@bastani/atomic 0.9.57 → 0.9.58

package/dist/builtin/subagents/src/extension/index.bundle.mjs

@@ -8939,6 +8939,13 @@ function subagentChildCardRecord(childId, taskId) {
   const id = childId ?? childIdsByTaskId.get(taskId);
   return id === undefined ? undefined : latest.get(id);
 }
+function observedSubagentRunIds(ownerSessionId) {
+  const runIds = new Set;
+  for (const record of latest.values())
+    if (record.owner?.sessionId === ownerSessionId)
+      runIds.add(record.runId);
+  return runIds;
+}
 function watchSubagentChildCard(card, invalidate, key, ownerSessionId) {
   invalidators.set(card, invalidate);
   const cards = watchers.get(key) ?? new Set;
@@ -10820,6 +10827,7 @@ ${launch.message}` : input.task,
       },
       wait: { kind: "background" },
       runtime: input.runtime,
+      ...input.heartbeatMs === undefined ? {} : { heartbeatMs: input.heartbeatMs },
       onTerminal: launch.onTerminal,
       ...input.schedule ? { schedule: input.schedule } : {},
       ...input.outputText ? { outputText: input.outputText } : {},
@@ -10834,7 +10842,8 @@ ${launch.message}` : input.task,
     agent: input.agent,
     task: input.intentTask ?? input.task,
     cwd: input.options.cwd ?? input.cwd,
-    ...input.options.modelRoute ? { routerSelection: input.options.modelRoute.routerSelection } : {}
+    ...input.options.modelRoute ? { routerSelection: input.options.modelRoute.routerSelection } : {},
+    ...input.heartbeatMs === undefined ? {} : { heartbeatMs: input.heartbeatMs }
   }, input.operation ?? `${input.options.runId}:${input.options.index ?? 0}`, (context) => {
     dispatched = true;
     queuedChildren.delete(context.ref.taskId);
@@ -11182,6 +11191,7 @@ async function runForegroundParallelTasks(input) {
         options: runOptions,
         runtime: input.runtime,
         wait: input.wait,
+        ...(task.heartbeatMs ?? input.heartbeatMs) === undefined ? {} : { heartbeatMs: task.heartbeatMs ?? input.heartbeatMs },
         schedule,
         onTerminal: (child) => {
           terminalChild = child;
@@ -11822,7 +11832,10 @@ import {
 import {
   SubagentControl
 } from "@bastani/atomic-natives";
-import { clampThinkingLevel } from "@bastani/pi-ai/compat";
+import {
+  clampThinkingLevel,
+  getSupportedThinkingLevels
+} from "@bastani/pi-ai/compat";
 
 // dist/builtin/subagents/src/agents/skills.ts
 import * as fs14 from "node:fs";
@@ -12563,6 +12576,23 @@ function createInProcessChildPromptBehavior(policy) {
 }
 
 // dist/builtin/subagents/src/runs/inprocess/runner.ts
+var SUBAGENT_SESSION_START_TIMEOUT_MS = 5 * 60000;
+function formatStartTimeout(ms) {
+  return ms % 60000 === 0 ? `${ms / 60000} min` : `${Math.max(1, Math.round(ms / 1000))}s`;
+}
+var DEFAULT_THINKING_LEVEL = "medium";
+function explicitEffortFor(model, requested, settings) {
+  if (getSupportedThinkingLevels(model).includes(requested))
+    return requested;
+  return clampThinkingLevel(model, settings.getModelThinkingLevel(model.provider, model.id) ?? settings.getDefaultThinkingLevel() ?? DEFAULT_THINKING_LEVEL);
+}
+function pinFallbackEffort(entry, requested, settings, resolve) {
+  const resolved = resolve?.(entry);
+  if (!resolved)
+    return entry;
+  const { baseModel } = splitKnownThinkingSuffix(entry.trim());
+  return `${baseModel}:${explicitEffortFor(resolved.model, resolved.thinkingLevel ?? requested, settings)}`;
+}
 function modelIdForCandidate(candidate, fallbackModel) {
   return candidate.modelId ?? (candidate.model ? `${candidate.model.provider}/${candidate.model.id}` : undefined) ?? (fallbackModel ? `${fallbackModel.provider}/${fallbackModel.id}` : undefined);
 }
@@ -13015,12 +13045,23 @@ class SubagentControlRuntime {
   delivered = new Set;
   deliveredEnvelopes = new Map;
   killedChildren = new Set;
+  childIds = new Map;
   nextAttemptId = 1;
   constructor(parent, sessionRoot, runningCapacity) {
     this.parent = { ...parent };
     this.sessionRoot = trustedSessionRoot(this.parent, sessionRoot);
     this.native = new SubagentControl(this.parent.path, runningCapacity);
   }
+  childIdOf(pathValue) {
+    return this.childIds.get(pathValue);
+  }
+  pathOfChildId(childId) {
+    let latest;
+    for (const [pathValue, id] of this.childIds)
+      if (id === childId)
+        latest = pathValue;
+    return latest;
+  }
   registerAgents(agents) {
     for (const agent of agents)
       this.native.registerAgent(agent.name);
@@ -13038,6 +13079,8 @@ class SubagentControlRuntime {
     if (!native.child)
       return refusal(native);
     const identity = native.child;
+    if (spec.childId !== undefined)
+      this.childIds.set(identity.path, spec.childId);
     const childModePolicy = resolveChildModePolicy(spec);
     const sessionDir = spec.sessionManager?.getSessionDir() || sessionDirectory(this.sessionRoot, identity.path);
     mkdirSync7(sessionDir, { recursive: true });
@@ -13180,9 +13223,34 @@ class SubagentControlRuntime {
     signals.interrupt.addEventListener("abort", interruptListener, {
       once: true
     });
+    const startTimeoutMs = admitted.spec.sessionStartTimeoutMs ?? SUBAGENT_SESSION_START_TIMEOUT_MS;
+    const startDeadline = new AbortController;
+    let startPhase = "opening its session";
+    const startTimer = setTimeout(() => startDeadline.abort(new Error(`Subagent session did not start within ${formatStartTimeout(startTimeoutMs)} (still ${startPhase}); it was stopped before its task ran.`)), startTimeoutMs);
+    const beforeStartDeadline = (phase, work, abandon) => {
+      startPhase = phase;
+      return new Promise((resolve, reject) => {
+        const expire = () => {
+          work.then((late) => abandon?.(late), () => {
+            return;
+          });
+          reject(startDeadline.signal.reason);
+        };
+        if (startDeadline.signal.aborted)
+          return expire();
+        startDeadline.signal.addEventListener("abort", expire, { once: true });
+        work.then((value) => {
+          startDeadline.signal.removeEventListener("abort", expire);
+          resolve(value);
+        }, (error) => {
+          startDeadline.signal.removeEventListener("abort", expire);
+          reject(error);
+        });
+      });
+    };
     try {
       const workflow = workflowMetadataFromContext(admitted.spec.parent?.orchestrationContext);
-      const sessionManager = admitted.sessionManager ? await SessionManager3.open(admitted.sessionManager.getSessionStore(), admitted.sessionManager.getSessionId(), admitted.policy.cwd) : await SessionManager3.create(admitted.policy.cwd, admitted.sessionDir, workflow ? { internal: true, workflow } : { internal: true });
+      const sessionManager = await beforeStartDeadline("opening its session", admitted.sessionManager ? SessionManager3.open(admitted.sessionManager.getSessionStore(), admitted.sessionManager.getSessionId(), admitted.policy.cwd) : SessionManager3.create(admitted.policy.cwd, admitted.sessionDir, workflow ? { internal: true, workflow } : { internal: true }));
       transcriptManager = sessionManager;
       taskHooks?.bindTranscript?.(sessionManager);
       const childSession = {
@@ -13192,7 +13260,7 @@ class SubagentControlRuntime {
       taskHooks?.bindChildSession?.(childSession);
       admitted.spec.onChildSession?.(childSession);
       if (workflow)
-        await sessionManager.markSessionInternal(workflow);
+        await beforeStartDeadline("opening its session", sessionManager.markSessionInternal(workflow));
       const launchLeafId = sessionManager.getLeafId();
       ownMessages = () => messagesSinceLaunch(sessionManager, launchLeafId);
       let created;
@@ -13218,6 +13286,13 @@ class SubagentControlRuntime {
         } : inherited.extensionBindings;
         const agentDir = inherited.agentDir ?? getAgentDir2();
         const settingsManager = inherited.settingsManager ?? SettingsManager.create(admitted.policy.cwd, agentDir);
+        const requested = admitted.spec.requestedThinking;
+        const startSuffix = splitKnownThinkingSuffix(candidate.modelId ?? "").thinkingSuffix.slice(1);
+        const startEffort = THINKING_LEVELS.find((level) => level === startSuffix) ?? requested;
+        const explicitEffort = requested === undefined || startEffort === undefined ? {} : {
+          ...inherited.model ? { thinkingLevel: explicitEffortFor(inherited.model, startEffort, settingsManager) } : {},
+          fallbackModels: (inherited.fallbackModels ?? settingsManager.getFallbackModels() ?? []).map((entry) => pinFallbackEffort(entry, requested, settingsManager, admitted.spec.resolveCandidateModel))
+        };
         const resourceLoader = new DefaultResourceLoader(inProcessChildResourceLoaderOptions({
           cwd: admitted.policy.cwd,
           agentDir,
@@ -13227,7 +13302,7 @@ class SubagentControlRuntime {
           resourceLoaderInheritanceSnapshot: admitted.spec.parent?.resourceLoaderInheritanceSnapshot,
           orchestrationContext: admitted.spec.parent?.orchestrationContext
         }));
-        await resourceLoader.reload();
+        await beforeStartDeadline("loading its resources", resourceLoader.reload());
         const selectedSkills = resolveSkillsFromCatalog([...admitted.policy.skills], resourceLoader.getSkillCatalog(), admitted.policy.cwd);
         skillReport = {
           ...selectedSkills.resolved.length > 0 ? { skills: selectedSkills.resolved.map((skill) => skill.name) } : {},
@@ -13239,26 +13314,28 @@ class SubagentControlRuntime {
         const skillInjection = buildSkillInjection(selectedSkills.resolved);
         const promptBehavior = createInProcessChildPromptBehavior(admitted.policy);
         const systemPromptTransform = createInProcessChildSystemPromptTransform(admitted.policy, skillInjection);
-        created = {
-          session: (await createAgentSession({
-            ...inherited,
-            extensionBindings,
-            customTools: admitted.policy.customTools ?? inherited.customTools,
-            resourceLoader,
-            resourceAdmissionSignal: AbortSignal.any([signals.abort, signals.interrupt]),
-            subagentPolicy: {
-              ...admitted.policy,
-              executionEnded: executionEnded.signal,
-              ...!childAllowsIntercom(inherited) ? { intercom: undefined, intercomGroup: undefined } : {}
-            },
-            settingsManager,
-            orchestrationContext: admitted.spec.parent?.orchestrationContext,
-            sessionManager,
-            systemPromptTransform,
-            initialContextTransform: promptBehavior.initialContextTransform
-          })).session
-        };
+        created = await beforeStartDeadline("creating its agent session, which includes any wait for a resource-admission permit", createAgentSession({
+          ...inherited,
+          ...explicitEffort,
+          extensionBindings,
+          customTools: admitted.policy.customTools ?? inherited.customTools,
+          resourceLoader,
+          resourceAdmissionSignal: AbortSignal.any([signals.abort, signals.interrupt, startDeadline.signal]),
+          subagentPolicy: {
+            ...admitted.policy,
+            executionEnded: executionEnded.signal,
+            ...!childAllowsIntercom(inherited) ? { intercom: undefined, intercomGroup: undefined } : {}
+          },
+          settingsManager,
+          orchestrationContext: admitted.spec.parent?.orchestrationContext,
+          sessionManager,
+          systemPromptTransform,
+          initialContextTransform: promptBehavior.initialContextTransform
+        }), (late) => void late.session.dispose().catch(() => {
+          return;
+        }));
       }
+      clearTimeout(startTimer);
       session = created.session;
       const transcriptSession = session;
       let observedStreamingMessage;
@@ -13490,6 +13567,7 @@ class SubagentControlRuntime {
         ...skillReport
       };
     } finally {
+      clearTimeout(startTimer);
       executionEnded.abort();
       signals.abort.removeEventListener("abort", abortListener);
       signals.interrupt.removeEventListener("abort", interruptListener);
@@ -13677,11 +13755,24 @@ function findSubagentControl(id) {
   if (direct)
     return direct;
   for (const control of controls.values()) {
+    if (control.pathOfChildId(id) !== undefined)
+      return control;
     if (control.native.listChildren().some((child) => child.path === id))
       return control;
   }
   return;
 }
+function subagentChildPath(control, id) {
+  return control.pathOfChildId(id) ?? id;
+}
+function subagentChildIdOfPath(id) {
+  for (const control of controls.values()) {
+    const childId = control.childIdOf(id);
+    if (childId !== undefined)
+      return childId;
+  }
+  return;
+}
 function listSubagentControls() {
   return [...controls.values()];
 }
@@ -14027,6 +14118,8 @@ async function runSingleInProcess(runtimeCwd, agent, task, options) {
   const resolvedCandidate = candidate ? options.resolveCandidateModel?.(candidate) : undefined;
   if (options.modelRoute && !resolvedCandidate)
     throw new Error("Subagent auto-selected model could not be resolved before execution. Retry explicitly with the current catalog.");
+  const requestedSuffix = splitKnownThinkingSuffix(options.modelOverride ?? agent.model ?? "").thinkingSuffix.slice(1);
+  const requestedThinking = options.modelRoute ? undefined : THINKING_LEVELS.find((level) => level === requestedSuffix);
   const orchestrationContext = workflowOrchestrationContext(options);
   const outputSnapshot = captureSingleOutputSnapshot(options.outputPath);
   const parent = {
@@ -14053,6 +14146,7 @@ ${task}`);
   const testSession = options.testSession ?? defaultTestSession();
   const spec = {
     taskName: agent.name,
+    childId: `${options.runId}:${options.index ?? 0}`,
     task,
     agent,
     cwd,
@@ -14066,6 +14160,11 @@ ${task}`);
     ...options.progressPath ? { progressPath: options.progressPath } : {},
     ...options.progressArtifactPath ? { progressArtifactPath: options.progressArtifactPath } : options.progressPath && artifactsDir ? { progressArtifactPath: options.progressPath } : {},
     ...artifactPaths?.outputPath ? { outputArtifactPath: artifactPaths.outputPath } : {},
+    ...options.sessionStartTimeoutMs === undefined ? {} : { sessionStartTimeoutMs: options.sessionStartTimeoutMs },
+    ...requestedThinking === undefined ? {} : {
+      requestedThinking,
+      ...options.resolveCandidateModel ? { resolveCandidateModel: options.resolveCandidateModel } : {}
+    },
     tools: agent.tools,
     mcpDirectTools: agent.mcpDirectTools,
     skills: options.skills ?? agent.skills,
@@ -15334,6 +15433,9 @@ async function recordReportedGoneSubagentChildren(pi, outcomes) {
   }
 }
 
+// dist/builtin/subagents/src/runs/foreground/subagent-executor.ts
+import { validateTaskHeartbeatMs } from "@bastani/atomic";
+
 // dist/builtin/subagents/src/agents/agent-management.ts
 import * as fs15 from "node:fs";
 import * as path16 from "node:path";
@@ -15892,34 +15994,52 @@ function handleSubagentControlNotice(input) {
 function canonicalChildren(control) {
   return [...control.listChildren()].sort((left, right) => left.path.localeCompare(right.path));
 }
+function settledCause(control, path, status) {
+  if (status !== "error" && status !== "interrupted")
+    return;
+  const cause = control.getDeliveredResult(path)?.cause?.replace(/\s+/g, " ").trim();
+  return cause || undefined;
+}
 function childLines(control, id) {
-  if (!control)
-    return [];
   if (id) {
-    const child = control.findChild(id);
+    const path = subagentChildPath(control, id);
+    const child = control.findChild(path);
     if (!child)
       return [];
-    const delivered = control.getDeliveredResult(id);
+    const delivered = control.getDeliveredResult(path);
+    const childId = control.childIdOf(path);
+    const cause = settledCause(control, path, child.status);
     return [
+      ...childId ? [`Id: ${childId}`] : [],
       `Child: ${child.path}`,
       `Parent: ${child.parentPath}`,
       `Task: ${child.taskName}`,
       `Depth: ${child.depth}`,
       `Status: ${child.status}`,
+      ...cause ? [`Cause: ${cause}`] : [],
       `Residency: ${child.loaded ? "loaded" : "cold"}`,
       ...delivered?.sessionFile ? [`Session: ${delivered.sessionFile}`] : []
     ];
   }
-  return control.listChildren().map((child) => `${child.path} — ${child.status} (${child.loaded ? "loaded" : "cold"})`);
+  return control.listChildren().map((child) => {
+    const childId = control.childIdOf(child.path);
+    const cause = settledCause(control, child.path, child.status);
+    return `${child.path} — ${child.status} (${child.loaded ? "loaded" : "cold"})${childId ? ` · id ${childId}` : ""}${cause ? ` · ${cause}` : ""}`;
+  });
 }
 function statusGroup(control, id) {
+  const path = id === undefined ? undefined : subagentChildPath(control, id);
   return {
     parentPath: control.parent.path,
-    children: canonicalChildren(control).filter((child) => !id || id === control.parent.path || child.path === id).map((child) => {
+    children: canonicalChildren(control).filter((child) => !path || path === control.parent.path || child.path === path).map((child) => {
       const delivered = control.getDeliveredResult(child.path);
       const metadata = control.getChildMetadata(child.path) ?? delivered;
+      const childId = control.childIdOf(child.path);
+      const failure = settledCause(control, child.path, child.status);
       return {
         ...child,
+        ...childId === undefined ? {} : { childId },
+        ...failure === undefined ? {} : { failure },
         sessionFile: delivered?.sessionFile,
         ...metadata?.model === undefined ? {} : { model: metadata.model },
         ...metadata?.thinking === undefined ? {} : { thinking: metadata.thinking }
@@ -15927,10 +16047,11 @@ function statusGroup(control, id) {
     })
   };
 }
-function inspectInProcessChildStatus(id) {
+function inspectInProcessChildStatus(id, ownsRun) {
+  const owned = (control) => ownsRun?.(control.parent.path) ?? true;
   if (id) {
     const control = findSubagentControl(id);
-    if (!control)
+    if (!control || !owned(control))
       return;
     const text = id === control.parent.path ? [`Parent: ${control.parent.path}`, ...childLines(control)].join(`
 `) : childLines(control, id).join(`
@@ -15942,7 +16063,7 @@ function inspectInProcessChildStatus(id) {
       details: { mode: "management", results: [], statusGroups: [statusGroup(control, id)] }
     };
   }
-  const controls = listSubagentControls();
+  const controls = listSubagentControls().filter(owned);
   const lines = controls.flatMap((control) => [`Parent: ${control.parent.path}`, ...childLines(control)]);
   if (lines.length === 0)
     return;
@@ -16257,7 +16378,8 @@ function topLevelSingleTask(item) {
     ...item.params.modelConstraints !== undefined ? { modelConstraints: item.params.modelConstraints } : {},
     ...item.params.taskNeeds !== undefined ? { taskNeeds: item.params.taskNeeds } : {},
     ...item.params.skill !== undefined ? { skill: item.params.skill } : {},
-    ...item.params.group !== undefined ? { group: item.params.group } : {}
+    ...item.params.group !== undefined ? { group: item.params.group } : {},
+    ...item.params.heartbeatMs !== undefined ? { heartbeatMs: item.params.heartbeatMs } : {}
   };
 }
 function existingParallelTask(item, task) {
@@ -16266,7 +16388,8 @@ function existingParallelTask(item, task) {
     ...task,
     cwd: originatingTaskCwd(callCwd, task.cwd),
     [BURST_TASK_DISCOVERY_CWD]: callCwd,
-    ...task.group === undefined && item.params.group !== undefined ? { group: item.params.group } : {}
+    ...task.group === undefined && item.params.group !== undefined ? { group: item.params.group } : {},
+    ...task.heartbeatMs === undefined && item.params.heartbeatMs !== undefined ? { heartbeatMs: item.params.heartbeatMs } : {}
   };
 }
 function mergeBurst(items) {
@@ -16992,6 +17115,7 @@ async function runParallelPath(data, deps) {
     const detachedNotificationScope = { orchestrationContext: ctx.orchestrationContext };
     const results = await runForegroundParallelTasks({
       wait: params.wait,
+      ...params.heartbeatMs === undefined ? {} : { heartbeatMs: params.heartbeatMs },
       schedule: params[BURST_EXECUTION_SCHEDULE],
       onDetachedExit: (index, result) => {
         if (workflowStageAcceptsDetachedNotification(detachedNotificationScope)) {
@@ -17340,6 +17464,7 @@ async function runSinglePath(data, deps) {
         intentTask: handoffTaskContext,
         options: runOptions,
         wait: params.wait,
+        ...params.heartbeatMs === undefined ? {} : { heartbeatMs: params.heartbeatMs },
         schedule: params[BURST_EXECUTION_SCHEDULE],
         runtime: deps.runtime,
         recordChild: createSubagentChildRecordSink(deps.pi, ctx),
@@ -17551,10 +17676,22 @@ function liveTaskIds(host) {
 function ownedRecords(ctx) {
   return readSubagentChildRecords(ctx.sessionManager.getEntries(), subagentChildOwner(ctx).sessionId);
 }
+async function subagentRunsLaunchedBy(ctx) {
+  const owner = subagentChildOwner(ctx).sessionId;
+  await subagentChildRecordsAppended(owner);
+  const runIds = observedSubagentRunIds(owner);
+  for (const record of ownedRecords(ctx).values())
+    runIds.add(record.runId);
+  return runIds;
+}
 function resolveRecord(records, id) {
   const direct = findSubagentChildRecord(records, id);
   if (direct)
     return direct;
+  const listedAs = subagentChildIdOfPath(id.trim());
+  const byPath = listedAs === undefined ? undefined : records.get(listedAs);
+  if (byPath)
+    return byPath;
   const inRun = [...records.values()].filter((record) => record.runId === id.trim());
   if (inRun.length === 1)
     return inRun[0];
@@ -17895,7 +18032,8 @@ async function handleManagementRequest(input) {
   if (action === "status") {
     const targetRunId = paramsWithResolvedCwd.id ?? paramsWithResolvedCwd.runId;
     const foreground = getForegroundControl(deps.state, targetRunId);
-    const live = inspectInProcessChildStatus(targetRunId) ?? (foreground ? foregroundStatusResult(foreground) : undefined);
+    const launched = await subagentRunsLaunchedBy(ctx);
+    const live = inspectInProcessChildStatus(targetRunId, (runId) => launched.has(runId)) ?? (foreground ? foregroundStatusResult(foreground) : undefined);
     if (awaitingResume.length > 0) {
       const liveText = live?.content.map((part) => part.type === "text" ? part.text : "").join(`
 `);
@@ -17926,6 +18064,22 @@ function inferExecutionMode(params) {
     return "parallel";
   return "single";
 }
+function invalidHeartbeatResult(params) {
+  try {
+    if (params.heartbeatMs !== undefined)
+      validateTaskHeartbeatMs(params.heartbeatMs);
+    for (const [index, task] of (params.tasks ?? []).entries())
+      if (task.heartbeatMs !== undefined)
+        validateTaskHeartbeatMs(task.heartbeatMs, `tasks[${index}].heartbeatMs`);
+    return;
+  } catch (error) {
+    return {
+      content: [{ type: "text", text: error instanceof Error ? error.message : String(error) }],
+      isError: true,
+      details: { mode: inferExecutionMode(params), results: [] }
+    };
+  }
+}
 function duplicateSubagentCallResult(params) {
   return {
     content: [
@@ -17999,7 +18153,7 @@ function createSubagentExecutor(rawDeps) {
     duplicateResult: duplicateSubagentCallResult
   });
   return {
-    execute: (id, params, signal, onUpdate, ctx) => params.action ? execute(id, params, signal, onUpdate, ctx) : executeWithBurstCollection(id, params, signal, onUpdate, ctx)
+    execute: async (id, params, signal, onUpdate, ctx) => params.action ? execute(id, params, signal, onUpdate, ctx) : invalidHeartbeatResult(params) ?? executeWithBurstCollection(id, params, signal, onUpdate, ctx)
   };
 }
 
@@ -19237,11 +19391,13 @@ function childStatusLines(child, expanded, theme, width) {
   return [
     heading,
     ...metadata,
+    ...child.childId ? [theme.fg("dim", `  Id: ${displayText(child.childId)}`)] : [],
     theme.fg("dim", `  Path: ${displayText(child.path)}`),
     theme.fg("dim", `  Parent: ${displayText(child.parentPath)} · Depth: ${child.depth}`),
     theme.fg("dim", `  Task: ${displayText(child.taskName)}`),
     theme.fg("dim", `  Residency: ${child.loaded ? "loaded" : "cold"}`),
     ...child.cause ? [theme.fg("muted", `  Cause: ${displayText(child.cause)}`)] : [],
+    ...child.failure ? [theme.fg("error", `  Error: ${displayText(child.failure)}`)] : [],
     ...child.sessionFile ? [theme.fg("dim", `  Session: ${displayText(child.sessionFile)}`)] : []
   ];
 }
@@ -19709,6 +19865,12 @@ var MaxOutputSchema = Type.Object({
   bytes: Type.Optional(Type.Number()),
   lines: Type.Optional(Type.Number())
 }, { additionalProperties: false });
+function heartbeatSchema(scope) {
+  return Type.Number({
+    minimum: 0,
+    description: `Heartbeat cadence in ms for ${scope} while it runs (default 15 min; 0 disables; min 60000).`
+  });
+}
 var GroupSchema = Type.Union([Type.String(), Type.Boolean()], {
   description: "Intercom group for spawned children. A named string joins that group; boolean `true` or the trimmed, case-insensitive string sentinel `true`/`auto` auto-generates one shared UUID group per parallel set. The names `true` and `auto` are reserved; use a different literal group name. Defaults to the current session/stage's group. Only applied when the child has intercom access; contact_supervisor still reaches the supervisor across groups."
 });
@@ -19727,7 +19889,8 @@ var TaskItem = Type.Object({
   modelConstraints: Type.Optional(ModelConstraintsSchema),
   taskNeeds: Type.Optional(TaskNeedsSchema),
   skill: Type.Optional(SkillOverride),
-  group: Type.Optional(GroupSchema)
+  group: Type.Optional(GroupSchema),
+  heartbeatMs: Type.Optional(heartbeatSchema("this task"))
 });
 var ControlOverrides = Type.Object({
   enabled: Type.Optional(Type.Boolean({ description: "Enable/disable subagent control attention tracking for this run" })),
@@ -19764,6 +19927,7 @@ var SubagentParams = Type.Object({
     budgetMs: Type.Optional(Type.Number({ description: "Foreground wait budget in milliseconds (kind='foreground' only)." }))
   }, { additionalProperties: false })),
   budgetMs: Type.Optional(Type.Number({ description: "Foreground wait budget in milliseconds." })),
+  heartbeatMs: Type.Optional(heartbeatSchema("each child; a task's own wins")),
   agent: Type.Optional(Type.String({ description: "Agent name (SINGLE mode) or target for management get/update/delete" })),
   task: Type.Optional(Type.String({ description: "Task (SINGLE mode, optional for self-contained agents)" })),
   action: Type.Optional(Type.String({

package/dist/builtin/web-access/CHANGELOG.md

@@ -4,6 +4,18 @@ All notable changes to this project will be documented in this file.
 
 ## [Unreleased]
 
+## [0.9.58] - 2026-11-01
+
+### Fixed
+
+- PDF text extraction no longer prints pdf.js font warnings such as `Warning: TT: undefined function: 32` over the interactive terminal UI when fetching PDFs with unusual embedded fonts. Extraction errors are still reported.
+
+## [0.9.58-alpha.1] - 2026-11-01
+
+### Fixed
+
+- PDF text extraction no longer prints pdf.js font warnings such as `Warning: TT: undefined function: 32` over the interactive terminal UI when fetching PDFs with unusual embedded fonts. Extraction errors are still reported.
+
 ## [0.9.52] - 2026-10-29
 
 ### Fixed

package/dist/builtin/web-access/index.bundle.mjs

@@ -72560,7 +72560,7 @@ async function extractPDFToMarkdown(buffer, url, options = {}) {
     filename
   } = options;
   const safeMaxPages = Number.isFinite(maxPages) ? Math.max(1, Math.floor(maxPages)) : DEFAULT_MAX_PAGES;
-  const pdf = await getDocumentProxy(new Uint8Array(buffer));
+  const pdf = await getDocumentProxy(new Uint8Array(buffer), { verbosity: PDF_JS_VERBOSITY_ERRORS });
   const metadata = await pdf.getMetadata();
   const metadataInfo = metadata.info && typeof metadata.info === "object" ? metadata.info : null;
   const metaTitle = typeof metadataInfo?.Title === "string" ? metadataInfo.Title : undefined;
@@ -72650,7 +72650,7 @@ function isPDF(url, contentType) {
     return false;
   }
 }
-var DEFAULT_MAX_PAGES = 100, DEFAULT_OUTPUT_DIR;
+var DEFAULT_MAX_PAGES = 100, PDF_JS_VERBOSITY_ERRORS = 0, DEFAULT_OUTPUT_DIR;
 var init_pdf_extract = __esm(() => {
   init_dist();
   DEFAULT_OUTPUT_DIR = join4(homedir4(), "Downloads");

package/dist/builtin/web-access/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/web-access",
-  "version": "0.9.57",
+  "version": "0.9.58",
   "private": true,
   "description": "Atomic extension for web search, URL fetching, GitHub repo cloning, PDF/video extraction. Fork of: https://github.com/nicobailon/pi-web-access",
   "contributors": [

package/dist/builtin/workflows/CHANGELOG.md

@@ -6,6 +6,48 @@ The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
 
 ## [Unreleased]
 
+## [0.9.58] - 2026-11-01
+
+### Added
+
+- `workflows.localWorkers.maxLoadPerCpu` (SDK `localWorkers.maxLoadPerCpu`): while the machine's load per CPU is above it, the stages of pooled children start one at a time, at most one every 5 seconds in the order they asked, instead of all at once while earlier ones are still starting. Running stages and their subagents and tool processes are not held. Default `3`; `0` turns it off. It is not saved with the root. ([#4146](https://github.com/bastani-inc/atomic/pull/4146))
+
+### Fixed
+
+- A message sent to a stage on a local worker (`workflows.localWorkers`) while its session was still starting no longer fails the stage with "generation closed before the stage finished; its tool calls were refused" when the stage has `fallbackModels`.
+- The chat of a stage on a local worker completes `/skill:` commands with that stage's skills, as an in-process stage chat does, instead of warning "Skill discovery unavailable: this stage host does not expose its session command metadata." While the stage's session is still starting, it says its skills are loading.
+- In the chat of a stage on a local worker, `/compact` compacts that stage's session and a `!` command runs on the worker in the stage's working directory, with its output streaming into the chat and Escape stopping it even while the pool is paused, as in the chat of an in-process stage. After `/compact` the open chat shows the compacted transcript. Before, `/compact` was sent to the model as plain text and a `!` command failed with "no live agent session on this stage". `/tasks` there lists the stage's agents and shells on its worker with their live status, opens their output and transcripts, sends them input and stops them, and the chat's task footer counts them, as in the chat of an in-process stage. An agent task's reply shows while it streams, tasks can be inspected and stopped while the pool is paused, and a stop that cannot reach the worker says so. Before, `/tasks` showed an empty list.
+- A stage on a local worker reads as queued, with what it is doing (`creating session 32s`), on its graph node, in the graph header, the stage list, `/workflow status` and the `workflow` tool's stage list until its session has started, instead of reading as running. Its chat says it is starting on its local worker instead of showing a blank transcript, streams the stage as soon as the session starts, and opens the live view by itself when the worker cannot serve one yet. A message sent meanwhile waits for the session instead of failing with "The connection to stage … was lost" and "1 message not delivered".
+- A root that runs children on local workers (`workflows.localWorkers`) no longer fails all of them with `Stale lease for run <id>` when its own process renews the root's lease late. A busy session, such as one that reloads its workflows and starts several other runs with in-process agent stages, could push the root's renewal past the pool's 90-second window. The pool treated that late renewal as a lost lease: it failed every running child, and every child the root started afterwards failed at once, while the root itself kept running. The pool now waits for the renewal and its workers keep their children; only a lease that was released or taken over by another holder ends them. Each run also renews its lease on its own, so a busy or blocked run no longer delays another run's renewal, and a renewal that fails without losing the lease, such as one that meets a briefly locked database, is retried on the next heartbeat instead of giving up the run. ([#4138](https://github.com/bastani-inc/atomic/pull/4138))
+- With `workflows.localWorkers` configured, a root starts its pool's coordinator process when it runs its first pooled `ctx.workflow` child, or when it resumes a pool that already ran. A root that only runs `ctx.task` stages and `ctx.tool` steps no longer starts one, so launching such runs no longer leaves an idle coordinator process for each of them. ([#4138](https://github.com/bastani-inc/atomic/pull/4138))
+- Stages on local workers (`workflows.localWorkers`) no longer fail with `Worker exited before delivery acknowledgement` while their worker is still running. On a busy machine a stage's startup delivery of queued Intercom messages could make no progress for a minute; the root then ended the worker's delivery request while the delivery kept going, and the next message reached a closed connection, failing the stage at `preparing message delivery` with an unknown, terminal failure. The delivery now keeps its connection to the worker until it finishes, and the stage still starts after a minute without progress, as an in-process stage does. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- A worker whose work lease ran out only because the machine was too busy to process its renewal in time keeps its child. Its next journal write, usage report or completion is no longer refused with `Stale work lease`, which ended the child's attempt and could fail the child as `killed`; a session it starts no longer fails with `Stale resource admission attempt`, and waits for the renewal instead. Only an attempt that recovery handed to another worker is refused. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- A fault of the pool itself that reaches a running pooled child, such as a journal write the database refused, replays the child from its last checkpoint (up to the usual three attempts) instead of failing its stage. A stage that does fail on a pool fault is classified with that fault as its failure code (`worker_lost`, `stale_work_lease`, …), recoverable and retryable, instead of `unknown`, and resume runs it again. A child whose attempt was fenced is no longer completed with the aborted run's result. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- A stage on a local worker that ended keeps the worker lease it ended under: its worker's later lease renewals for the child's other stages no longer show it as renewed after it ended. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- `/workflow pause <run>` and the `workflow` tool's `pause` pause a root's stages on local workers even while the root's own `ctx.tool` steps or in-process stages run, instead of reporting `No active stages to pause`. Pause, `pause --all` and resume no longer fail with `Worker control acknowledgement unresolved` when workers confirm slowly on a busy machine: the pause is recorded at once, workers pause their stages as they next check in, and the result says how many stages have not confirmed yet. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- Resuming a root that runs children on local workers runs again the children an earlier process failed on a fault of the pool itself (a stage that lost its worker's delivery or lease on a busy machine, recorded as `Worker exited before delivery acknowledgement`, `Stale work lease` or `Stale resource admission attempt`), instead of replaying their failure. This includes runs paused or failed by 0.9.57. Pausing a root whose children have durable progress now says it can be resumed from a new process once this one exits, instead of `Resume with /workflow resume on this live process … Cross-process resume requires durable checkpoint or pending prompt progress`. ([#4146](https://github.com/bastani-inc/atomic/pull/4146))
+
+## [0.9.58-alpha.1] - 2026-11-01
+
+### Added
+
+- `workflows.localWorkers.maxLoadPerCpu` (SDK `localWorkers.maxLoadPerCpu`): while the machine's load per CPU is above it, the stages of pooled children start one at a time, at most one every 5 seconds in the order they asked, instead of all at once while earlier ones are still starting. Running stages and their subagents and tool processes are not held. Default `3`; `0` turns it off. It is not saved with the root. ([#4146](https://github.com/bastani-inc/atomic/pull/4146))
+
+### Fixed
+
+- A message sent to a stage on a local worker (`workflows.localWorkers`) while its session was still starting no longer fails the stage with "generation closed before the stage finished; its tool calls were refused" when the stage has `fallbackModels`.
+- The chat of a stage on a local worker completes `/skill:` commands with that stage's skills, as an in-process stage chat does, instead of warning "Skill discovery unavailable: this stage host does not expose its session command metadata." While the stage's session is still starting, it says its skills are loading.
+- In the chat of a stage on a local worker, `/compact` compacts that stage's session and a `!` command runs on the worker in the stage's working directory, with its output streaming into the chat and Escape stopping it even while the pool is paused, as in the chat of an in-process stage. After `/compact` the open chat shows the compacted transcript. Before, `/compact` was sent to the model as plain text and a `!` command failed with "no live agent session on this stage". `/tasks` there lists the stage's agents and shells on its worker with their live status, opens their output and transcripts, sends them input and stops them, and the chat's task footer counts them, as in the chat of an in-process stage. An agent task's reply shows while it streams, tasks can be inspected and stopped while the pool is paused, and a stop that cannot reach the worker says so. Before, `/tasks` showed an empty list.
+- A stage on a local worker reads as queued, with what it is doing (`creating session 32s`), on its graph node, in the graph header, the stage list, `/workflow status` and the `workflow` tool's stage list until its session has started, instead of reading as running. Its chat says it is starting on its local worker instead of showing a blank transcript, streams the stage as soon as the session starts, and opens the live view by itself when the worker cannot serve one yet. A message sent meanwhile waits for the session instead of failing with "The connection to stage … was lost" and "1 message not delivered".
+- A root that runs children on local workers (`workflows.localWorkers`) no longer fails all of them with `Stale lease for run <id>` when its own process renews the root's lease late. A busy session, such as one that reloads its workflows and starts several other runs with in-process agent stages, could push the root's renewal past the pool's 90-second window. The pool treated that late renewal as a lost lease: it failed every running child, and every child the root started afterwards failed at once, while the root itself kept running. The pool now waits for the renewal and its workers keep their children; only a lease that was released or taken over by another holder ends them. Each run also renews its lease on its own, so a busy or blocked run no longer delays another run's renewal, and a renewal that fails without losing the lease, such as one that meets a briefly locked database, is retried on the next heartbeat instead of giving up the run. ([#4138](https://github.com/bastani-inc/atomic/pull/4138))
+- With `workflows.localWorkers` configured, a root starts its pool's coordinator process when it runs its first pooled `ctx.workflow` child, or when it resumes a pool that already ran. A root that only runs `ctx.task` stages and `ctx.tool` steps no longer starts one, so launching such runs no longer leaves an idle coordinator process for each of them. ([#4138](https://github.com/bastani-inc/atomic/pull/4138))
+- Stages on local workers (`workflows.localWorkers`) no longer fail with `Worker exited before delivery acknowledgement` while their worker is still running. On a busy machine a stage's startup delivery of queued Intercom messages could make no progress for a minute; the root then ended the worker's delivery request while the delivery kept going, and the next message reached a closed connection, failing the stage at `preparing message delivery` with an unknown, terminal failure. The delivery now keeps its connection to the worker until it finishes, and the stage still starts after a minute without progress, as an in-process stage does. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- A worker whose work lease ran out only because the machine was too busy to process its renewal in time keeps its child. Its next journal write, usage report or completion is no longer refused with `Stale work lease`, which ended the child's attempt and could fail the child as `killed`; a session it starts no longer fails with `Stale resource admission attempt`, and waits for the renewal instead. Only an attempt that recovery handed to another worker is refused. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- A fault of the pool itself that reaches a running pooled child, such as a journal write the database refused, replays the child from its last checkpoint (up to the usual three attempts) instead of failing its stage. A stage that does fail on a pool fault is classified with that fault as its failure code (`worker_lost`, `stale_work_lease`, …), recoverable and retryable, instead of `unknown`, and resume runs it again. A child whose attempt was fenced is no longer completed with the aborted run's result. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- A stage on a local worker that ended keeps the worker lease it ended under: its worker's later lease renewals for the child's other stages no longer show it as renewed after it ended. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- `/workflow pause <run>` and the `workflow` tool's `pause` pause a root's stages on local workers even while the root's own `ctx.tool` steps or in-process stages run, instead of reporting `No active stages to pause`. Pause, `pause --all` and resume no longer fail with `Worker control acknowledgement unresolved` when workers confirm slowly on a busy machine: the pause is recorded at once, workers pause their stages as they next check in, and the result says how many stages have not confirmed yet. ([#4145](https://github.com/bastani-inc/atomic/pull/4145))
+- Resuming a root that runs children on local workers runs again the children an earlier process failed on a fault of the pool itself (a stage that lost its worker's delivery or lease on a busy machine, recorded as `Worker exited before delivery acknowledgement`, `Stale work lease` or `Stale resource admission attempt`), instead of replaying their failure. This includes runs paused or failed by 0.9.57. Pausing a root whose children have durable progress now says it can be resumed from a new process once this one exits, instead of `Resume with /workflow resume on this live process … Cross-process resume requires durable checkpoint or pending prompt progress`. ([#4146](https://github.com/bastani-inc/atomic/pull/4146))
+
 ## [0.9.57] - 2026-10-31
 
 ### Changed

package/dist/builtin/workflows/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@bastani/workflows",
-  "version": "0.9.57",
+  "version": "0.9.58",
   "private": true,
   "description": "Atomic extension for multi-stage workflow authoring and execution.",
   "contributors": [