bloby-bot 0.70.4 → 0.70.6

package/supervisor/harnesses/codex.ts

@@ -8,7 +8,8 @@
  *
  * Lifecycle per live conversation:
  *   spawn → initialize → initialized → thread/start → turn/start (per
- *   user message; turn/steer to inject mid-turn) → turn/completed → idle
+ *   user message; messages arriving mid-turn are QUEUED and each gets its
+ *   own turn — see queueOrStart) → turn/completed → idle
  *   → endConversation → turn/interrupt (if needed) → kill subprocess
  *
  * Lifecycle per one-shot query: same as above, but the subprocess is killed
@@ -19,15 +20,22 @@
  * passed as `effort` on `turn/start`.
  *
  * Notes on parity with Claude harness:
- *   - System prompt → `baseInstructions` on `thread/start`
- *   - Sub-agents → not implemented (Codex has Skills, different model)
- *   - MCP servers → not wired yet (Codex has its own MCP layer)
- *   - Mid-turn input uses `turn/steer` (better than Claude's queue)
+ *   - System prompt → `developerInstructions` on `thread/start` (ADDITIVE —
+ *     `baseInstructions` would override codex's native apply_patch/shell spec)
+ *   - Mid-turn input is queued, one turn per message — same one-bot:response-
+ *     per-push contract the channel routing FIFO depends on. (`turn/steer`
+ *     merges inputs into the in-flight turn and breaks that contract.)
+ *   - Sub-agents → not implemented (codex collab tools are still experimental;
+ *     the collabAgentToolCall handlers below light up if that ever ships)
+ *   - MCP servers → MCP.json translated to `-c mcp_servers.*` spawn flags
+ *   - Pre-warm → spawn+initialize cached ahead of time (thread/start is issued
+ *     at claim time, so the warm process is reusable for any model/instructions)
  */
 
 import { spawn, type ChildProcessWithoutNullStreams } from 'child_process';
 import { createRequire } from 'module';
 import readline from 'readline';
+import crypto from 'crypto';
 import fs from 'fs';
 import path from 'path';
 import { log } from '../../shared/logger.js';
@@ -54,6 +62,21 @@ const VALID_EFFORTS = new Set(['none', 'minimal', 'low', 'medium', 'high', 'xhig
  * keeps emitting events and is never killed; only true silence trips recovery.
  */
 const TURN_WATCHDOG_MS = 5 * 60_000;
+/**
+ * Hard WALL-CLOCK cap for one-shot turns (pulse/cron, customer WhatsApp) —
+ * codex has no maxTurns equivalent, and the idle watchdog above never trips on
+ * an actively-emitting runaway turn (tool ping-pong). Claude one-shots abort at
+ * 5 minutes; mirror that so a runaway can't pin a customer slot forever.
+ */
+const ONE_SHOT_MAX_TURN_MS = 5 * 60_000;
+/**
+ * Micro-batch window for bot:token. Codex streams true per-word deltas — an
+ * order of magnitude more WS frames per reply than claude's per-block stream.
+ * Coalescing at ~60ms keeps the stream visually smooth while cutting frame
+ * volume. Every non-token event flushes first, so ordering (and the
+ * bot:response === streamed-text invariant) is preserved.
+ */
+const TOKEN_FLUSH_MS = 60;
 
 /**
  * Resolve the `codex` binary. We don't trust $PATH because Bloby may be
@@ -160,6 +183,19 @@ function parseModelString(model: string): { id: string; effort?: string } {
 type RpcResult<T = any> = { id: number; result?: T; error?: { code?: number; message: string } };
 type RpcNotification = { method: string; params?: any };
 
+/** Distinguishes a JSON-RPC error RESPONSE (server alive, request rejected)
+ *  from transport failures (timeout / closed / process exit) — turn/start
+ *  error containment keeps the conversation alive only for the former. */
+type RpcErrorKind = 'rpc-error' | 'timeout' | 'closed';
+function rpcError(message: string, kind: RpcErrorKind): Error {
+  const err = new Error(message);
+  (err as any).rpcKind = kind;
+  return err;
+}
+function isServerRejection(err: any): boolean {
+  return err?.rpcKind === 'rpc-error';
+}
+
 interface PendingRequest {
   resolve: (value: any) => void;
   reject: (err: Error) => void;
@@ -171,7 +207,7 @@ class CodexRpc {
   private pending = new Map<number, PendingRequest>();
   private nextId = 1;
   private notificationHandler: (n: RpcNotification) => void = () => {};
-  private closeHandler: (code: number | null) => void = () => {};
+  private closeHandler: (code: number | null, errMsg?: string) => void = () => {};
   private closed = false;
   private stderrBuf = '';
 
@@ -189,13 +225,18 @@ class CodexRpc {
     this.proc.on('exit', (code) => {
       if (this.closed) return;
       this.closed = true;
-      const err = new Error(`codex app-server exited (code=${code}). Stderr tail:\n${this.stderrBuf.trim().slice(-1000)}`);
+      const errMsg = `codex app-server exited (code=${code}). Stderr tail:\n${this.stderrBuf.trim().slice(-1000)}`;
+      const err = rpcError(errMsg, 'closed');
       for (const p of this.pending.values()) {
         clearTimeout(p.timer);
         p.reject(err);
       }
       this.pending.clear();
-      this.closeHandler(code);
+      // Pass the composed exit error along — mid-turn there are usually no
+      // pending requests, so this is the ONLY way crash detail reaches the
+      // conversation (the watchdog would otherwise ghost-fire 5 min later
+      // with a misleading "timed out" message).
+      this.closeHandler(code, errMsg);
     });
 
     this.proc.on('error', (err) => {
@@ -204,17 +245,22 @@ class CodexRpc {
       log.warn(`[codex-rpc] spawn error: ${err.message}`);
       for (const p of this.pending.values()) {
         clearTimeout(p.timer);
-        p.reject(err);
+        p.reject(rpcError(err.message, 'closed'));
       }
       this.pending.clear();
-      this.closeHandler(null);
+      this.closeHandler(null, `codex app-server failed to spawn: ${err.message}`);
     });
   }
 
   onNotification(handler: (n: RpcNotification) => void): void { this.notificationHandler = handler; }
-  onClose(handler: (code: number | null) => void): void { this.closeHandler = handler; }
+  onClose(handler: (code: number | null, errMsg?: string) => void): void { this.closeHandler = handler; }
 
   private onLine(line: string): void {
+    // close() already rejected everything pending; late stdout (buffered deltas,
+    // a turn/completed that landed the same instant as teardown) must not
+    // re-enter the event pipeline — a post-teardown turn/completed could drain
+    // pendingInputs into a dead rpc and wedge agentQueryActive forever.
+    if (this.closed) return;
     if (!line.trim()) return;
     let msg: any;
     try { msg = JSON.parse(line); } catch {
@@ -237,7 +283,7 @@ class CodexRpc {
       if (!pending) return;
       this.pending.delete(msg.id);
       clearTimeout(pending.timer);
-      if (msg.error) pending.reject(new Error(msg.error.message || 'RPC error'));
+      if (msg.error) pending.reject(rpcError(msg.error.message || 'RPC error', 'rpc-error'));
       else pending.resolve(msg.result);
       return;
     }
@@ -288,12 +334,12 @@ class CodexRpc {
   }
 
   request<T = any>(method: string, params?: any, timeoutMs = REQUEST_TIMEOUT_MS): Promise<T> {
-    if (this.closed || !this.proc) return Promise.reject(new Error('RPC connection closed'));
+    if (this.closed || !this.proc) return Promise.reject(rpcError('RPC connection closed', 'closed'));
     const id = this.nextId++;
     return new Promise<T>((resolve, reject) => {
       const timer = setTimeout(() => {
         this.pending.delete(id);
-        reject(new Error(`codex app-server: ${method} timed out after ${timeoutMs}ms`));
+        reject(rpcError(`codex app-server: ${method} timed out after ${timeoutMs}ms`, 'timeout'));
       }, timeoutMs);
       this.pending.set(id, { resolve, reject, timer });
       try {
@@ -301,7 +347,7 @@ class CodexRpc {
       } catch (err: any) {
         this.pending.delete(id);
         clearTimeout(timer);
-        reject(err);
+        reject(rpcError(err.message, 'closed'));
       }
     });
   }
@@ -320,7 +366,7 @@ class CodexRpc {
     this.closed = true;
     for (const p of this.pending.values()) {
       clearTimeout(p.timer);
-      p.reject(new Error('RPC connection closed'));
+      p.reject(rpcError('RPC connection closed', 'closed'));
     }
     this.pending.clear();
     const proc = this.proc;
@@ -339,11 +385,20 @@ class CodexRpc {
 
 /* ── Per-conversation state ────────────────────────────────────────────── */
 
+interface QueuedInput {
+  content: string;
+  savedFiles?: SavedFile[];
+  attachments?: AgentAttachment[];
+}
+
 interface CodexConversation {
   id: string;
   rpc: CodexRpc;
   threadId: string;
   effort?: string;
+  /** Original model string (with effort suffix) + names — kept for the post-teardown re-warm. */
+  model: string;
+  names?: { botName: string; humanName: string };
   onMessage: OnAgentMessage;
   /** Currently in-flight turn id (set on `turn/started`, cleared on `turn/completed`). */
   currentTurnId: string | null;
@@ -352,18 +407,32 @@ interface CodexConversation {
   currentMsgItemId: string | null;
   /** Streaming text accumulator for the current turn's agentMessage items. */
   fullText: string;
+  /** Per-itemId delta accumulation for the current turn. item/completed is the
+   *  authoritative final text per the docs — delta concatenation is not guaranteed
+   *  to match it, and a later agentMessage item may complete with NO deltas at all.
+   *  This map lets item/completed emit exactly the missing remainder per item. */
+  itemTexts: Map<string, string>;
   /** Tools/items used during the current turn, for the bot:turn-complete payload. */
   usedFileTools: boolean;
-  /**
-   * Queue of messages submitted via `pushMessage` that arrived while no turn
-   * was active *and* we hadn't yet returned from the previous turn — almost
-   * always empty, but covers a tight push-during-completed race.
-   */
-  pendingInputs: string[];
-  /** True once the harness has emitted the per-turn completion event. */
+  /** Messages pushed while a turn is active. Each is drained into its OWN turn on
+   *  turn/completed — one bot:response per push, mirroring claude's input queue.
+   *  (turn/steer would merge them into the in-flight turn and desync the channel
+   *  routing FIFO, which enqueues one target per push.) */
+  pendingInputs: QueuedInput[];
+  /** True while a turn is being processed. */
   busy: boolean;
   /** True for one-shot queries — the conversation ends after the first turn completes. */
   oneShot: boolean;
+  /** True once bot:error fired for the current turn — dedups the double-emit the
+   *  app-server produces on hard failures (a non-retry `error` notification followed
+   *  by `turn/completed {status:'failed'}` with the same message; live-verified). */
+  errorEmitted: boolean;
+  /** True once bot:done fired for this one-shot — teardown uses it to guarantee
+   *  exactly one bot:done on EVERY terminal path (slot-freeing consumers rely on it). */
+  doneEmitted: boolean;
+  /** Non-retry `error` notification stashed while a turn is active; surfaced by
+   *  turn/completed {failed} (its TurnError is authoritative, the stash is fallback). */
+  stashedError: { message: string; info: any } | null;
   /**
    * Latest context occupancy from `thread/tokenUsage/updated` (codex does NOT
    * report usage on `turn/completed` — Turn has no usage field). Emitted on
@@ -373,13 +442,127 @@ interface CodexConversation {
   lastContextWindow: number;
   /** Active per-turn watchdog timer (see TURN_WATCHDOG_MS). */
   turnWatchdog: NodeJS.Timeout | null;
+  /** Hard wall-clock cap for one-shot turns (see ONE_SHOT_MAX_TURN_MS). Non-resetting. */
+  hardTurnTimer: NodeJS.Timeout | null;
+  /** bot:token micro-batch buffer (see TOKEN_FLUSH_MS). */
+  tokenBuf: string;
+  tokenFlushTimer: NodeJS.Timeout | null;
 }
 
 const conversations = new Map<string, CodexConversation>();
 
-/* ── Helpers ───────────────────────────────────────────────────────────── */
+/** One-shot queries in their init window (token refresh + spawn + initialize +
+ *  thread/start) — not yet in `conversations`, so anyConversationBusy() can't see
+ *  them. Without this, a queued self-update / backend restart can fire mid-launch
+ *  and kill the just-spawning query. Mirrors claude's activeQueries registration. */
+const inFlightOneShots = new Set<string>();
+
+/* ── Event emission helpers ────────────────────────────────────────────── */
+
+function flushTokens(conv: CodexConversation): void {
+  if (conv.tokenFlushTimer) {
+    clearTimeout(conv.tokenFlushTimer);
+    conv.tokenFlushTimer = null;
+  }
+  if (conv.tokenBuf) {
+    const token = conv.tokenBuf;
+    conv.tokenBuf = '';
+    conv.onMessage('bot:token', { conversationId: conv.id, token });
+  }
+}
+
+/** Drop buffered tokens WITHOUT emitting. Teardown must discard, not flush:
+ *  on user-stop the frontend has already cleared its stream state, and a stray
+ *  trailing bot:token would re-open streaming with no bot:idle ever coming —
+ *  then corrupt the next turn's committedTextLength accounting. */
+function discardTokens(conv: CodexConversation): void {
+  if (conv.tokenFlushTimer) {
+    clearTimeout(conv.tokenFlushTimer);
+    conv.tokenFlushTimer = null;
+  }
+  conv.tokenBuf = '';
+}
+
+function emitToken(conv: CodexConversation, token: string): void {
+  conv.tokenBuf += token;
+  if (!conv.tokenFlushTimer) {
+    conv.tokenFlushTimer = setTimeout(() => {
+      conv.tokenFlushTimer = null;
+      flushTokens(conv);
+    }, TOKEN_FLUSH_MS);
+  }
+}
+
+/** Emit any non-token event, flushing buffered tokens first so ordering is
+ *  preserved (bot:tool commits the stream bubble; bot:response must equal the
+ *  streamed text the frontend already saw). */
+function emitEvent(conv: CodexConversation, type: string, data: any): void {
+  flushTokens(conv);
+  conv.onMessage(type, data);
+}
+
+/** Kinds we can branch on from CodexErrorInfo (string variant or `{ kind: {...} }`). */
+function errorInfoKind(info: any): string | undefined {
+  if (typeof info === 'string') return info;
+  if (info && typeof info === 'object') return Object.keys(info)[0];
+  return undefined;
+}
+
+/**
+ * Map codex's structured error onto a message that gives BLOBY's remedy.
+ * The raw upstream text actively misleads ("Please log out and sign in again"
+ * points at the codex CLI, not the dashboard re-auth flow).
+ */
+function humanizeCodexError(message?: string, info?: any, additionalDetails?: string | null): string {
+  const raw = message || 'Codex error';
+  switch (errorInfoKind(info)) {
+    case 'unauthorized':
+      return 'Codex session expired or unauthorized. Re-authenticate from the dashboard.';
+    case 'usageLimitExceeded':
+      return `Codex usage limit reached — ${raw}`;
+    case 'contextWindowExceeded':
+      return 'The conversation outgrew the model context window. Send your message again to continue in a fresh session.';
+    case 'serverOverloaded':
+      return 'OpenAI servers are overloaded right now — try again in a moment.';
+    case 'httpConnectionFailed':
+    case 'responseStreamConnectionFailed':
+    case 'responseStreamDisconnected': {
+      const status = (Object.values(info || {})[0] as any)?.httpStatusCode;
+      return `Connection to OpenAI failed${status ? ` (HTTP ${status})` : ''} — try again in a moment.`;
+    }
+    default:
+      return additionalDetails ? `${raw}\n${additionalDetails}` : raw;
+  }
+}
+
+/** Emit bot:error exactly once per turn (see errorEmitted). */
+function emitError(conv: CodexConversation, message?: string, info?: any, additionalDetails?: string | null): void {
+  const friendly = humanizeCodexError(message, info, additionalDetails);
+  if (conv.errorEmitted) {
+    log.info(`[codex] suppressing duplicate bot:error for conv=${conv.id}: ${friendly.slice(0, 120)}`);
+    return;
+  }
+  conv.errorEmitted = true;
+  emitEvent(conv, 'bot:error', { conversationId: conv.id, error: friendly });
+}
+
+/** Emit bot:done exactly once per one-shot (see doneEmitted). */
+function emitDone(conv: CodexConversation): void {
+  if (conv.doneEmitted) return;
+  conv.doneEmitted = true;
+  emitEvent(conv, 'bot:done', { conversationId: conv.id, usedFileTools: conv.usedFileTools });
+}
+
+/* ── Input building ────────────────────────────────────────────────────── */
+
+/** mediaTypes whose content we inline into the prompt — codex has no document
+ *  input type (verified against 0.138 UserInput), so this is the closest we get
+ *  to claude's native base64 document ingestion for text-like files. */
+const INLINE_TEXT_MEDIA = /^(text\/|application\/(json|xml|yaml|x-yaml|toml|csv|javascript|typescript))/;
+const INLINE_TEXT_MAX_BYTES = 48_000;
+const INLINE_TEXT_TOTAL_BUDGET = 96_000;
 
-function buildUserInput(text: string, savedFiles?: SavedFile[]): Array<Record<string, any>> {
+function buildUserInput(text: string, savedFiles?: SavedFile[], attachments?: AgentAttachment[]): Array<Record<string, any>> {
   const input: Array<Record<string, any>> = [];
 
   let promptText = text || '(attached files)';
@@ -387,19 +570,59 @@ function buildUserInput(text: string, savedFiles?: SavedFile[]): Array<Record<st
     const lines = savedFiles.map((f) => `- ${f.name} -> ${f.relPath}`);
     promptText += `\n\n[Attached files saved to disk]\n${lines.join('\n')}\nYou can read or reference these files using the paths above (relative to your cwd).`;
   }
+
+  // Inline text-like attachments (size-capped) so the model can answer about
+  // their contents immediately instead of shelling out to read them from disk.
+  if (attachments?.length) {
+    let budget = INLINE_TEXT_TOTAL_BUDGET;
+    for (const att of attachments) {
+      if (att.type !== 'file' || !INLINE_TEXT_MEDIA.test(att.mediaType || '')) continue;
+      const approxBytes = Math.floor((att.data?.length || 0) * 0.75);
+      if (approxBytes === 0 || approxBytes > INLINE_TEXT_MAX_BYTES || approxBytes > budget) continue;
+      try {
+        const decoded = Buffer.from(att.data, 'base64').toString('utf-8');
+        budget -= approxBytes;
+        promptText += `\n\n[Attached file content: ${att.name}]\n\`\`\`\n${decoded}\n\`\`\``;
+      } catch {}
+    }
+  }
+
   input.push({ type: 'text', text: promptText });
 
   // Codex understands `localImage` (path on disk) — Bloby's file-saver already
   // wrote attachments to disk, so we just point at the absolute path.
+  const savedImageCounts = new Map<string, number>();
   if (savedFiles?.length) {
     for (const f of savedFiles) {
-      if (f.type === 'image') input.push({ type: 'localImage', path: f.absPath });
+      if (f.type === 'image') {
+        input.push({ type: 'localImage', path: f.absPath });
+        savedImageCounts.set(f.name, (savedImageCounts.get(f.name) || 0) + 1);
+      }
+    }
+  }
+
+  // Fallback for image attachments that never made it to disk (file-saver
+  // failure): inline as a data URL so the agent still SEES the image — claude
+  // always gets the base64 inline, codex shouldn't silently go blind.
+  // Match by COUNT, not name presence: WhatsApp multi-image pushes share one
+  // attachment name, and each saved file covers exactly one of them.
+  if (attachments?.length) {
+    for (const att of attachments) {
+      if (att.type !== 'image' || !att.data) continue;
+      const remaining = savedImageCounts.get(att.name) || 0;
+      if (remaining > 0) {
+        savedImageCounts.set(att.name, remaining - 1);
+        continue;
+      }
+      input.push({ type: 'image', url: `data:${att.mediaType};base64,${att.data}` });
     }
   }
 
   return input;
 }
 
+/* ── Turn lifecycle ────────────────────────────────────────────────────── */
+
 function clearTurnWatchdog(conv: CodexConversation): void {
   if (conv.turnWatchdog) {
     clearTimeout(conv.turnWatchdog);
@@ -407,6 +630,13 @@ function clearTurnWatchdog(conv: CodexConversation): void {
   }
 }
 
+function clearHardTurnTimer(conv: CodexConversation): void {
+  if (conv.hardTurnTimer) {
+    clearTimeout(conv.hardTurnTimer);
+    conv.hardTurnTimer = null;
+  }
+}
+
 /**
  * Arm the per-turn watchdog. On fire, unstick the conversation the same way a
  * real `turn/completed` would (so the dashboard, `anyConversationBusy`, and the
@@ -420,11 +650,15 @@ function armTurnWatchdog(conv: CodexConversation): void {
     log.warn(`[codex] turn watchdog fired (${TURN_WATCHDOG_MS}ms) — conv=${conv.id}; unsticking + tearing down`);
     conv.busy = false;
     conv.currentTurnId = null;
-    conv.onMessage('bot:error', { conversationId: conv.id, error: 'Codex turn timed out — no response from app-server.' });
+    // Prefer a stashed non-retry error: if the server wedged right after
+    // emitting it (the turn/completed that normally surfaces it never came),
+    // the stash carries the real cause + the M4 remedy mapping.
+    const stash = conv.stashedError;
+    emitError(conv, stash?.message || 'Codex turn timed out — no response from app-server.', stash?.info);
     if (conv.oneShot) {
-      conv.onMessage('bot:done', { conversationId: conv.id, usedFileTools: conv.usedFileTools });
+      emitDone(conv);
     } else {
-      conv.onMessage('bot:turn-complete', {
+      emitEvent(conv, 'bot:turn-complete', {
         conversationId: conv.id,
         usedFileTools: conv.usedFileTools,
         contextTokens: conv.lastContextTokens || 0,
@@ -436,60 +670,98 @@ function armTurnWatchdog(conv: CodexConversation): void {
   }, TURN_WATCHDOG_MS);
 }
 
-async function startTurn(conv: CodexConversation, content: string, savedFiles?: SavedFile[]): Promise<void> {
-  const input = buildUserInput(content, savedFiles);
-  conv.busy = true;
+/** Hard wall-clock cap for one-shot turns — see ONE_SHOT_MAX_TURN_MS. */
+function armHardTurnTimer(conv: CodexConversation): void {
+  if (!conv.oneShot || conv.hardTurnTimer) return;
+  conv.hardTurnTimer = setTimeout(() => {
+    conv.hardTurnTimer = null;
+    log.warn(`[codex] one-shot hard cap fired (${ONE_SHOT_MAX_TURN_MS}ms) — conv=${conv.id}; interrupting`);
+    if (conv.currentTurnId) {
+      void conv.rpc.request('turn/interrupt', { threadId: conv.threadId, turnId: conv.currentTurnId }).catch(() => {});
+    }
+    conv.busy = false;
+    conv.currentTurnId = null;
+    emitError(conv, `Codex query hit the ${Math.round(ONE_SHOT_MAX_TURN_MS / 60_000)}-minute limit and was stopped.`);
+    teardownConversation(conv.id); // emits the guaranteed bot:done
+  }, ONE_SHOT_MAX_TURN_MS);
+}
+
+function resetTurnState(conv: CodexConversation): void {
   conv.fullText = '';
+  conv.currentMsgItemId = null;
+  conv.itemTexts = new Map();
   conv.usedFileTools = false;
-  conv.onMessage('bot:typing', { conversationId: conv.id });
+  conv.errorEmitted = false;
+  conv.stashedError = null;
+}
+
+async function startTurn(conv: CodexConversation, content: string, savedFiles?: SavedFile[], attachments?: AgentAttachment[]): Promise<void> {
+  const input = buildUserInput(content, savedFiles, attachments);
+  conv.busy = true;
+  resetTurnState(conv);
+  emitEvent(conv, 'bot:typing', { conversationId: conv.id });
   armTurnWatchdog(conv);
+  armHardTurnTimer(conv);
   try {
     const params: Record<string, any> = { threadId: conv.threadId, input };
     if (conv.effort) params.effort = conv.effort;
-    // turn/start resolves immediately with { turn }; seize the id now so a
-    // pushMessage arriving before the turn/started notification can steer
-    // instead of starting a second turn.
+    // turn/start resolves immediately with { turn }; seize the id now so an
+    // interrupt arriving before the turn/started notification still has a target.
     const res = await conv.rpc.request<{ turn?: { id?: string } }>('turn/start', params);
     if (res?.turn?.id) conv.currentTurnId = res.turn.id;
   } catch (err: any) {
     clearTurnWatchdog(conv);
     conv.busy = false;
     conv.currentTurnId = null;
-    conv.onMessage('bot:error', { conversationId: conv.id, error: `turn/start failed: ${err.message}` });
-    // turn/start produced no turn, so no turn/completed will arrive to clear the supervisor's
-    // agentQueryActive (set on bot:typing above). Left as-is, that wedges true forever:
-    // backend auto-heal is deferred indefinitely and chat is stuck showing "typing". Tear the
-    // conversation down so bot:conversation-ended fires (which, unlike bot:turn-complete, does
-    // NOT trigger a backend restart) — the next user message cold-starts a fresh thread.
-    teardownConversation(conv.id);
+    emitError(conv, `turn/start failed: ${err.message}`);
+    // turn/start produced no turn, so no turn/completed will arrive to clear the
+    // supervisor's agentQueryActive (set on bot:typing above).
+    //
+    // - One-shots and transport failures (timeout / process exit): tear down. A
+    //   60s timeout most plausibly means a hung app-server — keeping the conv
+    //   would just hang the next message too. teardown emits bot:conversation-
+    //   ended (clears agentQueryActive) and, for one-shots, the guaranteed bot:done.
+    // - A fast JSON-RPC REJECTION on a live conv means the server is alive (bad
+    //   effort value, transient thread error): keep the thread — its server-side
+    //   context (files read, compacted history) survives, matching claude's
+    //   per-turn error containment. bot:turn-complete unsticks the supervisor.
+    if (conv.oneShot || !isServerRejection(err)) {
+      teardownConversation(conv.id);
+    } else {
+      emitEvent(conv, 'bot:turn-complete', {
+        conversationId: conv.id,
+        usedFileTools: false,
+        contextTokens: conv.lastContextTokens || 0,
+        contextWindow: conv.lastContextWindow || 0,
+        idle: conv.pendingInputs.length === 0,
+      });
+      // Keep draining queued messages — each failure surfaces its own error,
+      // and the finite queue guarantees termination.
+      const next = conv.pendingInputs.shift();
+      if (next !== undefined) void startTurn(conv, next.content, next.savedFiles, next.attachments);
+    }
   }
 }
 
-async function steerOrQueue(conv: CodexConversation, content: string, savedFiles?: SavedFile[]): Promise<void> {
-  if (!conv.currentTurnId) {
-    // No active turn — start a fresh one.
-    await startTurn(conv, content, savedFiles);
+/**
+ * Queue-or-start: the codex side of claude's async input queue. While a turn
+ * is active, pushes are queued and drained ONE TURN PER MESSAGE on
+ * turn/completed — preserving the one-bot:response-per-push contract that
+ * channels/manager.ts's routing FIFO depends on (it enqueues exactly one
+ * routing target per push and consumes one per bot:response).
+ */
+function queueOrStart(conv: CodexConversation, content: string, savedFiles?: SavedFile[], attachments?: AgentAttachment[]): void {
+  if (conv.busy || conv.currentTurnId) {
+    conv.pendingInputs.push({ content, savedFiles, attachments });
+    // Mirror claude's pushMessage: every accepted push shows typing immediately.
+    emitEvent(conv, 'bot:typing', { conversationId: conv.id });
     return;
   }
-  // Active turn — inject mid-flight.
-  const input = buildUserInput(content, savedFiles);
-  try {
-    const res = await conv.rpc.request<{ turnId?: string }>('turn/steer', {
-      threadId: conv.threadId,
-      expectedTurnId: conv.currentTurnId,
-      input,
-    });
-    if (res?.turnId) conv.currentTurnId = res.turnId;
-    conv.onMessage('bot:typing', { conversationId: conv.id });
-  } catch (err: any) {
-    // expectedTurnId mismatch most likely means the turn just finished —
-    // retry as a fresh turn.
-    log.warn(`[codex] turn/steer failed (${err.message}); falling back to turn/start`);
-    if (!conv.currentTurnId) await startTurn(conv, content, savedFiles);
-    else conv.pendingInputs.push(content);
-  }
+  void startTurn(conv, content, savedFiles, attachments);
 }
 
+/* ── Notification handling ─────────────────────────────────────────────── */
+
 function handleNotification(conv: CodexConversation, n: { method: string; params?: any }): void {
   const p = n.params || {};
   // Any notification for this conv proves the app-server is alive and working —
@@ -498,9 +770,7 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
   switch (n.method) {
     case 'turn/started': {
       conv.currentTurnId = p.turn?.id || null;
-      conv.fullText = '';
-      conv.currentMsgItemId = null;
-      conv.usedFileTools = false;
+      resetTurnState(conv);
       break;
     }
 
@@ -511,52 +781,72 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
       // On a new itemId, insert a paragraph break so they don't run together (mirrors claude.ts).
       if (p.itemId && conv.currentMsgItemId && p.itemId !== conv.currentMsgItemId && conv.fullText && !conv.fullText.endsWith('\n')) {
         conv.fullText += '\n\n';
-        conv.onMessage('bot:token', { conversationId: conv.id, token: '\n\n' });
+        emitToken(conv, '\n\n');
+      }
+      if (p.itemId) {
+        conv.currentMsgItemId = p.itemId;
+        conv.itemTexts.set(p.itemId, (conv.itemTexts.get(p.itemId) || '') + delta);
       }
-      if (p.itemId) conv.currentMsgItemId = p.itemId;
       conv.fullText += delta;
-      conv.onMessage('bot:token', { conversationId: conv.id, token: delta });
+      emitToken(conv, delta);
       break;
     }
 
     case 'item/started': {
       const item = p.item || {};
-      // Surface tool-like items so the dashboard can show activity.
+      // Surface tool-like items so the dashboard can show activity. Names use
+      // claude's tool vocabulary (Bash/Edit/WebSearch/mcp__server__tool) so one
+      // UI label map serves both harnesses.
       switch (item.type) {
         case 'commandExecution':
-          conv.onMessage('bot:tool', {
+          emitEvent(conv, 'bot:tool', {
             conversationId: conv.id,
-            name: 'shell',
+            name: 'Bash',
             input: { command: item.command || item.commandLine || '' },
           });
           break;
         case 'mcpToolCall':
           // ThreadItem.mcpToolCall fields are `server` + `tool` (no toolName/name/input).
-          conv.onMessage('bot:tool', {
+          emitEvent(conv, 'bot:tool', {
             conversationId: conv.id,
-            name: item.tool ? (item.server ? `${item.server}/${item.tool}` : item.tool) : 'mcp_tool',
+            name: item.tool ? (item.server ? `mcp__${item.server}__${item.tool}` : item.tool) : 'mcp_tool',
             input: item.arguments || {},
           });
           break;
         case 'fileChange':
           conv.usedFileTools = true;
-          conv.onMessage('bot:tool', {
+          emitEvent(conv, 'bot:tool', {
             conversationId: conv.id,
-            name: 'file_change',
+            name: 'Edit',
             input: { changes: (item.changes || []).map((c: any) => c.path).filter(Boolean) },
           });
           break;
         case 'webSearch':
-          conv.onMessage('bot:tool', {
+          emitEvent(conv, 'bot:tool', {
             conversationId: conv.id,
-            name: 'web_search',
+            name: 'WebSearch',
             input: { query: item.query || '' },
           });
           break;
+        case 'reasoning':
+          // Activity signal during the (often long) thinking phase — without it
+          // the chat shows bare typing dots with zero events for the entire
+          // stretch and turns feel hung. The chat hooks dedup repeated
+          // same-name running tools, and channel chunk-flushes no-op on an
+          // empty buffer, so this is purely additive. Reasoning TEXT is never
+          // emitted as bot:token — that would corrupt the bot:response ===
+          // streamed-text contract and leak reasoning to channels.
+          emitEvent(conv, 'bot:tool', {
+            conversationId: conv.id,
+            name: 'thinking',
+            status: 'running',
+          });
+          break;
         case 'collabAgentToolCall':
-          // Codex's collaborating sub-agents (rarely enabled) → Bloby's sub-agent UX.
+          // Codex's collaborating sub-agents (experimental, not enabled by our
+          // initialize) → Bloby's sub-agent UX, if it ever lights up.
           if (item.tool === 'spawnAgent') {
-            conv.onMessage('bot:task-created', {
+            emitEvent(conv, 'bot:task-created', {
               conversationId: conv.id,
               taskId: item.id,
               description: item.prompt || 'sub-agent',
@@ -564,24 +854,70 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
             });
           }
           break;
-        // userMessage / agentMessage / reasoning — no tool-style event.
+        // userMessage / agentMessage — no tool-style event.
       }
       break;
     }
 
+    case 'turn/plan/updated': {
+      // Codex's native planning (its TodoWrite equivalent). Emit it in claude's
+      // TodoWrite shape so plans drive the same activity label + bubble/chunk
+      // boundaries — without this codex feels opaque exactly on the big agentic
+      // asks where claude feels alive. Statuses are camelCase in 0.138
+      // (pending | inProgress | completed) — normalize to claude's snake_case.
+      const steps = Array.isArray(p.plan) ? p.plan : [];
+      emitEvent(conv, 'bot:tool', {
+        conversationId: conv.id,
+        name: 'TodoWrite',
+        input: {
+          todos: steps.map((s: any) => ({
+            content: s.step || '',
+            status: s.status === 'inProgress' ? 'in_progress' : (s.status || 'pending'),
+          })),
+          ...(p.explanation ? { explanation: p.explanation } : {}),
+        },
+      });
+      break;
+    }
+
     case 'item/completed': {
       const item = p.item || {};
       if (item.type === 'fileChange') conv.usedFileTools = true;
-      // If a final agentMessage arrives without preceding deltas (rare), grab it now.
-      if (item.type === 'agentMessage' && !conv.fullText) {
-        const text = (item.content || []).map((c: any) => c.text || '').join('') || item.text || '';
-        if (text) {
-          conv.fullText = text;
-          conv.onMessage('bot:token', { conversationId: conv.id, token: text });
+      // item/completed carries the AUTHORITATIVE final text for an agentMessage.
+      // Reconcile per-itemId: an item that never streamed deltas (or whose final
+      // text extends beyond the concatenated deltas) gets the remainder emitted
+      // here — previously a turn-level `!conv.fullText` gate silently dropped any
+      // non-delta'd message once ANYTHING had streamed.
+      if (item.type === 'agentMessage') {
+        const finalText: string = (typeof item.text === 'string' && item.text)
+          ? item.text
+          : ((item.content || []).map((c: any) => c.text || '').join(''));
+        if (finalText) {
+          const itemId: string | undefined = item.id;
+          const streamed = itemId ? (conv.itemTexts.get(itemId) || '') : conv.fullText;
+          let remainder = '';
+          if (!streamed) {
+            remainder = finalText;
+          } else if (finalText.length > streamed.length && finalText.startsWith(streamed)) {
+            remainder = finalText.slice(streamed.length);
+          }
+          if (remainder) {
+            // Same paragraph-break rule as the delta path for a fresh item.
+            if (!streamed && conv.fullText && !conv.fullText.endsWith('\n')) {
+              conv.fullText += '\n\n';
+              emitToken(conv, '\n\n');
+            }
+            conv.fullText += remainder;
+            emitToken(conv, remainder);
+            if (itemId) {
+              conv.itemTexts.set(itemId, finalText);
+              conv.currentMsgItemId = itemId;
+            }
+          }
         }
       }
       if (item.type === 'collabAgentToolCall' && item.tool === 'spawnAgent') {
-        conv.onMessage('bot:task-done', {
+        emitEvent(conv, 'bot:task-done', {
           conversationId: conv.id,
           taskId: item.id,
           status: item.status,
@@ -593,11 +929,14 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
 
     case 'thread/tokenUsage/updated': {
       // Codex's only token-usage signal. ThreadTokenUsage = { total, last, modelContextWindow };
-      // `last` is the current prompt occupancy (mirrors Claude's input+cacheRead+cacheCreation),
-      // the right basis for the recycle compare in supervisor/index.ts (fraction*window, not lifetime).
+      // `last.inputTokens` is the current prompt occupancy — the right basis for
+      // the recycle compare in supervisor/index.ts (fraction*window, not lifetime).
+      // NB: codex's cachedInputTokens is a SUBSET of inputTokens (unlike
+      // Anthropic's disjoint cache counters) — summing them double-counts the
+      // cache and recycles sessions at ~35% real occupancy.
       const tu = p.tokenUsage || {};
       const last = tu.last || {};
-      conv.lastContextTokens = (last.inputTokens || 0) + (last.cachedInputTokens || 0);
+      conv.lastContextTokens = last.inputTokens || 0;
       if (typeof tu.modelContextWindow === 'number' && tu.modelContextWindow > 0) {
         conv.lastContextWindow = tu.modelContextWindow;
       }
@@ -609,29 +948,41 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
       const turnError = p.turn?.error;
 
       clearTurnWatchdog(conv);
+      clearHardTurnTimer(conv);
       conv.currentTurnId = null;
       conv.busy = false;
 
+      let failureKind: string | undefined;
       if (status === 'failed') {
-        conv.onMessage('bot:error', {
-          conversationId: conv.id,
-          error: turnError?.message || 'Codex turn failed.',
-        });
+        // Preserve streamed partials exactly like claude: if text already reached
+        // the user, commit it as the reply (the frontend's bot:error handler
+        // ERASES the uncommitted stream buffer); surface bot:error only when the
+        // turn died before producing anything.
+        const message = turnError?.message || conv.stashedError?.message || 'Codex turn failed.';
+        const info = turnError?.codexErrorInfo ?? conv.stashedError?.info;
+        failureKind = errorInfoKind(info);
+        if (conv.fullText) {
+          log.warn(`[codex] turn failed after streaming ${conv.fullText.length} chars (${message.slice(0, 160)}) — preserving partial as the reply`);
+          emitEvent(conv, 'bot:response', { conversationId: conv.id, content: conv.fullText });
+        } else {
+          emitError(conv, message, info, turnError?.additionalDetails);
+        }
       } else if (status === 'interrupted') {
         // Interrupted turns carry no final answer — stay silent.
       } else if (conv.fullText) {
-        conv.onMessage('bot:response', { conversationId: conv.id, content: conv.fullText });
+        emitEvent(conv, 'bot:response', { conversationId: conv.id, content: conv.fullText });
       }
+      conv.stashedError = null;
 
       if (conv.oneShot) {
-        conv.onMessage('bot:done', { conversationId: conv.id, usedFileTools: conv.usedFileTools });
+        emitDone(conv);
         teardownConversation(conv.id);
       } else {
         // Context-size signal for the orchestrator's proactive session recycling,
         // sourced from the cached `thread/tokenUsage/updated` values above. 0 if codex
         // never sent one this turn → falls back to codex's own in-thread auto-compaction.
         const idle = conv.pendingInputs.length === 0;
-        conv.onMessage('bot:turn-complete', {
+        emitEvent(conv, 'bot:turn-complete', {
           conversationId: conv.id,
           usedFileTools: conv.usedFileTools,
           contextTokens: conv.lastContextTokens || 0,
@@ -639,9 +990,30 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
           idle,
         });
 
-        // Drain any messages that were submitted while we were busy.
+        // An unauthorized/context-wall thread can't run further turns — tear it
+        // down so the NEXT message pre-flights cleanly (friendly dashboard
+        // message / fresh thread) instead of repeating the same failure forever.
+        if (failureKind === 'unauthorized' || failureKind === 'contextWindowExceeded') {
+          // Every queued message got bot:typing on push — give each a terminal
+          // event too, or it vanishes without feedback. (Direct emit: the
+          // per-turn errorEmitted dedup must not swallow these.)
+          const reason = failureKind === 'unauthorized' ? 'authentication expired' : 'context window full';
+          for (const dropped of conv.pendingInputs.splice(0)) {
+            const snippet = dropped.content.slice(0, 60) + (dropped.content.length > 60 ? '…' : '');
+            emitEvent(conv, 'bot:error', {
+              conversationId: conv.id,
+              error: `Codex session ended (${reason}) before your message "${snippet}" could run — please resend it.`,
+            });
+          }
+          log.warn(`[codex] tearing down conv=${conv.id} after ${failureKind} so the next message starts clean`);
+          teardownConversation(conv.id);
+          break;
+        }
+
+        // Drain any messages that were queued while we were busy — each gets
+        // its own turn (and thus its own bot:response).
         const next = conv.pendingInputs.shift();
-        if (next !== undefined) void startTurn(conv, next);
+        if (next !== undefined) void startTurn(conv, next.content, next.savedFiles, next.attachments);
       }
       break;
     }
@@ -653,8 +1025,19 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
         log.info(`[codex] transient error (will retry): ${p.error?.message || 'unknown'}`);
         break;
       }
-      const errMsg = p.error?.message || 'Codex error notification';
-      conv.onMessage('bot:error', { conversationId: conv.id, error: errMsg });
+      // Hard failures emit BOTH this notification AND turn/completed{failed}
+      // with the same TurnError (live-verified on 0.138.0). While a turn is
+      // active, stash it and let turn/completed be the single surface — its
+      // TurnError is authoritative per the protocol. Outside a turn there is
+      // no turn/completed coming, so emit directly.
+      const message = p.error?.message || 'Codex error notification';
+      const info = p.error?.codexErrorInfo;
+      if (conv.busy) {
+        log.info(`[codex] stashing mid-turn error for turn/completed: ${message.slice(0, 160)}`);
+        conv.stashedError = { message, info };
+      } else {
+        emitError(conv, message, info, p.error?.additionalDetails);
+      }
       break;
     }
 
@@ -672,13 +1055,110 @@ function handleNotification(conv: CodexConversation, n: { method: string; params
   }
 }
 
+/* ── Conversation teardown ─────────────────────────────────────────────── */
+
 function teardownConversation(conversationId: string): void {
   const conv = conversations.get(conversationId);
   if (!conv) return;
   clearTurnWatchdog(conv);
+  clearHardTurnTimer(conv);
   conversations.delete(conversationId);
+  discardTokens(conv);
   try { conv.rpc.close(); } catch {}
+  // bot:done guarantee: slot-freeing consumers (WhatsApp activeAgents, the
+  // scheduler) resolve ONLY on bot:done — claude guarantees it in a finally,
+  // so every codex terminal path (stop, settings:save teardown, init failure,
+  // crash) must too, or each failed customer message pins a slot until restart.
+  if (conv.oneShot) emitDone(conv);
   conv.onMessage('bot:conversation-ended', { conversationId });
+  // Re-warm for the next live conversation (mirrors claude's finally).
+  if (!conv.oneShot) void warmUpForLiveConversation(conv.model, conv.names);
+}
+
+/* ── Spawn + initialize (with pre-warm pool) ───────────────────────────── */
+
+/**
+ * Pre-warmed app-server: spawned + initialize handshake done. `thread/start`
+ * is deliberately NOT issued at warm time — it's a single fast local RPC and
+ * deferring it means the warm process is claimable for ANY model/instructions
+ * (claude's warmer must bake the full options in and misses whenever
+ * recentMessages differ; this design sidesteps that). Keyed on the MCP spawn
+ * flags, the only thing fixed at spawn time.
+ */
+interface CodexWarmEntry {
+  key: string;
+  rpc: CodexRpc;
+}
+
+let codexWarm: CodexWarmEntry | null = null;
+let codexWarmInflight: Promise<void> | null = null;
+
+function warmKeyFor(mcpArgs: string[]): string {
+  return crypto.createHash('sha256').update(JSON.stringify(mcpArgs)).digest('hex');
+}
+
+function claimWarmRpc(key: string): CodexRpc | null {
+  if (!codexWarm) return null;
+  if (codexWarm.key !== key) {
+    // MCP config changed since warm time — the flags are baked into the spawn.
+    try { codexWarm.rpc.close(); } catch {}
+    codexWarm = null;
+    return null;
+  }
+  const rpc = codexWarm.rpc;
+  codexWarm = null;
+  log.info('[codex] claimed pre-warmed app-server');
+  return rpc;
+}
+
+function discardCodexWarmup(): void {
+  if (codexWarm) {
+    try { codexWarm.rpc.close(); } catch {}
+    codexWarm = null;
+  }
+}
+
+export async function warmUpForLiveConversation(
+  _model: string,
+  _names?: { botName: string; humanName: string },
+): Promise<void> {
+  if (codexWarmInflight) return codexWarmInflight;
+  const mcpArgs = buildMcpConfigArgs();
+  const key = warmKeyFor(mcpArgs);
+  if (codexWarm?.key === key) return;
+
+  codexWarmInflight = (async () => {
+    try {
+      const token = await getCodexAccessToken();
+      if (!token) return; // not authed — nothing to warm
+      if (codexWarm && codexWarm.key !== key) discardCodexWarmup();
+      const rpc = new CodexRpc();
+      rpc.start(mcpArgs);
+      rpc.onClose(() => {
+        // Warm process died on its own — drop it from the cache so a claim
+        // never hands out a dead rpc.
+        if (codexWarm?.rpc === rpc) codexWarm = null;
+      });
+      await rpc.request('initialize', { clientInfo: CLIENT_INFO });
+      rpc.notify('initialized', {});
+      codexWarm = { key, rpc };
+      log.ok('[codex] app-server pre-warmed');
+    } catch (err: any) {
+      log.warn(`[codex] pre-warm skipped: ${err?.message || err}`);
+    } finally {
+      codexWarmInflight = null;
+    }
+  })();
+  return codexWarmInflight;
+}
+
+interface SpawnOpts {
+  oneShot: boolean;
+  /** Customer-facing runs (supportPrompt personas) get NO workspace skills —
+   *  mirrors claude's `skills: supportPrompt ? [] : ...` gate so internal ops
+   *  skills can't leak into the customer context. */
+  wantSkills: boolean;
+  names?: { botName: string; humanName: string };
 }
 
 async function spawnAndInitialize(
@@ -686,7 +1166,7 @@ async function spawnAndInitialize(
   model: string,
   onMessage: OnAgentMessage,
   instructions: string,
-  oneShot: boolean,
+  opts: SpawnOpts,
 ): Promise<CodexConversation | null> {
   // Pre-flight: confirm we have valid OAuth tokens before spending time spawning.
   const token = await getCodexAccessToken();
@@ -698,85 +1178,202 @@ async function spawnAndInitialize(
     return null;
   }
 
-  const { id: modelId, effort } = parseModelString(model);
-  const rpc = new CodexRpc();
-  rpc.start(buildMcpConfigArgs());
+  const { id: parsedModelId, effort: parsedEffort } = parseModelString(model);
+  const mcpArgs = buildMcpConfigArgs();
+
+  // One attempt with the pre-warmed process (live convs only — one-shots churn
+  // too fast to be worth re-warming for), falling back to a cold spawn if the
+  // warm one fails its thread/start.
+  const attempts: Array<'warm' | 'cold'> = [];
+  if (!opts.oneShot && codexWarm) attempts.push('warm');
+  attempts.push('cold');
+
+  for (const attempt of attempts) {
+    let rpc: CodexRpc;
+    if (attempt === 'warm') {
+      const claimed = claimWarmRpc(warmKeyFor(mcpArgs));
+      if (!claimed) continue;
+      rpc = claimed;
+    } else {
+      rpc = new CodexRpc();
+      rpc.start(mcpArgs);
+    }
 
-  const conv: CodexConversation = {
-    id: conversationId,
-    rpc,
-    threadId: '',
-    effort,
-    onMessage,
-    currentTurnId: null,
-    currentMsgItemId: null,
-    fullText: '',
-    usedFileTools: false,
-    pendingInputs: [],
-    busy: false,
-    oneShot,
-    lastContextTokens: 0,
-    lastContextWindow: 0,
-    turnWatchdog: null,
-  };
+    const conv: CodexConversation = {
+      id: conversationId,
+      rpc,
+      threadId: '',
+      effort: parsedEffort,
+      model,
+      names: opts.names,
+      onMessage,
+      currentTurnId: null,
+      currentMsgItemId: null,
+      fullText: '',
+      itemTexts: new Map(),
+      usedFileTools: false,
+      pendingInputs: [],
+      busy: false,
+      oneShot: opts.oneShot,
+      errorEmitted: false,
+      doneEmitted: false,
+      stashedError: null,
+      lastContextTokens: 0,
+      lastContextWindow: 0,
+      turnWatchdog: null,
+      hardTurnTimer: null,
+      tokenBuf: '',
+      tokenFlushTimer: null,
+    };
 
-  rpc.onNotification((n) => handleNotification(conv, n));
-  rpc.onClose(() => {
-    if (conversations.get(conversationId) === conv) {
+    rpc.onNotification((n) => handleNotification(conv, n));
+    rpc.onClose((code, errMsg) => {
+      if (conversations.get(conversationId) !== conv) return;
+      // App-server died out from under a live conversation. Surface it NOW —
+      // without this the user stares at frozen typing dots until the leaked
+      // watchdog ghost-fires 5 minutes later with a misleading "timed out"
+      // (and that orphaned watchdog could tear down a successor conversation
+      // started under the same convId in the meantime).
+      clearTurnWatchdog(conv);
+      clearHardTurnTimer(conv);
       conversations.delete(conversationId);
+      flushTokens(conv);
+      if (conv.busy) {
+        conv.busy = false;
+        conv.currentTurnId = null;
+        emitError(conv, errMsg || `codex app-server exited unexpectedly (code=${code}).`);
+        if (!conv.oneShot) {
+          emitEvent(conv, 'bot:turn-complete', {
+            conversationId: conv.id,
+            usedFileTools: conv.usedFileTools,
+            contextTokens: conv.lastContextTokens || 0,
+            contextWindow: conv.lastContextWindow || 0,
+            idle: true,
+          });
+        }
+      }
+      if (conv.oneShot) emitDone(conv);
       onMessage('bot:conversation-ended', { conversationId });
+      if (!conv.oneShot) void warmUpForLiveConversation(conv.model, conv.names);
+    });
+
+    try {
+      log.info(`[codex] init conversation ${conversationId} (model=${parsedModelId}${parsedEffort ? `, effort=${parsedEffort}` : ''}, ${attempt})`);
+      if (attempt === 'cold') {
+        await rpc.request('initialize', { clientInfo: CLIENT_INFO });
+        rpc.notify('initialized', {});
+      }
+
+      // Validate/repair the configured model + effort against the live catalog —
+      // a stale model id or retired effort tier otherwise fails the first message
+      // with a cryptic teardown (codex has no config-level model migrations).
+      const { modelId, effort } = await validateModelSelection(rpc, parsedModelId, parsedEffort);
+      conv.effort = effort;
+
+      // Context auto-compaction is ON by default in the codex app-server: when the
+      // thread's token count crosses the model's threshold it compacts history in
+      // place (emitting a `contextCompaction` item) and continues — no flag needed
+      // here. A manual trigger also exists (`thread/compact/start`) if we ever want
+      // to force it from the UI.
+      const startResult = await rpc.request<{ thread: { id: string } }>('thread/start', {
+        cwd: WORKSPACE_DIR,
+        model: modelId,
+        // Bloby's persona/workflow prompt rides developerInstructions (ADDITIVE),
+        // NOT baseInstructions. baseInstructions fully OVERRIDES codex's native base
+        // prompt — which carries the apply_patch FREEFORM spec + shell protocol the
+        // model needs to edit files. Leaving baseInstructions unset keeps that native
+        // scaffolding; developerInstructions layers Bloby's persona on top of it.
+        developerInstructions: instructions,
+        personality: 'pragmatic',
+        // Bloby's posture matches Claude's bypassPermissions — the bot is
+        // running on the user's own machine with their full consent. Skip the
+        // approval prompts and give it write access to the workspace + beyond.
+        approvalPolicy: 'never',
+        sandbox: 'danger-full-access',
+      });
+      conv.threadId = startResult.thread.id;
+      conversations.set(conversationId, conv);
+      log.ok(`[codex] thread started ${conv.threadId}`);
+
+      // Prime codex's per-thread skill cache with the workspace skills
+      // directory. Without this, codex only sees its system-scope skills and
+      // never discovers anything Bloby ships in `workspace/skills/*`. Fire and
+      // forget — failure here just means workspace skills won't be auto-routable
+      // for this thread, but the agent can still read SKILL.md files directly.
+      // Customer-facing personas skip it entirely (see SpawnOpts.wantSkills).
+      if (opts.wantSkills) primeWorkspaceSkills(rpc);
+
+      return conv;
+    } catch (err: any) {
+      rpc.close();
+      if (attempt === 'warm') {
+        log.warn(`[codex] warm claim failed (${err.message}) — retrying with a cold spawn`);
+        continue;
+      }
+      onMessage('bot:error', { conversationId, error: `Failed to initialize Codex: ${err.message}` });
+      return null;
     }
-  });
+  }
+  return null; // unreachable — 'cold' attempt always returns/errors above
+}
 
-  try {
-    log.info(`[codex] init conversation ${conversationId} (model=${modelId}${effort ? `, effort=${effort}` : ''})`);
-    await rpc.request('initialize', { clientInfo: CLIENT_INFO });
-    rpc.notify('initialized', {});
-    // Context auto-compaction is ON by default in the codex app-server: when the
-    // thread's token count crosses the model's threshold it compacts history in
-    // place (emitting a `contextCompaction` item) and continues — no flag needed
-    // here. A manual trigger also exists (`thread/compact/start`) if we ever want
-    // to force it from the UI.
-    const startResult = await rpc.request<{ thread: { id: string } }>('thread/start', {
-      cwd: WORKSPACE_DIR,
-      model: modelId,
-      // Bloby's persona/workflow prompt rides developerInstructions (ADDITIVE),
-      // NOT baseInstructions. baseInstructions fully OVERRIDES codex's native base
-      // prompt — which carries the apply_patch FREEFORM spec + shell protocol the
-      // model needs to edit files. Leaving baseInstructions unset keeps that native
-      // scaffolding; developerInstructions layers Bloby's persona on top of it.
-      developerInstructions: instructions,
-      personality: 'pragmatic',
-      // Bloby's posture matches Claude's bypassPermissions — the bot is
-      // running on the user's own machine with their full consent. Skip the
-      // approval prompts and give it write access to the workspace + beyond.
-      approvalPolicy: 'never',
-      sandbox: 'danger-full-access',
-    });
-    conv.threadId = startResult.thread.id;
-    conversations.set(conversationId, conv);
-    log.ok(`[codex] thread started ${conv.threadId}`);
-
-    // Prime codex's per-thread skill cache with the workspace skills
-    // directory. Without this, codex only sees its system-scope skills and
-    // never discovers anything Bloby ships in `workspace/skills/*`. Fire and
-    // forget — failure here just means workspace skills won't be auto-routable
-    // for this thread, but the agent can still read SKILL.md files directly.
-    primeWorkspaceSkills(rpc);
+/* ── Model catalog validation ──────────────────────────────────────────── */
+
+interface ValidatedModel { modelId: string; effort?: string }
+const modelValidationCache = new Map<string, ValidatedModel>();
 
-    return conv;
+/**
+ * Validate the configured model id + effort against `model/list` (local
+ * catalog, one fast RPC — cached per model string for the process lifetime).
+ * Unknown model → repair to the catalog default with a logged warning instead
+ * of letting thread/start or turn/start fail with a cryptic teardown. Unknown
+ * effort → the model's defaultReasoningEffort. If model/list itself fails,
+ * proceed unvalidated (today's behavior).
+ */
+async function validateModelSelection(rpc: CodexRpc, modelId: string, effort?: string): Promise<ValidatedModel> {
+  const cacheKey = `${modelId}:${effort || ''}`;
+  const cached = modelValidationCache.get(cacheKey);
+  if (cached) return cached;
+
+  let result: ValidatedModel = { modelId, effort };
+  try {
+    const res = await rpc.request<{ data?: any[] }>('model/list', { includeHidden: true });
+    const catalog = Array.isArray(res?.data) ? res.data : [];
+    if (catalog.length) {
+      let entry = catalog.find((m: any) => m.id === modelId || m.model === modelId);
+      if (!entry) {
+        const fallback = catalog.find((m: any) => m.isDefault) || catalog[0];
+        log.warn(`[codex] model "${modelId}" not in catalog — falling back to "${fallback.id}"`);
+        entry = fallback;
+        result.modelId = entry.id;
+        result.effort = undefined; // stale effort may not apply to the fallback
+      }
+      const supported: string[] = (entry.supportedReasoningEfforts || [])
+        .map((o: any) => (typeof o === 'string' ? o : o?.reasoningEffort))
+        .filter(Boolean);
+      if (result.effort && supported.length && !supported.includes(result.effort)) {
+        const repaired = entry.defaultReasoningEffort || undefined;
+        log.warn(`[codex] effort "${result.effort}" not supported by ${result.modelId} (supported: ${supported.join(', ')}) — using ${repaired || 'server default'}`);
+        result.effort = repaired;
+      }
+      // Only cache results actually validated against a catalog — caching the
+      // passthrough on a failed/empty model/list would disable the auto-repair
+      // for this model string for the whole process lifetime.
+      modelValidationCache.set(cacheKey, result);
+    }
   } catch (err: any) {
-    rpc.close();
-    onMessage('bot:error', { conversationId, error: `Failed to initialize Codex: ${err.message}` });
-    return null;
+    log.warn(`[codex] model/list validation skipped: ${err.message}`);
   }
+  return result;
 }
 
 // Codex discovers "repo"-scope skills under `<cwd>/.codex/skills` (verified
-// against 0.135.0 — a bare `<cwd>/skills` is NOT scanned, and `skills/list`
-// has no extra-root param). Bloby keeps the canonical skills in
-// `workspace/skills/<name>`, so we mirror each one into `.codex/skills/<name>`
-// as a symlink — single source of truth, discoverable by codex's native router.
+// against 0.135.0 — a bare `<cwd>/skills` is NOT scanned). Bloby keeps the
+// canonical skills in `workspace/skills/<name>`, so we mirror each one into
+// `.codex/skills/<name>` as a symlink — single source of truth, discoverable
+// by codex's native router. 0.138 added `skills/extraRoots/set` which could
+// replace this mirror — deliberately NOT adopted yet: the mirror is
+// e2e-verified working and the swap buys no user-visible change.
 // (Each SKILL.md needs YAML frontmatter or codex rejects it — see SKILL_FORMAT_MIGRATION.md.)
 const CODEX_SKILLS_ROOT = path.join(WORKSPACE_DIR, '.codex', 'skills');
 
@@ -826,12 +1423,25 @@ function toToml(v: any): string {
   return JSON.stringify(String(v)); // TOML basic string — JSON escaping is compatible
 }
 
+/** codex's config layer requires STRING values for args items, env values and
+ *  http_headers — a numeric `"PORT": 3000` in MCP.json (fine on claude) would
+ *  otherwise kill the app-server at config load, bricking EVERY codex spawn.
+ *  Coerce so non-string JSON values degrade to their string form instead. */
+function toTomlStringArray(arr: any[]): string {
+  return `[${arr.map((v) => JSON.stringify(String(v))).join(',')}]`;
+}
+function toTomlStringMap(obj: Record<string, any>): string {
+  return `{${Object.entries(obj).map(([k, v]) => `${JSON.stringify(k)}=${JSON.stringify(String(v))}`).join(',')}}`;
+}
+
 /**
  * Translate MCP.json into `codex app-server -c mcp_servers.<name>.<field>=<toml>`
  * spawn flags. Codex sources MCP from its own config layer rather than a per-query
  * param (verified against 0.135.0: a `-c mcp_servers.X.command=...` override shows
- * up in both mcpServerStatus/list and config/read). Only the stdio fields Bloby
- * uses (command/args/env) are translated; names must be TOML-bare-key safe.
+ * up in both mcpServerStatus/list and config/read). Stdio entries translate
+ * command/args/env; url entries (streamable HTTP — the Claude SDK's http/sse
+ * form) translate url + headers→http_headers + bearer_token_env_var, supported
+ * natively by the 0.138 binary. Names must be TOML-bare-key safe.
  */
 function buildMcpConfigArgs(): string[] {
   const servers = loadMcpServersForCodex();
@@ -840,10 +1450,22 @@ function buildMcpConfigArgs(): string[] {
   for (const [name, cfg] of Object.entries(servers)) {
     if (!/^[A-Za-z0-9_-]+$/.test(name)) { log.warn(`[codex] skipping MCP server "${name}" — name not TOML-bare-key safe`); continue; }
     const c: any = cfg || {};
-    if (!c.command) { log.warn(`[codex] skipping MCP server "${name}" — no command`); continue; }
-    args.push('-c', `mcp_servers.${name}.command=${toToml(c.command)}`);
-    if (Array.isArray(c.args) && c.args.length) args.push('-c', `mcp_servers.${name}.args=${toToml(c.args)}`);
-    if (c.env && typeof c.env === 'object' && Object.keys(c.env).length) args.push('-c', `mcp_servers.${name}.env=${toToml(c.env)}`);
+    if (c.command) {
+      args.push('-c', `mcp_servers.${name}.command=${toToml(c.command)}`);
+      if (Array.isArray(c.args) && c.args.length) args.push('-c', `mcp_servers.${name}.args=${toTomlStringArray(c.args)}`);
+      if (c.env && typeof c.env === 'object' && Object.keys(c.env).length) args.push('-c', `mcp_servers.${name}.env=${toTomlStringMap(c.env)}`);
+    } else if (typeof c.url === 'string' && c.url) {
+      args.push('-c', `mcp_servers.${name}.url=${toToml(c.url)}`);
+      if (c.headers && typeof c.headers === 'object' && Object.keys(c.headers).length) {
+        args.push('-c', `mcp_servers.${name}.http_headers=${toTomlStringMap(c.headers)}`);
+      }
+      if (typeof c.bearer_token_env_var === 'string' && c.bearer_token_env_var) {
+        args.push('-c', `mcp_servers.${name}.bearer_token_env_var=${toToml(c.bearer_token_env_var)}`);
+      }
+    } else {
+      log.warn(`[codex] skipping MCP server "${name}" — no command or url`);
+      continue;
+    }
     wired++;
   }
   if (wired) log.info(`[codex] wiring ${wired} MCP server(s) from MCP.json via -c overrides`);
@@ -860,12 +1482,11 @@ export function isConversationBusy(conversationId: string): boolean {
   return conversations.get(conversationId)?.busy ?? false;
 }
 
-/** True if ANY live conversation in this harness is mid-turn. Used by the supervisor to defer
- *  backend restarts during channel/Alexa turns (which don't set the dashboard's agentQueryActive). */
-/** Codex one-shot queries (startBlobyAgentQuery) reuse the conversations map and set conv.busy via
- *  startTurn, so anyConversationBusy() already covers them — nothing extra to track here. */
+/** True while any one-shot is in its init window (token refresh + spawn +
+ *  initialize + thread/start). Once the conv registers in `conversations` and
+ *  startTurn sets busy, anyConversationBusy() takes over seamlessly. */
 export function anyOneShotActive(): boolean {
-  return false;
+  return inFlightOneShots.size > 0;
 }
 
 export function anyConversationBusy(): boolean {
@@ -881,15 +1502,31 @@ export async function startConversation(
   recentMessages?: RecentMessage[],
 ): Promise<boolean> {
   if (conversations.has(conversationId)) endConversation(conversationId);
+  // Typing dots NOW — the supervisor awaits this whole handshake before the
+  // first pushMessage can run, and without an early signal the user stares at
+  // a dead chat through spawn + initialize + thread/start on every session
+  // start (boot, clear-context, every proactive recycle).
+  onMessage('bot:typing', { conversationId });
   const baseInstructions = await assembleBaseInstructions(names, recentMessages);
-  const conv = await spawnAndInitialize(conversationId, model, onMessage, baseInstructions, false);
-  return !!conv;
+  const conv = await spawnAndInitialize(conversationId, model, onMessage, baseInstructions, {
+    oneShot: false,
+    wantSkills: true,
+    names,
+  });
+  if (!conv) {
+    // bot:typing above set the supervisor's agentQueryActive; bot:error alone
+    // doesn't clear it — conversation-ended does (and is safe for a conv that
+    // never registered).
+    onMessage('bot:conversation-ended', { conversationId });
+    return false;
+  }
+  return true;
 }
 
 export function pushMessage(
   conversationId: string,
   content: string,
-  _attachments?: AgentAttachment[],
+  attachments?: AgentAttachment[],
   savedFiles?: SavedFile[],
 ): boolean {
   const conv = conversations.get(conversationId);
@@ -897,7 +1534,7 @@ export function pushMessage(
     log.warn(`[codex] pushMessage: no live conversation ${conversationId}`);
     return false;
   }
-  void steerOrQueue(conv, content, savedFiles);
+  queueOrStart(conv, content, savedFiles, attachments);
   return true;
 }
 
@@ -916,18 +1553,14 @@ export function endConversation(conversationId: string): void {
 
 export function endAllConversations(): void {
   for (const id of Array.from(conversations.keys())) endConversation(id);
+  // The pre-warmed app-server may hold pre-re-auth state — drop it (a fresh
+  // warm-up fires from the teardown re-warm path with current credentials).
+  discardCodexWarmup();
 }
 
 export async function stopSubAgentTask(_conversationId: string, _taskId: string): Promise<void> {
-  // Codex doesn't expose Claude-style sub-agent tasks. No-op for now.
-}
-
-export async function warmUpForLiveConversation(
-  _model: string,
-  _names?: { botName: string; humanName: string },
-): Promise<void> {
-  // No subprocess pre-warming yet — `codex app-server` startup is fast enough
-  // (~hundreds of ms). Re-evaluate if it becomes noticeable on the Pi.
+  // Codex doesn't expose Claude-style sub-agent tasks (collab tools are still
+  // experimental and not enabled by our initialize). No-op.
 }
 
 export async function startBlobyAgentQuery(
@@ -935,24 +1568,45 @@ export async function startBlobyAgentQuery(
   prompt: string,
   model: string,
   onMessage: OnAgentMessage,
-  _attachments?: AgentAttachment[],
+  attachments?: AgentAttachment[],
   savedFiles?: SavedFile[],
   names?: { botName: string; humanName: string },
   recentMessages?: RecentMessage[],
   supportPrompt?: string,
-  _maxTurns?: number,
+  _maxTurns?: number, // no codex equivalent — bounded by ONE_SHOT_MAX_TURN_MS instead
 ): Promise<void> {
-  if (conversations.has(conversationId)) endConversation(conversationId);
-  const baseInstructions = supportPrompt
-    ? supportPrompt
-    : await assembleBaseInstructions(names, recentMessages);
-  const conv = await spawnAndInitialize(conversationId, model, onMessage, baseInstructions, true);
-  if (!conv) return;
-  await startTurn(conv, prompt, savedFiles);
+  inFlightOneShots.add(conversationId);
+  try {
+    if (conversations.has(conversationId)) endConversation(conversationId);
+    onMessage('bot:typing', { conversationId });
+    // Support personas ride supportPrompt as the FULL instructions (SCRIPT.md
+    // governs them) — but the per-customer message buffer must still be
+    // appended, exactly like claude: every one-shot is a fresh thread, so
+    // recentMessages IS the agent's only memory of the conversation.
+    let baseInstructions = supportPrompt ?? await assembleBaseInstructions(names, recentMessages);
+    if (supportPrompt && recentMessages?.length) {
+      baseInstructions += `\n\n---\n# Recent Conversation\n${formatConversationHistory(recentMessages)}`;
+    }
+    const conv = await spawnAndInitialize(conversationId, model, onMessage, baseInstructions, {
+      oneShot: true,
+      wantSkills: !supportPrompt,
+      names,
+    });
+    if (!conv) {
+      // Init failed (no auth / spawn / thread-start error — bot:error already
+      // emitted). bot:done frees the caller's slot; without it each failed
+      // customer message pins one of the 5 WhatsApp agent slots until restart.
+      onMessage('bot:done', { conversationId, usedFileTools: false });
+      return;
+    }
+    await startTurn(conv, prompt, savedFiles, attachments);
+  } finally {
+    inFlightOneShots.delete(conversationId);
+  }
 }
 
 export function stopBlobyAgentQuery(conversationId: string): void {
-  endConversation(conversationId);
+  endConversation(conversationId); // teardown guarantees the one-shot's bot:done
 }
 
 // ── Workspace agent endpoint (POST /api/agent/query) ──────────────────────
@@ -996,49 +1650,82 @@ export async function runAgentQuery(req: AgentQueryRequest): Promise<AgentQueryR
   let resolvedThreadId = req.sessionId || '';
   let resolveTurn: (() => void) | null = null;
   let turnError: string | null = null;
+  const itemTexts = new Map<string, string>();
+
   const turnDone = new Promise<void>((r) => { resolveTurn = r; });
 
   rpc.onNotification((n) => {
     const p = n.params || {};
     switch (n.method) {
       case 'item/agentMessage/delta': {
-        if (typeof p.delta === 'string') fullText += p.delta;
+        if (typeof p.delta === 'string') {
+          fullText += p.delta;
+          if (p.itemId) itemTexts.set(p.itemId, (itemTexts.get(p.itemId) || '') + p.delta);
+        }
         break;
       }
       case 'item/started': {
         const item = p.item || {};
-        if (item.type === 'commandExecution') usedTools.add('shell');
-        else if (item.type === 'mcpToolCall') usedTools.add(item.tool || 'mcp_tool');
-        else if (item.type === 'fileChange') { usedTools.add('file_change'); usedFileTools = true; }
-        else if (item.type === 'webSearch') usedTools.add('web_search');
+        // Same claude-vocabulary normalization as the live path — agent-api
+        // callers see the identical toolsUsed names on both harnesses.
+        if (item.type === 'commandExecution') usedTools.add('Bash');
+        else if (item.type === 'mcpToolCall') usedTools.add(item.tool ? (item.server ? `mcp__${item.server}__${item.tool}` : item.tool) : 'mcp_tool');
+        else if (item.type === 'fileChange') { usedTools.add('Edit'); usedFileTools = true; }
+        else if (item.type === 'webSearch') usedTools.add('WebSearch');
         break;
       }
       case 'item/completed': {
         const item = p.item || {};
         if (item.type === 'fileChange') usedFileTools = true;
-        if (item.type === 'agentMessage' && !fullText) {
-          const text = (item.content || []).map((c: any) => c.text || '').join('') || item.text || '';
-          if (text) fullText = text;
+        // item/completed is authoritative per item — same per-itemId
+        // reconciliation as the live path: append items that never streamed
+        // deltas, and the remainder when the final text extends past them.
+        if (item.type === 'agentMessage' && item.id) {
+          const finalText: string = (typeof item.text === 'string' && item.text)
+            ? item.text
+            : ((item.content || []).map((c: any) => c.text || '').join(''));
+          if (finalText) {
+            const streamed = itemTexts.get(item.id) || '';
+            let remainder = '';
+            if (!streamed) {
+              remainder = finalText;
+            } else if (finalText.length > streamed.length && finalText.startsWith(streamed)) {
+              remainder = finalText.slice(streamed.length);
+            }
+            if (remainder) {
+              if (!streamed && fullText && !fullText.endsWith('\n')) fullText += '\n\n';
+              fullText += remainder;
+              itemTexts.set(item.id, finalText);
+            }
+          }
         }
         break;
       }
       case 'turn/completed': {
         const status = p.turn?.status || 'completed';
         if (status === 'failed') {
-          turnError = p.turn?.error?.message || 'Codex turn failed.';
+          turnError = humanizeCodexError(p.turn?.error?.message, p.turn?.error?.codexErrorInfo, p.turn?.error?.additionalDetails);
         }
         resolveTurn?.();
         break;
       }
       case 'error': {
         if (p.willRetry) break; // transient — codex retries itself
-        turnError = p.error?.message || 'Codex error';
+        turnError = humanizeCodexError(p.error?.message || 'Codex error', p.error?.codexErrorInfo, p.error?.additionalDetails);
         resolveTurn?.();
         break;
       }
     }
   });
 
+  // App-server crash mid-query: without this, no notification ever resolves
+  // turnDone and the caller burns the FULL timeout before getting a generic
+  // "timed out" — surface the real exit error immediately instead.
+  rpc.onClose((code, errMsg) => {
+    if (!turnError) turnError = errMsg || `codex app-server exited unexpectedly (code=${code}).`;
+    resolveTurn?.();
+  });
+
   const timeoutHandle = setTimeout(() => {
     if (!turnError) turnError = `Query timed out after ${timeout}ms.`;
     resolveTurn?.();
@@ -1049,31 +1736,41 @@ export async function runAgentQuery(req: AgentQueryRequest): Promise<AgentQueryR
     await rpc.request('initialize', { clientInfo: CLIENT_INFO });
     rpc.notify('initialized', {});
 
+    // Same stale-model/effort auto-repair as spawnAndInitialize — agent-api is
+    // otherwise the one codex entry point that fails raw on a retired model.
+    const validated = await validateModelSelection(rpc, model, effort);
+    model = validated.modelId;
+    effort = validated.effort;
+
+    // Same execution posture on EVERY thread path — resume included.
+    // ThreadResumeParams accepts all of these in 0.138; without them a resumed
+    // session silently dropped the caller's systemPrompt and fell back to
+    // config-default sandbox/approval settings.
+    const threadOverrides = {
+      cwd: WORKSPACE_DIR,
+      model,
+      ...(req.systemPrompt ? { developerInstructions: req.systemPrompt } : {}),
+      personality: 'pragmatic',
+      approvalPolicy: 'never',
+      sandbox: 'danger-full-access',
+    };
+
     if (req.sessionId) {
       // Resume an existing thread (if codex still has it). Caller must accept
       // failure here — we fall back to a fresh thread.
       try {
-        const r = await rpc.request<{ thread: { id: string } }>('thread/resume', { threadId: req.sessionId });
+        const r = await rpc.request<{ thread: { id: string } }>('thread/resume', {
+          threadId: req.sessionId,
+          ...threadOverrides,
+        });
         resolvedThreadId = r.thread.id;
       } catch (err: any) {
         log.warn(`[codex/agent-api] thread/resume failed (${err.message}); starting fresh thread`);
-        const r = await rpc.request<{ thread: { id: string } }>('thread/start', {
-          cwd: WORKSPACE_DIR,
-          model,
-          ...(req.systemPrompt ? { developerInstructions: req.systemPrompt } : {}),
-          approvalPolicy: 'never',
-          sandbox: 'danger-full-access',
-        });
+        const r = await rpc.request<{ thread: { id: string } }>('thread/start', threadOverrides);
         resolvedThreadId = r.thread.id;
       }
     } else {
-      const r = await rpc.request<{ thread: { id: string } }>('thread/start', {
-        cwd: WORKSPACE_DIR,
-        model,
-        ...(req.systemPrompt ? { developerInstructions: req.systemPrompt } : {}),
-        approvalPolicy: 'never',
-        sandbox: 'danger-full-access',
-      });
+      const r = await rpc.request<{ thread: { id: string } }>('thread/start', threadOverrides);
       resolvedThreadId = r.thread.id;
     }