@lifeaitools/clauth 1.19.4 → 1.30.2

package/cli/commands/agent-pool.js

@@ -1,1745 +1,1946 @@
-// agent-pool.js — warm Haiku worker pool for call_agent one-shot dispatch.
-//
-// Design (Gate B, WP-B1/B3/B4) — cold path:
-//   * Each `claude -p "<prompt>"` is inherently a fresh detached process — the
-//     prompt is passed as argv, so an idle pre-spawned process cannot be
-//     re-targeted. The cold "warm pool" therefore does two things:
-//       1. Resolves + caches the claude binary once (no per-call `--version`
-//          probe) and keeps a neutral, project-free cwd hot on disk.
-//       2. Optionally fires `poolSize` priming pings at startup so the OS file
-//          cache / model session is warm — the FIRST real call then returns in
-//          ~4-5s instead of the ~50-90s cold path.
-//
-// WP-P2 — TRUE persistent re-targetable warm workers:
-//   * The `claude` CLI (v2.1.170) supports `-p --input-format stream-json
-//     --output-format stream-json`: a single long-lived process that reads
-//     newline-delimited user-message envelopes from stdin and emits
-//     newline-delimited events (including a `result` per turn) on stdout. The
-//     SAME process (same session_id) answers MULTIPLE sequential prompts —
-//     proven: 3 prompts on one process, sid stable, 4.4s → 2.0s → 2.8s with NO
-//     per-call CLI boot. This is true re-targeting: a pre-spawned idle worker IS
-//     re-aimable at a new prompt over stdin.
-//   * `WarmWorker` wraps one such process. `AgentPool.dispatch()` prefers a free
-//     warm worker (sub-boot dispatch — latency is pure model inference, not the
-//     ~10-15s CLI cold-boot) and transparently FALLS BACK to the cold spawn path
-//     when warm workers are disabled, all busy, or unavailable. Fully backward
-//     compatible: the public dispatch()/job-record contract is unchanged.
-//   * Warm workers are health-checked and AUTO-RESTARTED on death; the reaper
-//     enforces per-turn deadlines on warm turns too. Model-tiered: each pool is
-//     pinned to one model (default Haiku); a Sonnet/Opus pool is a second
-//     AgentPool instance.
-//
-//   * Lean one-shot: neutral cwd + `--setting-sources user` + `--strict-mcp-config`
-//     strips the regen-root CLAUDE.md / SessionEnd hooks / lessons gate that
-//     otherwise contaminates replies ("Waived." instead of the real answer).
-//   * Auth: piggybacks the CLI's own login session. NEVER sets ANTHROPIC_API_KEY.
-//   * Reaping: a periodic sweep kills workers that exceed their timeout and
-//     frees concurrency slots so the queue self-heals.
-//
-// This module is intentionally self-contained and daemon-free so its core can
-// be unit/integration tested WITHOUT the clauth vault or boot.key.
-
-import { spawn } from "child_process";
-import { execSync } from "child_process";
-import os from "os";
-import path from "path";
-import fs from "fs";
-import crypto from "crypto";
-
-export const DEFAULT_MODEL = "claude-haiku-4-5";
-export const DEFAULT_TIMEOUT_MS = 120000;
-export const DEFAULT_POOL_SIZE = 2;
-// Warm workers are ON by default. Set CLAUTH_AGENT_WARM=0 to force the cold path.
-export const WARM_ENABLED_DEFAULT =
-  String(process.env.CLAUTH_AGENT_WARM ?? "1") !== "0";
-
-// ── WP-P7: hop-limited sub-agent delegation defaults ─────────────────────────
-// A parent call_agent worker may itself invoke call_agent SUB-agents, bounded by
-// a hop budget and routed through a SEPARATE execution lane (DelegationLane) so
-// it can NEVER consume the shared warm/interactive pool slots (anti-deadlock).
-export const DEFAULT_DELEGATION_MAX_HOPS = 2; // tree depth cap (hop 0 = top call)
-export const DEFAULT_DELEGATION_MAX_INFLIGHT = 4; // daemon-wide concurrent sub-agents
-
-const LEAN_SYSTEM_PROMPT =
-  "You are a stateless one-shot agent worker invoked via clauth call_agent. " +
-  "Answer the user's prompt directly, concisely, and completely. " +
-  "Do not editorialize, do not read project files unless the prompt asks, " +
-  "and do not emit meta-commentary about hooks, sessions, or governance. " +
-  "If a JSON answer is requested, return ONLY valid JSON.";
-
-// ── Tiered call_agent startup (bootstrap) ────────────────────────────────────
-// A spawned worker otherwise has NO identity and NO idea what it can reach (no
-// clauth creds, no Supabase, no repo). These three tiers fix that, additively:
-//
-//   * "none"     — today's lean fast path, byte-identical to LEAN_SYSTEM_PROMPT.
-//                  For trivial one-shots; warm ~2s. Neutral, project-free cwd.
-//   * "identity" — DEFAULT. Prepends a SHORT startup that tells the agent who it
-//                  is and HOW to reach access (clauth daemon, Supabase ref,
-//                  work_items RPCs, Bash/file tools). Neutral cwd. Low latency.
-//   * "full"     — for real repo/DB work: runs in the REPO cwd and points the
-//                  agent at .rdc/guides/agent-bootstrap.md (creds, git rules,
-//                  completion report) plus the identity/access block.
-//
-// IMPORTANT: the startup TELLS the agent how to fetch credentials; it NEVER
-// embeds a credential and NEVER sets ANTHROPIC_API_KEY.
-export const BOOTSTRAP_LEVELS = ["none", "identity", "full"];
-export const DEFAULT_BOOTSTRAP = "identity";
-export const DEFAULT_AGENT_REPO_ROOT =
-  process.env.CLAUTH_AGENT_REPO_ROOT || "C:/Dev/regen-root";
-
-// The clauth credential daemon every Regen agent uses.
-const CLAUTH_DAEMON_URL = "http://127.0.0.1:52437";
-const SUPABASE_PROJECT_REF = "uvojezuorjgqzmhhgluu";
-
-// Short identity/access block — a dozen lines so latency stays low. This is the
-// minimum every default agent must know: who it is + how to get access.
-const IDENTITY_SYSTEM_PROMPT =
-  "You are a Regen agent running on the operator's machine. " +
-  "ACCESS — discover service names through the clauth HTTP daemon at " +
-  `${CLAUTH_DAEMON_URL}: run \`curl -s ${CLAUTH_DAEMON_URL}/knowledge\` for ` +
-  "sanitized metadata, then fetch exactly one needed credential with " +
-  `\`curl -s ${CLAUTH_DAEMON_URL}/v/<service>\`. ` +
-  "Do not bulk-fetch secrets. NEVER print a credential to stdout — it is logged. " +
-  `The Supabase project ref is ${SUPABASE_PROJECT_REF}. ` +
-  "Work items live in Supabase and are accessed ONLY through RPCs " +
-  "(get_open_epics, insert_work_item, update_work_item_status) — never raw " +
-  "SQL CRUD against work_items. You have Bash plus file read/write tools. " +
-  "Answer the user's prompt directly and completely; if JSON is requested, " +
-  "return ONLY valid JSON.";
-
-/**
- * Build the "full" startup: identity/access block + a pointer to the project
- * bootstrap guide that the agent must read first. Runs in the repo cwd.
- * @param {string} [repoRoot]
- */
-function buildFullSystemPrompt(repoRoot = DEFAULT_AGENT_REPO_ROOT) {
-  const root = (repoRoot || DEFAULT_AGENT_REPO_ROOT).replace(/\\/g, "/");
-  return (
-    IDENTITY_SYSTEM_PROMPT +
-    " You are running in the repo at " +
-    root +
-    ". Before doing repo or database work, READ " +
-    root +
-    "/.rdc/guides/agent-bootstrap.md FIRST — it has the full credential, " +
-    "Supabase, git (commit to develop, never main, never force-push), and " +
-    "completion-report contract. Then read your role guide under " +
-    root +
-    "/.rdc/guides/ if one applies."
-  );
-}
-
-// ── Corpus handoff (Regen Agent corpus, matching @regen/claude-dispatch) ─────
-// Every Regen route agent dispatched through @regen/claude-dispatch gets a
-// `## Corpus` section assembled into its prompt body from the dispatch
-// envelope's `corpus` field. call_agent previously gave only a thin identity
-// system prompt and NO corpus. composeCorpus() reproduces the EXACT dispatch
-// format (packages/claude-dispatch/src/channels/cli.ts buildPrompt) so a
-// call_agent worker starts up with the same handoff a route agent gets.
-//
-// Shape (DispatchCorpus): { product?: string, design?: string, docs?: CorpusDocument[] }
-// CorpusDocument: { path: string, title?: string, content: string }
-//
-// CRITICAL — warm-reuse correctness: the corpus is assembled into the per-turn
-// PROMPT (the stdin/`-p` user message), NOT the worker's persistent
-// append-system-prompt. A warm worker's system prompt is fixed at spawn and
-// shared across all turns; corpus content is variable per call. Because the
-// corpus rides the turn's user message (composeCallAgentPrompt output) and not
-// the worker identity, a generic warm worker can serve a corpus call correctly
-// — the corpus is scoped to that single turn and never leaks into the next.
-
-// Per-document and total corpus byte caps so an oversized handoff cannot blow
-// the agent's context. Each doc/field is truncated to MAX_CORPUS_DOC_CHARS and
-// the assembled section to MAX_CORPUS_TOTAL_CHARS; truncations are logged (not
-// silently dropped). Overridable via env for ops tuning.
-export const MAX_CORPUS_DOC_CHARS = Number.isFinite(Number(process.env.CLAUTH_AGENT_CORPUS_DOC_CHARS))
-  ? Number(process.env.CLAUTH_AGENT_CORPUS_DOC_CHARS)
-  : 24000;
-export const MAX_CORPUS_TOTAL_CHARS = Number.isFinite(Number(process.env.CLAUTH_AGENT_CORPUS_TOTAL_CHARS))
-  ? Number(process.env.CLAUTH_AGENT_CORPUS_TOTAL_CHARS)
-  : 60000;
-
-const BOOTSTRAP_DOC_REL_PATH = ".rdc/guides/agent-bootstrap.md";
-
-/** Truncate a string to `cap` chars, appending a visible marker + logging. */
-function capText(text, cap, label) {
-  if (typeof text !== "string") return "";
-  if (text.length <= cap) return text;
-  const truncated = text.slice(0, cap);
-  console.error(
-    `[agent-pool] corpus ${label} truncated: ${text.length} → ${cap} chars (cap MAX_CORPUS_DOC_CHARS)`
-  );
-  return `${truncated}\n…[truncated ${text.length - cap} chars]`;
-}
-
-/**
- * Load the project bootstrap guide CONTENT from disk so a "full"-tier agent
- * gets the REAL bootstrap inline as a Supporting Doc — not just a "go read this
- * file" pointer. Returns a CorpusDocument or null if unreadable (never throws).
- * @param {string} [repoRoot]
- * @returns {{ path: string, title: string, content: string } | null}
- */
-export function loadBootstrapDoc(repoRoot = DEFAULT_AGENT_REPO_ROOT) {
-  const root = (repoRoot || DEFAULT_AGENT_REPO_ROOT).replace(/\\/g, "/");
-  const abs = path.join(root, BOOTSTRAP_DOC_REL_PATH);
-  try {
-    const content = fs.readFileSync(abs, "utf8");
-    if (!content || !content.trim()) return null;
-    return { path: BOOTSTRAP_DOC_REL_PATH, title: "Agent Bootstrap", content };
-  } catch (e) {
-    console.error(`[agent-pool] bootstrap doc unreadable (${abs}): ${e.message} — skipping`);
-    return null;
-  }
-}
-
-/**
- * Assemble the `## Corpus` section EXACTLY as @regen/claude-dispatch does
- * (buildPrompt). Returns the full section string beginning with "## Corpus\n",
- * or "" when nothing to emit. Order: Product MD, Design MD, Supporting Docs.
- *
- * Format (byte-for-byte with dispatch):
- *   ## Corpus
- *   ### Product MD\n<product>
- *
- *   ### Design MD\n<design>
- *
- *   ### Supporting Docs
- *   - <path> (<title>)\n```\n<content>\n```      ← per doc; title parenthetical
- *                                                  omitted if no title; docs
- *                                                  joined by "\n"
- * (corpusSections joined by "\n\n"; section is "## Corpus" + "\n" + body.)
- *
- * @param {{product?:string, design?:string, docs?:Array<{path:string,title?:string,content:string}>}|null|undefined} corpus
- * @returns {string}
- */
-export function composeCorpus(corpus) {
-  if (!corpus || typeof corpus !== "object") return "";
-  const corpusSections = [];
-  if (typeof corpus.product === "string" && corpus.product) {
-    corpusSections.push(`### Product MD\n${capText(corpus.product, MAX_CORPUS_DOC_CHARS, "product")}`);
-  }
-  if (typeof corpus.design === "string" && corpus.design) {
-    corpusSections.push(`### Design MD\n${capText(corpus.design, MAX_CORPUS_DOC_CHARS, "design")}`);
-  }
-  if (Array.isArray(corpus.docs) && corpus.docs.length > 0) {
-    const docs = corpus.docs
-      .filter((d) => d && typeof d.path === "string" && typeof d.content === "string")
-      .map((doc) => {
-        const titleLine = doc.title ? ` (${doc.title})` : "";
-        const content = capText(doc.content, MAX_CORPUS_DOC_CHARS, `doc ${doc.path}`);
-        return `- ${doc.path}${titleLine}\n\`\`\`\n${content}\n\`\`\``;
-      })
-      .join("\n");
-    if (docs) corpusSections.push(`### Supporting Docs\n${docs}`);
-  }
-  if (corpusSections.length === 0) return "";
-  let section = `## Corpus\n${corpusSections.join("\n\n")}`;
-  if (section.length > MAX_CORPUS_TOTAL_CHARS) {
-    console.error(
-      `[agent-pool] corpus total truncated: ${section.length} → ${MAX_CORPUS_TOTAL_CHARS} chars (cap MAX_CORPUS_TOTAL_CHARS)`
-    );
-    section = `${section.slice(0, MAX_CORPUS_TOTAL_CHARS)}\n…[corpus truncated]`;
-  }
-  return section;
-}
-
-/**
- * Build the effective corpus for a call: the caller's corpus (if any), with the
- * bootstrap doc auto-appended as a Supporting Doc when bootstrap === "full" and
- * the file is readable. Caller docs come first, then the bootstrap doc, matching
- * the merge contract (caller docs + bootstrap doc). Returns the assembled
- * `## Corpus` section string ("" when nothing to emit).
- *
- * @param {object} opts
- * @param {string} opts.bootstrap                 normalized level
- * @param {object} [opts.corpus]                  caller-supplied DispatchCorpus
- * @param {string} [opts.repoRoot]                repo root for the bootstrap doc
- * @param {function} [opts.loadDoc]               injectable loader (tests)
- * @returns {string}
- */
-export function buildCorpusSection({ bootstrap, corpus, repoRoot, loadDoc } = {}) {
-  const caller = corpus && typeof corpus === "object" ? corpus : null;
-  let effective = caller;
-  if (bootstrap === "full") {
-    const load = typeof loadDoc === "function" ? loadDoc : loadBootstrapDoc;
-    const bootDoc = load(repoRoot);
-    if (bootDoc) {
-      const callerDocs = Array.isArray(caller?.docs) ? caller.docs : [];
-      effective = {
-        ...(caller || {}),
-        docs: [...callerDocs, bootDoc],
-      };
-    }
-  }
-  return composeCorpus(effective);
-}
-
-/**
- * Normalize a requested bootstrap level to one of BOOTSTRAP_LEVELS.
- * Unknown/absent → DEFAULT_BOOTSTRAP ("identity").
- */
-export function normalizeBootstrap(level) {
-  if (typeof level !== "string") return DEFAULT_BOOTSTRAP;
-  const l = level.trim().toLowerCase();
-  return BOOTSTRAP_LEVELS.includes(l) ? l : DEFAULT_BOOTSTRAP;
-}
-
-/**
- * Compose the call_agent startup for a bootstrap level. Returns the
- * append-system-prompt text plus the cwd policy for that tier:
- *   - none/identity → neutral cwd (caller passes its neutral dir)
- *   - full          → the repo cwd (so the agent can read the bootstrap guide)
- *
- * @param {string} level                "none" | "identity" | "full"
- * @param {object} [opts]
- * @param {string} [opts.repoRoot]      repo root for the "full" tier
- * @param {string} [opts.extra]         optional caller append_system_prompt,
- *                                      concatenated after the tier text
- * @returns {{ level:string, appendSystemPrompt:string, useRepoCwd:boolean, repoRoot:string|null }}
- */
-export function composeBootstrap(level, { repoRoot, extra } = {}) {
-  const lvl = normalizeBootstrap(level);
-  let base;
-  let useRepoCwd = false;
-  let root = null;
-  if (lvl === "none") {
-    base = LEAN_SYSTEM_PROMPT;
-  } else if (lvl === "full") {
-    root = (repoRoot || DEFAULT_AGENT_REPO_ROOT).replace(/\\/g, "/");
-    base = buildFullSystemPrompt(root);
-    useRepoCwd = true;
-  } else {
-    base = IDENTITY_SYSTEM_PROMPT; // identity (default)
-  }
-  const appendSystemPrompt =
-    extra && typeof extra === "string" && extra.trim()
-      ? `${base}\n\n${extra.trim()}`
-      : base;
-  return { level: lvl, appendSystemPrompt, useRepoCwd, repoRoot: root };
-}
-
-function resolveClaudeBinary(explicit) {
-  const candidates = [
-    explicit,
-    process.env.CLAUDE_BIN,
-    path.join(os.homedir(), ".local", "bin", process.platform === "win32" ? "claude.exe" : "claude"),
-    path.join(process.env.APPDATA || "", "npm", "claude.cmd"),
-    path.join(process.env.APPDATA || "", "npm", "claude"),
-    "claude",
-  ].filter(Boolean);
-  for (const c of candidates) {
-    try {
-      execSync(`"${c}" --version`, { stdio: "ignore", timeout: 5000 });
-      return c;
-    } catch {
-      /* try next */
-    }
-  }
-  return null;
-}
-
-function makeJobId(prefix = "call") {
-  return `${prefix}-${Date.now()}-${crypto.randomBytes(4).toString("hex")}`;
-}
-
-/**
- * Build the persistent stream-json argv for a warm worker. Same lean isolation
- * as the cold buildArgs() (model, skip-permissions, user-only settings, strict
- * mcp, lean system prompt) but with the persistent stdin/stdout streaming
- * transport instead of a one-shot `-p "<prompt>"`.
- */
-function buildWarmArgs({ model, appendSystemPrompt }) {
-  return [
-    "-p",
-    "--input-format",
-    "stream-json",
-    "--output-format",
-    "stream-json",
-    "--verbose", // required by the CLI for stream-json output
-    "--model",
-    model,
-    "--dangerously-skip-permissions",
-    "--setting-sources",
-    "user",
-    "--strict-mcp-config",
-    "--append-system-prompt",
-    appendSystemPrompt || LEAN_SYSTEM_PROMPT,
-  ];
-}
-
-/**
- * WarmWorker — one long-lived `claude` process in stream-json mode that can be
- * RE-TARGETED with a new prompt per turn over stdin. This is the true persistent
- * re-targetable worker: no per-call CLI boot.
- *
- * Lifecycle: spawned → ready (process is alive and stdin is connected) → busy
- * (a turn is in flight) → idle again (after that turn's `result`). On process
- * death it is marked dead; the pool auto-restarts it. One turn at a time per
- * worker (the CLI session is single-threaded); the pool routes each call to a
- * free idle worker.
- *
- * READINESS NOTE: the `claude` CLI in `--input-format stream-json` mode emits
- * NOTHING (not even `system/init`) until the FIRST stdin user-message arrives —
- * `init` is part of the first turn's event stream, not a standalone boot signal.
- * So a warm worker is "ready" the moment its process is spawned with a live
- * stdin pipe. The first `send()` pays the CLI boot/init cost inside its turn;
- * every subsequent `send()` re-targets the already-booted process (pure
- * inference latency, no boot). That is the warm win: boot is paid ONCE, then the
- * hot process is re-aimed per call.
- */
-export class WarmWorker {
-  /**
-   * @param {object} opts
-   * @param {string} opts.binary           resolved claude binary
-   * @param {string} opts.model            model pinned for this worker
-   * @param {string} opts.cwd              neutral project-free cwd
-   * @param {string} [opts.appendSystemPrompt]
-   * @param {function} [opts.spawnImpl]    inject a fake spawn for tests
-   * @param {function} [opts.onExit]       called when the process dies
-   */
-  constructor(opts = {}) {
-    this.binary = opts.binary;
-    this.model = opts.model || DEFAULT_MODEL;
-    this.cwd = opts.cwd;
-    this.appendSystemPrompt = opts.appendSystemPrompt;
-    this._spawn = opts.spawnImpl || spawn;
-    this.onExit = typeof opts.onExit === "function" ? opts.onExit : () => {};
-
-    this.id = `warm-${crypto.randomBytes(4).toString("hex")}`;
-    this.proc = null;
-    this.sessionId = null;
-    this.state = "spawning"; // spawning | ready | busy | dead
-    this.spawnedAt = Date.now();
-    this.turns = 0;
-
-    this._buf = "";
-    this._current = null; // { resolve, stdout, deadline, jobId, settled }
-    this._readyResolve = null;
-    this._readyPromise = new Promise((r) => (this._readyResolve = r));
-  }
-
-  /** True iff the worker is alive and free to accept a new turn. */
-  isIdle() {
-    return this.state === "ready";
-  }
-  isBusy() {
-    return this.state === "busy";
-  }
-  isDead() {
-    return this.state === "dead";
-  }
-
-  /** Spawn the persistent process. Resolves once the worker is ready (or dead). */
-  async start() {
-    let proc;
-    try {
-      const isCmdShim = process.platform === "win32" && /\.cmd$/i.test(this.binary);
-      const args = buildWarmArgs({ model: this.model, appendSystemPrompt: this.appendSystemPrompt });
-      const command = isCmdShim ? "cmd" : this.binary;
-      const spawnArgs = isCmdShim ? ["/d", "/s", "/c", `"${this.binary}"`, ...args] : args;
-      proc = this._spawn(command, spawnArgs, {
-        cwd: this.cwd,
-        env: process.env, // inherits CLI login session; never sets ANTHROPIC_API_KEY
-        stdio: ["pipe", "pipe", "pipe"],
-        shell: false,
-        windowsHide: true,
-      });
-    } catch (e) {
-      this.state = "dead";
-      this.error = `spawn_failed: ${e.message}`;
-      this._readyResolve?.(false);
-      return false;
-    }
-
-    this.proc = proc;
-    this.pid = proc.pid;
-
-    proc.stdout?.on("data", (d) => this._onStdout(d));
-    proc.stderr?.on("data", (d) => {
-      this._stderr = `${this._stderr || ""}${d.toString()}`.slice(-20000);
-    });
-    proc.on("error", (e) => this._onDeath(`proc_error: ${e.message}`));
-    proc.on("close", (code) => this._onDeath(`exit_${code}`));
-
-    // In stream-json mode the CLI is silent until the first stdin message, so a
-    // freshly-spawned process with a live stdin pipe IS ready. Mark ready on the
-    // next tick (after spawn errors would have fired synchronously). The first
-    // send() pays the boot cost inside its turn; later sends are hot re-targets.
-    await new Promise((r) => setImmediate(r));
-    if (this.state === "spawning" && this.proc && this.proc.stdin && !this.isDead()) {
-      this.state = "ready";
-      this._readyResolve?.(true);
-    }
-    return await this._readyPromise;
-  }
-
-  _onStdout(d) {
-    this._buf += d.toString();
-    let idx;
-    while ((idx = this._buf.indexOf("\n")) >= 0) {
-      const line = this._buf.slice(0, idx).trim();
-      this._buf = this._buf.slice(idx + 1);
-      if (!line) continue;
-      let msg;
-      try {
-        msg = JSON.parse(line);
-      } catch {
-        continue; // ignore non-JSON noise
-      }
-      this._onEvent(msg);
-    }
-  }
-
-  _onEvent(msg) {
-    if (msg.session_id && !this.sessionId) this.sessionId = msg.session_id;
-    if (msg.type === "system" && msg.subtype === "init") {
-      if (this.state === "spawning") {
-        this.state = "ready";
-        this._readyResolve?.(true);
-      }
-      return;
-    }
-    if (msg.type === "result" && this._current) {
-      const cur = this._current;
-      if (cur.settled) return;
-      cur.settled = true;
-      if (cur.timer) clearTimeout(cur.timer); // C2: cancel internal timeout on normal settle
-      const text = typeof msg.result === "string" ? msg.result : cur.stdout;
-      this._current = null;
-      if (this.state === "busy") this.state = "ready";
-      cur.resolve({
-        ok: !msg.is_error,
-        package: (text || "").trim(),
-        is_error: !!msg.is_error,
-        session_id: this.sessionId,
-      });
-    }
-  }
-
-  _onDeath(reason) {
-    if (this.state === "dead") return;
-    this.state = "dead";
-    this.error = reason;
-    // Fail any in-flight turn so the caller (and concurrency slot) is released.
-    if (this._current && !this._current.settled) {
-      const cur = this._current;
-      cur.settled = true;
-      if (cur.timer) clearTimeout(cur.timer); // C2: cancel internal timeout on death
-      this._current = null;
-      cur.resolve({ ok: false, package: "", error: `worker_died: ${reason}`, dead: true });
-    }
-    // Unblock a start() still waiting on readiness.
-    if (this.state === "dead") this._readyResolve?.(false);
-    this.onExit(this, reason);
-  }
-
-  /**
-   * Re-target this warm worker with a new prompt. One turn at a time.
-   * Resolves with { ok, package, is_error } when the turn's `result` arrives,
-   * or { ok:false, ... } if the deadline elapses (caller/pool kills the worker).
-   * @returns {Promise<object>}
-   */
-  send(prompt, { timeout_ms } = {}) {
-    if (this.state !== "ready") {
-      return Promise.resolve({ ok: false, error: `worker_not_ready:${this.state}` });
-    }
-    return new Promise((resolve) => {
-      this.state = "busy";
-      this.turns++;
-      const ms = Number.isFinite(timeout_ms) ? timeout_ms : DEFAULT_TIMEOUT_MS;
-      const cur = { resolve, stdout: "", settled: false, startedAt: Date.now() };
-      cur.deadline = Date.now() + ms;
-      // C2 — own internal timeout. The pool reaper only covers turns that have a
-      // job record; the PRIMING turn (ensureWarmWorkers) has none, so a hung
-      // prime would otherwise leave `send()`'s promise unsettled forever and
-      // `await Promise.all(primes)` would never resolve. This timer settles the
-      // turn with a failure AND kills the wedged process independently of the
-      // reaper, hardening every send() (priming and normal) against a hung CLI.
-      cur.timer = setTimeout(() => {
-        if (cur.settled) return;
-        cur.settled = true;
-        // Detach so a late `result` event can't double-settle.
-        if (this._current === cur) this._current = null;
-        cur.resolve({ ok: false, package: "", error: "warm_send_timeout", timeout: true });
-        // The session is wedged mid-turn — kill the proc so the worker is
-        // recycled (onExit → circuit-breaker accounting / auto-restart).
-        this.kill("send_timeout");
-      }, ms);
-      if (cur.timer && cur.timer.unref) cur.timer.unref();
-      this._current = cur;
-      const envelope = {
-        type: "user",
-        message: { role: "user", content: [{ type: "text", text: prompt }] },
-      };
-      try {
-        this.proc.stdin.write(JSON.stringify(envelope) + "\n");
-      } catch (e) {
-        if (cur.timer) clearTimeout(cur.timer);
-        cur.settled = true;
-        this._current = null;
-        this.state = "ready";
-        resolve({ ok: false, error: `stdin_write_failed: ${e.message}` });
-      }
-    });
-  }
-
-  /** Deadline of the in-flight turn, or null if idle. Used by the pool reaper. */
-  currentDeadline() {
-    return this._current && !this._current.settled ? this._current.deadline : null;
-  }
-
-  kill(reason = "killed") {
-    try {
-      this.proc?.stdin?.end();
-    } catch {
-      /* ignore */
-    }
-    try {
-      this.proc?.kill("SIGKILL");
-    } catch {
-      /* already gone */
-    }
-    this._onDeath(reason);
-  }
-}
-
-/**
- * AgentPool — concurrency-limited, warm, model-targeted one-shot dispatcher.
- */
-export class AgentPool {
-  /**
-   * @param {object} opts
-   * @param {string} [opts.binary]       explicit claude binary path
-   * @param {number} [opts.poolSize]     max concurrent + prime count (default 2)
-   * @param {number} [opts.maxConcurrent] hard concurrency cap (default = poolSize)
-   * @param {string} [opts.defaultModel] default model alias
-   * @param {number} [opts.defaultTimeoutMs]
-   * @param {string} [opts.neutralCwd]   project-free working dir for lean runs
-   * @param {number} [opts.reaperIntervalMs] sweep cadence (default 5000)
-   * @param {number} [opts.maxJobs]      retained job history (default 200)
-   * @param {boolean} [opts.spawnImpl]   inject a fake spawn for tests
-   */
-  constructor(opts = {}) {
-    this.binary = opts.binary || resolveClaudeBinary(opts.binary);
-    this.poolSize = Number.isFinite(opts.poolSize) ? Math.max(1, opts.poolSize) : DEFAULT_POOL_SIZE;
-    this.maxConcurrent = Number.isFinite(opts.maxConcurrent)
-      ? Math.max(1, opts.maxConcurrent)
-      : this.poolSize;
-    this.defaultModel = opts.defaultModel || DEFAULT_MODEL;
-    this.defaultTimeoutMs = Number.isFinite(opts.defaultTimeoutMs)
-      ? opts.defaultTimeoutMs
-      : DEFAULT_TIMEOUT_MS;
-    this.neutralCwd = opts.neutralCwd || path.join(os.tmpdir(), "clauth-call-agent-neutral");
-    this.reaperIntervalMs = Number.isFinite(opts.reaperIntervalMs) ? opts.reaperIntervalMs : 5000;
-    this.maxJobs = Number.isFinite(opts.maxJobs) ? opts.maxJobs : 200;
-    this._spawn = opts.spawnImpl || spawn;
-
-    // WP-P2: true persistent re-targetable warm workers via stream-json.
-    // `warmEnabled` gates the whole warm path; when false the pool is exactly
-    // the legacy cold dispatcher. `warmSize` long-lived workers are kept hot.
-    this.warmEnabled = opts.warmEnabled != null ? !!opts.warmEnabled : WARM_ENABLED_DEFAULT;
-    this.warmSize = Number.isFinite(opts.warmSize) ? Math.max(0, opts.warmSize) : this.poolSize;
-    // When false, skip the per-worker priming turn (used by tests to isolate
-    // lifecycle behavior from prime-turn timing). Default: prime on (undefined).
-    this.warmPrimeTurn = opts.warmPrimeTurn;
-    // Warm workers serve the DEFAULT bootstrap tier ("identity"): they are
-    // pre-spawned with the identity startup prompt (who-am-I + how-to-get-access)
-    // on the neutral cwd. "none" and "full" calls carry a different prompt/cwd and
-    // are routed to the cold path in _drain(), so the warm prompt is the identity
-    // tier — not the bare LEAN prompt — to match what a default call expects.
-    this.appendSystemPrompt = opts.appendSystemPrompt || IDENTITY_SYSTEM_PROMPT;
-    this._warmWorkers = []; // WarmWorker[]
-    this._warmStarting = false; // H2: reentrancy guard for ensureWarmWorkers()
-
-    // C1 — boot-failure circuit breaker. A deterministic boot failure (bad
-    // CLAUTH_AGENT_MODEL, broken login/MCP) makes every warm worker die during
-    // its prime turn and respawn forever. We count CONSECUTIVE immediate deaths
-    // (lifeMs < IMMEDIATE_DEATH_MS); after `warmCircuitThreshold` of them we trip
-    // the breaker: disable warm for the daemon's life and fall back to the cold
-    // path permanently. A worker that lives past the immediate window resets the
-    // counter (genuine flapping isn't the same as a hard boot failure).
-    this.warmCircuitThreshold = Number.isFinite(opts.warmCircuitThreshold)
-      ? Math.max(1, opts.warmCircuitThreshold)
-      : 5;
-    this.immediateDeathMs = Number.isFinite(opts.immediateDeathMs)
-      ? opts.immediateDeathMs
-      : 2000;
-    this._consecutiveImmediateDeaths = 0;
-    this._warmCircuitTripped = false;
-    this._warmCircuitLogged = false;
-    // Base unit for the immediate-death respawn backoff (exponential, capped at
-    // 30s). Overridable so tests can exercise the breaker quickly.
-    this.warmBackoffBaseMs = Number.isFinite(opts.warmBackoffBaseMs) ? opts.warmBackoffBaseMs : 1000;
-
-    // WP-5: session-level warm worker pinning. A session (e.g. Studio editor)
-    // can claim a warm worker and reuse it across turns, preserving conversation
-    // context. Pinned workers are excluded from the general _acquireWarmWorker()
-    // pool so other callers cannot steal them mid-session.
-    this._sessionPins = new Map(); // Map<sessionId, WarmWorker>
-
-    this.jobs = new Map(); // jobId -> job record
-    this.active = 0;
-    this.queue = []; // pending { resolve, reject, args }
-    this.warm = false;
-    this._reaper = null;
-
-    try {
-      fs.mkdirSync(this.neutralCwd, { recursive: true });
-    } catch {
-      /* best effort */
-    }
-  }
-
-  available() {
-    return this.binary != null;
-  }
-
-  startReaper() {
-    if (this._reaper) return;
-    this._reaper = setInterval(() => this.reap(), this.reaperIntervalMs);
-    // Note: intentionally NOT unref'd. While a job is in flight the reaper is
-    // the only thing that can settle a hung worker, so it must keep the loop
-    // alive. The daemon owns lifecycle and calls shutdown()/stopReaper().
-  }
-
-  stopReaper() {
-    if (this._reaper) {
-      clearInterval(this._reaper);
-      this._reaper = null;
-    }
-  }
-
-  /**
-   * Kill any running job past its deadline, free its slot, drain the queue.
-   * Also evicts old terminal jobs beyond maxJobs.
-   */
-  reap() {
-    const now = Date.now();
-    let reaped = 0;
-    for (const job of this.jobs.values()) {
-      if (job.status === "running" && job.deadline && now > job.deadline) {
-        this._killJob(job, "timeout");
-        reaped++;
-      }
-    }
-    // evict oldest terminal jobs
-    if (this.jobs.size > this.maxJobs) {
-      const terminal = [...this.jobs.values()]
-        .filter((j) => j.status !== "running")
-        .sort((a, b) => (a.completed_at || 0) - (b.completed_at || 0));
-      let toRemove = this.jobs.size - this.maxJobs;
-      for (const j of terminal) {
-        if (toRemove-- <= 0) break;
-        this.jobs.delete(j.jobId);
-      }
-    }
-    this._drain();
-    return reaped;
-  }
-
-  _killJob(job, reason) {
-    if (job._proc && job.status === "running") {
-      try {
-        job._proc.kill("SIGKILL");
-      } catch {
-        /* already gone */
-      }
-    }
-    // Warm job past its deadline: the worker's session is wedged mid-turn. Kill
-    // the persistent worker (auto-restart re-tops the pool) so the slot frees.
-    if (job._worker && job.status === "running") {
-      try {
-        job._worker.kill(reason === "timeout" ? "turn_timeout" : reason);
-      } catch {
-        /* already gone */
-      }
-    }
-    if (job.status === "running") {
-      job.status = reason === "timeout" ? "timeout" : "killed";
-      job.completed_at = Date.now();
-      job.ms = job.completed_at - job.started_at;
-      job.error = reason;
-      // NOTE: do NOT decrement this.active here. settle() is the single source
-      // of truth for the concurrency-slot decrement (it guards with the
-      // `settled` flag). Decrementing here too double-counted every reaped /
-      // timed-out / shutdown-killed job, dropping active by 2 and letting
-      // _drain() launch more than maxConcurrent real workers.
-      if (job._settle) job._settle();
-    }
-  }
-
-  buildArgs({ prompt, model, appendSystemPrompt }) {
-    return [
-      "-p",
-      prompt,
-      "--model",
-      model || this.defaultModel,
-      "--dangerously-skip-permissions",
-      // Lean isolation: only the user's auth/settings, no project CLAUDE.md/hooks.
-      "--setting-sources",
-      "user",
-      "--strict-mcp-config",
-      "--append-system-prompt",
-      appendSystemPrompt || LEAN_SYSTEM_PROMPT,
-    ];
-  }
-
-  // ── WP-P2 warm worker management ───────────────────────────────────────────
-
-  /** Count of alive (ready or busy) warm workers. */
-  _liveWarmCount() {
-    return this._warmWorkers.filter((w) => !w.isDead()).length;
-  }
-
-  /**
-   * First idle warm worker matching `model`, or null. Warm workers are pinned
-   * to one model; a call requesting a different model must NOT be silently
-   * served by a mismatched warm worker — it falls back to a cold spawn that
-   * honors the requested model. Workers pinned to a session are excluded so
-   * general dispatch cannot steal a session's dedicated worker.
-   */
-  _acquireWarmWorker(model) {
-    const pinnedSet = new Set(this._sessionPins.values());
-    return this._warmWorkers.find((w) =>
-      w.isIdle() && (!model || w.model === model) && !pinnedSet.has(w)
-    ) || null;
-  }
-
-  // ── WP-5: session-level warm worker pinning ─────────────────────────────
-
-  /**
-   * Acquire (or re-acquire) a warm worker pinned to `sessionId`. Idempotent:
-   * calling with the same sessionId returns the already-pinned worker. A stale
-   * pin (dead worker) is transparently replaced. Returns null when no warm
-   * worker is available (caller should fall back or queue).
-   *
-   * @param {string} sessionId
-   * @param {object} [opts]
-   * @param {string} [opts.model]
-   * @returns {WarmWorker|null}
-   */
-  acquireForSession(sessionId, opts = {}) {
-    if (this._sessionPins.has(sessionId)) {
-      const worker = this._sessionPins.get(sessionId);
-      if (!worker.isDead()) return worker;
-      this._sessionPins.delete(sessionId); // stale pin, re-acquire below
-    }
-    const model = opts.model || this.defaultModel;
-    const worker = this._acquireWarmWorker(model);
-    if (!worker) return null;
-    this._sessionPins.set(sessionId, worker);
-    return worker;
-  }
-
-  /**
-   * Release a session's pin. The worker returns to the general pool (it is
-   * just no longer reserved). Returns the formerly-pinned worker, or null.
-   * @param {string} sessionId
-   * @returns {WarmWorker|null}
-   */
-  releaseSession(sessionId) {
-    const worker = this._sessionPins.get(sessionId);
-    this._sessionPins.delete(sessionId);
-    return worker || null;
-  }
-
-  /**
-   * Dispatch a prompt to the worker pinned for `sessionId`. Unlike general
-   * dispatch, this always targets the pinned worker (preserving conversation
-   * context). Returns a rejected-shape result if no pin exists or the worker
-   * is dead. If the worker is busy (prior turn still in flight), waits for it.
-   *
-   * @param {string} sessionId
-   * @param {object} [args]  { prompt, timeout_ms, ... }
-   * @returns {Promise<object>}
-   */
-  dispatchToSession(sessionId, args = {}) {
-    const worker = this._sessionPins.get(sessionId);
-    if (!worker || worker.isDead()) {
-      return Promise.resolve({ ok: false, error: "no_pinned_worker", sessionId });
-    }
-    const prompt = args.prompt;
-    if (!prompt || typeof prompt !== "string") {
-      return Promise.resolve({ ok: false, error: "prompt_required" });
-    }
-    return new Promise((resolve, reject) => {
-      const item = { args, resolve, reject };
-      const tryRun = () => {
-        if (worker.isDead()) {
-          return resolve({ ok: false, error: "pinned_worker_died", sessionId });
-        }
-        if (worker.isIdle()) {
-          this._runWarm(item, worker);
-        } else {
-          setTimeout(tryRun, 100);
-        }
-      };
-      tryRun();
-    });
-  }
-
-  /**
-   * Query the pin state for a session. Returns null if no pin exists.
-   * @param {string} sessionId
-   * @returns {object|null}
-   */
-  getSessionInfo(sessionId) {
-    const worker = this._sessionPins.get(sessionId);
-    if (!worker) return null;
-    return {
-      sessionId,
-      pinned: true,
-      alive: !worker.isDead(),
-      idle: worker.isIdle(),
-      model: worker.model,
-      turns: worker.turns,
-    };
-  }
-
-  /** Drop dead workers from the roster. */
-  _pruneDeadWarm() {
-    this._warmWorkers = this._warmWorkers.filter((w) => !w.isDead());
-  }
-
-  /**
-   * Boot `warmSize` persistent workers (idempotent — only tops up to target).
-   * Auto-restart is wired via each worker's onExit, which respawns a fresh
-   * worker if the pool still wants warm capacity and isn't shutting down.
-   */
-  async ensureWarmWorkers() {
-    if (!this.warmEnabled || !this.available() || this._shuttingDown) return 0;
-    if (this._warmCircuitTripped) return 0; // C1: breaker open — warm is permanently off
-    // H2 — reentrancy guard. Two overlapping ensureWarmWorkers() calls (e.g. a
-    // boot prime racing a manual top-up) would each independently compute the
-    // deficit BEFORE the other's workers register, and both spawn — overshooting
-    // warmSize. Honor `_warmStarting`: a second concurrent entrant returns 0 and
-    // lets the in-flight call do the provisioning.
-    if (this._warmStarting) return 0;
-    this._warmStarting = true;
-    try {
-      this._pruneDeadWarm();
-      const deficit = this.warmSize - this._liveWarmCount();
-      if (deficit <= 0) return 0;
-      const starts = [];
-      for (let i = 0; i < deficit; i++) {
-        const w = new WarmWorker({
-          binary: this.binary,
-          model: this.defaultModel,
-          cwd: this.neutralCwd,
-          appendSystemPrompt: this.appendSystemPrompt,
-          spawnImpl: this._spawn,
-          onExit: (worker, reason) => this._onWarmWorkerExit(worker, reason),
-        });
-        this._warmWorkers.push(w);
-        starts.push(w.start().catch(() => false));
-      }
-      const results = await Promise.all(starts);
-      this._pruneDeadWarm();
-      // Fire ONE priming turn per freshly-started worker so the CLI boot/init cost
-      // is paid during warm-up, not on the user's first real call. After this, the
-      // worker's session is hot and every dispatch is a pure-inference re-target.
-      // C2: send() now has its own internal timeout, so a hung prime can no longer
-      // wedge this Promise.all forever.
-      if (this.warmPrimeTurn !== false) {
-        const primes = [];
-        for (const w of this._warmWorkers) {
-          if (w.isIdle() && w.turns === 0) {
-            primes.push(w.send("Reply with exactly: OK", { timeout_ms: 60000 }).catch(() => null));
-          }
-        }
-        await Promise.all(primes);
-        this._pruneDeadWarm();
-      }
-      this.warm = this._warmWorkers.some((w) => !w.isDead());
-      return results.filter(Boolean).length;
-    } finally {
-      this._warmStarting = false;
-    }
-  }
-
-  /**
-   * Auto-restart on death: replace a dead worker if still under target. A short
-   * backoff prevents a tight respawn loop when a worker dies the instant it
-   * starts (e.g. a session that crashes immediately) — without it, each death
-   * would synchronously schedule another start that dies and reschedules,
-   * spinning CPU and spawning processes in a hot loop.
-   */
-  _onWarmWorkerExit(worker, reason) {
-    this._pruneDeadWarm();
-    this.warm = this._warmWorkers.some((w) => !w.isDead());
-
-    // Once shutting down, a dying worker is expected teardown — do not count it
-    // toward the boot-failure breaker and do not respawn. (Also prevents a
-    // late-firing backoff respawn's death from tripping the breaker post-shutdown.)
-    if (this._shuttingDown) return;
-
-    // C1/H1 — death classification. Measure how long THIS worker lived. An
-    // immediate death (lifeMs < immediateDeathMs) is the signature of a hard boot
-    // failure (bad model, login/MCP error) — the worker died during boot/prime.
-    // A worker that lived past the window died for a transient reason (a normal
-    // turn, an OOM, a kill) and must NOT clear protection a different flapping
-    // worker raised: we only reset the consecutive-immediate-death counter when a
-    // worker proves the boot path actually works by surviving the window.
-    const lifeMs = worker && worker.spawnedAt ? Date.now() - worker.spawnedAt : Infinity;
-    const immediate = lifeMs < this.immediateDeathMs;
-
-    if (immediate) {
-      this._consecutiveImmediateDeaths++;
-    } else {
-      // A healthy lifetime proves boot works → safe to forgive accumulated
-      // immediate-death backoff for the lineage.
-      this._consecutiveImmediateDeaths = 0;
-    }
-
-    // C1 — trip the breaker after N consecutive immediate deaths. Disable warm
-    // for the daemon's life, log ONCE, and stop respawning. dispatch() already
-    // falls back to the cold path whenever no warm worker is free, so flipping
-    // warmEnabled=false makes the cold fallback permanent. This protects the live
-    // daemon from endless `claude` churn on a misconfigured CLAUTH_AGENT_MODEL.
-    if (!this._warmCircuitTripped && this._consecutiveImmediateDeaths >= this.warmCircuitThreshold) {
-      this._warmCircuitTripped = true;
-      this.warmEnabled = false;
-      if (!this._warmCircuitLogged) {
-        this._warmCircuitLogged = true;
-        console.error(
-          `[agent-pool] WARM CIRCUIT BREAKER TRIPPED: ${this._consecutiveImmediateDeaths} ` +
-            `consecutive warm-worker boot failures (last: ${reason || worker?.error || "unknown"}). ` +
-            `Disabling warm workers for the rest of this process; call_agent now uses the cold ` +
-            `path exclusively. Check CLAUTH_AGENT_MODEL / CLI login / MCP config.`
-        );
-      }
-      return; // do not respawn — breaker is open
-    }
-
-    if (!this.warmEnabled || this._shuttingDown || !this.available()) return;
-    if (this._liveWarmCount() >= this.warmSize) return;
-
-    // H1 — backoff keyed to the CONSECUTIVE immediate-death count (lineage
-    // protection), not a shared mutable counter that any healthy exit zeroes.
-    // Each successive immediate death backs off longer (capped at 30s); a
-    // worker that lived past the window respawns immediately.
-    const delay = immediate
-      ? Math.min(30000, this.warmBackoffBaseMs * Math.pow(2, Math.max(0, this._consecutiveImmediateDeaths - 1)))
-      : 0;
-
-    const spawnReplacement = () => {
-      if (this._warmCircuitTripped) return;
-      if (!this.warmEnabled || this._shuttingDown || !this.available()) return;
-      if (this._liveWarmCount() >= this.warmSize) return;
-      const replacement = new WarmWorker({
-        binary: this.binary,
-        model: this.defaultModel,
-        cwd: this.neutralCwd,
-        appendSystemPrompt: this.appendSystemPrompt,
-        spawnImpl: this._spawn,
-        onExit: (w, r) => this._onWarmWorkerExit(w, r),
-      });
-      this._warmWorkers.push(replacement);
-      replacement
-        .start()
-        .then(() => {
-          this.warm = this._warmWorkers.some((w) => !w.isDead());
-        })
-        .catch(() => {});
-    };
-    if (delay > 0) {
-      const t = setTimeout(spawnReplacement, delay);
-      if (t.unref) t.unref();
-    } else {
-      spawnReplacement();
-    }
-  }
-
-  /**
-   * Pre-warm. With warm workers enabled (default), boots `warmSize` persistent
-   * re-targetable stream-json workers. Otherwise (legacy/cold), fires poolSize
-   * trivial one-shot pings so the OS file cache / model session is warm.
-   * Returns when warming settles (or immediately if unavailable).
-   */
-  async prime() {
-    if (!this.available()) return { warmed: 0, available: false, mode: "unavailable" };
-    if (this.warmEnabled) {
-      const warmed = await this.ensureWarmWorkers();
-      return { warmed, available: true, mode: "warm-workers", workers: this._liveWarmCount() };
-    }
-    const n = this.poolSize;
-    const pings = [];
-    for (let i = 0; i < n; i++) {
-      pings.push(
-        this.dispatch({
-          prompt: "Reply with exactly: OK",
-          timeout_ms: 30000,
-          _priming: true,
-        }).catch((e) => ({ ok: false, error: e.message }))
-      );
-    }
-    const results = await Promise.all(pings);
-    this.warm = results.some((r) => r && r.ok);
-    return { warmed: results.filter((r) => r && r.ok).length, available: true, mode: "cold-prime" };
-  }
-
-  /**
-   * Dispatch a one-shot. Resolves with a job record once the worker exits
-   * (sync semantics). Concurrency beyond maxConcurrent is queued.
-   *
-   * @returns {Promise<object>} job record { ok, jobId, package, model, ms, status, exit_code, error }
-   */
-  dispatch(args = {}) {
-    return new Promise((resolve, reject) => {
-      const item = { args, resolve, reject };
-      this.queue.push(item);
-      this._drain();
-    });
-  }
-
-  _drain() {
-    while (this.active < this.maxConcurrent && this.queue.length > 0) {
-      const item = this.queue.shift();
-      // Priming pings stay on the cold one-shot path (they exist only to warm
-      // the cold cache; with warm workers they are a no-op caller). Real calls
-      // prefer a free warm worker and fall back to cold when none is free.
-      //
-      // Bootstrap pinning (tiered startup): warm workers are pre-spawned with the
-      // DEFAULT ("identity") startup prompt and the neutral cwd. A call asking for
-      // a DIFFERENT bootstrap level — "none" (lean fast path) or "full" (repo cwd
-      // + bootstrap guide) — has a different system prompt and/or cwd, so it must
-      // NOT be served by a warm worker. Route it to the cold path, which honors
-      // args.appendSystemPrompt and args.cwd. Default/absent bootstrap → warm.
-      const bootstrapLevel = item.args.bootstrap || DEFAULT_BOOTSTRAP;
-      const warmEligible = bootstrapLevel === DEFAULT_BOOTSTRAP;
-      if (this.warmEnabled && !item.args._priming && warmEligible) {
-        // Only a warm worker pinned to the REQUESTED model may serve the call;
-        // a model mismatch must fall through to the cold path (which honors
-        // args.model) rather than be silently downgraded to the warm model.
-        const reqModel = item.args.model || this.defaultModel;
-        const worker = this._acquireWarmWorker(reqModel);
-        if (worker) {
-          this._runWarm(item, worker);
-          continue;
-        }
-      }
-      this._run(item);
-    }
-  }
-
-  /**
-   * Run a call on a persistent warm worker (re-targeting it with a new prompt).
-   * Produces the SAME public job record shape as the cold path so callers
-   * (runCallAgent, durable delivery, GET /call-agent/:jobId) are unchanged.
-   */
-  _runWarm(item, worker) {
-    const { args, resolve } = item;
-    const prompt = args.prompt;
-    if (!prompt || typeof prompt !== "string") {
-      return resolve({ ok: false, error: "prompt_required" });
-    }
-    const model = worker.model;
-    const timeout = Number.isFinite(args.timeout_ms) ? args.timeout_ms : this.defaultTimeoutMs;
-    const jobId = args.jobId || makeJobId("call");
-
-    const job = {
-      jobId,
-      status: "running",
-      model,
-      priming: false,
-      warm: true,
-      started_at: Date.now(),
-      deadline: Date.now() + timeout,
-      completed_at: null,
-      ms: null,
-      exit_code: null,
-      stdout: "",
-      stderr: "",
-      package: null,
-      error: null,
-      _proc: null,
-      _worker: worker,
-      _settle: null,
-    };
-    this.jobs.set(jobId, job);
-    this.active++;
-
-    let settled = false;
-    const settle = () => {
-      if (settled) return;
-      settled = true;
-      this.active = Math.max(0, this.active - 1);
-      const rec = {
-        ok: job.status === "completed",
-        jobId: job.jobId,
-        status: job.status,
-        model: job.model,
-        ms: job.ms,
-        exit_code: job.exit_code,
-        package: job.package,
-        error: job.error,
-        warm: true,
-      };
-      if (job.stderr && job.status !== "completed") rec.stderr = job.stderr.slice(-2000);
-      this._drain();
-      resolve(rec);
-    };
-    job._settle = () => {
-      // Reaper path: the turn blew its deadline. Kill the worker (its session is
-      // wedged); auto-restart re-tops the pool. Mark the job per its set status.
-      if (job.status === "running") {
-        job.status = "timeout";
-        job.completed_at = Date.now();
-        job.ms = job.completed_at - job.started_at;
-        job.error = "timeout";
-        try {
-          worker.kill("turn_timeout");
-        } catch {
-          /* ignore */
-        }
-      }
-      settle();
-    };
-
-    worker
-      .send(prompt, { timeout_ms: timeout })
-      .then((res) => {
-        if (settled) return; // reaper already settled this
-        job.completed_at = Date.now();
-        job.ms = job.completed_at - job.started_at;
-        if (res.ok) {
-          job.status = "completed";
-          job.exit_code = 0;
-          job.package = res.package;
-        } else {
-          job.status = "failed";
-          job.error = res.error || (res.is_error ? "agent_error" : "warm_failed");
-          if (res.package) job.package = res.package;
-        }
-        settle();
-      })
-      .catch((e) => {
-        if (settled) return;
-        job.status = "failed";
-        job.error = `warm_dispatch: ${e.message}`;
-        job.completed_at = Date.now();
-        job.ms = job.completed_at - job.started_at;
-        settle();
-      });
-  }
-
-  _run({ args, resolve, reject }) {
-    if (!this.available()) {
-      return resolve({
-        ok: false,
-        error: "binary_not_found",
-        message: "claude CLI not found in PATH or AppData/npm",
-      });
-    }
-    const prompt = args.prompt;
-    if (!prompt || typeof prompt !== "string") {
-      return resolve({ ok: false, error: "prompt_required" });
-    }
-    const model = args.model || this.defaultModel;
-    const timeout = Number.isFinite(args.timeout_ms) ? args.timeout_ms : this.defaultTimeoutMs;
-    const jobId = args.jobId || makeJobId(args._priming ? "warm" : "call");
-    const cwd = args.cwd || this.neutralCwd;
-
-    const job = {
-      jobId,
-      status: "running",
-      model,
-      priming: !!args._priming,
-      started_at: Date.now(),
-      deadline: Date.now() + timeout,
-      completed_at: null,
-      ms: null,
-      exit_code: null,
-      stdout: "",
-      stderr: "",
-      package: null,
-      error: null,
-      _proc: null,
-      _settle: null,
-    };
-    this.jobs.set(jobId, job);
-    this.active++;
-
-    let settled = false;
-    const settle = () => {
-      if (settled) return;
-      settled = true;
-      this.active = Math.max(0, this.active - 1);
-      // Build the public record (no internal handles).
-      const rec = {
-        ok: job.status === "completed",
-        jobId: job.jobId,
-        status: job.status,
-        model: job.model,
-        ms: job.ms,
-        exit_code: job.exit_code,
-        package: job.package,
-        error: job.error,
-      };
-      if (job.stderr && job.status !== "completed") rec.stderr = job.stderr.slice(-2000);
-      this._drain();
-      resolve(rec);
-    };
-    job._settle = () => settle();
-
-    const cmdArgs = this.buildArgs({
-      prompt,
-      model,
-      appendSystemPrompt: args.appendSystemPrompt,
-    });
-
-    let proc;
-    try {
-      const isCmdShim = process.platform === "win32" && /\.cmd$/i.test(this.binary);
-      const command = isCmdShim ? "cmd" : this.binary;
-      const spawnArgs = isCmdShim ? ["/d", "/s", "/c", `"${this.binary}"`, ...cmdArgs] : cmdArgs;
-      proc = this._spawn(command, spawnArgs, {
-        cwd,
-        env: process.env, // inherits CLI login session; no ANTHROPIC_API_KEY injected
-        stdio: ["ignore", "pipe", "pipe"],
-        shell: false,
-        windowsHide: true,
-      });
-    } catch (e) {
-      job.status = "failed";
-      job.error = `spawn_failed: ${e.message}`;
-      job.completed_at = Date.now();
-      job.ms = job.completed_at - job.started_at;
-      return settle();
-    }
-
-    job._proc = proc;
-    job.pid = proc.pid;
-
-    proc.stdout?.on("data", (d) => {
-      job.stdout = `${job.stdout}${d.toString()}`.slice(-200000);
-    });
-    proc.stderr?.on("data", (d) => {
-      job.stderr = `${job.stderr}${d.toString()}`.slice(-20000);
-    });
-    proc.on("error", (e) => {
-      if (job.status !== "running") return;
-      job.status = "failed";
-      job.error = `proc_error: ${e.message}`;
-      job.completed_at = Date.now();
-      job.ms = job.completed_at - job.started_at;
-      settle();
-    });
-    proc.on("close", (code) => {
-      if (job.status !== "running") return; // reaper/kill already settled
-      job.exit_code = code;
-      job.completed_at = Date.now();
-      job.ms = job.completed_at - job.started_at;
-      if (code === 0) {
-        job.status = "completed";
-        job.package = job.stdout.trim();
-      } else {
-        job.status = "failed";
-        job.error = `exit_${code}`;
-      }
-      settle();
-    });
-  }
-
-  getJob(jobId) {
-    const job = this.jobs.get(jobId);
-    if (!job) return null;
-    return {
-      jobId: job.jobId,
-      status: job.status,
-      model: job.model,
-      ms: job.ms,
-      exit_code: job.exit_code,
-      package: job.package,
-      error: job.error,
-      started_at: job.started_at,
-      completed_at: job.completed_at,
-    };
-  }
-
-  stats() {
-    let running = 0,
-      completed = 0,
-      failed = 0;
-    for (const j of this.jobs.values()) {
-      if (j.status === "running") running++;
-      else if (j.status === "completed") completed++;
-      else failed++;
-    }
-    const warmReady = this._warmWorkers.filter((w) => w.isIdle()).length;
-    const warmBusy = this._warmWorkers.filter((w) => w.isBusy()).length;
-    return {
-      active: this.active,
-      queued: this.queue.length,
-      maxConcurrent: this.maxConcurrent,
-      poolSize: this.poolSize,
-      // `warm` reflects REAL persistent workers when warm mode is on (WP-P2):
-      // true iff at least one live worker exists. Cold mode keeps legacy meaning.
-      warm: this.warmEnabled ? this._liveWarmCount() > 0 : this.warm,
-      warmEnabled: this.warmEnabled,
-      warmCircuitTripped: this._warmCircuitTripped,
-      warmWorkers: {
-        target: this.warmSize,
-        live: this._liveWarmCount(),
-        ready: warmReady,
-        busy: warmBusy,
-        consecutiveImmediateDeaths: this._consecutiveImmediateDeaths,
-      },
-      jobs: { running, completed, failed, total: this.jobs.size },
-    };
-  }
-
-  shutdown() {
-    this._shuttingDown = true;
-    this.stopReaper();
-    for (const job of this.jobs.values()) {
-      if (job.status === "running") this._killJob(job, "shutdown");
-    }
-    // Clear session pins before tearing down workers.
-    this._sessionPins.clear();
-    // Tear down persistent warm workers so no orphan `claude` processes linger.
-    for (const w of this._warmWorkers) {
-      try {
-        w.kill("shutdown");
-      } catch {
-        /* already gone */
-      }
-    }
-    this._warmWorkers = [];
-  }
-}
-
-// ── WP-P7: hop-limited sub-agent delegation ──────────────────────────────────
-//
-// THE DEADLOCK we must make impossible:
-//   The warm/interactive pool has only ~2 workers. A PARENT call_agent worker is
-//   itself a running `claude` process. If that parent calls call_agent and the
-//   sub-call needs a *pool* worker while all pool workers are held by parents
-//   blocked on their own sub-calls → classic resource deadlock. So delegated
-//   (sub-agent) calls MUST NOT touch the AgentPool. They run on this separate
-//   lane, which:
-//     1. NEVER references an AgentPool, warm worker, or its queue. It only
-//        cold-spawns its own `claude` one-shots, capped by `maxInFlight`. The
-//        interactive pool's latency/availability is therefore unaffected by
-//        delegation depth or fan-out.
-//     2. Enforces a strict HOP BUDGET (carried in agent_context.delegation):
-//        a call at hop >= max is REJECTED with a clean error and NO spawn, so a
-//        runaway delegation loop is impossible and tree depth is bounded.
-//     3. Enforces a daemon-wide in-flight cap so a wide fan-out can't exhaust
-//        resources.
-//
-// Hop accounting: the delegation context on an INCOMING call describes the call
-// itself — `{ hop: N, max: M }` means "this is a hop-N delegated call, budget M".
-// A top-level (interactive) call has no delegation context (hop is treated as 0).
-// When a worker at hop N spawns a sub-agent, the daemon computes the child's
-// context as `{ hop: N+1, max: M }` and rejects if `N+1 > M`. Equivalently, an
-// incoming delegated call is admitted iff `hop <= max`; hop is the depth of the
-// call being served, and a call at `hop > max` (or `hop == max` asking to go
-// deeper) is refused.
-
-/**
- * Normalize / validate a delegation marker from agent_context.delegation.
- * Returns { hop, max } with safe integer coercion, or null if absent/invalid.
- * A present-but-malformed marker normalizes to hop 0 with the default max so a
- * caller cannot smuggle a negative hop or a giant max past the budget.
- */
-export function normalizeDelegation(input, { defaultMax = DEFAULT_DELEGATION_MAX_HOPS } = {}) {
-  if (input == null) return null;
-  if (typeof input !== "object") return null;
-  let hop = Number(input.hop);
-  let max = Number(input.max);
-  if (!Number.isFinite(hop) || hop < 0) hop = 0;
-  hop = Math.floor(hop);
-  if (!Number.isFinite(max) || max < 0) max = defaultMax;
-  max = Math.floor(max);
-  return { hop, max };
-}
-
-/**
- * Decide whether an incoming DELEGATED call may proceed, and compute the child
- * delegation context the spawned sub-agent should carry if IT delegates further.
- *
- * @param {object|null} delegation  normalized { hop, max } of the incoming call
- * @param {object} [opts]
- * @param {number} [opts.defaultMax]
- * @returns {{ allow: boolean, hop: number, max: number, childContext?: object, error?: string }}
- */
-export function evaluateHopBudget(delegation, { defaultMax = DEFAULT_DELEGATION_MAX_HOPS } = {}) {
-  const d = normalizeDelegation(delegation, { defaultMax }) || { hop: 0, max: defaultMax };
-  // The call being served is at depth `hop`. It is admissible iff it does not
-  // exceed the budget. `hop > max` means a parent tried to spawn one level too
-  // deep — refuse with a clean error, no spawn.
-  if (d.hop > d.max) {
-    return {
-      allow: false,
-      hop: d.hop,
-      max: d.max,
-      error: "delegation_hop_exceeded",
-    };
-  }
-  // Child context for any sub-agent THIS call spawns: one hop deeper, same max.
-  return {
-    allow: true,
-    hop: d.hop,
-    max: d.max,
-    childContext: { hop: d.hop + 1, max: d.max },
-  };
-}
-
-/**
- * Single source of truth for the routing decision: should THIS call_agent
- * invocation go through the delegation lane instead of the interactive pool?
- *
- * The gate (`enabled`, from CLAUTH_AGENT_DELEGATION) is checked FIRST: when the
- * gate is OFF this returns false unconditionally and `agent_context.delegation`
- * is never even read — that is the byte-stable-off guarantee. A call is delegated
- * iff the gate is on AND the caller supplied an `agent_context.delegation`
- * marker (only a sub-agent invocation carries one).
- *
- * @param {object} args
- * @param {boolean} args.enabled        CLAUTH_AGENT_DELEGATION on?
- * @param {object}  [args.agent_context]
- * @returns {boolean}
- */
-export function shouldDelegate({ enabled, agent_context } = {}) {
-  if (!enabled) return false; // gate OFF → delegation marker ignored entirely
-  if (!agent_context || typeof agent_context !== "object") return false;
-  return agent_context.delegation != null;
-}
-
-/**
- * DelegationLane — a cold-spawn-only, concurrency-capped execution lane for
- * DELEGATED (sub-agent) call_agent invocations. Structurally distinct from
- * AgentPool: it holds no warm workers, owns no AgentPool reference, and shares
- * no queue/slot with the interactive pool. That separation is the anti-deadlock
- * guarantee — sub-agent spawns can never starve or hold an interactive slot.
- *
- * It deliberately reuses the SAME lean one-shot recipe (model, skip-permissions,
- * user-only settings, strict-mcp, lean system prompt) as AgentPool's cold path,
- * but on an independent set of processes with its own `maxInFlight` cap.
- */
-export class DelegationLane {
-  /**
-   * @param {object} opts
-   * @param {string} [opts.binary]        resolved claude binary
-   * @param {number} [opts.maxInFlight]   daemon-wide concurrent sub-agent cap
-   * @param {number} [opts.maxHops]       hop-budget default ceiling
-   * @param {string} [opts.defaultModel]
-   * @param {number} [opts.defaultTimeoutMs]
-   * @param {string} [opts.neutralCwd]
-   * @param {string} [opts.appendSystemPrompt]
-   * @param {function} [opts.spawnImpl]   inject a fake spawn for tests
-   */
-  constructor(opts = {}) {
-    this.binary = opts.binary || resolveClaudeBinary(opts.binary);
-    this.maxInFlight = Number.isFinite(opts.maxInFlight)
-      ? Math.max(1, opts.maxInFlight)
-      : DEFAULT_DELEGATION_MAX_INFLIGHT;
-    this.maxHops = Number.isFinite(opts.maxHops) ? Math.max(0, opts.maxHops) : DEFAULT_DELEGATION_MAX_HOPS;
-    this.defaultModel = opts.defaultModel || DEFAULT_MODEL;
-    this.defaultTimeoutMs = Number.isFinite(opts.defaultTimeoutMs) ? opts.defaultTimeoutMs : DEFAULT_TIMEOUT_MS;
-    this.neutralCwd = opts.neutralCwd || path.join(os.tmpdir(), "clauth-call-agent-delegation");
-    this.appendSystemPrompt = opts.appendSystemPrompt || LEAN_SYSTEM_PROMPT;
-    this._spawn = opts.spawnImpl || spawn;
-
-    this.active = 0; // live sub-agent cold spawns
-    this.totalSpawned = 0;
-    this.totalRejected = 0;
-    this._shuttingDown = false;
-    this._procs = new Set(); // live child procs for shutdown teardown
-
-    try {
-      fs.mkdirSync(this.neutralCwd, { recursive: true });
-    } catch {
-      /* best effort */
-    }
-  }
-
-  available() {
-    return this.binary != null;
-  }
-
-  buildArgs({ prompt, model, appendSystemPrompt }) {
-    return [
-      "-p",
-      prompt,
-      "--model",
-      model || this.defaultModel,
-      "--dangerously-skip-permissions",
-      "--setting-sources",
-      "user",
-      "--strict-mcp-config",
-      "--append-system-prompt",
-      appendSystemPrompt || this.appendSystemPrompt,
-    ];
-  }
-
-  /**
-   * Dispatch a DELEGATED sub-agent call. Enforces the hop budget and the
-   * in-flight cap, then cold-spawns a one-shot. Returns a job record shaped like
-   * AgentPool.dispatch's so callers are uniform. On rejection it resolves with
-   * `{ ok:false, error, delegated:true, ... }` and NEVER spawns.
-   *
-   * @param {object} args { prompt, model, timeout_ms, delegation:{hop,max}, jobId, appendSystemPrompt }
-   * @returns {Promise<object>}
-   */
-  dispatch(args = {}) {
-    return new Promise((resolve) => {
-      if (!this.available()) {
-        return resolve({ ok: false, error: "binary_not_found", delegated: true });
-      }
-      const prompt = args.prompt;
-      if (!prompt || typeof prompt !== "string") {
-        return resolve({ ok: false, error: "prompt_required", delegated: true });
-      }
-      // Hop budget — the single most important guard. A call past budget is
-      // refused cleanly with NO spawn, so a delegation loop cannot run away.
-      const budget = evaluateHopBudget(args.delegation, { defaultMax: this.maxHops });
-      if (!budget.allow) {
-        this.totalRejected++;
-        return resolve({
-          ok: false,
-          error: budget.error,
-          delegated: true,
-          hop: budget.hop,
-          max: budget.max,
-          message: `delegation hop ${budget.hop} exceeds max ${budget.max}`,
-        });
-      }
-      // Daemon-wide in-flight cap — a wide fan-out cannot exhaust resources.
-      if (this.active >= this.maxInFlight) {
-        this.totalRejected++;
-        return resolve({
-          ok: false,
-          error: "delegation_capacity",
-          delegated: true,
-          active: this.active,
-          maxInFlight: this.maxInFlight,
-          message: `delegation lane full (${this.active}/${this.maxInFlight})`,
-        });
-      }
-      if (this._shuttingDown) {
-        return resolve({ ok: false, error: "delegation_shutting_down", delegated: true });
-      }
-
-      const model = args.model || this.defaultModel;
-      const timeout = Number.isFinite(args.timeout_ms) ? args.timeout_ms : this.defaultTimeoutMs;
-      const jobId = args.jobId || makeJobId("deleg");
-      const cwd = args.cwd || this.neutralCwd;
-      const startedAt = Date.now();
-
-      const cmdArgs = this.buildArgs({ prompt, model, appendSystemPrompt: args.appendSystemPrompt });
-
-      let proc;
-      try {
-        const isCmdShim = process.platform === "win32" && /\.cmd$/i.test(this.binary);
-        const command = isCmdShim ? "cmd" : this.binary;
-        const spawnArgs = isCmdShim ? ["/d", "/s", "/c", `"${this.binary}"`, ...cmdArgs] : cmdArgs;
-        proc = this._spawn(command, spawnArgs, {
-          cwd,
-          env: process.env, // inherits CLI login; never sets ANTHROPIC_API_KEY
-          stdio: ["ignore", "pipe", "pipe"],
-          shell: false,
-          windowsHide: true,
-        });
-      } catch (e) {
-        return resolve({
-          ok: false,
-          error: `spawn_failed: ${e.message}`,
-          delegated: true,
-          jobId,
-        });
-      }
-
-      this.active++;
-      this.totalSpawned++;
-      this._procs.add(proc);
-
-      let stdout = "";
-      let stderr = "";
-      let settled = false;
-      // NOTE: intentionally NOT unref'd. While a sub-agent is in flight this
-      // timer is the only thing that can settle a hung one-shot, so it must keep
-      // the loop alive until it fires (it clears on normal close). The daemon
-      // owns process lifecycle; shutdown() kills all live procs explicitly.
-      const timer = setTimeout(() => {
-        if (settled) return;
-        try {
-          proc.kill("SIGKILL");
-        } catch {
-          /* already gone */
-        }
-        finish({ status: "timeout", error: "timeout" });
-      }, timeout);
-
-      const finish = (extra) => {
-        if (settled) return;
-        settled = true;
-        clearTimeout(timer);
-        this.active = Math.max(0, this.active - 1);
-        this._procs.delete(proc);
-        const ms = Date.now() - startedAt;
-        const status = extra.status;
-        resolve({
-          ok: status === "completed",
-          jobId,
-          status,
-          model,
-          ms,
-          exit_code: extra.exit_code ?? null,
-          package: status === "completed" ? stdout.trim() : null,
-          error: extra.error || null,
-          delegated: true,
-          hop: budget.childContext.hop, // depth of this spawned sub-agent
-          max: budget.max,
-          ...(stderr && status !== "completed" ? { stderr: stderr.slice(-2000) } : {}),
-        });
-      };
-
-      proc.stdout?.on("data", (d) => {
-        stdout = `${stdout}${d.toString()}`.slice(-200000);
-      });
-      proc.stderr?.on("data", (d) => {
-        stderr = `${stderr}${d.toString()}`.slice(-20000);
-      });
-      proc.on("error", (e) => finish({ status: "failed", error: `proc_error: ${e.message}` }));
-      proc.on("close", (code) => {
-        if (code === 0) finish({ status: "completed", exit_code: 0 });
-        else finish({ status: "failed", exit_code: code, error: `exit_${code}` });
-      });
-    });
-  }
-
-  stats() {
-    return {
-      active: this.active,
-      maxInFlight: this.maxInFlight,
-      maxHops: this.maxHops,
-      totalSpawned: this.totalSpawned,
-      totalRejected: this.totalRejected,
-    };
-  }
-
-  shutdown() {
-    this._shuttingDown = true;
-    for (const proc of this._procs) {
-      try {
-        proc.kill("SIGKILL");
-      } catch {
-        /* already gone */
-      }
-    }
-    this._procs.clear();
-  }
-}
-
-export {
-  LEAN_SYSTEM_PROMPT,
-  IDENTITY_SYSTEM_PROMPT,
-  resolveClaudeBinary,
-  buildWarmArgs,
-  buildFullSystemPrompt,
-  makeJobId,
-};
-
-// composeCorpus / buildCorpusSection / loadBootstrapDoc are exported inline
-// above (function declarations with `export`).
+// agent-pool.js — warm Haiku worker pool for call_agent one-shot dispatch.
+//
+// Design (Gate B, WP-B1/B3/B4) — cold path:
+//   * Each `claude -p "<prompt>"` is inherently a fresh detached process — the
+//     prompt is passed as argv, so an idle pre-spawned process cannot be
+//     re-targeted. The cold "warm pool" therefore does two things:
+//       1. Resolves + caches the claude binary once (no per-call `--version`
+//          probe) and keeps a neutral, project-free cwd hot on disk.
+//       2. Optionally fires `poolSize` priming pings at startup so the OS file
+//          cache / model session is warm — the FIRST real call then returns in
+//          ~4-5s instead of the ~50-90s cold path.
+//
+// WP-P2 — TRUE persistent re-targetable warm workers:
+//   * The `claude` CLI (v2.1.170) supports `-p --input-format stream-json
+//     --output-format stream-json`: a single long-lived process that reads
+//     newline-delimited user-message envelopes from stdin and emits
+//     newline-delimited events (including a `result` per turn) on stdout. The
+//     SAME process (same session_id) answers MULTIPLE sequential prompts —
+//     proven: 3 prompts on one process, sid stable, 4.4s → 2.0s → 2.8s with NO
+//     per-call CLI boot. This is true re-targeting: a pre-spawned idle worker IS
+//     re-aimable at a new prompt over stdin.
+//   * `WarmWorker` wraps one such process. `AgentPool.dispatch()` prefers a free
+//     warm worker (sub-boot dispatch — latency is pure model inference, not the
+//     ~10-15s CLI cold-boot) and transparently FALLS BACK to the cold spawn path
+//     when warm workers are disabled, all busy, or unavailable. Fully backward
+//     compatible: the public dispatch()/job-record contract is unchanged.
+//   * Warm workers are health-checked and AUTO-RESTARTED on death; the reaper
+//     enforces per-turn deadlines on warm turns too. Model-tiered: each pool is
+//     pinned to one model (default Haiku); a Sonnet/Opus pool is a second
+//     AgentPool instance.
+//
+//   * Lean one-shot: neutral cwd + `--setting-sources user` + `--strict-mcp-config`
+//     strips the regen-root CLAUDE.md / SessionEnd hooks / lessons gate that
+//     otherwise contaminates replies ("Waived." instead of the real answer).
+//   * Auth: piggybacks the CLI's own login session. NEVER sets ANTHROPIC_API_KEY.
+//   * Reaping: a periodic sweep kills workers that exceed their timeout and
+//     frees concurrency slots so the queue self-heals.
+//
+// This module is intentionally self-contained and daemon-free so its core can
+// be unit/integration tested WITHOUT the clauth vault or boot.key.
+
+import { spawn } from "child_process";
+import { execSync } from "child_process";
+import os from "os";
+import path from "path";
+import fs from "fs";
+import crypto from "crypto";
+
+export const DEFAULT_MODEL = "claude-haiku-4-5";
+export const DEFAULT_TIMEOUT_MS = 120000;
+export const DEFAULT_POOL_SIZE = 2;
+// Warm workers are ON by default. Set CLAUTH_AGENT_WARM=0 to force the cold path.
+export const WARM_ENABLED_DEFAULT =
+  String(process.env.CLAUTH_AGENT_WARM ?? "1") !== "0";
+
+// ── WP-P7: hop-limited sub-agent delegation defaults ─────────────────────────
+// A parent call_agent worker may itself invoke call_agent SUB-agents, bounded by
+// a hop budget and routed through a SEPARATE execution lane (DelegationLane) so
+// it can NEVER consume the shared warm/interactive pool slots (anti-deadlock).
+export const DEFAULT_DELEGATION_MAX_HOPS = 2; // tree depth cap (hop 0 = top call)
+export const DEFAULT_DELEGATION_MAX_INFLIGHT = 4; // daemon-wide concurrent sub-agents
+
+const LEAN_SYSTEM_PROMPT =
+  "You are a stateless one-shot agent worker invoked via clauth call_agent. " +
+  "Answer the user's prompt directly, concisely, and completely. " +
+  "Do not editorialize, do not read project files unless the prompt asks, " +
+  "and do not emit meta-commentary about hooks, sessions, or governance. " +
+  "If a JSON answer is requested, return ONLY valid JSON.";
+
+// ── Tiered call_agent startup (bootstrap) ────────────────────────────────────
+// A spawned worker otherwise has NO identity and NO idea what it can reach (no
+// clauth creds, no Supabase, no repo). These three tiers fix that, additively:
+//
+//   * "none"     — today's lean fast path, byte-identical to LEAN_SYSTEM_PROMPT.
+//                  For trivial one-shots; warm ~2s. Neutral, project-free cwd.
+//   * "identity" — DEFAULT. Prepends a SHORT startup that tells the agent who it
+//                  is and HOW to reach access (clauth daemon, Supabase ref,
+//                  work_items RPCs, Bash/file tools). Neutral cwd. Low latency.
+//   * "full"     — for real repo/DB work: runs in the REPO cwd and points the
+//                  agent at .rdc/guides/agent-bootstrap.md (creds, git rules,
+//                  completion report) plus the identity/access block.
+//
+// IMPORTANT: the startup TELLS the agent how to fetch credentials; it NEVER
+// embeds a credential and NEVER sets ANTHROPIC_API_KEY.
+export const BOOTSTRAP_LEVELS = ["none", "identity", "full"];
+export const DEFAULT_BOOTSTRAP = "identity";
+export const DEFAULT_AGENT_REPO_ROOT =
+  process.env.CLAUTH_AGENT_REPO_ROOT || "C:/Dev/regen-root";
+
+// The clauth credential daemon every Regen agent uses.
+const CLAUTH_DAEMON_URL = "http://127.0.0.1:52437";
+const SUPABASE_PROJECT_REF = "uvojezuorjgqzmhhgluu";
+
+// Short identity/access block — a dozen lines so latency stays low. This is the
+// minimum every default agent must know: who it is + how to get access.
+const IDENTITY_SYSTEM_PROMPT =
+  "You are a Regen agent running on the operator's machine. " +
+  "ACCESS — discover service names through the clauth HTTP daemon at " +
+  `${CLAUTH_DAEMON_URL}: run \`curl -s ${CLAUTH_DAEMON_URL}/knowledge\` for ` +
+  "sanitized metadata, then fetch exactly one needed credential with " +
+  `\`curl -s ${CLAUTH_DAEMON_URL}/v/<service>\`. ` +
+  "Do not bulk-fetch secrets. NEVER print a credential to stdout — it is logged. " +
+  `The Supabase project ref is ${SUPABASE_PROJECT_REF}. ` +
+  "Work items live in Supabase and are accessed ONLY through RPCs " +
+  "(get_open_epics, insert_work_item, update_work_item_status) — never raw " +
+  "SQL CRUD against work_items. You have Bash plus file read/write tools. " +
+  "Answer the user's prompt directly and completely; if JSON is requested, " +
+  "return ONLY valid JSON.";
+
+const EDITOR_SETTINGS_FILENAME = ".studio-editor-settings.json";
+
+/**
+ * Editable-prompts settings file — every Claude prompt this pool builds has a
+ * default baked into code; this loads the SAME-SHAPED override from
+ * `<repoRoot>/.studio-editor-settings.json` if present. Read fresh on every
+ * call (no caching) so an edit to the file takes effect on the next dispatch
+ * without a daemon restart. Never throws — a missing or malformed settings
+ * file must fall back to defaults silently, not break prompt construction.
+ * @param {string} repoRoot
+ * @returns {object}
+ */
+function loadEditorSettings(repoRoot) {
+  if (!repoRoot) return {};
+  try {
+    const settingsPath = path.join(repoRoot, EDITOR_SETTINGS_FILENAME);
+    if (!fs.existsSync(settingsPath)) return {};
+    const parsed = JSON.parse(fs.readFileSync(settingsPath, "utf8"));
+    return parsed && typeof parsed === "object" && !Array.isArray(parsed) ? parsed : {};
+  } catch (e) {
+    console.error(`[agent-pool] ${EDITOR_SETTINGS_FILENAME} malformed, using code defaults: ${e.message}`);
+    return {};
+  }
+}
+
+/**
+ * Build the "full" startup: identity/access block + a pointer to the project
+ * bootstrap guide that the agent must read first. Runs in the repo cwd.
+ *
+ * Editable: set `systemPrompt` in `<repoRoot>/.studio-editor-settings.json` to
+ * override this entire string. `{repoRoot}` and `{bootstrapPath}` tokens are
+ * substituted if present.
+ * @param {string} [repoRoot]
+ */
+function buildFullSystemPrompt(repoRoot = DEFAULT_AGENT_REPO_ROOT) {
+  const root = (repoRoot || DEFAULT_AGENT_REPO_ROOT).replace(/\\/g, "/");
+  const bootstrapPath = `${root}/${BOOTSTRAP_DOC_REL_PATH}`;
+  const settings = loadEditorSettings(root);
+  if (typeof settings.systemPrompt === "string" && settings.systemPrompt.trim()) {
+    return settings.systemPrompt
+      .replace(/\{repoRoot\}/g, root)
+      .replace(/\{bootstrapPath\}/g, bootstrapPath);
+  }
+  return (
+    IDENTITY_SYSTEM_PROMPT +
+    " You are running in the repo at " +
+    root +
+    ". Before doing repo or database work, READ " +
+    bootstrapPath +
+    " FIRST — it has the full credential, " +
+    "Supabase, git (commit to develop, never main, never force-push), and " +
+    "completion-report contract. Then read your role guide under " +
+    root +
+    "/.rdc/guides/ if one applies."
+  );
+}
+
+/**
+ * Lean system prompt for a Studio editor session's dedicated worker. Deliberately
+ * does NOT include buildFullSystemPrompt()'s clauth-credential / Supabase-RPC
+ * block — a studio-editor session is scoped to plain source-file edits, not
+ * general repo-or-database agent work, so that access-instruction boilerplate
+ * is pure noise for this task (it doesn't need to fetch a credential or touch
+ * Supabase to change a JSX string). Still editable via the SAME
+ * `.studio-editor-settings.json` `systemPrompt` override buildFullSystemPrompt()
+ * reads — set one there if a given repo's edits genuinely need DB/credential
+ * access.
+ * @param {string} [repoRoot]
+ */
+function buildStudioEditorSystemPrompt(repoRoot = DEFAULT_AGENT_REPO_ROOT) {
+  const root = (repoRoot || DEFAULT_AGENT_REPO_ROOT).replace(/\\/g, "/");
+  const settings = loadEditorSettings(root);
+  if (typeof settings.systemPrompt === "string" && settings.systemPrompt.trim()) {
+    return settings.systemPrompt
+      .replace(/\{repoRoot\}/g, root)
+      .replace(/\{bootstrapPath\}/g, `${root}/${BOOTSTRAP_DOC_REL_PATH}`);
+  }
+  return (
+    "You are a Regen Studio editor agent running on the operator's machine, in the repo at " +
+    root +
+    ". Your job is to make precise, minimal source-code edits per the instructions in " +
+    "each turn's prompt. You have Bash plus file read/write tools. Do not fetch " +
+    "credentials or touch Supabase/work_items — plain source edits never need either. " +
+    "After editing, output the requested JSON result exactly as specified."
+  );
+}
+
+// ── Corpus handoff (Regen Agent corpus, matching @regen/claude-dispatch) ─────
+// Every Regen route agent dispatched through @regen/claude-dispatch gets a
+// `## Corpus` section assembled into its prompt body from the dispatch
+// envelope's `corpus` field. call_agent previously gave only a thin identity
+// system prompt and NO corpus. composeCorpus() reproduces the EXACT dispatch
+// format (packages/claude-dispatch/src/channels/cli.ts buildPrompt) so a
+// call_agent worker starts up with the same handoff a route agent gets.
+//
+// Shape (DispatchCorpus): { product?: string, design?: string, docs?: CorpusDocument[] }
+// CorpusDocument: { path: string, title?: string, content: string }
+//
+// CRITICAL — warm-reuse correctness: the corpus is assembled into the per-turn
+// PROMPT (the stdin/`-p` user message), NOT the worker's persistent
+// append-system-prompt. A warm worker's system prompt is fixed at spawn and
+// shared across all turns; corpus content is variable per call. Because the
+// corpus rides the turn's user message (composeCallAgentPrompt output) and not
+// the worker identity, a generic warm worker can serve a corpus call correctly
+// — the corpus is scoped to that single turn and never leaks into the next.
+
+// Per-document and total corpus byte caps so an oversized handoff cannot blow
+// the agent's context. Each doc/field is truncated to MAX_CORPUS_DOC_CHARS and
+// the assembled section to MAX_CORPUS_TOTAL_CHARS; truncations are logged (not
+// silently dropped). Overridable via env for ops tuning.
+export const MAX_CORPUS_DOC_CHARS = Number.isFinite(Number(process.env.CLAUTH_AGENT_CORPUS_DOC_CHARS))
+  ? Number(process.env.CLAUTH_AGENT_CORPUS_DOC_CHARS)
+  : 24000;
+export const MAX_CORPUS_TOTAL_CHARS = Number.isFinite(Number(process.env.CLAUTH_AGENT_CORPUS_TOTAL_CHARS))
+  ? Number(process.env.CLAUTH_AGENT_CORPUS_TOTAL_CHARS)
+  : 60000;
+
+const BOOTSTRAP_DOC_REL_PATH = ".rdc/guides/agent-bootstrap.md";
+
+/** Truncate a string to `cap` chars, appending a visible marker + logging. */
+function capText(text, cap, label) {
+  if (typeof text !== "string") return "";
+  if (text.length <= cap) return text;
+  const truncated = text.slice(0, cap);
+  console.error(
+    `[agent-pool] corpus ${label} truncated: ${text.length} → ${cap} chars (cap MAX_CORPUS_DOC_CHARS)`
+  );
+  return `${truncated}\n…[truncated ${text.length - cap} chars]`;
+}
+
+/**
+ * Load the project bootstrap guide CONTENT from disk so a "full"-tier agent
+ * gets the REAL bootstrap inline as a Supporting Doc — not just a "go read this
+ * file" pointer. Returns a CorpusDocument or null if unreadable (never throws).
+ * @param {string} [repoRoot]
+ * @returns {{ path: string, title: string, content: string } | null}
+ */
+export function loadBootstrapDoc(repoRoot = DEFAULT_AGENT_REPO_ROOT) {
+  const root = (repoRoot || DEFAULT_AGENT_REPO_ROOT).replace(/\\/g, "/");
+  const abs = path.join(root, BOOTSTRAP_DOC_REL_PATH);
+  try {
+    const content = fs.readFileSync(abs, "utf8");
+    if (!content || !content.trim()) return null;
+    return { path: BOOTSTRAP_DOC_REL_PATH, title: "Agent Bootstrap", content };
+  } catch (e) {
+    console.error(`[agent-pool] bootstrap doc unreadable (${abs}): ${e.message} — skipping`);
+    return null;
+  }
+}
+
+/**
+ * Assemble the `## Corpus` section EXACTLY as @regen/claude-dispatch does
+ * (buildPrompt). Returns the full section string beginning with "## Corpus\n",
+ * or "" when nothing to emit. Order: Product MD, Design MD, Supporting Docs.
+ *
+ * Format (byte-for-byte with dispatch):
+ *   ## Corpus
+ *   ### Product MD\n<product>
+ *
+ *   ### Design MD\n<design>
+ *
+ *   ### Supporting Docs
+ *   - <path> (<title>)\n```\n<content>\n```      ← per doc; title parenthetical
+ *                                                  omitted if no title; docs
+ *                                                  joined by "\n"
+ * (corpusSections joined by "\n\n"; section is "## Corpus" + "\n" + body.)
+ *
+ * @param {{product?:string, design?:string, docs?:Array<{path:string,title?:string,content:string}>}|null|undefined} corpus
+ * @returns {string}
+ */
+export function composeCorpus(corpus) {
+  if (!corpus || typeof corpus !== "object") return "";
+  const corpusSections = [];
+  if (typeof corpus.product === "string" && corpus.product) {
+    corpusSections.push(`### Product MD\n${capText(corpus.product, MAX_CORPUS_DOC_CHARS, "product")}`);
+  }
+  if (typeof corpus.design === "string" && corpus.design) {
+    corpusSections.push(`### Design MD\n${capText(corpus.design, MAX_CORPUS_DOC_CHARS, "design")}`);
+  }
+  if (Array.isArray(corpus.docs) && corpus.docs.length > 0) {
+    const docs = corpus.docs
+      .filter((d) => d && typeof d.path === "string" && typeof d.content === "string")
+      .map((doc) => {
+        const titleLine = doc.title ? ` (${doc.title})` : "";
+        const content = capText(doc.content, MAX_CORPUS_DOC_CHARS, `doc ${doc.path}`);
+        return `- ${doc.path}${titleLine}\n\`\`\`\n${content}\n\`\`\``;
+      })
+      .join("\n");
+    if (docs) corpusSections.push(`### Supporting Docs\n${docs}`);
+  }
+  if (corpusSections.length === 0) return "";
+  let section = `## Corpus\n${corpusSections.join("\n\n")}`;
+  if (section.length > MAX_CORPUS_TOTAL_CHARS) {
+    console.error(
+      `[agent-pool] corpus total truncated: ${section.length} → ${MAX_CORPUS_TOTAL_CHARS} chars (cap MAX_CORPUS_TOTAL_CHARS)`
+    );
+    section = `${section.slice(0, MAX_CORPUS_TOTAL_CHARS)}\n…[corpus truncated]`;
+  }
+  return section;
+}
+
+/**
+ * Build the effective corpus for a call: the caller's corpus (if any), with the
+ * bootstrap doc auto-appended as a Supporting Doc when bootstrap === "full" and
+ * the file is readable. Caller docs come first, then the bootstrap doc, matching
+ * the merge contract (caller docs + bootstrap doc). Returns the assembled
+ * `## Corpus` section string ("" when nothing to emit).
+ *
+ * @param {object} opts
+ * @param {string} opts.bootstrap                 normalized level
+ * @param {object} [opts.corpus]                  caller-supplied DispatchCorpus
+ * @param {string} [opts.repoRoot]                repo root for the bootstrap doc
+ * @param {function} [opts.loadDoc]               injectable loader (tests)
+ * @returns {string}
+ */
+export function buildCorpusSection({ bootstrap, corpus, repoRoot, loadDoc } = {}) {
+  const caller = corpus && typeof corpus === "object" ? corpus : null;
+  let effective = caller;
+  if (bootstrap === "full") {
+    const load = typeof loadDoc === "function" ? loadDoc : loadBootstrapDoc;
+    const bootDoc = load(repoRoot);
+    if (bootDoc) {
+      const callerDocs = Array.isArray(caller?.docs) ? caller.docs : [];
+      effective = {
+        ...(caller || {}),
+        docs: [...callerDocs, bootDoc],
+      };
+    }
+  }
+  return composeCorpus(effective);
+}
+
+/**
+ * Normalize a requested bootstrap level to one of BOOTSTRAP_LEVELS.
+ * Unknown/absent → DEFAULT_BOOTSTRAP ("identity").
+ */
+export function normalizeBootstrap(level) {
+  if (typeof level !== "string") return DEFAULT_BOOTSTRAP;
+  const l = level.trim().toLowerCase();
+  return BOOTSTRAP_LEVELS.includes(l) ? l : DEFAULT_BOOTSTRAP;
+}
+
+/**
+ * Compose the call_agent startup for a bootstrap level. Returns the
+ * append-system-prompt text plus the cwd policy for that tier:
+ *   - none/identity → neutral cwd (caller passes its neutral dir)
+ *   - full          → the repo cwd (so the agent can read the bootstrap guide)
+ *
+ * @param {string} level                "none" | "identity" | "full"
+ * @param {object} [opts]
+ * @param {string} [opts.repoRoot]      repo root for the "full" tier
+ * @param {string} [opts.extra]         optional caller append_system_prompt,
+ *                                      concatenated after the tier text
+ * @returns {{ level:string, appendSystemPrompt:string, useRepoCwd:boolean, repoRoot:string|null }}
+ */
+export function composeBootstrap(level, { repoRoot, extra } = {}) {
+  const lvl = normalizeBootstrap(level);
+  let base;
+  let useRepoCwd = false;
+  let root = null;
+  if (lvl === "none") {
+    base = LEAN_SYSTEM_PROMPT;
+  } else if (lvl === "full") {
+    root = (repoRoot || DEFAULT_AGENT_REPO_ROOT).replace(/\\/g, "/");
+    base = buildFullSystemPrompt(root);
+    useRepoCwd = true;
+  } else {
+    base = IDENTITY_SYSTEM_PROMPT; // identity (default)
+  }
+  const appendSystemPrompt =
+    extra && typeof extra === "string" && extra.trim()
+      ? `${base}\n\n${extra.trim()}`
+      : base;
+  return { level: lvl, appendSystemPrompt, useRepoCwd, repoRoot: root };
+}
+
+function resolveClaudeBinary(explicit) {
+  const candidates = [
+    explicit,
+    process.env.CLAUDE_BIN,
+    path.join(os.homedir(), ".local", "bin", process.platform === "win32" ? "claude.exe" : "claude"),
+    path.join(process.env.APPDATA || "", "npm", "claude.cmd"),
+    path.join(process.env.APPDATA || "", "npm", "claude"),
+    "claude",
+  ].filter(Boolean);
+  for (const c of candidates) {
+    try {
+      execSync(`"${c}" --version`, { stdio: "ignore", timeout: 5000 });
+      return c;
+    } catch {
+      /* try next */
+    }
+  }
+  return null;
+}
+
+function makeJobId(prefix = "call") {
+  return `${prefix}-${Date.now()}-${crypto.randomBytes(4).toString("hex")}`;
+}
+
+/**
+ * Show or hide a process's console window by PID (Windows only). Uses the same
+ * user32.dll ShowWindowAsync/SetForegroundWindow P/Invoke pattern already used by
+ * clauth's terminal_show tool (serve.js showTerminalSession) — this is that same
+ * capability, applied to AgentPool-dispatched job windows, which previously had no
+ * window at all (spawned with windowsHide: true / CREATE_NO_WINDOW). Async and
+ * non-blocking (spawn, not spawnSync) so it never stalls the daemon's event loop.
+ */
+function toggleWindowForPid(pid, show) {
+  return new Promise((resolve) => {
+    if (process.platform !== "win32") {
+      resolve({ shown: false, reason: "unsupported_platform" });
+      return;
+    }
+    const nCmdShow = show ? 5 /* SW_SHOW */ : 0 /* SW_HIDE */;
+    const script = [
+      "Add-Type @'",
+      "using System;",
+      "using System.Runtime.InteropServices;",
+      "public class ClauthAgentWindow {",
+      "  [DllImport(\"user32.dll\")] public static extern bool ShowWindowAsync(IntPtr hWnd, int nCmdShow);",
+      "  [DllImport(\"user32.dll\")] public static extern bool SetForegroundWindow(IntPtr hWnd);",
+      "}",
+      "'@",
+      `$p = Get-Process -Id ${Number(pid)} -ErrorAction SilentlyContinue`,
+      "if ($p -and $p.MainWindowHandle -ne 0) {",
+      `  [ClauthAgentWindow]::ShowWindowAsync($p.MainWindowHandle, ${nCmdShow}) | Out-Null`,
+      show ? "  [ClauthAgentWindow]::SetForegroundWindow($p.MainWindowHandle) | Out-Null" : "",
+      "  'ok'",
+      "} else {",
+      "  'no-window'",
+      "}",
+    ].filter(Boolean).join("\n");
+    let proc;
+    try {
+      proc = spawn("powershell.exe", ["-NoProfile", "-ExecutionPolicy", "Bypass", "-Command", script], {
+        stdio: ["ignore", "pipe", "pipe"],
+        windowsHide: true,
+      });
+    } catch (e) {
+      resolve({ shown: false, reason: `spawn_failed: ${e.message}` });
+      return;
+    }
+    let stdout = "";
+    proc.stdout?.on("data", (d) => { stdout += d.toString(); });
+    proc.on("error", () => resolve({ shown: false, reason: "toggle_process_error" }));
+    proc.on("close", () => {
+      resolve(stdout.includes("ok") ? { shown: show, reason: "ok" } : { shown: false, reason: "no_window" });
+    });
+  });
+}
+
+/**
+ * Build the persistent stream-json argv for a warm worker. Same lean isolation
+ * as the cold buildArgs() (model, skip-permissions, user-only settings, strict
+ * mcp, lean system prompt) but with the persistent stdin/stdout streaming
+ * transport instead of a one-shot `-p "<prompt>"`.
+ */
+function buildWarmArgs({ model, appendSystemPrompt }) {
+  return [
+    "-p",
+    "--input-format",
+    "stream-json",
+    "--output-format",
+    "stream-json",
+    "--verbose", // required by the CLI for stream-json output
+    "--model",
+    model,
+    "--dangerously-skip-permissions",
+    "--setting-sources",
+    "user",
+    "--strict-mcp-config",
+    "--append-system-prompt",
+    appendSystemPrompt || LEAN_SYSTEM_PROMPT,
+  ];
+}
+
+/**
+ * WarmWorker — one long-lived `claude` process in stream-json mode that can be
+ * RE-TARGETED with a new prompt per turn over stdin. This is the true persistent
+ * re-targetable worker: no per-call CLI boot.
+ *
+ * Lifecycle: spawned → ready (process is alive and stdin is connected) → busy
+ * (a turn is in flight) → idle again (after that turn's `result`). On process
+ * death it is marked dead; the pool auto-restarts it. One turn at a time per
+ * worker (the CLI session is single-threaded); the pool routes each call to a
+ * free idle worker.
+ *
+ * READINESS NOTE: the `claude` CLI in `--input-format stream-json` mode emits
+ * NOTHING (not even `system/init`) until the FIRST stdin user-message arrives —
+ * `init` is part of the first turn's event stream, not a standalone boot signal.
+ * So a warm worker is "ready" the moment its process is spawned with a live
+ * stdin pipe. The first `send()` pays the CLI boot/init cost inside its turn;
+ * every subsequent `send()` re-targets the already-booted process (pure
+ * inference latency, no boot). That is the warm win: boot is paid ONCE, then the
+ * hot process is re-aimed per call.
+ */
+export class WarmWorker {
+  /**
+   * @param {object} opts
+   * @param {string} opts.binary           resolved claude binary
+   * @param {string} opts.model            model pinned for this worker
+   * @param {string} opts.cwd              neutral project-free cwd
+   * @param {string} [opts.appendSystemPrompt]
+   * @param {function} [opts.spawnImpl]    inject a fake spawn for tests
+   * @param {function} [opts.onExit]       called when the process dies
+   * @param {boolean} [opts.allowVisibleWindow] spawn with a real (immediately-hidden)
+   *   console window instead of CREATE_NO_WINDOW, so showJob() can later bring it to
+   *   the foreground. Only set for session-pinned dedicated workers (see
+   *   AgentPool.acquireForSession's cwd branch) -- the general shared warm pool never
+   *   needs this and shouldn't pay the extra spawn overhead for every boot.
+   */
+  constructor(opts = {}) {
+    this.binary = opts.binary;
+    this.model = opts.model || DEFAULT_MODEL;
+    this.cwd = opts.cwd;
+    this.appendSystemPrompt = opts.appendSystemPrompt;
+    this._spawn = opts.spawnImpl || spawn;
+    this.onExit = typeof opts.onExit === "function" ? opts.onExit : () => {};
+    this.allowVisibleWindow = !!opts.allowVisibleWindow;
+
+    this.id = `warm-${crypto.randomBytes(4).toString("hex")}`;
+    this.proc = null;
+    this.sessionId = null;
+    this.state = "spawning"; // spawning | ready | busy | dead
+    this.spawnedAt = Date.now();
+    this.turns = 0;
+
+    this._buf = "";
+    this._current = null; // { resolve, stdout, deadline, jobId, settled }
+    this._readyResolve = null;
+    this._readyPromise = new Promise((r) => (this._readyResolve = r));
+  }
+
+  /** True iff the worker is alive and free to accept a new turn. */
+  isIdle() {
+    return this.state === "ready";
+  }
+  isBusy() {
+    return this.state === "busy";
+  }
+  isDead() {
+    return this.state === "dead";
+  }
+
+  /** Spawn the persistent process. Resolves once the worker is ready (or dead). */
+  async start() {
+    let proc;
+    try {
+      const isCmdShim = process.platform === "win32" && /\.cmd$/i.test(this.binary);
+      const args = buildWarmArgs({ model: this.model, appendSystemPrompt: this.appendSystemPrompt });
+      const command = isCmdShim ? "cmd" : this.binary;
+      const spawnArgs = isCmdShim ? ["/d", "/s", "/c", `"${this.binary}"`, ...args] : args;
+      proc = this._spawn(command, spawnArgs, {
+        cwd: this.cwd,
+        env: process.env, // inherits CLI login session; never sets ANTHROPIC_API_KEY
+        stdio: ["pipe", "pipe", "pipe"],
+        shell: false,
+        windowsHide: !this.allowVisibleWindow,
+      });
+    } catch (e) {
+      this.state = "dead";
+      this.error = `spawn_failed: ${e.message}`;
+      this._readyResolve?.(false);
+      return false;
+    }
+
+    this.proc = proc;
+    this.pid = proc.pid;
+    if (this.allowVisibleWindow && process.platform === "win32" && proc.pid) {
+      toggleWindowForPid(proc.pid, false).catch(() => {});
+    }
+
+    proc.stdout?.on("data", (d) => this._onStdout(d));
+    proc.stderr?.on("data", (d) => {
+      this._stderr = `${this._stderr || ""}${d.toString()}`.slice(-20000);
+    });
+    proc.on("error", (e) => this._onDeath(`proc_error: ${e.message}`));
+    proc.on("close", (code) => this._onDeath(`exit_${code}`));
+
+    // In stream-json mode the CLI is silent until the first stdin message, so a
+    // freshly-spawned process with a live stdin pipe IS ready. Mark ready on the
+    // next tick (after spawn errors would have fired synchronously). The first
+    // send() pays the boot cost inside its turn; later sends are hot re-targets.
+    await new Promise((r) => setImmediate(r));
+    if (this.state === "spawning" && this.proc && this.proc.stdin && !this.isDead()) {
+      this.state = "ready";
+      this._readyResolve?.(true);
+    }
+    return await this._readyPromise;
+  }
+
+  _onStdout(d) {
+    this._buf += d.toString();
+    let idx;
+    while ((idx = this._buf.indexOf("\n")) >= 0) {
+      const line = this._buf.slice(0, idx).trim();
+      this._buf = this._buf.slice(idx + 1);
+      if (!line) continue;
+      let msg;
+      try {
+        msg = JSON.parse(line);
+      } catch {
+        continue; // ignore non-JSON noise
+      }
+      this._onEvent(msg);
+    }
+  }
+
+  _onEvent(msg) {
+    if (msg.session_id && !this.sessionId) this.sessionId = msg.session_id;
+    if (msg.type === "system" && msg.subtype === "init") {
+      if (this.state === "spawning") {
+        this.state = "ready";
+        this._readyResolve?.(true);
+      }
+      return;
+    }
+    if (msg.type === "result" && this._current) {
+      const cur = this._current;
+      if (cur.settled) return;
+      cur.settled = true;
+      if (cur.timer) clearTimeout(cur.timer); // C2: cancel internal timeout on normal settle
+      const text = typeof msg.result === "string" ? msg.result : cur.stdout;
+      this._current = null;
+      if (this.state === "busy") this.state = "ready";
+      cur.resolve({
+        ok: !msg.is_error,
+        package: (text || "").trim(),
+        is_error: !!msg.is_error,
+        session_id: this.sessionId,
+      });
+    }
+  }
+
+  _onDeath(reason) {
+    if (this.state === "dead") return;
+    this.state = "dead";
+    this.error = reason;
+    // Fail any in-flight turn so the caller (and concurrency slot) is released.
+    if (this._current && !this._current.settled) {
+      const cur = this._current;
+      cur.settled = true;
+      if (cur.timer) clearTimeout(cur.timer); // C2: cancel internal timeout on death
+      this._current = null;
+      cur.resolve({ ok: false, package: "", error: `worker_died: ${reason}`, dead: true });
+    }
+    // Unblock a start() still waiting on readiness.
+    if (this.state === "dead") this._readyResolve?.(false);
+    this.onExit(this, reason);
+  }
+
+  /**
+   * Re-target this warm worker with a new prompt. One turn at a time.
+   * Resolves with { ok, package, is_error } when the turn's `result` arrives,
+   * or { ok:false, ... } if the deadline elapses (caller/pool kills the worker).
+   * @returns {Promise<object>}
+   */
+  send(prompt, { timeout_ms } = {}) {
+    if (this.state !== "ready") {
+      return Promise.resolve({ ok: false, error: `worker_not_ready:${this.state}` });
+    }
+    return new Promise((resolve) => {
+      this.state = "busy";
+      this.turns++;
+      const ms = Number.isFinite(timeout_ms) ? timeout_ms : DEFAULT_TIMEOUT_MS;
+      const cur = { resolve, stdout: "", settled: false, startedAt: Date.now() };
+      cur.deadline = Date.now() + ms;
+      // C2 — own internal timeout. The pool reaper only covers turns that have a
+      // job record; the PRIMING turn (ensureWarmWorkers) has none, so a hung
+      // prime would otherwise leave `send()`'s promise unsettled forever and
+      // `await Promise.all(primes)` would never resolve. This timer settles the
+      // turn with a failure AND kills the wedged process independently of the
+      // reaper, hardening every send() (priming and normal) against a hung CLI.
+      cur.timer = setTimeout(() => {
+        if (cur.settled) return;
+        cur.settled = true;
+        // Detach so a late `result` event can't double-settle.
+        if (this._current === cur) this._current = null;
+        cur.resolve({ ok: false, package: "", error: "warm_send_timeout", timeout: true });
+        // The session is wedged mid-turn — kill the proc so the worker is
+        // recycled (onExit → circuit-breaker accounting / auto-restart).
+        this.kill("send_timeout");
+      }, ms);
+      if (cur.timer && cur.timer.unref) cur.timer.unref();
+      this._current = cur;
+      const envelope = {
+        type: "user",
+        message: { role: "user", content: [{ type: "text", text: prompt }] },
+      };
+      try {
+        this.proc.stdin.write(JSON.stringify(envelope) + "\n");
+      } catch (e) {
+        if (cur.timer) clearTimeout(cur.timer);
+        cur.settled = true;
+        this._current = null;
+        this.state = "ready";
+        resolve({ ok: false, error: `stdin_write_failed: ${e.message}` });
+      }
+    });
+  }
+
+  /** Deadline of the in-flight turn, or null if idle. Used by the pool reaper. */
+  currentDeadline() {
+    return this._current && !this._current.settled ? this._current.deadline : null;
+  }
+
+  kill(reason = "killed") {
+    try {
+      this.proc?.stdin?.end();
+    } catch {
+      /* ignore */
+    }
+    try {
+      this.proc?.kill("SIGKILL");
+    } catch {
+      /* already gone */
+    }
+    this._onDeath(reason);
+  }
+}
+
+/**
+ * AgentPool — concurrency-limited, warm, model-targeted one-shot dispatcher.
+ */
+export class AgentPool {
+  /**
+   * @param {object} opts
+   * @param {string} [opts.binary]       explicit claude binary path
+   * @param {number} [opts.poolSize]     max concurrent + prime count (default 2)
+   * @param {number} [opts.maxConcurrent] hard concurrency cap (default = poolSize)
+   * @param {string} [opts.defaultModel] default model alias
+   * @param {number} [opts.defaultTimeoutMs]
+   * @param {string} [opts.neutralCwd]   project-free working dir for lean runs
+   * @param {number} [opts.reaperIntervalMs] sweep cadence (default 5000)
+   * @param {number} [opts.maxJobs]      retained job history (default 200)
+   * @param {boolean} [opts.spawnImpl]   inject a fake spawn for tests
+   */
+  constructor(opts = {}) {
+    this.binary = opts.binary || resolveClaudeBinary(opts.binary);
+    this.poolSize = Number.isFinite(opts.poolSize) ? Math.max(1, opts.poolSize) : DEFAULT_POOL_SIZE;
+    this.maxConcurrent = Number.isFinite(opts.maxConcurrent)
+      ? Math.max(1, opts.maxConcurrent)
+      : this.poolSize;
+    this.defaultModel = opts.defaultModel || DEFAULT_MODEL;
+    this.defaultTimeoutMs = Number.isFinite(opts.defaultTimeoutMs)
+      ? opts.defaultTimeoutMs
+      : DEFAULT_TIMEOUT_MS;
+    this.neutralCwd = opts.neutralCwd || path.join(os.tmpdir(), "clauth-call-agent-neutral");
+    this.reaperIntervalMs = Number.isFinite(opts.reaperIntervalMs) ? opts.reaperIntervalMs : 5000;
+    this.maxJobs = Number.isFinite(opts.maxJobs) ? opts.maxJobs : 200;
+    this._spawn = opts.spawnImpl || spawn;
+
+    // WP-P2: true persistent re-targetable warm workers via stream-json.
+    // `warmEnabled` gates the whole warm path; when false the pool is exactly
+    // the legacy cold dispatcher. `warmSize` long-lived workers are kept hot.
+    this.warmEnabled = opts.warmEnabled != null ? !!opts.warmEnabled : WARM_ENABLED_DEFAULT;
+    this.warmSize = Number.isFinite(opts.warmSize) ? Math.max(0, opts.warmSize) : this.poolSize;
+    // When false, skip the per-worker priming turn (used by tests to isolate
+    // lifecycle behavior from prime-turn timing). Default: prime on (undefined).
+    this.warmPrimeTurn = opts.warmPrimeTurn;
+    // Warm workers serve the DEFAULT bootstrap tier ("identity"): they are
+    // pre-spawned with the identity startup prompt (who-am-I + how-to-get-access)
+    // on the neutral cwd. "none" and "full" calls carry a different prompt/cwd and
+    // are routed to the cold path in _drain(), so the warm prompt is the identity
+    // tier — not the bare LEAN prompt — to match what a default call expects.
+    this.appendSystemPrompt = opts.appendSystemPrompt || IDENTITY_SYSTEM_PROMPT;
+    this._warmWorkers = []; // WarmWorker[]
+    this._warmStarting = false; // H2: reentrancy guard for ensureWarmWorkers()
+
+    // C1 — boot-failure circuit breaker. A deterministic boot failure (bad
+    // CLAUTH_AGENT_MODEL, broken login/MCP) makes every warm worker die during
+    // its prime turn and respawn forever. We count CONSECUTIVE immediate deaths
+    // (lifeMs < IMMEDIATE_DEATH_MS); after `warmCircuitThreshold` of them we trip
+    // the breaker: disable warm for the daemon's life and fall back to the cold
+    // path permanently. A worker that lives past the immediate window resets the
+    // counter (genuine flapping isn't the same as a hard boot failure).
+    this.warmCircuitThreshold = Number.isFinite(opts.warmCircuitThreshold)
+      ? Math.max(1, opts.warmCircuitThreshold)
+      : 5;
+    this.immediateDeathMs = Number.isFinite(opts.immediateDeathMs)
+      ? opts.immediateDeathMs
+      : 2000;
+    this._consecutiveImmediateDeaths = 0;
+    this._warmCircuitTripped = false;
+    this._warmCircuitLogged = false;
+    // Base unit for the immediate-death respawn backoff (exponential, capped at
+    // 30s). Overridable so tests can exercise the breaker quickly.
+    this.warmBackoffBaseMs = Number.isFinite(opts.warmBackoffBaseMs) ? opts.warmBackoffBaseMs : 1000;
+
+    // WP-5: session-level warm worker pinning. A session (e.g. Studio editor)
+    // can claim a warm worker and reuse it across turns, preserving conversation
+    // context. Pinned workers are excluded from the general _acquireWarmWorker()
+    // pool so other callers cannot steal them mid-session.
+    this._sessionPins = new Map(); // Map<sessionId, WarmWorker>
+
+    this.jobs = new Map(); // jobId -> job record
+    this.active = 0;
+    this.queue = []; // pending { resolve, reject, args }
+    this.warm = false;
+    this._reaper = null;
+
+    try {
+      fs.mkdirSync(this.neutralCwd, { recursive: true });
+    } catch {
+      /* best effort */
+    }
+  }
+
+  available() {
+    return this.binary != null;
+  }
+
+  startReaper() {
+    if (this._reaper) return;
+    this._reaper = setInterval(() => this.reap(), this.reaperIntervalMs);
+    // Note: intentionally NOT unref'd. While a job is in flight the reaper is
+    // the only thing that can settle a hung worker, so it must keep the loop
+    // alive. The daemon owns lifecycle and calls shutdown()/stopReaper().
+  }
+
+  stopReaper() {
+    if (this._reaper) {
+      clearInterval(this._reaper);
+      this._reaper = null;
+    }
+  }
+
+  /**
+   * Kill any running job past its deadline, free its slot, drain the queue.
+   * Also evicts old terminal jobs beyond maxJobs.
+   */
+  reap() {
+    const now = Date.now();
+    let reaped = 0;
+    for (const job of this.jobs.values()) {
+      if (job.status === "running" && job.deadline && now > job.deadline) {
+        this._killJob(job, "timeout");
+        reaped++;
+      }
+    }
+    // evict oldest terminal jobs
+    if (this.jobs.size > this.maxJobs) {
+      const terminal = [...this.jobs.values()]
+        .filter((j) => j.status !== "running")
+        .sort((a, b) => (a.completed_at || 0) - (b.completed_at || 0));
+      let toRemove = this.jobs.size - this.maxJobs;
+      for (const j of terminal) {
+        if (toRemove-- <= 0) break;
+        this.jobs.delete(j.jobId);
+      }
+    }
+    this._drain();
+    return reaped;
+  }
+
+  _killJob(job, reason) {
+    if (job._proc && job.status === "running") {
+      try {
+        job._proc.kill("SIGKILL");
+      } catch {
+        /* already gone */
+      }
+    }
+    // Warm job past its deadline: the worker's session is wedged mid-turn. Kill
+    // the persistent worker (auto-restart re-tops the pool) so the slot frees.
+    if (job._worker && job.status === "running") {
+      try {
+        job._worker.kill(reason === "timeout" ? "turn_timeout" : reason);
+      } catch {
+        /* already gone */
+      }
+    }
+    if (job.status === "running") {
+      job.status = reason === "timeout" ? "timeout" : "killed";
+      job.completed_at = Date.now();
+      job.ms = job.completed_at - job.started_at;
+      job.error = reason;
+      // NOTE: do NOT decrement this.active here. settle() is the single source
+      // of truth for the concurrency-slot decrement (it guards with the
+      // `settled` flag). Decrementing here too double-counted every reaped /
+      // timed-out / shutdown-killed job, dropping active by 2 and letting
+      // _drain() launch more than maxConcurrent real workers.
+      if (job._settle) job._settle();
+    }
+  }
+
+  buildArgs({ prompt, model, appendSystemPrompt }) {
+    return [
+      "-p",
+      prompt,
+      "--model",
+      model || this.defaultModel,
+      "--dangerously-skip-permissions",
+      // Lean isolation: only the user's auth/settings, no project CLAUDE.md/hooks.
+      "--setting-sources",
+      "user",
+      "--strict-mcp-config",
+      "--append-system-prompt",
+      appendSystemPrompt || LEAN_SYSTEM_PROMPT,
+    ];
+  }
+
+  // ── WP-P2 warm worker management ───────────────────────────────────────────
+
+  /** Count of alive (ready or busy) warm workers. */
+  _liveWarmCount() {
+    return this._warmWorkers.filter((w) => !w.isDead()).length;
+  }
+
+  /**
+   * First idle warm worker matching `model`, or null. Warm workers are pinned
+   * to one model; a call requesting a different model must NOT be silently
+   * served by a mismatched warm worker — it falls back to a cold spawn that
+   * honors the requested model. Workers pinned to a session are excluded so
+   * general dispatch cannot steal a session's dedicated worker.
+   */
+  _acquireWarmWorker(model) {
+    const pinnedSet = new Set(this._sessionPins.values());
+    return this._warmWorkers.find((w) =>
+      w.isIdle() && (!model || w.model === model) && !pinnedSet.has(w)
+    ) || null;
+  }
+
+  // ── WP-5: session-level warm worker pinning ─────────────────────────────
+
+  /**
+   * Acquire (or re-acquire) a warm worker pinned to `sessionId`. Idempotent:
+   * calling with the same sessionId returns the already-pinned worker. A stale
+   * pin (dead worker) is transparently replaced. Returns null when no warm
+   * worker is available (caller should fall back or queue).
+   *
+   * IMPORTANT — cwd: a WarmWorker's working directory is fixed at process spawn
+   * and CANNOT be changed later. The shared general warm pool (ensureWarmWorkers)
+   * always boots workers in this.neutralCwd (a throwaway tmp dir) so a generic
+   * one-shot call_agent turn never accidentally sees a real repo. That means a
+   * worker pulled from the shared pool cannot reliably reach files by relative
+   * path for a session that needs a specific repo. When `opts.cwd` is given
+   * (e.g. the editor session's repoRoot), this boots a DEDICATED worker rooted
+   * there instead of reusing the shared neutral-cwd pool — it is never returned
+   * to the general pool on release (see releaseSession below), so it can never
+   * leak the wrong cwd into an unrelated one-shot dispatch.
+   *
+   * @param {string} sessionId
+   * @param {object} [opts]
+   * @param {string} [opts.model]
+   * @param {string} [opts.cwd]
+   * @returns {WarmWorker|null}
+   */
+  acquireForSession(sessionId, opts = {}) {
+    if (this._sessionPins.has(sessionId)) {
+      const worker = this._sessionPins.get(sessionId);
+      if (!worker.isDead()) return worker;
+      this._sessionPins.delete(sessionId); // stale pin, re-acquire below
+    }
+    const model = opts.model || this.defaultModel;
+    const cwd = typeof opts.cwd === "string" && opts.cwd.trim() ? opts.cwd.trim() : null;
+
+    let worker;
+    if (cwd) {
+      // This worker genuinely runs IN the repo (unlike the shared neutral-cwd pool), so
+      // it's told it's running at `cwd` — but with the LEAN studio-editor prompt, not
+      // buildFullSystemPrompt()'s general clauth-credential/Supabase-RPC block. A studio
+      // session is scoped to plain source-file edits; it never needs to fetch a
+      // credential or touch Supabase, so that boilerplate is unnecessary noise here.
+      worker = new WarmWorker({
+        binary: this.binary,
+        model,
+        cwd,
+        appendSystemPrompt: buildStudioEditorSystemPrompt(cwd),
+        spawnImpl: this._spawn,
+        onExit: (w, reason) => this._onWarmWorkerExit(w, reason),
+        allowVisibleWindow: true,
+      });
+      worker._dedicatedSession = true;
+      this._warmWorkers.push(worker);
+      // Fire-and-forget: dispatchToSession's tryRun() loop already polls isIdle()/isDead(),
+      // so callers don't need to await boot here.
+      worker.start().catch(() => {});
+    } else {
+      worker = this._acquireWarmWorker(model);
+      if (!worker) return null;
+    }
+    this._sessionPins.set(sessionId, worker);
+    return worker;
+  }
+
+  /**
+   * Release a session's pin. A general-pool worker (no dedicated cwd) returns
+   * to the shared pool — it is just no longer reserved. A dedicated worker
+   * (booted with a session-specific cwd) is killed and removed instead of
+   * being returned to the pool, since its cwd would be wrong for any other
+   * session or general one-shot dispatch that later acquired it.
+   * @param {string} sessionId
+   * @returns {WarmWorker|null}
+   */
+  releaseSession(sessionId) {
+    const worker = this._sessionPins.get(sessionId);
+    this._sessionPins.delete(sessionId);
+    if (worker && worker._dedicatedSession) {
+      worker.kill("session_released");
+      this._warmWorkers = this._warmWorkers.filter((w) => w !== worker);
+    }
+    return worker || null;
+  }
+
+  /**
+   * Dispatch a prompt to the worker pinned for `sessionId`. Unlike general
+   * dispatch, this always targets the pinned worker (preserving conversation
+   * context). Returns a rejected-shape result if no pin exists or the worker
+   * is dead. If the worker is busy (prior turn still in flight), waits for it.
+   *
+   * @param {string} sessionId
+   * @param {object} [args]  { prompt, timeout_ms, ... }
+   * @returns {Promise<object>}
+   */
+  dispatchToSession(sessionId, args = {}) {
+    const worker = this._sessionPins.get(sessionId);
+    if (!worker || worker.isDead()) {
+      return Promise.resolve({ ok: false, error: "no_pinned_worker", sessionId });
+    }
+    const prompt = args.prompt;
+    if (!prompt || typeof prompt !== "string") {
+      return Promise.resolve({ ok: false, error: "prompt_required" });
+    }
+    return new Promise((resolve, reject) => {
+      const item = { args, resolve, reject };
+      const tryRun = () => {
+        if (worker.isDead()) {
+          return resolve({ ok: false, error: "pinned_worker_died", sessionId });
+        }
+        if (worker.isIdle()) {
+          this._runWarm(item, worker);
+        } else {
+          setTimeout(tryRun, 100);
+        }
+      };
+      tryRun();
+    });
+  }
+
+  /**
+   * Query the pin state for a session. Returns null if no pin exists.
+   * @param {string} sessionId
+   * @returns {object|null}
+   */
+  getSessionInfo(sessionId) {
+    const worker = this._sessionPins.get(sessionId);
+    if (!worker) return null;
+    return {
+      sessionId,
+      pinned: true,
+      alive: !worker.isDead(),
+      idle: worker.isIdle(),
+      model: worker.model,
+      turns: worker.turns,
+    };
+  }
+
+  /** Drop dead workers from the roster. */
+  _pruneDeadWarm() {
+    this._warmWorkers = this._warmWorkers.filter((w) => !w.isDead());
+  }
+
+  /**
+   * Boot `warmSize` persistent workers (idempotent — only tops up to target).
+   * Auto-restart is wired via each worker's onExit, which respawns a fresh
+   * worker if the pool still wants warm capacity and isn't shutting down.
+   */
+  async ensureWarmWorkers() {
+    if (!this.warmEnabled || !this.available() || this._shuttingDown) return 0;
+    if (this._warmCircuitTripped) return 0; // C1: breaker open — warm is permanently off
+    // H2 — reentrancy guard. Two overlapping ensureWarmWorkers() calls (e.g. a
+    // boot prime racing a manual top-up) would each independently compute the
+    // deficit BEFORE the other's workers register, and both spawn — overshooting
+    // warmSize. Honor `_warmStarting`: a second concurrent entrant returns 0 and
+    // lets the in-flight call do the provisioning.
+    if (this._warmStarting) return 0;
+    this._warmStarting = true;
+    try {
+      this._pruneDeadWarm();
+      const deficit = this.warmSize - this._liveWarmCount();
+      if (deficit <= 0) return 0;
+      const starts = [];
+      for (let i = 0; i < deficit; i++) {
+        const w = new WarmWorker({
+          binary: this.binary,
+          model: this.defaultModel,
+          cwd: this.neutralCwd,
+          appendSystemPrompt: this.appendSystemPrompt,
+          spawnImpl: this._spawn,
+          onExit: (worker, reason) => this._onWarmWorkerExit(worker, reason),
+        });
+        this._warmWorkers.push(w);
+        starts.push(w.start().catch(() => false));
+      }
+      const results = await Promise.all(starts);
+      this._pruneDeadWarm();
+      // Fire ONE priming turn per freshly-started worker so the CLI boot/init cost
+      // is paid during warm-up, not on the user's first real call. After this, the
+      // worker's session is hot and every dispatch is a pure-inference re-target.
+      // C2: send() now has its own internal timeout, so a hung prime can no longer
+      // wedge this Promise.all forever.
+      if (this.warmPrimeTurn !== false) {
+        const primes = [];
+        for (const w of this._warmWorkers) {
+          if (w.isIdle() && w.turns === 0) {
+            primes.push(w.send("Reply with exactly: OK", { timeout_ms: 60000 }).catch(() => null));
+          }
+        }
+        await Promise.all(primes);
+        this._pruneDeadWarm();
+      }
+      this.warm = this._warmWorkers.some((w) => !w.isDead());
+      return results.filter(Boolean).length;
+    } finally {
+      this._warmStarting = false;
+    }
+  }
+
+  /**
+   * Auto-restart on death: replace a dead worker if still under target. A short
+   * backoff prevents a tight respawn loop when a worker dies the instant it
+   * starts (e.g. a session that crashes immediately) — without it, each death
+   * would synchronously schedule another start that dies and reschedules,
+   * spinning CPU and spawning processes in a hot loop.
+   */
+  _onWarmWorkerExit(worker, reason) {
+    this._pruneDeadWarm();
+    this.warm = this._warmWorkers.some((w) => !w.isDead());
+
+    // Once shutting down, a dying worker is expected teardown — do not count it
+    // toward the boot-failure breaker and do not respawn. (Also prevents a
+    // late-firing backoff respawn's death from tripping the breaker post-shutdown.)
+    if (this._shuttingDown) return;
+
+    // C1/H1 — death classification. Measure how long THIS worker lived. An
+    // immediate death (lifeMs < immediateDeathMs) is the signature of a hard boot
+    // failure (bad model, login/MCP error) — the worker died during boot/prime.
+    // A worker that lived past the window died for a transient reason (a normal
+    // turn, an OOM, a kill) and must NOT clear protection a different flapping
+    // worker raised: we only reset the consecutive-immediate-death counter when a
+    // worker proves the boot path actually works by surviving the window.
+    const lifeMs = worker && worker.spawnedAt ? Date.now() - worker.spawnedAt : Infinity;
+    const immediate = lifeMs < this.immediateDeathMs;
+
+    if (immediate) {
+      this._consecutiveImmediateDeaths++;
+    } else {
+      // A healthy lifetime proves boot works → safe to forgive accumulated
+      // immediate-death backoff for the lineage.
+      this._consecutiveImmediateDeaths = 0;
+    }
+
+    // C1 — trip the breaker after N consecutive immediate deaths. Disable warm
+    // for the daemon's life, log ONCE, and stop respawning. dispatch() already
+    // falls back to the cold path whenever no warm worker is free, so flipping
+    // warmEnabled=false makes the cold fallback permanent. This protects the live
+    // daemon from endless `claude` churn on a misconfigured CLAUTH_AGENT_MODEL.
+    if (!this._warmCircuitTripped && this._consecutiveImmediateDeaths >= this.warmCircuitThreshold) {
+      this._warmCircuitTripped = true;
+      this.warmEnabled = false;
+      if (!this._warmCircuitLogged) {
+        this._warmCircuitLogged = true;
+        console.error(
+          `[agent-pool] WARM CIRCUIT BREAKER TRIPPED: ${this._consecutiveImmediateDeaths} ` +
+            `consecutive warm-worker boot failures (last: ${reason || worker?.error || "unknown"}). ` +
+            `Disabling warm workers for the rest of this process; call_agent now uses the cold ` +
+            `path exclusively. Check CLAUTH_AGENT_MODEL / CLI login / MCP config.`
+        );
+      }
+      return; // do not respawn — breaker is open
+    }
+
+    if (!this.warmEnabled || this._shuttingDown || !this.available()) return;
+    if (this._liveWarmCount() >= this.warmSize) return;
+
+    // H1 — backoff keyed to the CONSECUTIVE immediate-death count (lineage
+    // protection), not a shared mutable counter that any healthy exit zeroes.
+    // Each successive immediate death backs off longer (capped at 30s); a
+    // worker that lived past the window respawns immediately.
+    const delay = immediate
+      ? Math.min(30000, this.warmBackoffBaseMs * Math.pow(2, Math.max(0, this._consecutiveImmediateDeaths - 1)))
+      : 0;
+
+    const spawnReplacement = () => {
+      if (this._warmCircuitTripped) return;
+      if (!this.warmEnabled || this._shuttingDown || !this.available()) return;
+      if (this._liveWarmCount() >= this.warmSize) return;
+      const replacement = new WarmWorker({
+        binary: this.binary,
+        model: this.defaultModel,
+        cwd: this.neutralCwd,
+        appendSystemPrompt: this.appendSystemPrompt,
+        spawnImpl: this._spawn,
+        onExit: (w, r) => this._onWarmWorkerExit(w, r),
+      });
+      this._warmWorkers.push(replacement);
+      replacement
+        .start()
+        .then(() => {
+          this.warm = this._warmWorkers.some((w) => !w.isDead());
+        })
+        .catch(() => {});
+    };
+    if (delay > 0) {
+      const t = setTimeout(spawnReplacement, delay);
+      if (t.unref) t.unref();
+    } else {
+      spawnReplacement();
+    }
+  }
+
+  /**
+   * Pre-warm. With warm workers enabled (default), boots `warmSize` persistent
+   * re-targetable stream-json workers. Otherwise (legacy/cold), fires poolSize
+   * trivial one-shot pings so the OS file cache / model session is warm.
+   * Returns when warming settles (or immediately if unavailable).
+   */
+  async prime() {
+    if (!this.available()) return { warmed: 0, available: false, mode: "unavailable" };
+    if (this.warmEnabled) {
+      const warmed = await this.ensureWarmWorkers();
+      return { warmed, available: true, mode: "warm-workers", workers: this._liveWarmCount() };
+    }
+    const n = this.poolSize;
+    const pings = [];
+    for (let i = 0; i < n; i++) {
+      pings.push(
+        this.dispatch({
+          prompt: "Reply with exactly: OK",
+          timeout_ms: 30000,
+          _priming: true,
+        }).catch((e) => ({ ok: false, error: e.message }))
+      );
+    }
+    const results = await Promise.all(pings);
+    this.warm = results.some((r) => r && r.ok);
+    return { warmed: results.filter((r) => r && r.ok).length, available: true, mode: "cold-prime" };
+  }
+
+  /**
+   * Dispatch a one-shot. Resolves with a job record once the worker exits
+   * (sync semantics). Concurrency beyond maxConcurrent is queued.
+   *
+   * @returns {Promise<object>} job record { ok, jobId, package, model, ms, status, exit_code, error }
+   */
+  dispatch(args = {}) {
+    return new Promise((resolve, reject) => {
+      const item = { args, resolve, reject };
+      this.queue.push(item);
+      this._drain();
+    });
+  }
+
+  _drain() {
+    while (this.active < this.maxConcurrent && this.queue.length > 0) {
+      const item = this.queue.shift();
+      // Priming pings stay on the cold one-shot path (they exist only to warm
+      // the cold cache; with warm workers they are a no-op caller). Real calls
+      // prefer a free warm worker and fall back to cold when none is free.
+      //
+      // Bootstrap pinning (tiered startup): warm workers are pre-spawned with the
+      // DEFAULT ("identity") startup prompt and the neutral cwd. A call asking for
+      // a DIFFERENT bootstrap level — "none" (lean fast path) or "full" (repo cwd
+      // + bootstrap guide) — has a different system prompt and/or cwd, so it must
+      // NOT be served by a warm worker. Route it to the cold path, which honors
+      // args.appendSystemPrompt and args.cwd. Default/absent bootstrap → warm.
+      const bootstrapLevel = item.args.bootstrap || DEFAULT_BOOTSTRAP;
+      const warmEligible = bootstrapLevel === DEFAULT_BOOTSTRAP;
+      if (this.warmEnabled && !item.args._priming && warmEligible) {
+        // Only a warm worker pinned to the REQUESTED model may serve the call;
+        // a model mismatch must fall through to the cold path (which honors
+        // args.model) rather than be silently downgraded to the warm model.
+        const reqModel = item.args.model || this.defaultModel;
+        const worker = this._acquireWarmWorker(reqModel);
+        if (worker) {
+          this._runWarm(item, worker);
+          continue;
+        }
+      }
+      this._run(item);
+    }
+  }
+
+  /**
+   * Run a call on a persistent warm worker (re-targeting it with a new prompt).
+   * Produces the SAME public job record shape as the cold path so callers
+   * (runCallAgent, durable delivery, GET /call-agent/:jobId) are unchanged.
+   */
+  _runWarm(item, worker) {
+    const { args, resolve } = item;
+    const prompt = args.prompt;
+    if (!prompt || typeof prompt !== "string") {
+      return resolve({ ok: false, error: "prompt_required" });
+    }
+    const model = worker.model;
+    const timeout = Number.isFinite(args.timeout_ms) ? args.timeout_ms : this.defaultTimeoutMs;
+    const jobId = args.jobId || makeJobId("call");
+
+    const job = {
+      jobId,
+      status: "running",
+      model,
+      priming: false,
+      warm: true,
+      pid: worker.pid || null,
+      started_at: Date.now(),
+      deadline: Date.now() + timeout,
+      completed_at: null,
+      ms: null,
+      exit_code: null,
+      stdout: "",
+      stderr: "",
+      package: null,
+      error: null,
+      _proc: null,
+      _worker: worker,
+      _settle: null,
+    };
+    this.jobs.set(jobId, job);
+    this.active++;
+
+    let settled = false;
+    const settle = () => {
+      if (settled) return;
+      settled = true;
+      this.active = Math.max(0, this.active - 1);
+      const rec = {
+        ok: job.status === "completed",
+        jobId: job.jobId,
+        status: job.status,
+        model: job.model,
+        ms: job.ms,
+        exit_code: job.exit_code,
+        package: job.package,
+        error: job.error,
+        warm: true,
+      };
+      if (job.stderr && job.status !== "completed") rec.stderr = job.stderr.slice(-2000);
+      this._drain();
+      resolve(rec);
+    };
+    job._settle = () => {
+      // Reaper path: the turn blew its deadline. Kill the worker (its session is
+      // wedged); auto-restart re-tops the pool. Mark the job per its set status.
+      if (job.status === "running") {
+        job.status = "timeout";
+        job.completed_at = Date.now();
+        job.ms = job.completed_at - job.started_at;
+        job.error = "timeout";
+        try {
+          worker.kill("turn_timeout");
+        } catch {
+          /* ignore */
+        }
+      }
+      settle();
+    };
+
+    worker
+      .send(prompt, { timeout_ms: timeout })
+      .then((res) => {
+        if (settled) return; // reaper already settled this
+        job.completed_at = Date.now();
+        job.ms = job.completed_at - job.started_at;
+        if (res.ok) {
+          job.status = "completed";
+          job.exit_code = 0;
+          job.package = res.package;
+        } else {
+          job.status = "failed";
+          job.error = res.error || (res.is_error ? "agent_error" : "warm_failed");
+          if (res.package) job.package = res.package;
+        }
+        settle();
+      })
+      .catch((e) => {
+        if (settled) return;
+        job.status = "failed";
+        job.error = `warm_dispatch: ${e.message}`;
+        job.completed_at = Date.now();
+        job.ms = job.completed_at - job.started_at;
+        settle();
+      });
+  }
+
+  _run({ args, resolve, reject }) {
+    if (!this.available()) {
+      return resolve({
+        ok: false,
+        error: "binary_not_found",
+        message: "claude CLI not found in PATH or AppData/npm",
+      });
+    }
+    const prompt = args.prompt;
+    if (!prompt || typeof prompt !== "string") {
+      return resolve({ ok: false, error: "prompt_required" });
+    }
+    const model = args.model || this.defaultModel;
+    const timeout = Number.isFinite(args.timeout_ms) ? args.timeout_ms : this.defaultTimeoutMs;
+    const jobId = args.jobId || makeJobId(args._priming ? "warm" : "call");
+    const cwd = args.cwd || this.neutralCwd;
+
+    const job = {
+      jobId,
+      status: "running",
+      model,
+      priming: !!args._priming,
+      started_at: Date.now(),
+      deadline: Date.now() + timeout,
+      completed_at: null,
+      ms: null,
+      exit_code: null,
+      stdout: "",
+      stderr: "",
+      package: null,
+      error: null,
+      _proc: null,
+      _settle: null,
+    };
+    this.jobs.set(jobId, job);
+    this.active++;
+
+    let settled = false;
+    const settle = () => {
+      if (settled) return;
+      settled = true;
+      this.active = Math.max(0, this.active - 1);
+      // Build the public record (no internal handles).
+      const rec = {
+        ok: job.status === "completed",
+        jobId: job.jobId,
+        status: job.status,
+        model: job.model,
+        ms: job.ms,
+        exit_code: job.exit_code,
+        package: job.package,
+        error: job.error,
+      };
+      if (job.stderr && job.status !== "completed") rec.stderr = job.stderr.slice(-2000);
+      this._drain();
+      resolve(rec);
+    };
+    job._settle = () => settle();
+
+    const cmdArgs = this.buildArgs({
+      prompt,
+      model,
+      appendSystemPrompt: args.appendSystemPrompt,
+    });
+
+    let proc;
+    try {
+      const isCmdShim = process.platform === "win32" && /\.cmd$/i.test(this.binary);
+      const command = isCmdShim ? "cmd" : this.binary;
+      const spawnArgs = isCmdShim ? ["/d", "/s", "/c", `"${this.binary}"`, ...cmdArgs] : cmdArgs;
+      proc = this._spawn(command, spawnArgs, {
+        cwd,
+        env: process.env, // inherits CLI login session; no ANTHROPIC_API_KEY injected
+        stdio: ["ignore", "pipe", "pipe"],
+        shell: false,
+        windowsHide: true,
+      });
+    } catch (e) {
+      job.status = "failed";
+      job.error = `spawn_failed: ${e.message}`;
+      job.completed_at = Date.now();
+      job.ms = job.completed_at - job.started_at;
+      return settle();
+    }
+
+    job._proc = proc;
+    job.pid = proc.pid;
+    // REVERTED (2026-07-02): a visible window here is a dead end — this job's stdin/stdout
+    // are piped to this Node process for programmatic stream-json parsing, not connected to
+    // the console, so "showing" it just displays an empty, non-interactive window. Real
+    // interactive access needs a genuinely separate terminal (see codevelop.js's wt.exe
+    // pattern) attached to a session-pinned warm worker, not this ephemeral one-shot job.
+
+    proc.stdout?.on("data", (d) => {
+      job.stdout = `${job.stdout}${d.toString()}`.slice(-200000);
+    });
+    proc.stderr?.on("data", (d) => {
+      job.stderr = `${job.stderr}${d.toString()}`.slice(-20000);
+    });
+    proc.on("error", (e) => {
+      if (job.status !== "running") return;
+      job.status = "failed";
+      job.error = `proc_error: ${e.message}`;
+      job.completed_at = Date.now();
+      job.ms = job.completed_at - job.started_at;
+      settle();
+    });
+    proc.on("close", (code) => {
+      if (job.status !== "running") return; // reaper/kill already settled
+      job.exit_code = code;
+      job.completed_at = Date.now();
+      job.ms = job.completed_at - job.started_at;
+      if (code === 0) {
+        job.status = "completed";
+        job.package = job.stdout.trim();
+      } else {
+        job.status = "failed";
+        job.error = `exit_${code}`;
+      }
+      settle();
+    });
+  }
+
+  getJob(jobId) {
+    const job = this.jobs.get(jobId);
+    if (!job) return null;
+    return {
+      jobId: job.jobId,
+      status: job.status,
+      model: job.model,
+      ms: job.ms,
+      exit_code: job.exit_code,
+      package: job.package,
+      error: job.error,
+      started_at: job.started_at,
+      completed_at: job.completed_at,
+      pid: job.pid || null,
+    };
+  }
+
+  /**
+   * Bring a running job's console window to the foreground (un-hides it — the window
+   * exists but was hidden immediately after spawn, see _run()'s post-spawn hide call).
+   * No-op with a clear reason if the job is unknown, has no pid, or already finished.
+   */
+  async showJob(jobId) {
+    const job = this.jobs.get(jobId);
+    if (!job) return { shown: false, reason: "job_not_found" };
+    if (!job.pid) return { shown: false, reason: "no_pid" };
+    if (job.status !== "running") return { shown: false, reason: "job_not_running" };
+    return toggleWindowForPid(job.pid, true);
+  }
+
+  /** Re-hide a job's console window (e.g. operator is done looking at it). */
+  async hideJob(jobId) {
+    const job = this.jobs.get(jobId);
+    if (!job) return { shown: false, reason: "job_not_found" };
+    if (!job.pid) return { shown: false, reason: "no_pid" };
+    return toggleWindowForPid(job.pid, false);
+  }
+
+  stats() {
+    let running = 0,
+      completed = 0,
+      failed = 0;
+    for (const j of this.jobs.values()) {
+      if (j.status === "running") running++;
+      else if (j.status === "completed") completed++;
+      else failed++;
+    }
+    const warmReady = this._warmWorkers.filter((w) => w.isIdle()).length;
+    const warmBusy = this._warmWorkers.filter((w) => w.isBusy()).length;
+    return {
+      active: this.active,
+      queued: this.queue.length,
+      maxConcurrent: this.maxConcurrent,
+      poolSize: this.poolSize,
+      // `warm` reflects REAL persistent workers when warm mode is on (WP-P2):
+      // true iff at least one live worker exists. Cold mode keeps legacy meaning.
+      warm: this.warmEnabled ? this._liveWarmCount() > 0 : this.warm,
+      warmEnabled: this.warmEnabled,
+      warmCircuitTripped: this._warmCircuitTripped,
+      warmWorkers: {
+        target: this.warmSize,
+        live: this._liveWarmCount(),
+        ready: warmReady,
+        busy: warmBusy,
+        consecutiveImmediateDeaths: this._consecutiveImmediateDeaths,
+      },
+      jobs: { running, completed, failed, total: this.jobs.size },
+    };
+  }
+
+  shutdown() {
+    this._shuttingDown = true;
+    this.stopReaper();
+    for (const job of this.jobs.values()) {
+      if (job.status === "running") this._killJob(job, "shutdown");
+    }
+    // Clear session pins before tearing down workers.
+    this._sessionPins.clear();
+    // Tear down persistent warm workers so no orphan `claude` processes linger.
+    for (const w of this._warmWorkers) {
+      try {
+        w.kill("shutdown");
+      } catch {
+        /* already gone */
+      }
+    }
+    this._warmWorkers = [];
+  }
+}
+
+// ── WP-P7: hop-limited sub-agent delegation ──────────────────────────────────
+//
+// THE DEADLOCK we must make impossible:
+//   The warm/interactive pool has only ~2 workers. A PARENT call_agent worker is
+//   itself a running `claude` process. If that parent calls call_agent and the
+//   sub-call needs a *pool* worker while all pool workers are held by parents
+//   blocked on their own sub-calls → classic resource deadlock. So delegated
+//   (sub-agent) calls MUST NOT touch the AgentPool. They run on this separate
+//   lane, which:
+//     1. NEVER references an AgentPool, warm worker, or its queue. It only
+//        cold-spawns its own `claude` one-shots, capped by `maxInFlight`. The
+//        interactive pool's latency/availability is therefore unaffected by
+//        delegation depth or fan-out.
+//     2. Enforces a strict HOP BUDGET (carried in agent_context.delegation):
+//        a call at hop >= max is REJECTED with a clean error and NO spawn, so a
+//        runaway delegation loop is impossible and tree depth is bounded.
+//     3. Enforces a daemon-wide in-flight cap so a wide fan-out can't exhaust
+//        resources.
+//
+// Hop accounting: the delegation context on an INCOMING call describes the call
+// itself — `{ hop: N, max: M }` means "this is a hop-N delegated call, budget M".
+// A top-level (interactive) call has no delegation context (hop is treated as 0).
+// When a worker at hop N spawns a sub-agent, the daemon computes the child's
+// context as `{ hop: N+1, max: M }` and rejects if `N+1 > M`. Equivalently, an
+// incoming delegated call is admitted iff `hop <= max`; hop is the depth of the
+// call being served, and a call at `hop > max` (or `hop == max` asking to go
+// deeper) is refused.
+
+/**
+ * Normalize / validate a delegation marker from agent_context.delegation.
+ * Returns { hop, max } with safe integer coercion, or null if absent/invalid.
+ * A present-but-malformed marker normalizes to hop 0 with the default max so a
+ * caller cannot smuggle a negative hop or a giant max past the budget.
+ */
+export function normalizeDelegation(input, { defaultMax = DEFAULT_DELEGATION_MAX_HOPS } = {}) {
+  if (input == null) return null;
+  if (typeof input !== "object") return null;
+  let hop = Number(input.hop);
+  let max = Number(input.max);
+  if (!Number.isFinite(hop) || hop < 0) hop = 0;
+  hop = Math.floor(hop);
+  if (!Number.isFinite(max) || max < 0) max = defaultMax;
+  max = Math.floor(max);
+  return { hop, max };
+}
+
+/**
+ * Decide whether an incoming DELEGATED call may proceed, and compute the child
+ * delegation context the spawned sub-agent should carry if IT delegates further.
+ *
+ * @param {object|null} delegation  normalized { hop, max } of the incoming call
+ * @param {object} [opts]
+ * @param {number} [opts.defaultMax]
+ * @returns {{ allow: boolean, hop: number, max: number, childContext?: object, error?: string }}
+ */
+export function evaluateHopBudget(delegation, { defaultMax = DEFAULT_DELEGATION_MAX_HOPS } = {}) {
+  const d = normalizeDelegation(delegation, { defaultMax }) || { hop: 0, max: defaultMax };
+  // The call being served is at depth `hop`. It is admissible iff it does not
+  // exceed the budget. `hop > max` means a parent tried to spawn one level too
+  // deep — refuse with a clean error, no spawn.
+  if (d.hop > d.max) {
+    return {
+      allow: false,
+      hop: d.hop,
+      max: d.max,
+      error: "delegation_hop_exceeded",
+    };
+  }
+  // Child context for any sub-agent THIS call spawns: one hop deeper, same max.
+  return {
+    allow: true,
+    hop: d.hop,
+    max: d.max,
+    childContext: { hop: d.hop + 1, max: d.max },
+  };
+}
+
+/**
+ * Single source of truth for the routing decision: should THIS call_agent
+ * invocation go through the delegation lane instead of the interactive pool?
+ *
+ * The gate (`enabled`, from CLAUTH_AGENT_DELEGATION) is checked FIRST: when the
+ * gate is OFF this returns false unconditionally and `agent_context.delegation`
+ * is never even read — that is the byte-stable-off guarantee. A call is delegated
+ * iff the gate is on AND the caller supplied an `agent_context.delegation`
+ * marker (only a sub-agent invocation carries one).
+ *
+ * @param {object} args
+ * @param {boolean} args.enabled        CLAUTH_AGENT_DELEGATION on?
+ * @param {object}  [args.agent_context]
+ * @returns {boolean}
+ */
+export function shouldDelegate({ enabled, agent_context } = {}) {
+  if (!enabled) return false; // gate OFF → delegation marker ignored entirely
+  if (!agent_context || typeof agent_context !== "object") return false;
+  return agent_context.delegation != null;
+}
+
+/**
+ * DelegationLane — a cold-spawn-only, concurrency-capped execution lane for
+ * DELEGATED (sub-agent) call_agent invocations. Structurally distinct from
+ * AgentPool: it holds no warm workers, owns no AgentPool reference, and shares
+ * no queue/slot with the interactive pool. That separation is the anti-deadlock
+ * guarantee — sub-agent spawns can never starve or hold an interactive slot.
+ *
+ * It deliberately reuses the SAME lean one-shot recipe (model, skip-permissions,
+ * user-only settings, strict-mcp, lean system prompt) as AgentPool's cold path,
+ * but on an independent set of processes with its own `maxInFlight` cap.
+ */
+export class DelegationLane {
+  /**
+   * @param {object} opts
+   * @param {string} [opts.binary]        resolved claude binary
+   * @param {number} [opts.maxInFlight]   daemon-wide concurrent sub-agent cap
+   * @param {number} [opts.maxHops]       hop-budget default ceiling
+   * @param {string} [opts.defaultModel]
+   * @param {number} [opts.defaultTimeoutMs]
+   * @param {string} [opts.neutralCwd]
+   * @param {string} [opts.appendSystemPrompt]
+   * @param {function} [opts.spawnImpl]   inject a fake spawn for tests
+   */
+  constructor(opts = {}) {
+    this.binary = opts.binary || resolveClaudeBinary(opts.binary);
+    this.maxInFlight = Number.isFinite(opts.maxInFlight)
+      ? Math.max(1, opts.maxInFlight)
+      : DEFAULT_DELEGATION_MAX_INFLIGHT;
+    this.maxHops = Number.isFinite(opts.maxHops) ? Math.max(0, opts.maxHops) : DEFAULT_DELEGATION_MAX_HOPS;
+    this.defaultModel = opts.defaultModel || DEFAULT_MODEL;
+    this.defaultTimeoutMs = Number.isFinite(opts.defaultTimeoutMs) ? opts.defaultTimeoutMs : DEFAULT_TIMEOUT_MS;
+    this.neutralCwd = opts.neutralCwd || path.join(os.tmpdir(), "clauth-call-agent-delegation");
+    this.appendSystemPrompt = opts.appendSystemPrompt || LEAN_SYSTEM_PROMPT;
+    this._spawn = opts.spawnImpl || spawn;
+
+    this.active = 0; // live sub-agent cold spawns
+    this.totalSpawned = 0;
+    this.totalRejected = 0;
+    this._shuttingDown = false;
+    this._procs = new Set(); // live child procs for shutdown teardown
+
+    try {
+      fs.mkdirSync(this.neutralCwd, { recursive: true });
+    } catch {
+      /* best effort */
+    }
+  }
+
+  available() {
+    return this.binary != null;
+  }
+
+  buildArgs({ prompt, model, appendSystemPrompt }) {
+    return [
+      "-p",
+      prompt,
+      "--model",
+      model || this.defaultModel,
+      "--dangerously-skip-permissions",
+      "--setting-sources",
+      "user",
+      "--strict-mcp-config",
+      "--append-system-prompt",
+      appendSystemPrompt || this.appendSystemPrompt,
+    ];
+  }
+
+  /**
+   * Dispatch a DELEGATED sub-agent call. Enforces the hop budget and the
+   * in-flight cap, then cold-spawns a one-shot. Returns a job record shaped like
+   * AgentPool.dispatch's so callers are uniform. On rejection it resolves with
+   * `{ ok:false, error, delegated:true, ... }` and NEVER spawns.
+   *
+   * @param {object} args { prompt, model, timeout_ms, delegation:{hop,max}, jobId, appendSystemPrompt }
+   * @returns {Promise<object>}
+   */
+  dispatch(args = {}) {
+    return new Promise((resolve) => {
+      if (!this.available()) {
+        return resolve({ ok: false, error: "binary_not_found", delegated: true });
+      }
+      const prompt = args.prompt;
+      if (!prompt || typeof prompt !== "string") {
+        return resolve({ ok: false, error: "prompt_required", delegated: true });
+      }
+      // Hop budget — the single most important guard. A call past budget is
+      // refused cleanly with NO spawn, so a delegation loop cannot run away.
+      const budget = evaluateHopBudget(args.delegation, { defaultMax: this.maxHops });
+      if (!budget.allow) {
+        this.totalRejected++;
+        return resolve({
+          ok: false,
+          error: budget.error,
+          delegated: true,
+          hop: budget.hop,
+          max: budget.max,
+          message: `delegation hop ${budget.hop} exceeds max ${budget.max}`,
+        });
+      }
+      // Daemon-wide in-flight cap — a wide fan-out cannot exhaust resources.
+      if (this.active >= this.maxInFlight) {
+        this.totalRejected++;
+        return resolve({
+          ok: false,
+          error: "delegation_capacity",
+          delegated: true,
+          active: this.active,
+          maxInFlight: this.maxInFlight,
+          message: `delegation lane full (${this.active}/${this.maxInFlight})`,
+        });
+      }
+      if (this._shuttingDown) {
+        return resolve({ ok: false, error: "delegation_shutting_down", delegated: true });
+      }
+
+      const model = args.model || this.defaultModel;
+      const timeout = Number.isFinite(args.timeout_ms) ? args.timeout_ms : this.defaultTimeoutMs;
+      const jobId = args.jobId || makeJobId("deleg");
+      const cwd = args.cwd || this.neutralCwd;
+      const startedAt = Date.now();
+
+      const cmdArgs = this.buildArgs({ prompt, model, appendSystemPrompt: args.appendSystemPrompt });
+
+      let proc;
+      try {
+        const isCmdShim = process.platform === "win32" && /\.cmd$/i.test(this.binary);
+        const command = isCmdShim ? "cmd" : this.binary;
+        const spawnArgs = isCmdShim ? ["/d", "/s", "/c", `"${this.binary}"`, ...cmdArgs] : cmdArgs;
+        proc = this._spawn(command, spawnArgs, {
+          cwd,
+          env: process.env, // inherits CLI login; never sets ANTHROPIC_API_KEY
+          stdio: ["ignore", "pipe", "pipe"],
+          shell: false,
+          windowsHide: true,
+        });
+      } catch (e) {
+        return resolve({
+          ok: false,
+          error: `spawn_failed: ${e.message}`,
+          delegated: true,
+          jobId,
+        });
+      }
+
+      this.active++;
+      this.totalSpawned++;
+      this._procs.add(proc);
+
+      let stdout = "";
+      let stderr = "";
+      let settled = false;
+      // NOTE: intentionally NOT unref'd. While a sub-agent is in flight this
+      // timer is the only thing that can settle a hung one-shot, so it must keep
+      // the loop alive until it fires (it clears on normal close). The daemon
+      // owns process lifecycle; shutdown() kills all live procs explicitly.
+      const timer = setTimeout(() => {
+        if (settled) return;
+        try {
+          proc.kill("SIGKILL");
+        } catch {
+          /* already gone */
+        }
+        finish({ status: "timeout", error: "timeout" });
+      }, timeout);
+
+      const finish = (extra) => {
+        if (settled) return;
+        settled = true;
+        clearTimeout(timer);
+        this.active = Math.max(0, this.active - 1);
+        this._procs.delete(proc);
+        const ms = Date.now() - startedAt;
+        const status = extra.status;
+        resolve({
+          ok: status === "completed",
+          jobId,
+          status,
+          model,
+          ms,
+          exit_code: extra.exit_code ?? null,
+          package: status === "completed" ? stdout.trim() : null,
+          error: extra.error || null,
+          delegated: true,
+          hop: budget.childContext.hop, // depth of this spawned sub-agent
+          max: budget.max,
+          ...(stderr && status !== "completed" ? { stderr: stderr.slice(-2000) } : {}),
+        });
+      };
+
+      proc.stdout?.on("data", (d) => {
+        stdout = `${stdout}${d.toString()}`.slice(-200000);
+      });
+      proc.stderr?.on("data", (d) => {
+        stderr = `${stderr}${d.toString()}`.slice(-20000);
+      });
+      proc.on("error", (e) => finish({ status: "failed", error: `proc_error: ${e.message}` }));
+      proc.on("close", (code) => {
+        if (code === 0) finish({ status: "completed", exit_code: 0 });
+        else finish({ status: "failed", exit_code: code, error: `exit_${code}` });
+      });
+    });
+  }
+
+  stats() {
+    return {
+      active: this.active,
+      maxInFlight: this.maxInFlight,
+      maxHops: this.maxHops,
+      totalSpawned: this.totalSpawned,
+      totalRejected: this.totalRejected,
+    };
+  }
+
+  shutdown() {
+    this._shuttingDown = true;
+    for (const proc of this._procs) {
+      try {
+        proc.kill("SIGKILL");
+      } catch {
+        /* already gone */
+      }
+    }
+    this._procs.clear();
+  }
+}
+
+export {
+  LEAN_SYSTEM_PROMPT,
+  IDENTITY_SYSTEM_PROMPT,
+  resolveClaudeBinary,
+  buildWarmArgs,
+  buildFullSystemPrompt,
+  buildStudioEditorSystemPrompt,
+  loadEditorSettings,
+  makeJobId,
+};
+
+// composeCorpus / buildCorpusSection / loadBootstrapDoc are exported inline
+// above (function declarations with `export`).