@arhen/pi-core-subagent 1.3.29 → 1.3.31

package/README.md

@@ -147,7 +147,17 @@ On completion the extension commits the child's changes (the child is told not t
 git merge --no-ff subagents/<run>/<task>
 ```
 
-Merged branches + their worktree dirs are cleaned automatically after the run. Failed/canceled tasks keep the branch (partial work survives for manual merging) but drop the worktree dir. Crash leftovers are swept at session start — worktree dirs are removed, branches are kept. Non-git repos fall back to in-place edits.
+Isolation follows the toolset the child actually receives: explicit `tools: ["bash", "edit", "write"]` earns a worktree even without `write: true`, and an agent file that narrows the child to read-only gets no branch at all.
+
+Cleanup, in order of trust:
+
+| When | What |
+|---|---|
+| Session start | Registered worktrees can't be live yet — an interrupted child's uncommitted work is committed, the branch kept, the dir dropped. Then merged branches are reaped and dirs git no longer tracks are removed. |
+| After a run | Merged branches + their dirs, never touching a branch a live run owns (checkout state re-checked per branch). |
+| Task failed/canceled | Partial work is committed first, so the branch keeps it; the dir is dropped. |
+
+Non-git repos fall back to in-place edits.
 
 ## Graph mode — `needs`
 

package/package.json

@@ -1,6 +1,6 @@
 {
 	"name": "@arhen/pi-core-subagent",
-	"version": "1.3.29",
+	"version": "1.3.31",
 	"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/index.ts

@@ -18,7 +18,7 @@ import type { ExtensionAPI, ExtensionContext } from "@earendil-works/pi-coding-a
 import { Text, truncateToWidth } from "@earendil-works/pi-tui";
 import { compactLines, formatUsage, makeSummary, statusIcon, taskLine, truncateText } from "./format.ts";
 import { waveNotation } from "./graph.ts";
-import { cloneRun, SubagentManager } from "./manager.ts";
+import { cloneRun, type ParkedMsg, SubagentManager } from "./manager.ts";
 import { createPeekPane, type PeekTask } from "./peek.ts";
 import {
 	AwaitParam,
@@ -30,7 +30,7 @@ import {
 	type SubagentParamsShape,
 } from "./schemas.ts";
 import { type RunDetails, type RunSnapshot, TERMINAL } from "./types.ts";
-import { repoRoot, sweepStale } from "./worktree.ts";
+import { cleanupMerged, reapDeadWorktrees, repoRoot, sweepStale } from "./worktree.ts";
 
 export default function (pi: ExtensionAPI) {
 	const manager = new SubagentManager(pi);
@@ -111,15 +111,31 @@ export default function (pi: ExtensionAPI) {
 	pi.on("session_start", async (_event, ctx) => {
 		await manager.restoreFromSidecar(ctx);
 		// Crash leftovers: remove stale worktree dirs (branches survive for merging).
-		sweepStale(ctx.cwd);
+		// Also reap branches the leader merged in a previous session (H1: cleanup can't
+		// fire at run end — the leader merges after).
+		const roots = new Set<string>();
+		const cwdRoot = repoRoot(ctx.cwd);
+		if (cwdRoot) roots.add(cwdRoot);
 		for (const run of manager.listRuns()) {
 			for (const task of run.tasks) {
 				if (task.branch) {
 					const root = repoRoot(task.cwd);
-					if (root) sweepStale(root);
+					if (root) roots.add(root);
 				}
 			}
 		}
+		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);
+				sweepStale(root);
+			} catch {
+				/* recovery is best-effort — never block session start */
+			}
+		}
 	});
 	pi.on("session_shutdown", async (_event, ctx) => {
 		if (ctx?.hasUI) {
@@ -158,13 +174,34 @@ export default function (pi: ExtensionAPI) {
 			const typed = params as SubagentParamsShape;
 			const details = manager.startInBackground(typed, ctx);
 			if (typed.autoAwait) {
-				// awaitRun wakes on every child→leader message (ask/notify/done) — that's the
-				// slice-loop feature. autoAwait wants the final result: re-park until terminal.
+				// awaitRun wakes on every child→leader message (ask/notify/done). Re-park
+				// until terminal — but surface an ask_parent: the child is waiting on the
+				// leader, so break out, reply via reply_subagent, then await again.
 				let run = details.run;
+				// Each park gets a FRESH msgs array — accumulate, or every wake but the
+				// last is lost (they were consumed by the park, never sent as followUp).
+				const intercom: ParkedMsg[] = [];
 				while (!TERMINAL.includes(run.status)) {
-					run = (await manager.awaitRun(details.run.id))?.run ?? run;
+					const awaited = await manager.awaitRun(details.run.id);
+					if (!awaited) break; // run gone (session shutdown) — stop, no busy-spin
+					if (awaited.run) run = awaited.run;
+					intercom.push(...awaited.intercom);
+					if (awaited.intercom.some((m) => m.kind === "ask")) break;
 				}
-				return { content: [{ type: "text", text: makeSummary(run) }], details: { run } };
+				const asked = intercom.find((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.`
+						: "",
+				]
+					.filter(Boolean)
+					.join("\n");
+				return { content: [{ type: "text", text }], details: { run } };
 			}
 			return {
 				content: [

package/src/manager.ts

@@ -1,7 +1,7 @@
 /** SubagentManager: run lifecycle, child sessions, intercom, persistence, widget plumbing. */
-import { existsSync, readFileSync } from "node:fs";
+import { existsSync, readFileSync, realpathSync } from "node:fs";
 import { writeFile } from "node:fs/promises";
-import { join } from "node:path";
+import { join, relative } from "node:path";
 import type { ThinkingLevel } from "@earendil-works/pi-agent-core";
 import type { Api, AssistantMessage, Model } from "@earendil-works/pi-ai";
 import {
@@ -48,10 +48,8 @@ import {
 	cleanupMerged,
 	commitWorktree,
 	createWorktree,
-	removeByBranch,
 	removeWorktree,
 	repoRoot,
-	sweepStale,
 	type Worktree,
 } from "./worktree.ts";
 
@@ -62,6 +60,10 @@ 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.
 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. */
+const WRITE_CAPABLE = ["bash", "edit", "write"];
+/** Task ids become git refs + filesystem paths. */
+const SAFE_TASK_ID = /^[A-Za-z0-9_-]{1,64}$/;
 const WIDGET_THROTTLE_MS = 150;
 
 // ── helpers ──────────────────────────────────────────────────────────────
@@ -84,6 +86,14 @@ function aggregateUsage(tasks: TaskSnapshot[]): UsageStats {
 	}
 	return total;
 }
+/** realpath when possible; the raw path otherwise (cwd may not exist yet). */
+function safeRealPath(path: string): string {
+	try {
+		return realpathSync(path);
+	} catch {
+		return path;
+	}
+}
 function getParentSessionFile(ctx: ExtensionContext): string | undefined {
 	try {
 		return ctx.sessionManager.getSessionFile?.();
@@ -218,6 +228,9 @@ export class SubagentManager {
 		{ abort: () => void; dispose: () => void; touchWatchdog: () => void; steer: (message: string) => void }
 	>();
 	private mailboxes: Mailbox = createMailbox();
+	/** Live worktrees by `${runId}:${taskId}` — lets cancel drop dirs and keeps
+	 *  cleanup from touching a branch that a running child owns. */
+	private liveWorktrees = new Map<string, Worktree>();
 	private runControllers = new Map<string, AbortController>();
 	private widgetTimers = new Map<string, ReturnType<typeof setTimeout>>(); // per-run stream throttle
 	private widgetRuns: RunSnapshot[] = [];
@@ -500,7 +513,18 @@ export class SubagentManager {
 				// 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 })) {
-					return "Your question was delivered to the parent (they're waiting on this run). Keep working; the answer arrives via the pending reply.";
+					// 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.notifyParent(run, "asked", { taskId: task.id, question });
 				// M3: a waiting child is not stalled — keep the watchdog fed until the reply.
@@ -516,13 +540,13 @@ export class SubagentManager {
 			},
 			onNotifyParent: (_taskId, message, level) => {
 				this.emit("subagent:intercom", { runId: run.id, taskId: task.id, kind: "notify", level, message });
+				// Parked leader gets it through the wait; otherwise queue it. `awaited` must
+				// NOT gate this — between two parks the leader is awaited but listening.
 				if (this.collectParked(run.id, { kind: "notify", taskId: task.id, agent: task.agent, text: message })) return;
-				if (!run.awaited) {
-					try {
-						this.pi.sendUserMessage(`[Subagent ${task.agent}] ${message}`, { deliverAs: "followUp" });
-					} catch {
-						/* parent mid-stream */
-					}
+				try {
+					this.pi.sendUserMessage(`[Subagent ${task.agent}] ${message}`, { deliverAs: "followUp" });
+				} catch {
+					/* parent mid-stream */
 				}
 			},
 			onSendMessage: (_taskId, to, text) => {
@@ -535,12 +559,10 @@ export class SubagentManager {
 						message: text,
 					});
 					if (this.collectParked(run.id, { kind: "notify", taskId: task.id, agent: task.agent, text })) return true;
-					if (!run.awaited) {
-						try {
-							this.pi.sendUserMessage(`[Subagent ${task.agent}] ${text}`, { deliverAs: "followUp" });
-						} catch {
-							/* parent mid-stream */
-						}
+					try {
+						this.pi.sendUserMessage(`[Subagent ${task.agent}] ${text}`, { deliverAs: "followUp" });
+					} catch {
+						/* parent mid-stream */
 					}
 					return true;
 				}
@@ -649,21 +671,42 @@ export class SubagentManager {
 		const file = resolveAgentFile(input.agent, input.task, task.cwd, getAgentDir());
 		const prompt = file?.body ?? input.prompt?.trim();
 		const thinking = input.thinking;
-		const baseTools = file?.tools ?? input.tools ?? (input.write ? WRITE_TOOLS : READONLY_TOOLS);
+		// Trust boundary: a file can NARROW the toolset (intersect with the leader's
+		// intent) but never widen it — a repo-planted agent file can't grant write.
+		const allowedTools = input.write ? WRITE_TOOLS : READONLY_TOOLS;
+		const fileTools = file?.tools?.filter((t) => allowedTools.includes(t));
+		const baseTools = fileTools?.length ? fileTools : (input.tools ?? allowedTools);
 		const tools = [...baseTools, ...(run.allowIntercom ? CHILD_TALK_TOOLS : [])];
+		// Isolation follows the DELIVERED toolset, never the raw request: explicit
+		// tools: [bash] without write:true still gets a worktree, and a file that
+		// narrowed the child to read-only never gets the commit/merge ceremony.
+		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 (input.write) {
+		if (canWrite) {
 			try {
 				wt = createWorktree(task.cwd, run.id, task.id);
 			} catch {
 				wt = undefined; // git failure → in-place
 			}
 		}
-		const childCwd = wt?.path ?? task.cwd;
+		// 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.
@@ -818,15 +861,30 @@ export class SubagentManager {
 					// Commit the child's changes, then report the branch + diff so the
 					// leader can review and merge (PR-style). The worktree dir stays
 					// until the branch is merged — cleanupMerged removes both then.
-					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,
-					);
+					// Commit/diff failures must NOT downgrade a completed task or destroy
+					// its work: the error is reported, the status stays completed.
+					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,
+						);
+					} catch (commitErr) {
+						this.updateTask(
+							run,
+							task,
+							{
+								branch: wt.branch,
+								error: `Worktree commit failed (changes remain in ${wt.path}): ${commitErr instanceof Error ? commitErr.message : String(commitErr)}`,
+							},
+							ctx,
+							onUpdate,
+						);
+					}
 				}
 			}
 		} catch (err) {
@@ -861,9 +919,17 @@ export class SubagentManager {
 			watchdog.dispose();
 			if (timeout) clearTimeout(timeout);
 			child?.dispose();
-			// Failed/aborted: the branch keeps any partial work for manual merging, the
-			// checkout dir is removed so it can't linger on disk.
+			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.
 			if (wt && task.status !== "completed") {
+				await new Promise((r) => setTimeout(r, 250));
+				try {
+					commitWorktree(wt, `subagent ${task.agent} (partial, ${task.status})`);
+				} catch {
+					/* nothing to commit */
+				}
 				this.updateTask(run, task, { branch: wt.branch }, ctx, onUpdate);
 				removeWorktree(wt);
 			}
@@ -898,13 +964,25 @@ export class SubagentManager {
 				? params.tasks!
 				: params.chain!;
 		if (inputs.length > MAX_TASKS) throw new Error(`Too many subagent tasks (${inputs.length}). Max is ${MAX_TASKS}.`);
+		// Task ids become git refs + filesystem paths — refuse anything unsafe.
+		// Explicit ids are checked against each other; generated ones are checked
+		// against explicit ones so a collision can't silently fall back to in-place.
 		const ids = new Set<string>();
 		for (const input of inputs) {
 			if (input.id !== undefined) {
+				if (!SAFE_TASK_ID.test(input.id)) {
+					throw new Error(`Unsafe task id: "${input.id}" (allowed: letters, digits, _ and - only).`);
+				}
 				if (ids.has(input.id)) throw new Error(`Duplicate task id: ${input.id}`);
 				ids.add(input.id);
 			}
 		}
+		for (let i = 0; i < inputs.length; i++) {
+			const generated = `task_${i + 1}`;
+			if (inputs[i]?.id === undefined && ids.has(generated)) {
+				throw new Error(`Generated task id ${generated} collides with an explicit id — rename the explicit id.`);
+			}
+		}
 		const edges = resolveNeeds(inputs, mode);
 
 		const run: RunSnapshot = {
@@ -968,14 +1046,30 @@ export class SubagentManager {
 		for (const task of run.tasks) {
 			if (TERMINAL.includes(task.status)) settled.add(task.id); // canceled before start
 		}
+		// id → input, immune to filtered-array index drift (C4).
+		const inputById = new Map(run.tasks.map((t, i) => [t.id, inputs[i]]));
 
 		const { skipped } = await runWaveScheduler(
 			run.tasks.filter((t) => !TERMINAL.includes(t.status)),
 			run.mode === "single" ? 1 : run.concurrency,
 			outputs,
 			settled,
-			async (task, index) => {
-				const input = inputs[index]!;
+			async (task) => {
+				// The scheduler passes the index into the FILTERED list — never use it
+				// against the unfiltered inputs. Look the input up by task id instead.
+				const input = inputById.get(task.id);
+				if (!input) {
+					// Impossible unless ids drift from inputs — fail loudly instead of
+					// leaving the task queued forever (hasActiveRun would never clear).
+					this.updateTask(
+						run,
+						task,
+						{ status: "failed", error: `No input for task ${task.id}`, endedAt: Date.now() },
+						ctx,
+						onUpdate,
+					);
+					return;
+				}
 				await this.runChild(
 					run,
 					task,
@@ -1031,14 +1125,26 @@ export class SubagentManager {
 		for (const task of run.tasks) this.mailboxes.close(`${run.id}:${task.id}`);
 		this.persist(ctx);
 		// Branches merged by the leader since the run ended: drop worktree dir + branch.
-		for (const task of run.tasks) {
-			if (task.branch) {
+		// Once per repo, never for a branch another live run owns, never fatal — a
+		// throw here would re-settle an already-finished run as failed.
+		try {
+			const roots = new Set<string>();
+			for (const task of run.tasks) {
+				if (!task.branch) continue;
 				const root = repoRoot(task.cwd);
-				if (root) cleanupMerged(root);
+				if (root) roots.add(root);
 			}
+			for (const root of roots) cleanupMerged(root, { skipBranches: this.liveBranches() });
+		} catch {
+			/* cleanup is best-effort; the run outcome must stand */
 		}
 	}
 
+	/** Branches owned by worktrees of still-running children. */
+	liveBranches(): Set<string> {
+		return new Set(Array.from(this.liveWorktrees.values(), (wt) => wt.branch));
+	}
+
 	/** Spawn a run that keeps executing after this call returns. Every run is background. */
 	startInBackground(params: SubagentParamsShape, ctx: ExtensionContext): RunDetails {
 		const { run, inputs } = this.createRun(params, ctx);
@@ -1113,7 +1219,10 @@ export class SubagentManager {
 			task.error = task.error || "Canceled by subagent_cancel"; // never overwrite a real error
 			task.endedAt = Date.now();
 			aborted += 1;
-			if (task.branch) removeByBranch(task.cwd, task.branch); // dir only; branch keeps partial work
+			// The branch is recorded here so the leader can still merge partial work;
+			// runChild's finally commits + drops the dir (it owns the live worktree).
+			const wt = this.liveWorktrees.get(`${runId}:${task.id}`);
+			if (wt) task.branch = wt.branch;
 		}
 		run.status = "aborted";
 		run.endedAt = Date.now();
@@ -1168,7 +1277,7 @@ export class SubagentManager {
 			// IN the await result, no steering queue, no turn boundary needed.
 			this.parked.set(runId, { msgs, wake: () => resolve(cloneRun(run)) });
 		});
-		if (timeoutMs) {
+		if (timeoutMs !== undefined && timeoutMs > 0) {
 			return Promise.race([
 				settled.then((r) => {
 					finish();

package/src/worktree.ts

@@ -2,21 +2,26 @@
  *  Worktrees live inside `<repo>/.git/subagents/<runId>/<taskId>` so the child's
  *  ancestor walk still finds the project AGENTS.md chain. node_modules is
  *  symlinked from the main tree. The extension commits the child's changes on
- *  completion; the leader reviews and merges the branch manually; merged
- *  branches are cleaned automatically, crash leftovers are swept at session
- *  start (dir removed, branch kept — the work survives). */
+ *  completion; the leader reviews and merges the branch manually.
+ *
+ *  Cleanup, in order of trust: `reapDeadWorktrees` (session start — nothing can
+ *  be live yet, so every registered worktree is a crash leftover: commit its
+ *  work, drop the dir, keep the branch), `cleanupMerged` (merged branches, never
+ *  touching a checked-out one), `sweepStale` (dirs git no longer knows about). */
 
 import { execFileSync } from "node:child_process";
-import { existsSync, readdirSync, rmSync, symlinkSync } from "node:fs";
+import { existsSync, readdirSync, realpathSync, rmSync, symlinkSync } from "node:fs";
 import { join } from "node:path";
 
 export interface Worktree {
 	root: string; // repo root (main tree)
 	path: string; // worktree checkout dir
 	branch: string; // subagents/<runId>/<taskId>
-	base: string; // branch HEAD was created from
+	base: string; // SHA the branch was created from
 }
 
+const BRANCH_PREFIX = "subagents/";
+
 function git(root: string, args: string[]): string {
 	return execFileSync("git", ["-C", root, ...args], { encoding: "utf8" }).trim();
 }
@@ -35,6 +40,25 @@ 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;
+	try {
+		return realpathSync(a) === realpathSync(b);
+	} catch {
+		return false;
+	}
+}
+
 /** Repo root for cwd, or undefined when not a git repo (or cwd doesn't exist). */
 export function repoRoot(cwd: string): string | undefined {
 	if (!existsSync(cwd)) return undefined;
@@ -50,12 +74,12 @@ export function createWorktree(cwd: string, runId: string, taskId: string): Work
 	const root = repoRoot(cwd);
 	if (!root) return undefined;
 	const path = join(root, ".git", "subagents", runId, taskId);
-	const branch = `subagents/${runId}/${taskId}`;
+	const branch = `${BRANCH_PREFIX}${runId}/${taskId}`;
 	let base: string;
 	try {
-		base = git(root, ["rev-parse", "--abbrev-ref", "HEAD"]);
+		base = git(root, ["rev-parse", "HEAD"]); // SHA — detached HEAD stays correct
 	} catch {
-		return undefined; // detached or broken repo — fall back to in-place
+		return undefined; // broken repo — fall back to in-place
 	}
 	git(root, ["worktree", "add", "-b", branch, path, "HEAD"]);
 	// Deps follow the child into the worktree; anything else the task needs is
@@ -71,14 +95,22 @@ export function createWorktree(cwd: string, runId: string, taskId: string): Work
 	return { root, path, branch, base };
 }
 
-/** Commit all child changes. No-op when the worktree is already clean. */
+/**
+ * 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.
+ */
 export function commitWorktree(wt: Worktree, message: string): void {
-	if (gitIn(wt.path, ["status", "--porcelain"]).length === 0) return;
-	gitIn(wt.path, ["add", "-A"]);
-	gitIn(wt.path, ["commit", "-m", message, "--no-verify"]);
+	commitIn(wt.path, message);
+}
+
+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"]);
 }
 
-/** Diffstat + changed files of the branch vs its base. */
+/** Diffstat + changed files of the branch vs its base SHA. */
 export function branchDiff(wt: Worktree): { stat: string; files: string[] } {
 	const files = git(wt.root, ["diff", "--name-only", `${wt.base}...${wt.branch}`])
 		.split("\n")
@@ -89,34 +121,59 @@ export function branchDiff(wt: Worktree): { stat: string; files: string[] } {
 
 /** Remove the worktree dir. The branch is KEPT (the work survives for merging). */
 export function removeWorktree(wt: Worktree): void {
-	try {
-		git(wt.root, ["worktree", "remove", "--force", wt.path]);
-	} catch {
-		/* already gone */
-	}
+	dropDir(wt.root, wt.path);
 }
 
 /** Remove a worktree dir by branch name (cancel paths that didn't keep a Worktree). */
 export function removeByBranch(cwd: string, branch: string): void {
+	if (!branch.startsWith(BRANCH_PREFIX)) return;
 	const root = repoRoot(cwd);
 	if (!root) return;
-	const path = join(root, ".git", "subagents", branch.slice("subagents/".length));
+	dropDir(root, join(root, ".git", "subagents", branch.slice(BRANCH_PREFIX.length)));
+}
+
+function dropDir(root: string, path: string): void {
 	try {
 		git(root, ["worktree", "remove", "--force", path]);
 	} catch {
-		/* already gone */
+		if (existsSync(path)) rmSync(path, { recursive: true, force: true });
+	}
+	prune(root);
+}
+
+/** Drop git's stale worktree admin entries (they pile up under .git/worktrees). */
+function prune(root: string): void {
+	try {
+		git(root, ["worktree", "prune"]);
+	} catch {
+		/* ignore */
 	}
 }
 
-/** Delete branch + worktree for branches already merged into `target`. */
-export function cleanupMerged(root: string, target = "HEAD"): number {
+/** Is this branch checked out in some worktree right now? Checked fresh, per call. */
+function isCheckedOut(root: string, branch: string): boolean {
+	return worktreeBranches(root).includes(branch);
+}
+
+/**
+ * Delete branch + worktree for branches already merged into `target`.
+ * SAFETY: a branch checked out in a LIVE worktree is skipped — a fresh branch's
+ * tip equals its base until the child commits, so it looks "merged". The
+ * registration is re-checked immediately before each removal (a concurrent run
+ * may have created its worktree after the first listing).
+ */
+export function cleanupMerged(root: string, opts: { skipBranches?: Set<string>; target?: string } = {}): number {
+	root = realpathSync(root);
+	const target = opts.target ?? "HEAD";
 	const merged = git(root, ["branch", "--merged", target])
 		.split("\n")
 		.map((b) => b.trim().replace(/^[+*]\s*/, ""));
 	let cleaned = 0;
 	for (const branch of merged) {
-		if (!branch.startsWith("subagents/")) continue;
-		const path = join(root, ".git", "subagents", branch.slice("subagents/".length));
+		if (!branch.startsWith(BRANCH_PREFIX)) 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));
 		if (existsSync(path)) {
 			try {
 				git(root, ["worktree", "remove", "--force", path]);
@@ -126,30 +183,88 @@ export function cleanupMerged(root: string, target = "HEAD"): number {
 		}
 		if (gitOk(root, ["branch", "-d", branch])) cleaned += 1;
 	}
+	prune(root);
 	return cleaned;
 }
 
-/** Remove worktree dirs of branches that no longer exist (crash leftovers). */
+/** 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 [];
+	}
+}
+
+/**
+ * Session-start recovery: every registered subagent worktree is a crash leftover
+ * (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 {
+	root = realpathSync(root);
+	const sub = join(root, ".git", "subagents");
+	if (!existsSync(sub)) return 0;
+	let reaped = 0;
+	for (const path of worktreePaths(root)) {
+		if (!isInside(path, sub)) continue; // not ours
+		try {
+			commitIn(path, "subagent (recovered after interrupted session)");
+		} catch {
+			/* nothing committable */
+		}
+		dropDir(root, path);
+		reaped += 1;
+	}
+	return reaped;
+}
+
+/**
+ * Remove worktree dirs that git no longer knows about (partial-crash leftovers).
+ * Registered worktrees are never touched here — `reapDeadWorktrees` owns those,
+ * and a live child's checkout must survive.
+ */
 export function sweepStale(root: string): void {
+	root = realpathSync(root);
 	const sub = join(root, ".git", "subagents");
 	if (!existsSync(sub)) return;
-	const branches = new Set(
-		git(root, ["branch", "--list", "subagents/*"])
-			.split("\n")
-			.map((b) => b.trim().replace(/^[+*]\s*/, "")),
-	);
+	const registered = worktreePaths(root);
 	for (const runDir of readDirs(sub)) {
 		for (const taskDir of readDirs(join(sub, runDir))) {
-			const branch = `subagents/${runDir}/${taskDir}`;
-			if (branches.has(branch)) continue; // live branch, dir may be an active worktree
-			try {
-				git(root, ["worktree", "remove", "--force", join(sub, runDir, taskDir)]);
-			} catch {
-				// dir not a registered worktree (partial crash) — plain rm
-				rmSync(join(sub, runDir, taskDir), { recursive: true, force: true });
-			}
+			const dir = join(sub, runDir, taskDir);
+			if (registered.some((p) => samePath(p, dir))) continue; // live/registered worktree
+			rmSync(dir, { recursive: true, force: true });
 		}
 	}
+	prune(root);
+}
+
+function worktreePaths(root: string): string[] {
+	try {
+		return git(root, ["worktree", "list", "--porcelain"])
+			.split("\n")
+			.filter((l) => l.startsWith("worktree "))
+			.map((l) => unquote(l.slice("worktree ".length).trim()));
+	} catch {
+		return [];
+	}
+}
+
+function isInside(path: string, dir: string): boolean {
+	const p = safeReal(path);
+	const d = safeReal(dir);
+	return p === d || p.startsWith(`${d}/`);
+}
+
+function safeReal(path: string): string {
+	try {
+		return realpathSync(path);
+	} catch {
+		return path;
+	}
 }
 
 function readDirs(dir: string): string[] {