pi-better-subagents 0.14.0 → 0.15.0

package/child-incidents.ts

@@ -18,6 +18,7 @@ import { DISPOSITION_TOOL, describeOpenTargets, foldToolEnd, foldToolStart,
     newIncidentModel, readDispositionRequest, resolveDisposition, type IncidentModel, type IncidentSink } from "./incident-model.ts";
 
 const { createBashToolDefinition } = PiCodingAgent;
+import { PERMISSION_RESOURCES } from "./shared-permission-blocker.ts";
 
 // Each intent field also admits `null`, which means "not declared" (withoutAbsentIntent). Models send
 // optional fields as explicit null; accepting it in the schema means Pi hands execute the same raw
@@ -133,6 +134,7 @@ export function failureDispositionTool(cwd: () => string) {
             "expected: the failure was intentional. open: it still needs the parent's action (this asks for attention).",
             "targets: incident ids, the attemptId/operationId you declared on bash, or tool call ids. Unknown or already-disposed incidents are rejected.",
             "Never claim recovery without a successful later attempt as evidence.",
+            "For open only, permissionResource optionally reports a permission blocker on the failed attempt; this is agent-reported, not a proven policy denial or remote authorization outcome.",
         ].join(" "),
         promptSnippet: "Classify handled tool failures (recovered, superseded, expected) or flag one as open for the parent.",
         parameters: {
@@ -141,7 +143,8 @@ export function failureDispositionTool(cwd: () => string) {
                 disposition: { type: "string", enum: ["recovered", "superseded", "expected", "open"] },
                 targets: { type: "array", items: { type: "string", minLength: 1, maxLength: 200 }, minItems: 1, maxItems: 20 },
                 reason: { type: "string", minLength: 1, maxLength: 400 },
-                evidence: { type: "string", minLength: 1, maxLength: 200, description: "attemptId or tool call id of the successful later attempt; required for recovered and superseded." },
+                evidence: { ...nullable({ type: "string", minLength: 1, maxLength: 200 }), description: "attemptId or tool call id of the successful later attempt; required for recovered and superseded." },
+                permissionResource: { ...nullable({ type: "string", enum: [...PERMISSION_RESOURCES] }), description: "Optional permission resource for open only. Bound to an unresolved failed worker attempt; remote outcome remains unknown." },
             },
             required: ["disposition", "targets", "reason"],
             additionalProperties: false,
@@ -158,7 +161,8 @@ export function failureDispositionTool(cwd: () => string) {
             const incidents = outcome.event.incidents ?? [];
             return {
                 content: [{ type: "text" as const, text: `Recorded ${parsed.request.disposition} for ${incidents.length} incident${incidents.length === 1 ? "" : "s"}: ${incidents.join(", ")}.` }],
-                details: { disposition: parsed.request.disposition, incidents },
+                details: { disposition: parsed.request.disposition, incidents,
+                    ...(outcome.event.permissionBlockers ? { permissionBlockers: outcome.event.permissionBlockers } : {}) },
             };
         },
     };

package/docs/failure-observations.md

@@ -66,6 +66,40 @@ Trust is decided once per scan, when the scan begins. A running trusted run whos
 
 Unconfined children keep the exact-retry rule, and this is intentional: an unconfined child can rewrite its own log and metadata, so the parent cannot trust intent or dispositions it reports. Background tasks, which the parent agent launches itself, accept `operation_id` and `expected_exit_codes` directly (see the background-tasks README).
 
+## Permission Blockers
+
+A trusted worker can attach `permissionResource` to an `open` disposition of
+an unresolved failed tool attempt. Supported categories include
+`credential-files`, `process-inspection`, `project-files`,
+`outside-project-files`, `runtime-control-files`, `command-execution`,
+`network-access`, and `unknown`. For example:
+
+```json
+{
+  "disposition": "open",
+  "targets": ["failed-attempt-id"],
+  "reason": "Local credential-cache access needs human review",
+  "permissionResource": "credential-files"
+}
+```
+
+The parent validates the failed-attempt reference and worker provenance before
+exposing actionable reports in `subagent_result` and `subagent_output` details.
+The report is explicitly **agent-reported**, with **remote outcome unknown**.
+It does not prove sandbox causation, expired login, an IAM denial, a foreground
+restriction, or successful validation. Raw `EPERM` output is not classified
+into this contract. Failed attempts remain failed and history is retained after
+later recovery.
+
+Inspecting an actionable report can permission-hold the active goal. Ordinary
+questions, unrelated completions, and settings changes do not release the hold.
+The human's `/goal resume` or resume shortcut confirms one same-operation,
+same-scope retry; it does not grant additional access or authorize product tests
+or mutations. Changed worker settings require a fresh worker because running
+workers retain their launch policy. See
+[ADR 0012](../../../docs/adr/0012-actionable-permission-blockers.md) for the
+transport limitations and bounded-turn recovery contract.
+
 ## Evidence health
 
 The collector scans complete structured records independently of the finite progress tail. Missing terminal logs, unreadable records, and detected truncation are shown as **observation incomplete**. Failure classification uses structured error and exit fields; output text and domain-specific status codes are not failure signals.

package/docs/usage.md

@@ -67,7 +67,7 @@ is present; without this extension it uses adaptive guidance.
   `subagent_spawn` starts a detached `pi -p` child and returns immediately,
   leaving the session free for the human. When the child finishes, it sends a
   lightweight trigger; the foreground calls `subagent_result` and presents the
-  result (as a `followUp`, never cutting into work in progress). The foreground
+  result (as a `followUp` by default; optional busy steering is described below). The foreground
   is nudged once, at completion — never on a wait/poll loop.
 - **Subagents are autonomous; communication is one-way (parent → child).** The
   parent front-loads everything the child needs into the spawn; the child runs to
@@ -247,7 +247,7 @@ given.
   its output is durable in a log file.
 - **Completions trigger one batched turn that fetches durable results.** Ordinary
   terminal callbacks stay in the harness queue while the foreground agent is
-  busy. Once Pi is idle, they accumulate for 100 ms and share one
+  busy by default. Once Pi is idle, they accumulate for 100 ms and share one
   `pi.sendMessage(..., { deliverAs: "followUp", triggerTurn: true })` notification
   with background-task completions from the same Pi host. Each row contains only
   source, id, label, terminal status, and the `subagent_result` lookup. Full
@@ -260,6 +260,23 @@ given.
   pending records retryable, and `/reload` recovers records explicitly marked
   pending. `callback:false` sends no model message; read the durable result later
   with `subagent_result`.
+- **Optional busy steering.** In `/harness-settings`, set **Completions while
+  busy** to **Steer active run** to receive ordinary completions during
+  foreground work. Changes autosave to the current session branch and apply
+  immediately; **Ctrl+S** saves the choice as the user default for future
+  sessions, including standalone callback packages. **Wait until idle** is the
+  default. The former `PI_BETTER_CALLBACK_WHILE_BUSY` environment variable is
+  no longer supported. Completions spanning multiple accumulation windows during
+  one tool call stay together until the final active tool ends, then share one
+  compact `steer` notification without triggering a new run. Parallel and nested
+  tools are tracked together by both extensions. While busy with no active tool,
+  the normal accumulation window applies. Manual compaction and branch
+  summarization hold callbacks until an agent run starts or Pi becomes idle.
+  Idle delivery remains
+  `{ deliverAs: "followUp", triggerTurn: true }`. The same byte budget, queued overflow,
+  dedupe, delivery receipts, retry, ownership checks, and `callback:false` rules
+  apply to both modes; a completion handed off by steer is not sent again at
+  idle. Urgent health/failure alerts are unchanged.
 - **The prompt guidelines forbid polling.** The foreground agent is told, in the
   tool guidelines, that spawning is done and it must not loop on `output`/`result`
   or sleep to wait.
@@ -362,6 +379,14 @@ Pi SDK 0.82.1 or newer; it is not a replacement user-facing Pi command.
 
 ### Git-mutating subagents and linked worktrees
 
+Confined bash commands default to `GIT_OPTIONAL_LOCKS=0`, unless the inherited
+environment or the call supplies a value. This keeps inventory commands such as
+`git status` from creating an optional `index.lock` in the parent repository's
+linked-worktree metadata. On macOS, Outside project = Write can permit creation
+there but refuse unlink, leaving an empty lock even when status exits zero.
+The default skips optional index refreshes; it does not disable required locks
+for `git add` or commits, grant metadata removal, or remove existing locks.
+
 A sandboxed subagent that will mutate Git should set **`git_clone_workspace:true`**
 on `subagent_spawn`. The parent prepares a fresh, self-contained Git clone whose
 `.git/` directory lives **inside the sandbox writable root**, then runs the child

package/failures.ts

@@ -201,7 +201,7 @@ export function collectRunFailures(id: string, cwd: string, terminal = false): F
                 scan = { offset: 0, head, identity, model: newIncidentModel(false) };
             } else {
                 trustUnknownSince.delete(id);
-                scan = { offset: 0, head, identity, model: newIncidentModel(trust === "trusted") };
+                scan = { offset: 0, head, identity, model: newIncidentModel(trust === "trusted", id) };
                 // The gap is settled after this scan has folded the log (settleMetadataGap), and again after
                 // every later scan, once no leftover of the exact-rule scan that trust would have read
                 // differently is still unresolved (#332).

package/incident-model.ts

@@ -10,6 +10,7 @@
  */
 import { failureIdentity, findIncident, validateDisposition, INCIDENT_DISPOSITIONS,
     REJECTED_INTENT_CATEGORY, readCommandIntent, withoutAbsentIntent, type CommandIntent, type FailureEvent, type FailureState, type IncidentDisposition } from "./shared-failure-observations.ts";
+import { isPermissionResource, type PermissionBlocker, type PermissionResource } from "./shared-permission-blocker.ts";
 
 export const DISPOSITION_TOOL = "failure_disposition";
 
@@ -75,6 +76,7 @@ interface Finished extends Attempt {
 export interface IncidentModel {
     /** Honour operationId / attemptId / expectedExitCodes / dispositions (confined runtime only). */
     structuredIntent: boolean;
+    runId?: string;
     sequence: number;
     open: Map<string, Attempt>;
     finished: Map<string, Finished>;
@@ -96,8 +98,8 @@ export interface IncidentModel {
  * (intent bash + failure_disposition). Anywhere else the intent fields and disposition
  * requests are ignored and the exact-retry identity rule applies.
  */
-export function newIncidentModel(structuredIntent = false): IncidentModel {
-    return { structuredIntent, sequence: 0, open: new Map(), finished: new Map(), attemptIds: new Map(), namedAttempts: new Set(), operationIds: new Map(),
+export function newIncidentModel(structuredIntent = false, runId?: string): IncidentModel {
+    return { structuredIntent, ...(runId ? { runId } : {}), sequence: 0, open: new Map(), finished: new Map(), attemptIds: new Map(), namedAttempts: new Set(), operationIds: new Map(),
         failed: new Map(), incidentSequence: new Map(), pendingDispositions: new Map() };
 }
 export interface IncidentSink {
@@ -114,8 +116,11 @@ export interface DispositionRequest {
     targets: string[];
     reason: string;
     evidence?: string;
+    permissionResource?: PermissionResource;
 }
 export function readDispositionRequest(args: unknown): { request?: DispositionRequest; error?: string } {
+    if (!args || typeof args !== "object" || Array.isArray(args) || Object.keys(args).some(key =>
+        !["disposition", "targets", "reason", "evidence", "permissionResource"].includes(key))) return { error: "Unknown disposition fields" };
     const input = (args && typeof args === "object" ? args : {}) as Record<string, unknown>;
     const disposition = input.disposition;
     if (typeof disposition !== "string" || !INCIDENT_DISPOSITIONS.includes(disposition as IncidentDisposition)) {
@@ -126,9 +131,16 @@ export function readDispositionRequest(args: unknown): { request?: DispositionRe
         return { error: "targets must name 1-20 incidents, attempt ids, or operation ids" };
     }
     if (typeof input.reason !== "string" || !input.reason.trim()) return { error: "reason is required" };
-    if (input.evidence !== undefined && (typeof input.evidence !== "string" || !input.evidence.trim())) return { error: "evidence must be an attempt id" };
+    // The SDK may drop optional nulls; the session/process log retains them. Both mean absent.
+    const evidence = input.evidence ?? undefined;
+    const permissionResource = input.permissionResource ?? undefined;
+    if (evidence !== undefined && (typeof evidence !== "string" || !evidence.trim())) return { error: "evidence must be an attempt id" };
+    if (permissionResource !== undefined && (disposition !== "open" || !isPermissionResource(permissionResource))) {
+        return { error: "permissionResource requires open and a supported permission resource" };
+    }
     return { request: { disposition: disposition as IncidentDisposition, targets: [...targets] as string[], reason: input.reason.trim(),
-        ...(typeof input.evidence === "string" ? { evidence: input.evidence.trim() } : {}) } };
+        ...(permissionResource !== undefined ? { permissionResource: permissionResource as PermissionResource } : {}),
+        ...(typeof evidence === "string" ? { evidence: evidence.trim() } : {}) } };
 }
 
 function attemptRef(model: IncidentModel, ref: string): Finished | undefined {
@@ -146,6 +158,9 @@ function attemptLabel(attempt: Finished): string {
  * recovery additionally needs the same operation. Everything else fails closed.
  */
 export function resolveDisposition(model: IncidentModel, state: FailureState, request: DispositionRequest, requestId: string): { event?: FailureEvent; error?: string } {
+    if (request.permissionResource !== undefined && (!model.structuredIntent || request.disposition !== "open" || !isPermissionResource(request.permissionResource))) {
+        return { error: "Permission reports require trusted worker open dispositions" };
+    }
     const incidents: string[] = [];
     for (const target of request.targets) {
         let id: string | undefined;
@@ -189,6 +204,11 @@ export function resolveDisposition(model: IncidentModel, state: FailureState, re
     }
     const event: FailureEvent = { id: `disposition:${requestId}`, operation: "incident-disposition", kind: "disposition",
         disposition: request.disposition, incidents, reason: request.reason, ...(evidence ? { evidence } : {}) };
+    if (request.permissionResource !== undefined) {
+        event.permissionBlockers = incidents.map(id => ({ version: 1, kind: "permission-blocker", context: "worker",
+            resource: request.permissionResource!, basis: "agent-reported", operation: findIncident(state, id)?.operation ?? "unknown",
+            remoteOutcome: "unknown", incidentId: id, ...(model.runId ? { runId: model.runId } : {}) } satisfies PermissionBlocker));
+    }
     const error = validateDisposition(state, event);
     return error ? { error } : { event };
 }

package/index.ts

@@ -13,7 +13,7 @@
 
 import { execSync } from "node:child_process";
 import { writeFileSync, mkdirSync, renameSync, statSync } from "node:fs";
-import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-agent";
+import type { ExtensionAPI, ExtensionCommandContext, ExtensionContext } from "@earendil-works/pi-coding-agent";
 import * as PiCodingAgent from "@earendil-works/pi-coding-agent";
 import * as PiTui from "@earendil-works/pi-tui";
 import { CustomEditor } from "@earendil-works/pi-coding-agent";
@@ -2283,6 +2283,27 @@ export default function (pi: ExtensionAPI) {
         },
         enrich: createLaunchEnricher(),
     });
+    const openSettings = async (ctx: ExtensionCommandContext): Promise<void> => {
+        if (ctx.mode === "tui" && ctx.hasUI && typeof ctx.ui.custom === "function") {
+            await ctx.ui.custom<void>((tui, theme, _kb, done) => createSubagentSettingsPage(theme, {
+                get: () => ({
+                    mode: activeDelegationMode(), maxConcurrent: activeConcurrencyCap(),
+                    modeSource: delegationOverride ? "session" : "config",
+                    capSource: capOverride !== undefined ? "session" : "config",
+                    defaultMode: normalizeDelegationMode(loadConfig().delegationMode),
+                    defaultCap: normalizeConcurrencyCap(loadConfig().maxConcurrent),
+                }),
+                changeMode, changeCap, save: saveSettings, reset: resetSettings,
+            }, () => { refreshBackgroundWorkNavigator(ctx); tui.requestRender(); }, () => done()));
+        } else {
+            ctx.ui.notify(`Delegation mode: ${activeDelegationMode()}. Concurrent subagents: ${activeConcurrencyCap()}. Use /subagents mode <mode> or /subagents cap <number>; /subagents save persists both.`, "info");
+        }
+    };
+    const contribution = { id: "subagents", label: "Subagents", command: "/subagents settings", open: openSettings };
+    const registerSettings = () => pi.events?.emit("harness-settings:register", contribution);
+    const unsubscribeSettingsRequest = pi.events?.on?.("harness-settings:request", registerSettings);
+    registerSettings();
+
     if (typeof pi.registerCommand === "function") {
         agentOperations.registerCommands(pi);
         pi.registerCommand("subagents", {
@@ -2291,20 +2312,7 @@ export default function (pi: ExtensionAPI) {
             async handler(args, ctx) {
                 const tokens = args.trim().split(/\s+/).filter(Boolean);
                 if (tokens.length === 0 || (tokens.length === 1 && tokens[0] === "settings")) {
-                    if (ctx.mode === "tui" && ctx.hasUI && typeof ctx.ui.custom === "function") {
-                        await ctx.ui.custom<void>((tui, theme, _kb, done) => createSubagentSettingsPage(theme, {
-                            get: () => ({
-                                mode: activeDelegationMode(), maxConcurrent: activeConcurrencyCap(),
-                                modeSource: delegationOverride ? "session" : "config",
-                                capSource: capOverride !== undefined ? "session" : "config",
-                                defaultMode: normalizeDelegationMode(loadConfig().delegationMode),
-                                defaultCap: normalizeConcurrencyCap(loadConfig().maxConcurrent),
-                            }),
-                            changeMode, changeCap, save: saveSettings, reset: resetSettings,
-                        }, () => { refreshBackgroundWorkNavigator(ctx); tui.requestRender(); }, () => done()));
-                    } else {
-                        ctx.ui.notify(`Delegation mode: ${activeDelegationMode()}. Concurrent subagents: ${activeConcurrencyCap()}. Use /subagents mode <mode> or /subagents cap <number>; /subagents save persists both.`, "info");
-                    }
+                    await openSettings(ctx);
                     return;
                 }
                 const requested = tokens[1];
@@ -2346,18 +2354,23 @@ export default function (pi: ExtensionAPI) {
     // ---- live-status lifecycle -----------------------------------------
     pi.on("agent_start", async (_event, ctx) => {
         setCallbackBatchContext(pi, ctx);
+        getCallbackBatcher(pi).setForegroundRunning(true);
         uiCtx = ctx;
         mainAgentStartedAt = Date.now();
         mainAgentTools.clear();
         refreshBackgroundWorkNavigator(ctx);
     });
 
+    pi.on("agent_end", () => { getCallbackBatcher(pi).setForegroundRunning(false); });
+
     pi.on("tool_execution_start", async (event, ctx) => {
+        getCallbackBatcher(pi).toolStarted(event.toolCallId);
         mainAgentTools.set(event.toolCallId, event.toolName);
         refreshBackgroundWorkNavigator(ctx);
     });
 
     pi.on("tool_execution_end", async (event, ctx) => {
+        await getCallbackBatcher(pi).toolEnded(event.toolCallId);
         mainAgentTools.delete(event.toolCallId);
         refreshBackgroundWorkNavigator(ctx);
     });
@@ -2378,6 +2391,7 @@ export default function (pi: ExtensionAPI) {
     });
 
     pi.on("agent_settled", async (_event, ctx) => {
+        getCallbackBatcher(pi).setForegroundRunning(false);
         setCallbackBatchContext(pi, ctx);
         mainAgentStartedAt = undefined;
         mainAgentTools.clear();
@@ -2385,6 +2399,7 @@ export default function (pi: ExtensionAPI) {
     });
 
     pi.on("session_tree", async (_event, ctx) => {
+        setCallbackBatchContext(pi, ctx);
         agentSessionSettings.restore(ctx);
         try { noteCatalogHost(catalogHostFrom(ctx)); } catch { /* catalog inspection stays undecided */ }
         restoreDelegationMode(ctx);
@@ -2446,6 +2461,7 @@ export default function (pi: ExtensionAPI) {
 
     // Tear down the timer and clear the widget when the session ends.
     pi.on("session_shutdown", async (_event, ctx) => {
+        if (typeof unsubscribeSettingsRequest === "function") unsubscribeSettingsRequest();
         // The foreground Pi session owns this work. Stop current-session
         // running/orphaned children before dropping the session origin, so they
         // cannot become live process groups with no coordinator.

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "pi-better-subagents",
-  "version": "0.14.0",
+  "version": "0.15.0",
   "description": "Pi extension for detached, sandboxed subagent runs that keep the foreground session free.",
   "license": "MIT",
   "type": "module",

package/shared-callback-batcher.ts

@@ -1,6 +1,126 @@
 // Generated from packages/callback-batcher/index.ts. Do not edit directly.
+import { closeSync, mkdirSync, openSync, readFileSync, renameSync, rmSync, writeFileSync } from "node:fs";
+import { randomUUID } from "node:crypto";
+import { join } from "node:path";
+import { getAgentDir } from "@earendil-works/pi-coding-agent";
+
 export type CallbackSource = "subagent" | "background-task";
 export type CallbackDetailTool = "subagent_result" | "bg_task_status";
+export type CallbackDeliveryMode = "hold" | "steer";
+
+export interface CallbackSettingsContext {
+  sessionManager?: { getBranch?(): readonly unknown[]; getSessionId?(): string };
+  ui?: { notify(message: string, type: "error"): void };
+}
+
+export interface CallbackSettingsSeams {
+  agentDir?: () => string;
+}
+
+export interface CallbackSettings {
+  mode: CallbackDeliveryMode;
+  source: "session" | "default";
+}
+
+export const CALLBACK_SETTINGS_ENTRY = "pi-better-callback-settings";
+const CALLBACK_PREFERENCES_FILE = "pi-better-callback-preferences.json";
+
+function isDeliveryMode(mode: unknown): mode is CallbackDeliveryMode {
+  return mode === "hold" || mode === "steer";
+}
+
+function validCallbackPreferences(value: unknown): value is { version: 1; mode: CallbackDeliveryMode } {
+  if (!value || typeof value !== "object" || Array.isArray(value)) return false;
+  const data = value as Record<string, unknown>;
+  return Object.keys(data).length === 2 && Object.hasOwn(data, "version") && Object.hasOwn(data, "mode")
+    && data.version === 1 && isDeliveryMode(data.mode);
+}
+
+function callbackPreferencesPath(seams: CallbackSettingsSeams): string {
+  return join((seams.agentDir ?? getAgentDir)(), "extensions", CALLBACK_PREFERENCES_FILE);
+}
+
+/** Branch history is authoritative; abandoned branches and invalid entries are ignored. */
+export function getCallbackSettings(
+  ctx: CallbackSettingsContext,
+  seams: CallbackSettingsSeams = {},
+): CallbackSettings {
+  const branch = ctx.sessionManager?.getBranch?.() ?? [];
+  for (let i = branch.length - 1; i >= 0; i--) {
+    const entry = branch[i] as { type?: unknown; customType?: unknown; data?: unknown } | null;
+    if (entry?.type === "custom" && entry.customType === CALLBACK_SETTINGS_ENTRY && validCallbackPreferences(entry.data)) {
+      return { mode: entry.data.mode, source: "session" };
+    }
+  }
+  return getSessionDefault(ctx, seams);
+}
+
+function getSessionDefault(ctx: CallbackSettingsContext, seams: CallbackSettingsSeams): CallbackSettings {
+  const sessionId = ctx.sessionManager?.getSessionId?.();
+  const defaults = globalState().sessionDefaults ??= new WeakMap<object, { sessionId?: string; mode: CallbackDeliveryMode }>();
+  const initial = ctx.sessionManager ? defaults.get(ctx.sessionManager) : undefined;
+  if (initial && initial.sessionId === sessionId) return { mode: initial.mode, source: "default" };
+  const snapshot = (mode: CallbackDeliveryMode): CallbackSettings => {
+    if (ctx.sessionManager) defaults.set(ctx.sessionManager, { sessionId, mode });
+    return { mode, source: "default" };
+  };
+  const path = callbackPreferencesPath(seams);
+  let text: string;
+  try {
+    text = readFileSync(path, "utf8");
+  } catch (error) {
+    if ((error as NodeJS.ErrnoException).code === "ENOENT") return snapshot("hold");
+    throw new Error(`Cannot read callback default at ${path}: ${String(error)}`, { cause: error });
+  }
+  try {
+    const data: unknown = JSON.parse(text);
+    if (!validCallbackPreferences(data)) throw new Error('Expected { version: 1, mode: "hold" | "steer" }');
+    return snapshot(data.mode);
+  } catch (error) {
+    throw new Error(`Invalid callback default at ${path}: ${String(error)}`, { cause: error });
+  }
+}
+
+/** Saves only the future-session default; never changes the active branch or batcher. */
+export function saveCallbackDefault(mode: CallbackDeliveryMode, seams: CallbackSettingsSeams = {}): void {
+  if (!isDeliveryMode(mode)) throw new Error("Invalid callback delivery mode. Nothing saved.");
+  const dir = join((seams.agentDir ?? getAgentDir)(), "extensions");
+  const path = join(dir, CALLBACK_PREFERENCES_FILE);
+  const tmp = join(dir, `.${CALLBACK_PREFERENCES_FILE}.${randomUUID()}.tmp`);
+  let created = false;
+  try {
+    mkdirSync(dir, { recursive: true });
+    const fd = openSync(tmp, "wx", 0o600);
+    created = true;
+    try { writeFileSync(fd, `${JSON.stringify({ version: 1, mode })}\n`); }
+    finally { closeSync(fd); }
+    renameSync(tmp, path);
+  } catch (error) {
+    throw new Error(`Cannot save callback default at ${path}: ${String(error)}`, { cause: error });
+  } finally {
+    if (created) rmSync(tmp, { force: true });
+  }
+}
+
+/** Persist first so a failed append cannot become a transient delivery-mode change. */
+export function changeCallbackSetting(
+  host: CallbackBatchHost & { appendEntry(customType: string, data: unknown): void },
+  ctx: CallbackSettingsContext & { isIdle(): boolean },
+  mode: CallbackDeliveryMode,
+): void {
+  if (!isDeliveryMode(mode)) throw new Error("Invalid callback delivery mode. Nothing changed.");
+  const data = { version: 1, mode };
+  try { host.appendEntry(CALLBACK_SETTINGS_ENTRY, data); }
+  catch (error) {
+    // Pi advances its in-memory branch before disk persistence. Invalidate
+    // our payload so that failed entry cannot become a restored setting.
+    data.version = 0;
+    throw error;
+  }
+  const batcher = bindCallbackBatcher(host, ctx);
+  batcher.setWhileBusy(mode);
+  batcher.setAvailability(() => ctx.isIdle());
+}
 
 export interface CallbackBatchHost {
   /** Older Pi shares this object; newer Pi supplies per-extension wrappers. */
@@ -70,6 +190,8 @@ export interface CallbackBatcherOptions {
   retryMs?: number;
   /** Ordinary callbacks stay here, not in Pi's follow-up queue, while unavailable. */
   isAvailable?: () => boolean;
+  /** Internal initial mode. Session/default settings are restored with context. */
+  whileBusy?: CallbackDeliveryMode;
   /** UTF-8 byte cap for one sendMessage payload. Defaults to 2 KiB. */
   maxBytes?: number;
 }
@@ -81,6 +203,14 @@ export interface CallbackBatcher {
   setHost(host: CallbackBatchHost): void;
   /** Update the session predicate and schedule an asynchronous drain when available. */
   setAvailability(isAvailable: () => boolean): void;
+  /** Change delivery immediately without clearing pending callbacks or handoff receipts. */
+  setWhileBusy(mode: CallbackDeliveryMode): void;
+  /** Non-idle context operations are not foreground agent runs. */
+  setForegroundRunning(running: boolean): void;
+  /** Shared IDs make duplicate notifications from both extensions idempotent. */
+  toolStarted(toolCallId: string): void;
+  /** Await the final active tool's handoff before Pi checks its steering queue. */
+  toolEnded(toolCallId: string): Promise<boolean>;
   deliverUrgent(event: UrgentCallbackEvent): boolean | Promise<boolean>;
   cancel(): void;
   pendingCount(): number;
@@ -104,6 +234,7 @@ interface PendingEvent {
 interface SharedCallbackBatcherState {
   byHost: WeakMap<object, CallbackBatcher>;
   bySession?: WeakMap<object, CallbackBatcher>;
+  sessionDefaults?: WeakMap<object, { sessionId?: string; mode: CallbackDeliveryMode }>;
 }
 
 const GLOBAL_STATE_KEY = Symbol.for("@1aboveio/pi-better-harness/callback-batcher");
@@ -131,7 +262,6 @@ const RETRIEVAL_FOOTER =
 
 export const CALLBACK_BATCH_WINDOW_ENV = "PI_BETTER_CALLBACK_BATCH_MS";
 export const DEFAULT_CALLBACK_BATCH_WINDOW_MS = DEFAULT_WINDOW_MS;
-
 export function resolveCallbackBatchWindowMs(
   value: unknown = process.env[CALLBACK_BATCH_WINDOW_ENV],
 ): number {
@@ -436,26 +566,31 @@ export function createCallbackBatcher(
 ): CallbackBatcher {
   const windowMs = options.windowMs ?? resolveCallbackBatchWindowMs();
   const retryMs = Math.max(0, options.retryMs ?? DEFAULT_RETRY_MS);
+  let whileBusy = options.whileBusy ?? "hold";
   const maxBytes = callbackBatchBudget(options.maxBytes);
   const pending = new Map<string, PendingEvent>();
   const inFlight = new Set<string>();
   const urgentInFlight = new Set<string>();
   const handedOff = new Map<string, number>();
+  const activeTools = new Set<string>();
   let sequence = 0;
   let timer: ReturnType<typeof setTimeout> | undefined;
   let flushPromise: Promise<boolean> | undefined;
   let isAvailable = options.isAvailable ?? (() => true);
   let gated = options.isAvailable !== undefined;
   let awaitingRun = false;
+  let cancelled = false;
+  let foregroundRunning = false;
 
-  const available = (): boolean => {
+  const deliveryMode = (): "followUp" | "steer" | undefined => {
+    if (cancelled) return undefined;
     try {
       if (!isAvailable()) {
         awaitingRun = false;
-        return false;
+        return whileBusy === "steer" && foregroundRunning && activeTools.size === 0 ? "steer" : undefined;
       }
-      return !awaitingRun;
-    } catch { return false; }
+      return !awaitingRun ? "followUp" : undefined;
+    } catch { return undefined; }
   };
 
   const cancelTimer = (): void => {
@@ -466,7 +601,7 @@ export function createCallbackBatcher(
   const schedule = (delayMs: number): void => {
     if (timer || pending.size === 0) return;
     // Some idle transitions (manual compaction) have no agent_settled event.
-    const delay = available() ? delayMs : retryMs;
+    const delay = deliveryMode() ? delayMs : retryMs;
     timer = setTimeout(() => {
       timer = undefined;
       void api.flush();
@@ -485,7 +620,8 @@ export function createCallbackBatcher(
 
   const performFlush = async (): Promise<boolean> => {
     cancelTimer();
-    if (!available()) {
+    const mode = deliveryMode();
+    if (!mode) {
       schedule(retryMs);
       return false;
     }
@@ -539,14 +675,16 @@ export function createCallbackBatcher(
     try {
       // Pi may defer the requested run while settled handlers finish, leaving
       // isIdle() true. Permit no second handoff until a run starts.
-      awaitingRun = gated;
+      awaitingRun = mode === "followUp" && gated;
       await host.sendMessage(
         {
           customType: "background-completion-batch",
           content: packed.text,
           display: true,
         },
-        { deliverAs: "followUp", triggerTurn: true },
+        // Newer Pi requires triggerTurn for steering too; false only appends
+        // a custom message at the turn boundary without continuing the agent.
+        { deliverAs: mode, triggerTurn: true },
       );
     } catch {
       awaitingRun = false;
@@ -632,15 +770,42 @@ export function createCallbackBatcher(
     setAvailability(check) {
       gated = true;
       isAvailable = check;
-      available();
+      cancelled = false;
+      deliveryMode();
       cancelTimer();
       schedule(windowMs);
     },
+    setWhileBusy(mode) {
+      if (!isDeliveryMode(mode)) throw new Error("Invalid callback delivery mode. Nothing changed.");
+      whileBusy = mode;
+      cancelTimer();
+      schedule(windowMs);
+    },
+    setForegroundRunning(running) {
+      foregroundRunning = running;
+      cancelTimer();
+      schedule(windowMs);
+    },
+    toolStarted(toolCallId) {
+      activeTools.add(toolCallId);
+    },
+    async toolEnded(toolCallId) {
+      const ended = activeTools.delete(toolCallId);
+      if (activeTools.size > 0 || whileBusy !== "steer" || cancelled) return false;
+      if (!ended) return flushPromise ?? false;
+      // A timer may still be completing a deferred flush. Recheck after it so
+      // the tool boundary cannot lose the pending batch to that older attempt.
+      if (flushPromise) await flushPromise;
+      return pending.size > 0 ? api.flush() : true;
+    },
     deliverUrgent,
     cancel() {
       cancelTimer();
       pending.clear();
       awaitingRun = false;
+      activeTools.clear();
+      foregroundRunning = false;
+      cancelled = true;
     },
     pendingCount() {
       return pending.size;
@@ -671,11 +836,39 @@ export function cancelCallbackBatch(host: CallbackBatchHost): void {
   batcher?.cancel();
 }
 
-/** Called on session_start, agent_start and agent_settled by both extensions. */
+/** Restore current branch/default mode without resetting foreground tools or receipts. */
 export function setCallbackBatchContext(
   host: CallbackBatchHost,
-  ctx: { isIdle(): boolean; sessionManager?: object },
+  ctx: CallbackSettingsContext & { isIdle(): boolean },
+  seams: CallbackSettingsSeams = {},
 ): void {
+  const batcher = bindCallbackBatcher(host, ctx);
+  let mode: CallbackDeliveryMode = "hold";
+  try {
+    const settings = getCallbackSettings(ctx, seams);
+    if (settings.source === "session") {
+      // Capture the initial default even when an override currently hides it,
+      // so navigating to an unconfigured branch cannot inherit a later save.
+      try { getSessionDefault(ctx, seams); }
+      catch {
+        if (ctx.sessionManager) {
+          const defaults = globalState().sessionDefaults ??= new WeakMap();
+          defaults.set(ctx.sessionManager, { sessionId: ctx.sessionManager.getSessionId?.(), mode: "hold" });
+        }
+      }
+    }
+    mode = settings.mode;
+  } catch (error) {
+    try { ctx.ui?.notify(`Callback delivery restored to hold: ${String(error)}`, "error"); }
+    catch { /* A failing UI must not retain the prior session's delivery mode. */ }
+  }
+  batcher.setWhileBusy(mode);
+  // Schedule rather than await delivery inside agent_settled: its handlers are
+  // still settling the previous run, and sendMessage may start a new one.
+  batcher.setAvailability(() => ctx.isIdle());
+}
+
+function bindCallbackBatcher(host: CallbackBatchHost, ctx: CallbackSettingsContext): CallbackBatcher {
   const state = globalState();
   const bySession = state.bySession ??= new WeakMap<object, CallbackBatcher>();
   const shared = ctx.sessionManager ? bySession.get(ctx.sessionManager) : undefined;
@@ -683,9 +876,7 @@ export function setCallbackBatchContext(
   if (ctx.sessionManager) bySession.set(ctx.sessionManager, batcher);
   state.byHost.set(host.events ?? host, batcher);
   batcher.setHost(host);
-  // Schedule rather than await delivery inside agent_settled: its handlers are
-  // still settling the previous run, and sendMessage may start a new one.
-  batcher.setAvailability(() => ctx.isIdle());
+  return batcher;
 }
 
 function globalState(): SharedCallbackBatcherState {

package/shared-failure-observations.ts

@@ -2,6 +2,7 @@
 import { appendFileSync, closeSync, existsSync, openSync, readFileSync, mkdirSync, statSync } from "node:fs";
 import { dirname } from "node:path";
 import { createHash } from "node:crypto";
+import { isPermissionBlocker, permissionBlockerKey, type PermissionBlocker } from "./shared-permission-blocker.ts";
 
 /**
  * Explicit, append-only incident dispositions (#315).
@@ -30,6 +31,7 @@ export interface FailureEvent {
   disposition?: IncidentDisposition;
   /** Disposition events only: why the incidents are classified this way. */
   reason?: string;
+  permissionBlockers?: PermissionBlocker[];
 }
 export interface DispositionRecord {
   eventId: string;
@@ -38,6 +40,7 @@ export interface DispositionRecord {
   reason: string;
   evidence?: string;
   at: number;
+  permissionBlockers?: PermissionBlocker[];
 }
 export interface FailureObservation {
   id: string;
@@ -56,6 +59,8 @@ export interface FailureObservation {
   resolvedAt?: number;
   /** Latest explicit disposition applied to this incident. */
   disposition?: DispositionRecord;
+  /** Retained even after recovery; active surfaces filter by current actionability. */
+  permissionBlockers?: PermissionBlocker[];
 }
 export interface FailureState {
   version: 1;
@@ -247,12 +252,26 @@ export function validateDisposition(state: FailureState, event: FailureEvent): s
     if (current.status !== "unresolved") return `Incident ${id} is already disposed (${failureLabel(current).toLowerCase()})`;
     if (event.disposition === "open" && current.disposition?.disposition === "open") return `Incident ${id} is already open`;
   }
+  if (event.permissionBlockers !== undefined) {
+    if (event.disposition !== "open" || !Array.isArray(event.permissionBlockers) || !event.permissionBlockers.length ||
+      event.permissionBlockers.length > 20 || !event.permissionBlockers.every(isPermissionBlocker)) return "Invalid permission blocker metadata";
+    const bound = new Set<string>();
+    for (const blocker of event.permissionBlockers) {
+      const current = blocker.incidentId ? findIncident(state, blocker.incidentId) : undefined;
+      if (!current || !ids.includes(current.id) || current.category !== "tool" || current.operation !== blocker.operation ||
+        blocker.context !== "worker" || blocker.basis !== "agent-reported" || blocker.remoteOutcome !== "unknown" ||
+        blocker.policySnapshotId !== undefined || bound.has(current.id)) return "Permission blocker must name an unresolved failed worker attempt";
+      bound.add(current.id);
+    }
+    if (bound.size !== ids.length) return "Permission blocker metadata must cover every target";
+  }
   return undefined;
 }
 
 /** Pure transition. Lifecycle is deliberately not an input or an output. */
 export function reduceFailure(state: FailureState, event: FailureEvent, observedAt: number): FailureState {
   if (state.seen.includes(event.id)) return state;
+  if (event.permissionBlockers !== undefined && event.kind !== "disposition") return state;
   // Invalid dispositions fail closed: nothing changes and the id is not consumed.
   if (event.kind === "disposition" && validateDisposition(state, event)) return state;
   const next: FailureState = { ...state, seen: [...state.seen, event.id],
@@ -265,11 +284,13 @@ export function reduceFailure(state: FailureState, event: FailureEvent, observed
     const at = event.at ?? observedAt;
     const record: DispositionRecord = { eventId: event.id, disposition: event.disposition!, incidents: [...event.incidents!],
       reason: text(event.reason, ""), ...(event.evidence ? { evidence: text(event.evidence, "") } : {}), at };
+    if (event.permissionBlockers) record.permissionBlockers = event.permissionBlockers.map(x => ({ ...x }));
     for (const id of record.incidents) {
       const [key, current] = Object.entries(next.observations).find(([, item]) => item.id === id)!;
       const status = record.disposition === "recovered" ? "resolved" : record.disposition === "superseded" ? "superseded"
         : record.disposition === "expected" ? "expected" : current.status;
       next.observations[key] = { ...current, status, disposition: record,
+        ...(record.permissionBlockers ? { permissionBlockers: record.permissionBlockers.filter(x => x.incidentId === id) } : {}),
         ...(status === "resolved" || status === "superseded" ? { resolvedAt: at } : {}) };
       if (status === "resolved" || status === "superseded") next.resolved = { ...next.resolved, [id]: at };
     }
@@ -298,6 +319,7 @@ export function reduceFailure(state: FailureState, event: FailureEvent, observed
     lastObservedAt: observedAt, lastSequence: next.seen.length, at: event.at,
     count: active ? previous.count + 1 : 1,
     ...(active && previous.disposition ? { disposition: previous.disposition } : {}),
+    ...(active && previous.permissionBlockers ? { permissionBlockers: previous.permissionBlockers } : {}),
   };
   return next;
 }
@@ -315,6 +337,16 @@ export function activeFailures(state: FailureState): FailureObservation[] {
 export function actionableFailures(state: FailureState): FailureObservation[] {
   return activeFailures(state).filter(needsAction);
 }
+/** Only currently actionable reports, not closed history or arbitrary tool-result claims. */
+export function actionablePermissionBlockers(state: FailureState): PermissionBlocker[] {
+  const unique = new Map<string, PermissionBlocker>();
+  for (const incident of actionableFailures(state)) {
+    for (const blocker of incident.permissionBlockers ?? []) {
+      if (isPermissionBlocker(blocker)) unique.set(permissionBlockerKey(blocker), { ...blocker });
+    }
+  }
+  return [...unique.values()];
+}
 function closedOrder(a: FailureObservation, b: FailureObservation): number {
   return (b.lastSequence ?? 0) - (a.lastSequence ?? 0) || b.lastObservedAt - a.lastObservedAt;
 }
@@ -449,12 +481,20 @@ export function shortEvidence(evidence: string): string {
 function failureRow(x: FailureObservation, detail: IncidentDetail = "compact"): string {
   const time = x.at === undefined ? `observed ${new Date(x.firstObservedAt).toISOString()}` : new Date(x.at).toISOString();
   const compact = detail !== "full";
+  const permission = Boolean(x.permissionBlockers?.length);
+  let annotation = "";
+  if (x.permissionBlockers?.length) {
+    const resources = x.permissionBlockers.map(b => b.resource).join(", ");
+    annotation = ` · Worker permission blocker: ${resources} (agent-reported; remote outcome unknown; not a foreground denial). No remote authorization or successful validation inferred.`;
+    if (needsAction(x)) annotation += " Changed worker settings require a fresh launch.";
+  }
   // Full rows (raw evidence) show the journal summary verbatim; compact rows unwrap and cap it.
-  const summary = compact ? capUtf8(dropLoneSurrogates(unwrapToolResultText(x.summary)).replace(/\s+/g, " ").trim(), COMPACT_EXCERPT_BYTES) : x.summary;
+  const summary = compact && permission ? "Permission access failure" : compact ? capUtf8(dropLoneSurrogates(unwrapToolResultText(x.summary)).replace(/\s+/g, " ").trim(), COMPACT_EXCERPT_BYTES) : x.summary;
   const reason = x.disposition ? text(x.disposition.reason, "") : "";
-  const disposition = x.disposition ? ` · ${x.disposition.disposition}: ${compact ? capUtf8(reason, COMPACT_EXCERPT_BYTES) : reason}` : "";
-  const evidence = x.evidence ? text(x.evidence, "") : "";
-  return `${failureLabel(x)} · ${time} · ${summary}${x.count > 1 ? ` (${x.count} occurrences)` : ""}${disposition}${evidence ? ` · evidence: ${compact ? shortEvidence(evidence) : evidence}` : ""}`;
+  const disposition = x.disposition ? ` · ${x.disposition.disposition}${compact && permission ? "" : `: ${compact ? capUtf8(reason, COMPACT_EXCERPT_BYTES) : reason}`}` : "";
+  const evidence = x.evidence && !(compact && permission) ? text(x.evidence, "") : "";
+  const proof = !compact && permission && x.disposition?.evidence ? ` · disposition evidence: ${text(x.disposition.evidence, "")}` : "";
+  return `${failureLabel(x)} · ${time} · ${summary}${x.count > 1 ? ` (${x.count} occurrences)` : ""}${disposition}${evidence ? ` · evidence: ${compact ? shortEvidence(evidence) : evidence}` : ""}${proof}${annotation}`;
 }
 
 /**
@@ -573,13 +613,15 @@ export interface FailureIncidentPage {
  */
 export function failureRevision(state: FailureState): string {
   return failureIdentity(Object.values(state.observations)
-    .map((item) => [item.id, item.status, item.category, item.count, item.lastSequence ?? 0, item.summary, item.evidence ?? ""])
+    .map((item) => [item.id, item.status, item.category, item.count, item.lastSequence ?? 0, item.summary, item.evidence ?? "",
+      ...(item.permissionBlockers ? [item.permissionBlockers.map(permissionBlockerKey).sort()] : [])])
     .sort((a, b) => String(a[0]).localeCompare(String(b[0]))));
 }
 
 /** Digest of the rows a cursor pages. Scope and detail are part of it: the same incidents render different bytes. */
 function incidentRevision(state: FailureState, scope: IncidentScope = "actionable", detail: IncidentDetail = "compact"): string {
-  return failureIdentity(scope, detail, scopedFailures(state, scope).map((item) => [item.id, item.status, item.count, item.summary, item.evidence ?? ""])).slice(0, 16);
+  return failureIdentity(scope, detail, scopedFailures(state, scope).map((item) => [item.id, item.status, item.count, item.summary, item.evidence ?? "",
+    ...(item.permissionBlockers ? [item.permissionBlockers.map(permissionBlockerKey).sort()] : [])])).slice(0, 16);
 }
 
 /** Cursors carry a digest of their resource/scope, not the scope text. */
@@ -999,6 +1041,8 @@ function readStoredState(path: string): FailureState {
           !Number.isFinite(row.observedAt) || Math.abs(row.observedAt) > 8.64e15 ||
           (e.at !== undefined && (!Number.isFinite(e.at) || Math.abs(e.at) > 8.64e15)) ||
           (e.incidents !== undefined && (!Array.isArray(e.incidents) || !e.incidents.every((id: unknown) => typeof id === "string"))) ||
+          (e.permissionBlockers !== undefined && (e.kind !== "disposition" ||
+            (!state.seen.includes(e.id) && validateDisposition(state, e) !== undefined))) ||
           [e.summary, e.category, e.evidence, e.reason].some((v) => v !== undefined && typeof v !== "string")) throw new Error("invalid record");
       state = reduceFailure(state, e, row.observedAt);
     } catch { state = storageProblem(state, "Failure journal contains unreadable records; observations may be incomplete"); }
@@ -1021,6 +1065,7 @@ export function observeFailures(path: string, events: readonly FailureEvent[], n
     }
     // Rejected dispositions are never journaled or marked seen.
     if (raw.kind === "disposition" && validateDisposition(state, raw)) continue;
+    if (raw.permissionBlockers !== undefined && raw.kind !== "disposition") continue;
     const event = { ...raw, ...(raw.summary ? { summary: text(raw.summary, "") } : {}),
       ...(raw.evidence ? { evidence: text(raw.evidence, "") } : {}), ...(raw.reason ? { reason: text(raw.reason, "") } : {}) };
     try {

package/shared-permission-blocker.ts

@@ -0,0 +1,48 @@
+// Generated from packages/failure-observations/permission-blocker.ts. Do not edit directly.
+/** Privacy-safe wire contract. Context and identity must be assigned by a trusted adapter. */
+export const PERMISSION_RESOURCES = Object.freeze([
+  "credential-files", "process-inspection", "project-files", "outside-project-files",
+  "runtime-control-files", "command-execution", "network-access", "unknown",
+] as const);
+export type PermissionResource = typeof PERMISSION_RESOURCES[number];
+export interface PermissionBlocker {
+  version: 1;
+  kind: "permission-blocker";
+  context: "worker" | "foreground";
+  resource: PermissionResource;
+  basis: "policy-refusal" | "os-permission-error" | "agent-reported";
+  /** A stable hash or bounded logical identifier, never a command or path. */
+  operation: string;
+  remoteOutcome: "unknown" | "not-started";
+  incidentId?: string;
+  runId?: string;
+  policySnapshotId?: string;
+}
+const FIELDS = new Set(["version", "kind", "context", "resource", "basis", "operation", "remoteOutcome", "incidentId", "runId", "policySnapshotId"]);
+const LOGICAL_ID = /^[A-Za-z0-9][A-Za-z0-9._:-]{0,199}$/;
+// Responses SDK references join call_id and item_id with "|"; these are opaque evidence ids,
+// not caller-chosen logical operations. Keep the same bound and exclude paths and prose.
+const INCIDENT_REFERENCE = /^[A-Za-z0-9][A-Za-z0-9._:|\-]{0,199}$/;
+export function isPermissionResource(value: unknown): value is PermissionResource {
+  return typeof value === "string" && (PERMISSION_RESOURCES as readonly string[]).includes(value);
+}
+/** Pure and strict: no coercion, unknown fields, paths, argv, or free-form prose. */
+export function isPermissionBlocker(value: unknown): value is PermissionBlocker {
+  if (!value || typeof value !== "object" || Array.isArray(value)) return false;
+  const prototype = Object.getPrototypeOf(value);
+  if (prototype !== Object.prototype && prototype !== null) return false;
+  if (Reflect.ownKeys(value).some(key => typeof key !== "string" || !FIELDS.has(key))) return false;
+  const v = value as Record<string, unknown>;
+  if (v.version !== 1 || v.kind !== "permission-blocker" || !["worker", "foreground"].includes(v.context as string) ||
+    !isPermissionResource(v.resource) || !["policy-refusal", "os-permission-error", "agent-reported"].includes(v.basis as string) ||
+    !["unknown", "not-started"].includes(v.remoteOutcome as string) ||
+    typeof v.operation !== "string" || v.operation.length > 64 || !LOGICAL_ID.test(v.operation)) return false;
+  if (Object.hasOwn(v, "incidentId") && (typeof v.incidentId !== "string" || !INCIDENT_REFERENCE.test(v.incidentId))) return false;
+  return ["runId", "policySnapshotId"].every(key => !Object.hasOwn(v, key) ||
+    (typeof v[key] === "string" && LOGICAL_ID.test(v[key] as string)));
+}
+/** Order-independent identity without I/O or a platform-specific crypto dependency. */
+export function permissionBlockerKey(blocker: PermissionBlocker): string {
+  return JSON.stringify([blocker.version, blocker.kind, blocker.context, blocker.resource, blocker.basis,
+    blocker.operation, blocker.remoteOutcome, blocker.incidentId ?? null, blocker.runId ?? null, blocker.policySnapshotId ?? null]);
+}

package/shared-task-sandbox.ts

@@ -118,7 +118,8 @@ export function createTaskBashOperations(
             const { trackDetachedChildPid, untrackDetachedChildPid } = await import(pathToFileURL(join(sdkUtils, "shell.js")).href);
             if (shell.commandTransport === "stdin") throw new Error("Sandbox: this shell cannot be confined by the available backend.");
             const scratch = plan.policy.runtimeWrite?.[0];
-            const taskEnv = { ...process.env, ...options.env,
+            // Git inventory otherwise creates optional locks in an outside gitdir it cannot unlink (#419).
+            const taskEnv = { GIT_OPTIONAL_LOCKS: "0", ...process.env, ...options.env,
                 ...(scratch ? { TMPDIR: scratch, TMP: scratch, TEMP: scratch } : {}) };
             if (options.signal?.aborted) throw new Error("aborted");
             if (options.timeout !== undefined && (!Number.isFinite(options.timeout) || options.timeout <= 0 || options.timeout * 1000 > 2147483647)) {

package/shared-task-tools.ts

@@ -4,7 +4,7 @@
  *
  * - Guarded tools are harness adapters that go through the task's guarded file
  *   operations, so the file rules govern them. Today: apply_patch.
- * - Trusted tools are third-party tools admitted by name AND owning package.
+ * - Trusted tools are extension tools admitted by name AND owning package.
  *   They run in the child Pi process, outside the file rules; a human ticks each
  *   one in /sandbox. Network tools are refused while Network is Off.
  */
@@ -36,15 +36,17 @@ export function defaultSubagentTools(): SubagentToolSettings {
 
 /** Tools that reach the network; refused while the profile's Network access is Off. */
 export function isNetworkTool(name: string): boolean {
-    return ["web_fetch", "web_search", "firecrawl_scrape", "firecrawl_extract", "mcp", "mcpScript", "remote_bash"].includes(name) ||
+    return ["web_fetch", "web_search", "firecrawl_scrape", "firecrawl_extract", "mcp", "mcpScript", ...SSH_TOOL_NAMES].includes(name) ||
         name.startsWith("mcp__");
 }
 
 /** Names never offered as trusted: builtins and the guarded adapters. */
 export const RESERVED_TOOL_NAMES = Object.freeze(["read", "write", "edit", "bash", "grep", "find", "ls", "powershell", "apply_patch", "process_list"]);
 
-/** Packages that are this harness: their tools are never offered as trusted tools. */
+/** Harness tools stay excluded except for the explicitly supported SSH tools. */
 const HARNESS_PACKAGES = /(^|[/:])pi-better-[a-z-]+$/;
+const SSH_PACKAGE = /(^|[/:])pi-better-ssh$/;
+const SSH_TOOL_NAMES = ["remote_bash", "ssh_profile", "ssh_mux"];
 
 /** Strict decoding: a malformed tool list must not silently admit anything. */
 export function parseSubagentTools(value: unknown): SubagentToolSettings {
@@ -72,6 +74,8 @@ export interface ToolSource {
 export interface DiscoveredTool {
     name: string;
     package: string;
+    /** Explicit SDK provenance for display only; never a package or admission identity. */
+    source?: "builtin";
     /**
      * What to load and admit: the package root directory, or, for an extension
      * file with no package manifest, that file itself. Loading a manifest-less
@@ -80,7 +84,7 @@ export interface DiscoveredTool {
     root: string;
 }
 
-/** The owning package of a registered tool, or undefined for builtins and inline registrations. */
+/** The owning package of a registered tool, or undefined for synthetic core/inline sources. */
 export function toolPackage(tool: ToolSource): DiscoveredTool | undefined {
     const info = tool.sourceInfo;
     if (!info?.path || info.path.startsWith("<")) return undefined;
@@ -93,14 +97,19 @@ export function toolPackage(tool: ToolSource): DiscoveredTool | undefined {
     return { name: tool.name, package: pkg, root: base };
 }
 
-/** Candidate trusted tools in a running Pi, excluding builtins, guarded names and this harness. */
+/** Candidate trusted tools, excluding core builtins, guarded names and non-SSH harness tools. */
 export function discoverTrustedTools(tools: readonly ToolSource[]): DiscoveredTool[] {
     const found: DiscoveredTool[] = [];
     for (const tool of tools) {
         if (RESERVED_TOOL_NAMES.includes(tool.name)) continue;
         const owner = toolPackage(tool);
-        if (!owner || [owner.package, owner.root].some((id) => HARNESS_PACKAGES.test(id.replace(/\/+$/, "")))) continue;
-        if (!found.some((t) => t.name === owner.name && t.package === owner.package)) found.push(owner);
+        if (!owner) continue;
+        const harnessOwners = [owner.package, owner.root].map((id) => id.replace(/\/+$/, ""))
+            .filter((id) => HARNESS_PACKAGES.test(id));
+        if (harnessOwners.length && !(SSH_TOOL_NAMES.includes(tool.name) && harnessOwners.every((id) => SSH_PACKAGE.test(id)))) continue;
+        if (!found.some((t) => t.name === owner.name && t.package === owner.package)) {
+            found.push(tool.sourceInfo?.source === "builtin" ? { ...owner, source: "builtin" } : owner);
+        }
     }
     return found.sort((a, b) => a.name.localeCompare(b.name) || a.package.localeCompare(b.package));
 }

package/tools.ts

@@ -42,7 +42,10 @@ import {
     unknownRunError,
     type SubagentToolSession,
 } from "./output-payload.ts";
-import { readOutputControls } from "./shared-log-utils.ts";
+import { budgetFor, readOutputControls } from "./shared-log-utils.ts";
+import { readRunFailures } from "./failures.ts";
+import { actionablePermissionBlockers } from "./shared-failure-observations.ts";
+import type { PermissionBlocker } from "./shared-permission-blocker.ts";
 import { describeTiming } from "./timing.ts";
 import {
     extractChildEventFactsFromLog,
@@ -112,6 +115,23 @@ const INCLUDE_DESCRIPTION = "Explicit opt-in facts omitted by default: \"cost\"
 /** pi's tool-result text shape. */
 export const text = (t: string) => ({ content: [{ type: "text" as const, text: t }] });
 
+/** Metadata has its own bounded allowance; it never embeds raw evidence or unbounded history. */
+function permissionDetails(id: string, maxBytes: number): { permissionBlockers?: PermissionBlocker[]; permissionBlockersOmitted?: number } {
+    const all = actionablePermissionBlockers(readRunFailures(id));
+    if (!all.length) return {};
+    const permissionBlockers: PermissionBlocker[] = [];
+    const allowance = Math.min(4096, maxBytes);
+    if (Buffer.byteLength(JSON.stringify({ permissionBlockers, permissionBlockersOmitted: all.length })) > allowance) {
+        throw new Error("Permission blocker metadata cannot fit this budget; retry with a larger max_bytes.");
+    }
+    for (const blocker of all) {
+        if (permissionBlockers.length === 32 || Buffer.byteLength(JSON.stringify({ permissionBlockers: [...permissionBlockers, blocker],
+            permissionBlockersOmitted: all.length - permissionBlockers.length - 1 })) > allowance) break;
+        permissionBlockers.push(blocker);
+    }
+    return { permissionBlockers, ...(all.length > permissionBlockers.length ? { permissionBlockersOmitted: all.length - permissionBlockers.length } : {}) };
+}
+
 const SUBAGENT_RESULT_PREVIEW_LINES = 8;
 
 function resultTextContent(result: unknown): string {
@@ -409,7 +429,9 @@ export function subagentOutputTool(Type: TypeModule, baseSession: SubagentToolSe
             if (access.kind === "denied") return text(access.payload);
             const scopeKey = requestScopeKey(p, session);
             const healthLine = diagnosticLines(access.meta);
-            return text(assembleSubagentOutput(p.id, access.meta, p, healthLine, scopeKey));
+            const result = text(assembleSubagentOutput(p.id, access.meta, p, healthLine, scopeKey));
+            const details = permissionDetails(p.id, budgetFor("log", p.maxBytes));
+            return Object.keys(details).length ? { ...result, details } : result;
         },
     } as ToolDefinition;
 }
@@ -451,20 +473,24 @@ export function subagentResultTool(Type: TypeModule, baseSession: SubagentToolSe
             }
             if (access.kind === "denied") return subagentResultText(access.payload);
             const meta = access.meta;
+            const resultText = (body: string) => {
+                const result = subagentResultText(body);
+                return { ...result, details: { ...result.details, ...permissionDetails(p.id, budgetFor("answer", p.maxBytes)) } };
+            };
             const scopeKey = requestScopeKey(p, session);
             const st = effectiveStatus(meta);
             const healthLine = diagnosticLines(meta);
             if (!isFinalResultStatus(st)) {
                 if (st === "orphaned") {
-                    return subagentResultText(assembleOrphanedResult(p.id, meta, healthLine, p, scopeKey));
+                    return resultText(assembleOrphanedResult(p.id, meta, healthLine, p, scopeKey));
                 }
-                return subagentResultText(assembleRunningResult(p.id, meta, p, scopeKey, describeTiming({ ...meta, status: "running" })?.line));
+                return resultText(assembleRunningResult(p.id, meta, p, scopeKey, describeTiming({ ...meta, status: "running" })?.line));
             }
             const body = buildSubagentResultPayload(p.id, p, healthLine, scopeKey);
             if (body === null) {
-                return subagentResultText(assembleRunningResult(p.id, meta, p, scopeKey, describeTiming({ ...meta, status: "running" })?.line));
+                return resultText(assembleRunningResult(p.id, meta, p, scopeKey, describeTiming({ ...meta, status: "running" })?.line));
             }
-            return subagentResultText(body);
+            return resultText(body);
         },
     } as ToolDefinition;
 }