@arhen/pi-core-subagent 1.3.31 → 1.3.33

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@arhen/pi-core-subagent",
-	"version": "1.3.31",
+	"version": "1.3.33",
 	"type": "module",
 	"description": "pi extension: fast in-process subagents with a dependency-graph scheduler (needs edges gate tasks and carry upstream output into dependent prompts), plus background runs, intercom and agent-to-agent mailbox. Leader defines agents inline.",
 	"license": "MIT",

package/src/format.ts

@@ -196,6 +196,23 @@ export class SubagentsWidget implements Component {
 	}
 }
 /** Blocking-call summary: full text, because the model asked for it. */
+/** Where a write child's edits went: a branch to merge, or straight into the tree. */
+function worktreeLine(task: TaskSnapshot): string {
+	const parts: string[] = [];
+	if (task.branch) {
+		const files = task.changedFiles?.length
+			? ` (${task.changedFiles.length} file(s): ${truncateText(task.changedFiles.join(", "), 160)})`
+			: "";
+		parts.push(`Branch: ${task.branch}${files} — merge with \`git merge --no-ff ${task.branch}\` after review.`);
+	} else if (task.isolation === "in-place") {
+		parts.push(
+			`Applied IN PLACE (no branch) — ${task.isolationReason ?? "worktree unavailable"}. Review the working tree directly.`,
+		);
+	}
+	if (task.worktreeError) parts.push(`Worktree: ${task.worktreeError}`);
+	return parts.length ? `\n${parts.join("\n")}` : "";
+}
+
 export function makeSummary(run: RunSnapshot): string {
 	const succeeded = run.tasks.filter((t) => t.status === "completed").length;
 	const failed = run.tasks.filter((t) => t.status === "failed").length;
@@ -210,7 +227,7 @@ export function makeSummary(run: RunSnapshot): string {
 		// Edges are named so the leader can compare what it delegated against what came back.
 		const edge = task.needs?.length ? ` (${task.id}, needs ${task.needs.join(", ")})` : ` (${task.id})`;
 		lines.push(
-			`\n## ${task.agent}${edge} ${statusIcon(task.status)}${task.error ? `\nError: ${task.error}` : `\n${truncateText(task.finalText || "(no output)")}`}${task.branch ? `\nBranch: ${task.branch}${task.changedFiles?.length ? ` (${task.changedFiles.length} file(s): ${truncateText(task.changedFiles.join(", "), 160)})` : ""} — merge with \`git merge --no-ff ${task.branch}\` after review.` : ""}`,
+			`\n## ${task.agent}${edge} ${statusIcon(task.status)}${task.error ? `\nError: ${task.error}` : `\n${truncateText(task.finalText || "(no output)")}`}${worktreeLine(task)}`,
 		);
 	}
 	// Ceiling on the WHOLE summary — 16 tasks × 24KB would otherwise flood the parent context.

package/src/index.ts

@@ -30,7 +30,7 @@ import {
 	type SubagentParamsShape,
 } from "./schemas.ts";
 import { type RunDetails, type RunSnapshot, TERMINAL } from "./types.ts";
-import { cleanupMerged, reapDeadWorktrees, repoRoot, sweepStale } from "./worktree.ts";
+import { cleanupMerged, ownerAlive, reapDeadWorktrees, repoRoot, sweepStale } from "./worktree.ts";
 
 export default function (pi: ExtensionAPI) {
 	const manager = new SubagentManager(pi);
@@ -126,11 +126,11 @@ export default function (pi: ExtensionAPI) {
 		}
 		for (const root of roots) {
 			try {
-				// Nothing of ours can be live at session start: every registered subagent
-				// worktree is a crash leftover — commit its work, keep the branch, drop the
-				// dir. Then reap merged branches and dirs git no longer tracks.
-				reapDeadWorktrees(root);
-				cleanupMerged(root);
+				// A registered subagent worktree is a crash leftover UNLESS another pi
+				// session still owns it (pid marker) — commit its work, keep the branch,
+				// drop the dir. Then reap merged branches and dirs git no longer tracks.
+				reapDeadWorktrees(root, ownerAlive);
+				cleanupMerged(root, { skipBranches: manager.liveBranches() });
 				sweepStale(root);
 			} catch {
 				/* recovery is best-effort — never block session start */
@@ -188,15 +188,23 @@ export default function (pi: ExtensionAPI) {
 					intercom.push(...awaited.intercom);
 					if (awaited.intercom.some((m) => m.kind === "ask")) break;
 				}
-				const asked = intercom.find((m) => m.kind === "ask");
+				// EVERY ask must surface: siblings that asked in the same wake got no
+				// followUp notice (the park swallowed it), so showing only the first
+				// leaves the rest blocked until their 10-minute timeout.
+				const asks = intercom.filter((m) => m.kind === "ask");
 				const heard = intercom.filter((m) => m.kind !== "ask");
 				const text = [
 					makeSummary(run),
 					heard.length > 0
 						? `\nIntercom while waiting:\n${heard.map((m) => `- [${m.kind}] ${m.agent} (${m.taskId}): ${truncateText(m.text)}`).join("\n")}`
 						: "",
-					asked
-						? `\nA child is waiting for your answer (${asked.agent}, ${asked.taskId}): ${asked.text}\nReply with reply_subagent(runId: "${run.id}", taskId: "${asked.taskId}", message: ...), then await_subagent again for the result.`
+					asks.length > 0
+						? `\n${asks.length} child(ren) waiting for your answer:\n${asks
+								.map(
+									(a) =>
+										`- ${a.agent} (${a.taskId}): ${a.text}\n  reply_subagent(runId: "${run.id}", taskId: "${a.taskId}", message: ...)`,
+								)
+								.join("\n")}\nAnswer each, then await_subagent again for the result.`
 						: "",
 				]
 					.filter(Boolean)
@@ -321,9 +329,12 @@ export default function (pi: ExtensionAPI) {
 				`Run ${run.id} — ${run.status}`,
 				...tasks.map((t) => {
 					const wt = t.branch
-						? `\nBranch: ${t.branch}\n${t.diffStat || "(no changes committed)"}\nMerge after review: \`git merge --no-ff ${t.branch}\``
-						: "";
-					return `\n## ${t.agent} ${statusIcon(t.status)}\nGoal: ${truncateText(t.task, 300)}\n${t.error ? `Error: ${t.error}` : t.finalText || "(no output yet)"}${wt}\n${formatUsage(t.usage)}`;
+						? `\nBranch: ${t.branch}\n${t.diffStat || "(no diff available)"}\nMerge after review: \`git merge --no-ff ${t.branch}\``
+						: t.isolation === "in-place"
+							? `\nApplied IN PLACE (no branch) — ${t.isolationReason ?? "worktree unavailable"}. The changes are already in your working tree.`
+							: "";
+					const wtErr = t.worktreeError ? `\nWorktree: ${t.worktreeError}` : "";
+					return `\n## ${t.agent} ${statusIcon(t.status)}\nGoal: ${truncateText(t.task, 300)}\n${t.error ? `Error: ${t.error}` : t.finalText || "(no output yet)"}${wt}${wtErr}\n${formatUsage(t.usage)}`;
 				}),
 			].join("\n");
 			return { content: [{ type: "text", text: truncateText(text) }], details: { run: cloneRun(run) } };

package/src/manager.ts

@@ -1,7 +1,7 @@
 /** SubagentManager: run lifecycle, child sessions, intercom, persistence, widget plumbing. */
-import { existsSync, readFileSync, realpathSync } from "node:fs";
+import { existsSync, mkdirSync, readFileSync, realpathSync } from "node:fs";
 import { writeFile } from "node:fs/promises";
-import { join, relative } from "node:path";
+import { join, relative, sep } from "node:path";
 import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
 import type { Api, AssistantMessage, Model } from "@earendil-works/pi-ai";
 import {
@@ -45,6 +45,7 @@ import {
 } from "./types.ts";
 import {
 	branchDiff,
+	claimWorktree,
 	cleanupMerged,
 	commitWorktree,
 	createWorktree,
@@ -58,6 +59,10 @@ export const MAX_CONCURRENCY = 8;
 /** No default wall-clock cap: a subagent runs until its task is done, it stalls, or the user aborts. */
 const DEFAULT_RUNTIME_MS = 0;
 const DEFAULT_STALL_MS = 180_000; // 3 min: long model thinking streams emit no events, but they're not stalled.
+/** Cap on a child's wait for reply_subagent — an ignored question must not pin the run open forever. */
+const PARENT_REPLY_TIMEOUT_MS = 600_000; // 10 min
+/** Intercom messages buffered per park before the followUp path takes over. */
+const PARKED_MSG_CAP = 24;
 const READONLY_TOOLS = ["read", "grep", "find", "ls"];
 const WRITE_TOOLS = ["read", "grep", "find", "ls", "bash", "edit", "write"];
 /** Tools that can mutate the tree — their presence is what earns a worktree. */
@@ -221,7 +226,10 @@ export interface ParkedMsg {
 
 export class SubagentManager {
 	private runs = new Map<string, RunSnapshot>();
-	private settlers = new Map<string, (run: RunSnapshot) => void>();
+	/** Runs that are still settleable (presence = not yet settled). */
+	private settlers = new Map<string, true>();
+	/** Everyone parked on a run — a set, so re-parking can't build a closure chain. */
+	private settleWaiters = new Map<string, Set<(run: RunSnapshot) => void>>();
 	private pendingReplies = new Map<string, PendingReply>();
 	private liveChildren = new Map<
 		string,
@@ -296,8 +304,22 @@ export class SubagentManager {
 			child.dispose();
 		}
 		this.liveChildren.clear();
+		// Release anyone parked on a run before the maps go — dropping waiters would
+		// leave their promises pending forever (autoAwait / await_subagent hang).
+		for (const [runId, waiters] of this.settleWaiters) {
+			const run = this.runs.get(runId);
+			for (const waiter of waiters) waiter(run ? cloneRun(run) : ({ id: runId, status: "aborted" } as RunSnapshot));
+		}
+		for (const pending of this.pendingReplies.values()) {
+			pending.resolve("(session ended — stop work immediately)");
+		}
+		this.parked.clear();
+		// Ownership markers stay on disk; the next session reaps those dirs (commit,
+		// keep branch, drop dir) once this pid is gone.
+		this.liveWorktrees.clear();
 		this.runs.clear();
 		this.settlers.clear();
+		this.settleWaiters.clear();
 		this.pendingReplies.clear();
 		this.runControllers.clear();
 		this.mailboxes = createMailbox();
@@ -508,31 +530,36 @@ export class SubagentManager {
 	private makeChildHandlers(run: RunSnapshot, task: TaskSnapshot, ctx: ExtensionContext): ChildHandlers {
 		return {
 			onAskParent: async (_taskId, question) => {
+				const key = `${run.id}:${task.id}`;
+				// A tool call already in flight can reach here AFTER the task ended
+				// (abort/timeout/cancel). Reviving it would leave a "running" task in a
+				// finished run — hasActiveRun() then never clears.
+				if (TERMINAL.includes(task.status)) {
+					return "(your task has already ended — stop work and return immediately)";
+				}
 				this.updateTask(run, task, { status: "awaiting_parent" }, ctx);
-				this.liveChildren.get(`${run.id}:${task.id}`)?.touchWatchdog();
-				// While the leader is parked in await_subagent, the question rides the wait
-				// instead of the steering queue — no boundary needed, no starvation.
-				if (this.collectParked(run.id, { kind: "ask", taskId: task.id, agent: task.agent, text: question })) {
-					// The parked leader gets the question inside its await result and answers
-					// with reply_subagent — so the pending entry MUST exist here too, or the
-					// reply lands nowhere and the child waits on an answer that never comes.
-					const keepAlive = setInterval(() => this.liveChildren.get(`${run.id}:${task.id}`)?.touchWatchdog(), 30_000);
-					try {
-						const reply = await this.awaitParentReply(run.id, task.id);
-						this.updateTask(run, task, { status: "running" }, ctx);
-						this.liveChildren.get(`${run.id}:${task.id}`)?.touchWatchdog();
-						return reply;
-					} finally {
-						clearInterval(keepAlive);
-					}
+				this.liveChildren.get(key)?.touchWatchdog();
+				// While the leader is parked in await_subagent the question rides the wait
+				// (no steering queue, no turn boundary); otherwise it goes out as a notice.
+				// Either way the pending reply entry must exist, or reply_subagent has
+				// nowhere to land and the child waits on an answer that never comes.
+				if (!this.collectParked(run.id, { kind: "ask", taskId: task.id, agent: task.agent, text: question })) {
+					this.notifyParent(run, "asked", { taskId: task.id, question });
 				}
-				this.notifyParent(run, "asked", { taskId: task.id, question });
-				// M3: a waiting child is not stalled — keep the watchdog fed until the reply.
-				const keepAlive = setInterval(() => this.liveChildren.get(`${run.id}:${task.id}`)?.touchWatchdog(), 30_000);
+				// A waiting child is not stalled — keep the watchdog fed until the reply.
+				// But the wait is BOUNDED: an unanswered question would otherwise keep the
+				// run non-terminal forever (widget never clears, run never settles).
+				const keepAlive = setInterval(() => this.liveChildren.get(key)?.touchWatchdog(), 30_000);
 				try {
-					const reply = await this.awaitParentReply(run.id, task.id);
+					const reply = await this.awaitParentReply(run.id, task.id, PARENT_REPLY_TIMEOUT_MS);
+					// Cancel wins over a reply that arrived in the same tick: never move a
+					// terminal task back to "running" (that would let a canceled task be
+					// reported as completed).
+					if (TERMINAL.includes(task.status)) {
+						return "(your task was canceled while you waited — stop work and return immediately)";
+					}
 					this.updateTask(run, task, { status: "running" }, ctx);
-					this.liveChildren.get(`${run.id}:${task.id}`)?.touchWatchdog();
+					this.liveChildren.get(key)?.touchWatchdog();
 					return reply;
 				} finally {
 					clearInterval(keepAlive);
@@ -572,16 +599,35 @@ export class SubagentManager {
 			onPollMailbox: (taskId) => this.mailboxes.poll(`${run.id}:${taskId}`),
 		};
 	}
-	private awaitParentReply(runId: string, taskId: string): Promise<string> {
+	private awaitParentReply(runId: string, taskId: string, timeoutMs = 0): Promise<string> {
+		const key = `${runId}:${taskId}`;
 		return new Promise<string>((resolve) => {
-			this.pendingReplies.set(`${runId}:${taskId}`, { resolve });
+			// Identity-tagged: two asks from one child must not delete each other's
+			// entry (the loser would hang until its own timer).
+			const entry: PendingReply = {
+				resolve: (message) => {
+					if (timer) clearTimeout(timer);
+					if (this.pendingReplies.get(key) === entry) this.pendingReplies.delete(key);
+					resolve(message);
+				},
+			};
+			const timer =
+				timeoutMs > 0
+					? setTimeout(
+							() =>
+								entry.resolve(
+									"The parent did not answer in time. Proceed autonomously with your best judgment and state the assumption you made in your final answer.",
+								),
+							timeoutMs,
+						)
+					: undefined;
+			this.pendingReplies.set(key, entry);
 		});
 	}
 	deliverReply(runId: string, taskId: string, message: string): boolean {
 		const pending = this.pendingReplies.get(`${runId}:${taskId}`);
 		if (!pending) return false;
-		this.pendingReplies.delete(`${runId}:${taskId}`);
-		pending.resolve(message);
+		pending.resolve(message); // clears its own entry + timer
 		return true;
 	}
 
@@ -683,33 +729,9 @@ export class SubagentManager {
 		const canWrite = baseTools.some((t) => WRITE_CAPABLE.includes(t));
 
 		// Write agents run in an isolated git worktree (branch subagents/<run>/<task>);
-		// non-git repos fall back to in-place. The worktree is created BEFORE session
-		// start so the child's cwd + AGENTS.md context chain are the worktree's.
-		let wt: Worktree | undefined;
-		if (canWrite) {
-			try {
-				wt = createWorktree(task.cwd, run.id, task.id);
-			} catch {
-				wt = undefined; // git failure → in-place
-			}
-		}
-		// Map a per-task cwd subpath into the worktree so relative paths stay correct.
-		// If the mapping can't be trusted (cwd outside the repo via symlink), drop the
-		// worktree rather than run in the main tree while reporting a branch.
-		let childCwd = wt?.path ?? task.cwd;
-		if (wt) {
-			const rel = relative(wt.root, safeRealPath(task.cwd));
-			if (rel.startsWith("..")) {
-				removeWorktree(wt);
-				wt = undefined;
-			} else if (rel && rel !== ".") {
-				childCwd = join(wt.path, rel);
-			}
-		}
-		if (wt) this.liveWorktrees.set(`${run.id}:${task.id}`, wt);
-
-		// Model + thinking resolve against the pi model registry; a bad request
-		// fails the TASK with a helpful message, not the whole run.
+		// Model + thinking resolve against the pi model registry BEFORE any worktree
+		// exists — a bad request fails the TASK with a helpful message and can't leak a
+		// checkout past this early return.
 		let model: Model<Api> | undefined;
 		try {
 			model = resolveChildModel(ctx, file?.model ?? input.model);
@@ -729,6 +751,54 @@ export class SubagentManager {
 			return;
 		}
 
+		// Write agents run in an isolated git worktree (branch subagents/<run>/<task>);
+		// non-git repos fall back to in-place. Created BEFORE session start so the
+		// child's cwd + AGENTS.md context chain are the worktree's.
+		let wt: Worktree | undefined;
+		let isolationReason: string | undefined;
+		if (canWrite) {
+			try {
+				wt = createWorktree(task.cwd, run.id, task.id);
+				if (!wt) isolationReason = "not a git repository";
+			} catch (err) {
+				wt = undefined; // git failure → in-place
+				isolationReason = `git worktree add failed: ${err instanceof Error ? err.message : String(err)}`;
+			}
+		}
+		// Map a per-task cwd subpath into the worktree so relative paths stay correct.
+		// Both sides go through realpath — a symlinked root would otherwise look
+		// "outside" the repo. If the mapping can't be trusted, drop the worktree AND
+		// reset the cwd (never point the child at a dir that was just removed).
+		let childCwd = wt?.path ?? task.cwd;
+		if (wt) {
+			const rel = relative(safeRealPath(wt.root), safeRealPath(task.cwd));
+			if (rel === ".." || rel.startsWith(`..${sep}`)) {
+				removeWorktree(wt);
+				wt = undefined;
+				childCwd = task.cwd;
+				isolationReason = "task cwd is outside the repository";
+			} else if (rel && rel !== ".") {
+				childCwd = join(wt.path, rel);
+				// The subpath may be gitignored/untracked, so it won't exist in a fresh
+				// checkout — create it rather than fail session start.
+				try {
+					mkdirSync(childCwd, { recursive: true });
+				} catch {
+					childCwd = wt.path;
+				}
+			}
+		}
+		if (canWrite) {
+			// Never let isolation lapse quietly: the leader must know its edits landed
+			// straight in the working tree with no branch to review.
+			task.isolation = wt ? "worktree" : "in-place";
+			task.isolationReason = wt ? undefined : (isolationReason ?? "worktree unavailable");
+		}
+		if (wt) {
+			claimWorktree(wt); // pid marker: another pi session must not reap this
+			this.liveWorktrees.set(`${run.id}:${task.id}`, wt);
+		}
+
 		this.updateTask(
 			run,
 			task,
@@ -746,6 +816,9 @@ export class SubagentManager {
 			onUpdate,
 		);
 
+		// Set once the dir must outlive this call: committed work awaiting the
+		// leader's merge, or a commit failure whose work exists ONLY in the dir.
+		let keepWorktreeDir = false;
 		let child: Awaited<ReturnType<typeof createAgentSession>>["session"] | undefined;
 		let unsubscribe: (() => void) | undefined;
 		let timeout: ReturnType<typeof setTimeout> | undefined;
@@ -863,28 +936,51 @@ export class SubagentManager {
 					// until the branch is merged — cleanupMerged removes both then.
 					// Commit/diff failures must NOT downgrade a completed task or destroy
 					// its work: the error is reported, the status stays completed.
+					let committed: "committed" | "empty" | undefined;
 					try {
-						commitWorktree(wt, `subagent ${task.agent}: ${truncateText(input.task, 60)}`);
-						const { stat, files } = branchDiff(wt);
-						this.updateTask(
-							run,
-							task,
-							{ branch: wt.branch, diffStat: stat || undefined, changedFiles: files.length ? files : undefined },
-							ctx,
-							onUpdate,
-						);
+						committed = commitWorktree(wt, `subagent ${task.agent}: ${truncateText(input.task, 60)}`);
+						// Only a real commit is worth a branch: an empty one would send the
+						// leader off to review and merge nothing.
+						keepWorktreeDir = committed === "committed";
 					} catch (commitErr) {
+						// Never drop a checkout whose work isn't on the branch — it would be
+						// unreachable once the base-tip branch is reaped as "merged".
+						keepWorktreeDir = true;
 						this.updateTask(
 							run,
 							task,
 							{
 								branch: wt.branch,
-								error: `Worktree commit failed (changes remain in ${wt.path}): ${commitErr instanceof Error ? commitErr.message : String(commitErr)}`,
+								worktreeError: `commit failed (uncommitted changes remain in ${wt.path}): ${commitErr instanceof Error ? commitErr.message : String(commitErr)}`,
 							},
 							ctx,
 							onUpdate,
 						);
 					}
+					// Diff separately: a diff failure must not be reported as a lost commit.
+					if (committed === "committed") {
+						try {
+							const { stat, files } = branchDiff(wt);
+							this.updateTask(
+								run,
+								task,
+								{ branch: wt.branch, diffStat: stat || undefined, changedFiles: files.length ? files : undefined },
+								ctx,
+								onUpdate,
+							);
+						} catch (diffErr) {
+							this.updateTask(
+								run,
+								task,
+								{
+									branch: wt.branch,
+									worktreeError: `committed, but the diff could not be read: ${diffErr instanceof Error ? diffErr.message : String(diffErr)}`,
+								},
+								ctx,
+								onUpdate,
+							);
+						}
+					}
 				}
 			}
 		} catch (err) {
@@ -919,20 +1015,27 @@ export class SubagentManager {
 			watchdog.dispose();
 			if (timeout) clearTimeout(timeout);
 			child?.dispose();
-			this.liveWorktrees.delete(key);
 			// Failed/aborted: let the aborted child's last writes land (its tools may
 			// still be unwinding), commit whatever partial work exists so the branch
-			// really keeps it, then drop the checkout dir.
+			// really keeps it, then drop the checkout dir. A commit FAILURE keeps the
+			// dir — dropping it would make the work unreachable.
 			if (wt && task.status !== "completed") {
 				await new Promise((r) => setTimeout(r, 250));
+				let partial: "committed" | "empty" | undefined;
 				try {
-					commitWorktree(wt, `subagent ${task.agent} (partial, ${task.status})`);
+					partial = commitWorktree(wt, `subagent ${task.agent} (partial, ${task.status})`);
 				} catch {
-					/* nothing to commit */
+					keepWorktreeDir = true; // work exists only in the dir — keep it
+				}
+				// A branch is only worth reporting when it actually carries something.
+				if (partial === "committed" || keepWorktreeDir) {
+					this.updateTask(run, task, { branch: wt.branch }, ctx, onUpdate);
 				}
-				this.updateTask(run, task, { branch: wt.branch }, ctx, onUpdate);
-				removeWorktree(wt);
 			}
+			if (wt && !keepWorktreeDir) removeWorktree(wt);
+			// Released only after the dir is gone: while it exists, the branch must stay
+			// in liveBranches() so cleanup can't reap it.
+			this.liveWorktrees.delete(key);
 		}
 	}
 
@@ -1020,7 +1123,7 @@ export class SubagentManager {
 		});
 		this.turnActivity = true;
 		this.runs.set(run.id, run);
-		this.settlers.set(run.id, () => {});
+		this.settlers.set(run.id, true);
 		this.runControllers.set(run.id, new AbortController());
 		for (const task of run.tasks) this.mailboxes.open(`${run.id}:${task.id}`);
 		this.emit("subagent:run-created", { run: cloneRun(run) });
@@ -1087,13 +1190,14 @@ export class SubagentManager {
 		// Broken-upstream tasks are detected by the scheduler; mark them after the wave.
 		for (const s of skipped) {
 			const task = run.tasks.find((t) => t.id === s.id);
-			if (task) {
+			if (task && !TERMINAL.includes(task.status)) {
 				this.updateTask(
 					run,
 					task,
 					{
 						status: "aborted",
-						error: `Skipped: upstream task(s) did not complete: ${s.needs.join(", ")}`,
+						// Don't overwrite a real reason (e.g. "Canceled by subagent_cancel").
+						error: task.error || `Skipped: upstream task(s) did not complete: ${s.needs.join(", ")}`,
 						endedAt: Date.now(),
 					},
 					ctx,
@@ -1101,6 +1205,19 @@ export class SubagentManager {
 				);
 			}
 		}
+		// Belt and braces: the wave loop breaks out when no frontier is ready, which
+		// would otherwise leave tasks queued inside a terminal run — hasActiveRun()
+		// then never clears and the widget stays pinned.
+		for (const task of run.tasks) {
+			if (TERMINAL.includes(task.status)) continue;
+			this.updateTask(
+				run,
+				task,
+				{ status: "aborted", error: task.error || "Never ran: no runnable wave", endedAt: Date.now() },
+				ctx,
+				onUpdate,
+			);
+		}
 
 		const failed = run.tasks.some((t) => t.status === "failed");
 		const aborted = run.tasks.some((t) => t.status === "aborted") || Boolean(signal?.aborted);
@@ -1195,6 +1312,8 @@ export class SubagentManager {
 		task.status = "aborted";
 		task.error = task.error || "Canceled from peek";
 		task.endedAt = Date.now();
+		this.pendingReplies.get(`${runId}:${taskId}`)?.resolve("(task canceled by the parent — stop work now)");
+		this.pendingReplies.delete(`${runId}:${taskId}`);
 		this.liveChildren.get(`${runId}:${taskId}`)?.abort();
 		this.mailboxes.close(`${runId}:${taskId}`);
 		if (ctx) this.flushWidget(run, ctx);
@@ -1208,6 +1327,14 @@ export class SubagentManager {
 		if (TERMINAL.includes(run.status)) return { aborted: 0 }; // never corrupt a finished run
 		let aborted = 0;
 		this.runControllers.get(runId)?.abort();
+		// Release children parked in ask_parent first — an unresolved wait would keep
+		// the child alive past the abort.
+		for (const [key, pending] of this.pendingReplies) {
+			if (key.startsWith(`${runId}:`)) {
+				this.pendingReplies.delete(key);
+				pending.resolve("(run canceled by the parent — stop work and return what you have)");
+			}
+		}
 		for (const [key, child] of this.liveChildren) {
 			if (key.startsWith(`${runId}:`)) {
 				child.abort();
@@ -1233,24 +1360,42 @@ export class SubagentManager {
 		return { aborted };
 	}
 
-	/** Settle-and-delete: awaiters resolve once; no leak, no closure chain. */
+	/** Settle-and-delete: every awaiter resolves once, then the set is dropped. */
 	private settleRun(runId: string, run: RunSnapshot): void {
-		const s = this.settlers.get(runId);
-		if (!s) return;
+		if (!this.settlers.has(runId)) return;
 		this.settlers.delete(runId);
-		s(cloneRun(run));
+		const waiters = this.settleWaiters.get(runId);
+		this.settleWaiters.delete(runId);
+		if (!waiters) return;
+		const snapshot = cloneRun(run);
+		for (const waiter of waiters) waiter(snapshot);
 	}
 
 	/** Child→leader messages collected while the parent is parked in await_subagent. */
-	private parked = new Map<string, { msgs: ParkedMsg[]; wake: () => void }>();
+	/** Every awaiter parked on a run — a SET, so two concurrent awaits can't
+	 *  overwrite each other's buffer and silently swallow one side's intercom. */
+	private parked = new Map<string, Set<{ msgs: ParkedMsg[]; wake: () => void }>>();
 
 	/** While the parent is parked on this run, deliver the message through the wait instead of the steering queue. */
 	private collectParked(runId: string, msg: ParkedMsg): boolean {
-		const p = this.parked.get(runId);
-		if (!p) return false;
-		if (p.msgs.length < 24) p.msgs.push(msg);
-		p.wake(); // resolve the parked await early — the leader breathes on every message
-		return true;
+		const parked = this.parked.get(runId);
+		if (!parked || parked.size === 0) return false;
+		let delivered = false;
+		for (const p of parked) {
+			if (p.msgs.length < PARKED_MSG_CAP) {
+				p.msgs.push(msg);
+				delivered = true;
+			} else if (msg.kind === "ask") {
+				// An unanswered ask blocks a child for 10 minutes — it must never be the
+				// message that gets dropped by the cap.
+				p.msgs[p.msgs.length - 1] = msg;
+				delivered = true;
+			}
+			p.wake(); // resolve the parked await early — the leader breathes on every message
+		}
+		// Not buffered anywhere → report undelivered so the caller falls back to a
+		// followUp notice instead of assuming the leader saw it.
+		return delivered;
 	}
 
 	awaitRun(
@@ -1259,8 +1404,12 @@ export class SubagentManager {
 	): Promise<{ run: RunSnapshot | undefined; intercom: ParkedMsg[] } | undefined> {
 		const run = this.runs.get(runId);
 		if (!run) return Promise.resolve(undefined);
+		let entry: { msgs: ParkedMsg[]; wake: () => void } | undefined;
 		const finish = (): void => {
-			this.parked.delete(runId);
+			const parked = this.parked.get(runId);
+			if (!parked || !entry) return;
+			parked.delete(entry); // only our own park — a sibling await keeps receiving
+			if (parked.size === 0) this.parked.delete(runId);
 		};
 		if (TERMINAL.includes(run.status)) {
 			run.awaited = true;
@@ -1268,24 +1417,43 @@ export class SubagentManager {
 		}
 		const msgs: ParkedMsg[] = [];
 		const settled = new Promise<RunSnapshot | undefined>((resolve) => {
-			const prev = this.settlers.get(runId);
-			this.settlers.set(runId, (r) => {
-				prev?.(r);
+			// Waiters are a SET, not a chain: the autoAwait loop re-parks on every
+			// child message, and wrapping the previous settler each time grew an
+			// unbounded closure chain (each holding a snapshot clone).
+			const waiter = (r: RunSnapshot) => {
+				this.settleWaiters.get(runId)?.delete(waiter);
 				resolve(r);
-			});
+			};
+			let waiters = this.settleWaiters.get(runId);
+			if (!waiters) {
+				waiters = new Set();
+				this.settleWaiters.set(runId, waiters);
+			}
+			waiters.add(waiter);
 			// A child→leader message while parked wakes the wait: the leader gets it
 			// IN the await result, no steering queue, no turn boundary needed.
-			this.parked.set(runId, { msgs, wake: () => resolve(cloneRun(run)) });
+			entry = { msgs, wake: () => waiter(cloneRun(run)) };
+			let parked = this.parked.get(runId);
+			if (!parked) {
+				parked = new Set();
+				this.parked.set(runId, parked);
+			}
+			parked.add(entry);
 		});
 		if (timeoutMs !== undefined && timeoutMs > 0) {
+			let timedOut = false;
 			return Promise.race([
 				settled.then((r) => {
 					finish();
-					run.awaited = true;
+					// Only mark awaited when this call actually hands the run back to the
+					// leader. A slice that already timed out is abandoned — setting it here
+					// would suppress the run's completion notice the leader still needs.
+					if (!timedOut) run.awaited = true;
 					return { run: r, intercom: msgs };
 				}),
 				new Promise<{ run: RunSnapshot | undefined; intercom: ParkedMsg[] } | undefined>((resolve) => {
 					const timer = setTimeout(() => {
+						timedOut = true;
 						finish();
 						resolve(this.runs.get(runId) ? { run: cloneRun(this.runs.get(runId)!), intercom: msgs } : undefined);
 					}, timeoutMs);

package/src/types.ts

@@ -45,6 +45,13 @@ export interface TaskSnapshot {
 	branch?: string;
 	diffStat?: string;
 	changedFiles?: string[];
+	/** How a write child's edits were applied. "in-place" means NO branch: the
+	 *  changes are already in the leader's tree — always surfaced, never silent. */
+	isolation?: "worktree" | "in-place";
+	isolationReason?: string;
+	/** Worktree commit/diff trouble. Kept apart from `error` so a completed task
+	 *  still reports its answer. */
+	worktreeError?: string;
 }
 
 export interface RunSnapshot {

package/src/worktree.ts

@@ -10,8 +10,9 @@
  *  touching a checked-out one), `sweepStale` (dirs git no longer knows about). */
 
 import { execFileSync } from "node:child_process";
-import { existsSync, readdirSync, realpathSync, rmSync, symlinkSync } from "node:fs";
-import { join } from "node:path";
+import { existsSync, readdirSync, readFileSync, realpathSync, rmSync, symlinkSync, writeFileSync } from "node:fs";
+import { hostname, uptime } from "node:os";
+import { join, resolve } from "node:path";
 
 export interface Worktree {
 	root: string; // repo root (main tree)
@@ -22,13 +23,33 @@ export interface Worktree {
 
 const BRANCH_PREFIX = "subagents/";
 
+/** Identity + signing fallbacks: a machine without user.email (CI, fresh box) or
+ *  with commit.gpgsign set must not fail — or worse, block on a passphrase prompt. */
+const COMMIT_CONFIG = ["-c", "commit.gpgsign=false", "-c", "user.name=pi subagent", "-c", "user.email=subagent@local"];
+const GIT_TIMEOUT_MS = 120_000;
+const GIT_MAX_BUFFER = 32 * 1024 * 1024;
+
 function git(root: string, args: string[]): string {
-	return execFileSync("git", ["-C", root, ...args], { encoding: "utf8" }).trim();
+	return gitRaw(root, args).trim();
+}
+
+/** Untrimmed git output — required for -z parsing, where a path may end in a space. */
+function gitRaw(root: string, args: string[]): string {
+	return execFileSync("git", ["-C", root, ...args], {
+		encoding: "utf8",
+		timeout: GIT_TIMEOUT_MS,
+		maxBuffer: GIT_MAX_BUFFER,
+	});
 }
 
 /** Run git directly inside a directory (worktree ops). */
 function gitIn(dir: string, args: string[]): string {
-	return execFileSync("git", [...args], { cwd: dir, encoding: "utf8" }).trim();
+	return execFileSync("git", [...args], {
+		cwd: dir,
+		encoding: "utf8",
+		timeout: GIT_TIMEOUT_MS,
+		maxBuffer: GIT_MAX_BUFFER,
+	}).trim();
 }
 
 function gitOk(root: string, args: string[]): boolean {
@@ -40,15 +61,6 @@ function gitOk(root: string, args: string[]): boolean {
 	}
 }
 
-/** git C-quotes porcelain paths containing spaces/specials — undo that. */
-function unquote(path: string): string {
-	if (!path.startsWith('"') || !path.endsWith('"')) return path;
-	return path
-		.slice(1, -1)
-		.replace(/\\([0-7]{3})/g, (_, o) => String.fromCharCode(Number.parseInt(o, 8)))
-		.replace(/\\(.)/g, (_, c) => ({ n: "\n", t: "\t", r: "\r" })[c as string] ?? c);
-}
-
 /** Compare paths through realpath so /var vs /private/var can't diverge. */
 function samePath(a: string, b: string): boolean {
 	if (a === b) return true;
@@ -73,15 +85,22 @@ export function repoRoot(cwd: string): string | undefined {
 export function createWorktree(cwd: string, runId: string, taskId: string): Worktree | undefined {
 	const root = repoRoot(cwd);
 	if (!root) return undefined;
-	const path = join(root, ".git", "subagents", runId, taskId);
-	const branch = `${BRANCH_PREFIX}${runId}/${taskId}`;
+	// --git-common-dir, not "<root>/.git": inside a linked worktree or a submodule
+	// `.git` is a FILE, and joining it would make `worktree add` fail (silently
+	// dropping isolation).
+	const container = subagentsDir(root);
+	if (!container) return undefined;
 	let base: string;
 	try {
 		base = git(root, ["rev-parse", "HEAD"]); // SHA — detached HEAD stays correct
 	} catch {
 		return undefined; // broken repo — fall back to in-place
 	}
-	git(root, ["worktree", "add", "-b", branch, path, "HEAD"]);
+	const path = join(container, runId, taskId);
+	const branch = `${BRANCH_PREFIX}${runId}/${taskId}`;
+	// Branch from the recorded SHA, not "HEAD" — a concurrent commit in the main
+	// tree between the two would otherwise skew every later diff against base.
+	git(root, ["worktree", "add", "-b", branch, path, base]);
 	// Deps follow the child into the worktree; anything else the task needs is
 	// project content already checked out there.
 	const nm = join(root, "node_modules");
@@ -98,16 +117,33 @@ export function createWorktree(cwd: string, runId: string, taskId: string): Work
 /**
  * Commit the child's changes. Stages first, then commits only when something is
  * actually staged — an untracked node_modules symlink must not fake "dirty" and
- * turn into a failed empty commit.
+ * turn into a failed empty commit. Returns "committed" | "empty"; a real git
+ * failure THROWS, and callers must not delete the checkout in that case (the
+ * work would become unreachable once the base-tip branch is reaped as merged).
  */
-export function commitWorktree(wt: Worktree, message: string): void {
-	commitIn(wt.path, message);
+export function commitWorktree(wt: Worktree, message: string): "committed" | "empty" {
+	return commitIn(wt.path, message, wt.branch);
 }
 
-function commitIn(dir: string, message: string): void {
-	gitIn(dir, ["add", "-A", "--", ".", ":(exclude)node_modules"]); // never stage the dep symlink
-	if (gitIn(dir, ["diff", "--cached", "--name-only"]).length === 0) return; // nothing to commit
-	gitIn(dir, ["commit", "-m", message, "--no-verify"]);
+function commitIn(dir: string, message: string, expectBranch?: string): "committed" | "empty" {
+	// A child that detached HEAD or switched branches would commit somewhere the
+	// leader is never told about — refuse rather than report a branch without the work.
+	if (expectBranch) {
+		// `symbolic-ref` EXITS NON-ZERO on a detached HEAD — catch it rather than
+		// letting the raw git failure masquerade as a commit error.
+		let head = "detached";
+		try {
+			head = gitIn(dir, ["symbolic-ref", "--quiet", "--short", "HEAD"]) || "detached";
+		} catch {
+			head = "detached";
+		}
+		if (head !== expectBranch) throw new Error(`worktree HEAD is "${head}", expected ${expectBranch}`);
+	}
+	// Exclude the root dep symlink and any nested node_modules the child created.
+	gitIn(dir, ["add", "-A", "--", ".", ":(exclude)node_modules", ":(exclude,glob)**/node_modules/**"]);
+	if (gitIn(dir, ["diff", "--cached", "--name-only"]).length === 0) return "empty";
+	gitIn(dir, [...COMMIT_CONFIG, "commit", "-m", message, "--no-verify"]);
+	return "committed";
 }
 
 /** Diffstat + changed files of the branch vs its base SHA. */
@@ -129,7 +165,28 @@ export function removeByBranch(cwd: string, branch: string): void {
 	if (!branch.startsWith(BRANCH_PREFIX)) return;
 	const root = repoRoot(cwd);
 	if (!root) return;
-	dropDir(root, join(root, ".git", "subagents", branch.slice(BRANCH_PREFIX.length)));
+	const container = subagentsDir(root);
+	if (container) dropDir(root, join(container, branch.slice(BRANCH_PREFIX.length)));
+}
+
+/** `<git-common-dir>/subagents` — where our worktrees live for this repo.
+ *  `--path-format` needs git ≥ 2.31; fall back to resolving the relative form. */
+function subagentsDir(root: string): string | undefined {
+	try {
+		return join(git(root, ["rev-parse", "--path-format=absolute", "--git-common-dir"]), "subagents");
+	} catch {
+		try {
+			return join(resolve(root, git(root, ["rev-parse", "--git-common-dir"])), "subagents");
+		} catch {
+			return undefined;
+		}
+	}
+}
+
+/** Marker path lives BESIDE the checkout, never inside it — a file in the worktree
+ *  would be staged by `add -A`, committed into the branch, and merged into main. */
+function ownerFile(path: string): string {
+	return `${path}.owner`;
 }
 
 function dropDir(root: string, path: string): void {
@@ -138,6 +195,7 @@ function dropDir(root: string, path: string): void {
 	} catch {
 		if (existsSync(path)) rmSync(path, { recursive: true, force: true });
 	}
+	rmSync(ownerFile(path), { force: true }); // marker lives beside the dir
 	prune(root);
 }
 
@@ -150,9 +208,10 @@ function prune(root: string): void {
 	}
 }
 
-/** Is this branch checked out in some worktree right now? Checked fresh, per call. */
+/** Is this branch checked out right now? "Unknown" counts as YES (never delete blind). */
 function isCheckedOut(root: string, branch: string): boolean {
-	return worktreeBranches(root).includes(branch);
+	const branches = worktreeBranches(root);
+	return branches === undefined || branches.includes(branch);
 }
 
 /**
@@ -169,34 +228,36 @@ export function cleanupMerged(root: string, opts: { skipBranches?: Set<string>;
 		.split("\n")
 		.map((b) => b.trim().replace(/^[+*]\s*/, ""));
 	let cleaned = 0;
+	const container = subagentsDir(root);
 	for (const branch of merged) {
-		if (!branch.startsWith(BRANCH_PREFIX)) continue;
+		if (!branch.startsWith(BRANCH_PREFIX) || !container) continue;
 		if (opts.skipBranches?.has(branch)) continue; // owned by a live run
 		if (isCheckedOut(root, branch)) continue; // fresh re-check, not a stale snapshot
-		const path = join(root, ".git", "subagents", branch.slice(BRANCH_PREFIX.length));
+		const path = join(container, branch.slice(BRANCH_PREFIX.length));
+		if (existsSync(path) && ownerAlive(path)) continue; // another session's live checkout
+		// Branch FIRST: `-d` refuses anything not truly merged, so a stale "merged"
+		// listing can no longer cost us a checkout that still holds work.
+		if (!gitOk(root, ["branch", "-d", branch])) continue;
 		if (existsSync(path)) {
 			try {
 				git(root, ["worktree", "remove", "--force", path]);
 			} catch {
-				continue;
+				/* branch is gone; a leftover dir is swept later */
 			}
 		}
-		if (gitOk(root, ["branch", "-d", branch])) cleaned += 1;
+		rmSync(ownerFile(path), { force: true });
+		cleaned += 1;
 	}
 	prune(root);
 	return cleaned;
 }
 
-/** Branch names currently checked out in any worktree (incl. the main one). */
-function worktreeBranches(root: string): string[] {
-	try {
-		return git(root, ["worktree", "list", "--porcelain"])
-			.split("\n")
-			.filter((l) => l.startsWith("branch "))
-			.map((l) => l.slice("branch refs/heads/".length).trim());
-	} catch {
-		return [];
-	}
+/** Branch names currently checked out in any worktree, or undefined when git
+ *  couldn't be asked — callers MUST treat undefined as "unknown", never as "none",
+ *  or they will happily delete live checkouts. */
+function worktreeBranches(root: string): string[] | undefined {
+	const listing = worktreeListing(root);
+	return listing?.filter((l) => l.startsWith("branch ")).map((l) => l.slice("branch refs/heads/".length));
 }
 
 /**
@@ -204,17 +265,20 @@ function worktreeBranches(root: string): string[] {
  * (nothing of ours can be live yet). Commit whatever the dead child left so the
  * branch keeps it, then drop the dir. Branches always survive.
  */
-export function reapDeadWorktrees(root: string): number {
+export function reapDeadWorktrees(root: string, isLive: (path: string) => boolean = () => false): number {
 	root = realpathSync(root);
-	const sub = join(root, ".git", "subagents");
-	if (!existsSync(sub)) return 0;
+	const sub = subagentsDir(root);
+	if (!sub || !existsSync(sub)) return 0;
+	const registered = worktreePaths(root);
+	if (!registered) return 0; // listing failed — touch nothing
 	let reaped = 0;
-	for (const path of worktreePaths(root)) {
+	for (const path of registered) {
 		if (!isInside(path, sub)) continue; // not ours
+		if (isLive(path)) continue; // another pi session owns it
 		try {
 			commitIn(path, "subagent (recovered after interrupted session)");
 		} catch {
-			/* nothing committable */
+			continue; // git failed — never drop a dir whose work isn't on the branch
 		}
 		dropDir(root, path);
 		reaped += 1;
@@ -222,6 +286,51 @@ export function reapDeadWorktrees(root: string): number {
 	return reaped;
 }
 
+/**
+ * Ownership marker: a live worktree gets `<dir>/.subagent-owner` holding the
+ * owning pid. Another pi session must not reap a checkout whose owner is alive.
+ */
+export function claimWorktree(wt: Worktree): void {
+	try {
+		writeFileSync(
+			ownerFile(wt.path),
+			JSON.stringify({ pid: process.pid, host: hostname(), boot: bootId(), at: Date.now() }),
+		);
+	} catch {
+		/* best-effort: worst case another session reaps it after a crash */
+	}
+}
+
+/**
+ * True when the worktree is claimed by a process that still exists HERE.
+ * Guards against pid reuse across reboots (boot id) and other hosts (hostname);
+ * EPERM means the pid exists under another user — alive, not reapable.
+ */
+export function ownerAlive(path: string): boolean {
+	let marker: { pid?: number; host?: string; boot?: string };
+	try {
+		marker = JSON.parse(readFileSync(ownerFile(path), "utf8"));
+	} catch {
+		return false; // no marker (or unreadable) → nobody claims it
+	}
+	const pid = marker.pid;
+	if (!pid || !Number.isFinite(pid) || pid <= 0) return false;
+	if (marker.host !== hostname()) return true; // another machine's checkout — never ours to reap
+	if (marker.boot !== bootId()) return false; // pre-reboot pid: reuse is near-certain
+	if (pid === process.pid) return true;
+	try {
+		process.kill(pid, 0);
+		return true;
+	} catch (err) {
+		return (err as NodeJS.ErrnoException)?.code === "EPERM"; // exists, other user
+	}
+}
+
+/** Stable per-boot id, so a recycled pid from before a reboot can't look alive. */
+function bootId(): string {
+	return String(Math.floor(Date.now() / 1000 - Math.floor(uptime())));
+}
+
 /**
  * Remove worktree dirs that git no longer knows about (partial-crash leftovers).
  * Registered worktrees are never touched here — `reapDeadWorktrees` owns those,
@@ -229,27 +338,42 @@ export function reapDeadWorktrees(root: string): number {
  */
 export function sweepStale(root: string): void {
 	root = realpathSync(root);
-	const sub = join(root, ".git", "subagents");
-	if (!existsSync(sub)) return;
+	const sub = subagentsDir(root);
+	if (!sub || !existsSync(sub)) return;
 	const registered = worktreePaths(root);
+	// Unknown registration = every dir might be live. Deleting here would be the
+	// single most destructive thing this module can do; bail instead.
+	if (!registered) return;
 	for (const runDir of readDirs(sub)) {
 		for (const taskDir of readDirs(join(sub, runDir))) {
 			const dir = join(sub, runDir, taskDir);
 			if (registered.some((p) => samePath(p, dir))) continue; // live/registered worktree
+			if (ownerAlive(dir)) continue; // claimed by a running session
 			rmSync(dir, { recursive: true, force: true });
+			rmSync(ownerFile(dir), { force: true });
 		}
 	}
 	prune(root);
 }
 
-function worktreePaths(root: string): string[] {
+/** Registered worktree paths, or undefined when the listing failed (see above). */
+function worktreePaths(root: string): string[] | undefined {
+	const listing = worktreeListing(root);
+	return listing?.filter((l) => l.startsWith("worktree ")).map((l) => l.slice("worktree ".length));
+}
+
+/** `worktree list --porcelain -z`; `-z` needs git ≥ 2.36, so fall back to the
+ *  newline form (which C-quotes exotic paths — those simply won't match, and a
+ *  non-match is safe: it only ever means "treat as live"). */
+function worktreeListing(root: string): string[] | undefined {
 	try {
-		return git(root, ["worktree", "list", "--porcelain"])
-			.split("\n")
-			.filter((l) => l.startsWith("worktree "))
-			.map((l) => unquote(l.slice("worktree ".length).trim()));
+		return gitRaw(root, ["worktree", "list", "--porcelain", "-z"]).split("\0").filter(Boolean);
 	} catch {
-		return [];
+		try {
+			return git(root, ["worktree", "list", "--porcelain"]).split("\n").filter(Boolean);
+		} catch {
+			return undefined; // unknown — callers must bail out
+		}
 	}
 }