapsed = Date.now() - stats.startedAt;
 		if (elapsed < SLOW_RESULT_SCAN_MS) return;
+		if (deps.resultScanLogging === "off") return;
+		// A scan that inspected and scheduled nothing is a quiet no-op (e.g. the
+		// healthy periodic rescan while no async runs are pending). Under
+		// "activity", skip it so empty scans do not burn context tokens in the
+		// session transcript.
+		if (deps.resultScanLogging === "activity" && stats.files === 0 && stats.scheduled === 0) return;
 		console.error(`Subagent result scan inspected ${stats.files} indexed result file(s), scheduled ${stats.scheduled} in ${elapsed}ms (${resultsDir}).`);
 	};
 	const indexedResultCandidates = (observed: ReadonlySet<string>): string[] => {

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

@@ -85,7 +85,7 @@ import { readChildToolDiagnosticError } from "../shared/tool-availability.ts";
 import { collectDynamicResults, DynamicFanoutError, materializeDynamicParallelStep, validateDynamicCollection } from "../shared/dynamic-fanout.ts";
 import { claimRunFanoutBatch, getRunFanoutBudgetSnapshot } from "../shared/run-fanout-budget.ts";
 import { nestedSummaryFromAsyncStatus, projectNestedEvents, resolveNestedAsyncDir, writeNestedEvent } from "../shared/nested-events.ts";
-import { formatModelAttemptNote, isRetryableModelFailure } from "../shared/model-fallback.ts";
+import { formatModelAttemptNote, isRetryableModelFailure, recordRetryableModelFailure } from "../shared/model-fallback.ts";
 import {
 	SUBAGENT_STARTUP_RETRY_DELAYS_MS,
 	formatSubagentExtensionConflictError,
@@ -1432,8 +1432,8 @@ async function runSingleStepInner(
 		});
 	}
 
-	const candidates = step.modelCandidates && step.modelCandidates.length > 0
-		? step.modelCandidates
+	const candidates = step.modelCandidates !== undefined
+		? step.modelCandidates.length > 0 ? step.modelCandidates : [undefined]
 		: step.model
 			? [step.model]
 			: [undefined];
@@ -1769,7 +1769,9 @@ async function runSingleStepInner(
 			finalResult.finalOutput = startupError;
 			break modelAttemptsLoop;
 		}
-		if (!isRetryableModelFailure(error) || modelIndex === candidates.length - 1) break modelAttemptsLoop;
+		const retryableModelFailure = isRetryableModelFailure(error);
+		if (retryableModelFailure) recordRetryableModelFailure(candidate ?? run.model ?? step.model, error);
+		if (!retryableModelFailure || modelIndex === candidates.length - 1) break modelAttemptsLoop;
 		attemptNotes.push(formatModelAttemptNote(attempt, candidates[modelIndex + 1]));
 		modelIndex += 1;
 		startupAttemptIndex = 0;
@@ -3330,6 +3332,7 @@ async function runSubagent(
 				startedAt: step.startedAt ?? overallStartTime,
 				lastActivityAt,
 				currentTool: step.currentTool,
+				thinking: step.thinking,
 				now,
 			}));
 			if (idleState === "needs_attention") {

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

@@ -86,6 +86,8 @@ export interface SubagentWaitParams {
 	all?: boolean;
 	/** Give up after this many milliseconds. Defaults to 30 minutes. */
 	timeoutMs?: number;
+	/** False keeps a blocking wait open through idle attention; supervisor/contact requests still stop the wait. */
+	stopOnAttention?: boolean;
 }
 
 /** Minimal event-bus surface wait subscribes to (matches pi.events). */
@@ -110,6 +112,8 @@ export interface SubagentWaitDeps {
 	stopOnAttention?: boolean;
 	/** Internal auto-drain mode surfaces failed terminal subagent runs as errors. */
 	failOnFailedRuns?: boolean;
+	/** Internal auto-drain mode surfaces actionable attention as an error. */
+	failOnAttention?: boolean;
 	/** Arm a durable exact-target wait subscription in a long-lived interactive runtime. */
 	subscribe?: (input: { targetKind: "async" | "foreground"; runId: string; requestedId: string; timeoutMs: number }) => { token: string; expiresAt: number };
 	/** Injectable provider protocol surfaces for deterministic tests. */
@@ -453,7 +457,7 @@ async function waitForDetachedForegroundRun(
 			return result(`Remembered foreground run "${run.runId}" disappeared before a terminal child result was recorded. Completion cannot be confirmed; do not launch a replacement without checking the originating child session.`, true);
 		}
 		const pending = current.children.filter((child) => initialDetachedIndices.has(child.index) && child.status === "detached");
-		const attention = deps.stopOnAttention === false ? [] : foregroundChildrenNeedingAttention(current, initialDetachedIndices);
+		const attention = foregroundChildrenNeedingAttention(current, initialDetachedIndices);
 		if (attention.length > 0) return formatForegroundAttention(current, attention, now() - startedAt);
 		if (pending.length === 0) {
 			const outcome = summarizeForegroundChildren(current, initialDetachedIndices);
@@ -555,11 +559,11 @@ export async function waitForSubagents(
 	const initialProviderIds = new Set(providerActive.map(backgroundWorkIdentity));
 	const initialProviderNames = new Set(providerActive.map((item) => item.provider));
 	const initialCount = initialAsyncIds.size + initialProviderIds.size;
-	const stopOnAttention = deps.stopOnAttention !== false;
+	const stopOnAttention = params.stopOnAttention ?? deps.stopOnAttention !== false;
 	let attention = active.filter((run) => needsAttention(run));
 
 	const isDone = (): boolean => {
-		if (stopOnAttention && attention.some((run) => initialAsyncIds.has(run.id))) return true;
+		if (attention.some((run) => initialAsyncIds.has(run.id) && (stopOnAttention || hasSupervisorTool(run)))) return true;
 		const activeAsyncIds = new Set(active.map((run) => run.id));
 		const activeProviderIds = new Set(providerActive.map(backgroundWorkIdentity));
 		if (waitForAll) {
@@ -643,7 +647,7 @@ export async function waitForSubagents(
 		const status = relevantAttention.length > 0 ? "attention required" : "done";
 		return result(
 			`Waited ${elapsed} for ${scope}; ${status}.${outcome}${resumeGuidance}${attentionNote} Completion/control events have been observed; inspect status if a notification is not visible yet.`,
-			deps.failOnFailedRuns === true && failedAsyncCount > 0,
+			(deps.failOnFailedRuns === true && failedAsyncCount > 0) || (deps.failOnAttention === true && relevantAttention.length > 0),
 			completions,
 		);
 	}
@@ -660,7 +664,7 @@ export async function waitForSubagents(
 		: `${finishedCount} of ${initialCount} ${subject} finished`;
 	return result(
 		`Waited ${elapsed}; ${progress}.${outcome}${resumeGuidance}${attentionNote}${remainder} Relevant completion/control events have been observed; inspect status if a notification is not visible yet.`,
-		deps.failOnFailedRuns === true && failedAsyncCount > 0,
+		(deps.failOnFailedRuns === true && failedAsyncCount > 0) || (deps.failOnAttention === true && relevantAttention.length > 0),
 		completions,
 	);
 }

package/src/runs/background/terminal-run-index.ts

@@ -68,11 +68,16 @@ function removeInvalidMarker(marker: string): void {
 	}
 }
 
-function markerFiles(dir: string): string[] {
+interface MarkerFile {
+	dir: string;
+	name: string;
+}
+
+function markerFiles(dir: string): MarkerFile[] {
 	try {
 		return fs.readdirSync(dir, { withFileTypes: true })
 			.filter((entry) => entry.isFile() && entry.name.endsWith(".json"))
-			.map((entry) => path.join(dir, entry.name));
+			.map((entry) => ({ dir, name: entry.name }));
 	} catch (error) {
 		const code = (error as NodeJS.ErrnoException).code;
 		if (code === "ENOENT" || code === "ENOTDIR") return [];
@@ -93,12 +98,16 @@ function sessionDirs(asyncDirRoot: string, sessionId: string | undefined): strin
 	}
 }
 
+function recentMarkerFiles(dirs: string[], limit: number): string[] {
+	return dirs.flatMap(markerFiles)
+		.sort((left, right) => right.name.localeCompare(left.name))
+		.slice(0, limit)
+		.map((marker) => path.join(marker.dir, marker.name));
+}
+
 export function readRecentTerminalRunIndex(asyncDirRoot: string, options: { sessionId?: string; limit?: number } = {}): string[] {
 	const limit = options.limit === undefined ? Number.POSITIVE_INFINITY : Math.max(0, Math.floor(options.limit));
-	const candidates = sessionDirs(asyncDirRoot, options.sessionId)
-		.flatMap(markerFiles)
-		.sort((left, right) => path.basename(right).localeCompare(path.basename(left)))
-		.slice(0, limit);
+	const candidates = recentMarkerFiles(sessionDirs(asyncDirRoot, options.sessionId), limit);
 	const runIds: string[] = [];
 	const seen = new Set<string>();
 	for (const marker of candidates) {

package/src/runs/background/wait-tool.ts

@@ -16,6 +16,7 @@ In an interactive chat, do not call this merely to wait: return control to the u
 • { all: true } — wait for every async run and provider item that was active when the call began.
 • { id: "..." } — wait for one async or remembered detached foreground subagent run (id or prefix).
 • { id: "...", nonBlocking: true } — resolve the prefix once, persist an exact-run wake subscription, and return immediately. The originating interactive session wakes on completion, failure, attention, reconciliation failure, or timeout.
+• { stopOnAttention: false } — for blocking waits only, keep waiting through idle or long-thinking attention; supervisor/contact requests still stop the wait.
 • { timeoutMs: 600000 } — stop waiting after N ms; active work keeps running.
 
 Non-blocking subscriptions are visible in subagent status and differ from disabling waitTool: waitTool.enabled=false returns immediately without registering any future wake. Provider jobs are session-scoped and identified exactly, so replacing one job with another cannot hide a completion. Provider extensions must be explicitly loaded in this process. In a child agent, keep \`subagent_wait\` in the child tool allowlist and load each provider through the agent's extensions or subagentOnlyExtensions; this tool never loads providers or grants tools itself.${enabled ? "" : "\n\nConfigured behavior: subagent_wait is disabled by config.waitTool or PI_SUBAGENT_WAIT_TOOL_ENABLED and returns immediately without blocking."}`,

package/src/runs/foreground/execution.ts

@@ -74,6 +74,7 @@ import {
 	buildModelCandidates,
 	formatModelAttemptNote,
 	isRetryableModelFailure,
+	recordRetryableModelFailure,
 } from "../shared/model-fallback.ts";
 import {
 	SUBAGENT_STARTUP_RETRY_DELAYS_MS,
@@ -872,6 +873,7 @@ async function runSingleAttempt(
 				startedAt: startTime,
 				lastActivityAt: progress.lastActivityAt,
 				currentTool: progress.currentTool,
+				thinking: resolvedThinking,
 				now,
 			});
 			if (idleState === "needs_attention") {
@@ -1860,7 +1862,9 @@ async function runSyncCompletionInner(
 				attempt.error = startupError;
 				break modelAttemptsLoop;
 			}
-			if (!isRetryableModelFailure(result.error) || modelIndex === modelsToTry.length - 1) break modelAttemptsLoop;
+			const retryableModelFailure = isRetryableModelFailure(result.error);
+			if (retryableModelFailure) recordRetryableModelFailure(result.model ?? candidate, result.error);
+			if (!retryableModelFailure || modelIndex === modelsToTry.length - 1) break modelAttemptsLoop;
 			attemptNotes.push(formatModelAttemptNote(attempt, modelsToTry[modelIndex + 1]));
 			break;
 		}