package.json

@@ -4,12 +4,14 @@
   "private": true,
   "type": "module",
   "scripts": {
-    "test": "node --test tests/domain.test.js tests/analysis.test.js tests/vision-service.test.js tests/vision-config.test.js tests/hermes-agent-service.test.js tests/agent-gateway.acceptance.test.js tests/staging-health.test.js tests/service-worker-runtime.test.js tests/training-ingest.test.js tests/ci-workflow.test.js tests/product-decisions.test.js tests/release-demo.test.js tests/selfhost-bootstrap.test.js tests/staging-config.test.js",
+    "test": "node --test tests/domain.test.js tests/analysis.test.js tests/vision-service.test.js tests/vision-config.test.js tests/hermes-agent-service.test.js tests/agent-queue.test.js tests/agent-turn-timeout.test.js tests/inference-queue.test.js tests/inference-queue-cancellation.test.js tests/inference-zero-call.test.js tests/agent-gateway.acceptance.test.js tests/staging-health.test.js tests/service-worker-runtime.test.js tests/training-ingest.test.js tests/ci-workflow.test.js tests/product-decisions.test.js tests/release-demo.test.js tests/selfhost-bootstrap.test.js tests/staging-config.test.js",
     "test:ui": "node tests/ui.acceptance.mjs",
     "test:photo": "node tests/photo-first.acceptance.mjs",
     "test:sleek": "node tests/sleek-chat.acceptance.mjs",
+    "test:queue-acceptance": "node tests/inference-queue.acceptance.mjs",
+    "test:shutdown-acceptance": "node tests/server-shutdown.acceptance.mjs",
     "test:staging-smoke": "node tests/staging.acceptance.mjs",
-    "check:syntax": "node --check app.js && node --check server.mjs && node --check service-worker.js && node --check src/analysis.js && node --check src/domain.js && node --check src/hermes-agent-service.js && node --check src/vision-config.js && node --check src/vision-service.js && node --check scripts/record_release_demo.mjs && node --check tests/staging.acceptance.mjs && bash -n scripts/bootstrap_selfhost_smolvlm.sh && bash -n scripts/run_selfhost_smolvlm.sh && python3 -m py_compile scripts/ingest_training_photo.py scripts/build_release.py scripts/deploy_staging.py",
+    "check:syntax": "node --check app.js && node --check server.mjs && node --check service-worker.js && node --check src/analysis.js && node --check src/domain.js && node --check src/hermes-agent-service.js && node --check src/inference-queue.js && node --check src/vision-config.js && node --check src/vision-service.js && node --check scripts/record_release_demo.mjs && node --check tests/inference-queue.acceptance.mjs && node --check tests/server-shutdown.acceptance.mjs && node --check tests/staging.acceptance.mjs && bash -n scripts/bootstrap_selfhost_smolvlm.sh && bash -n scripts/run_selfhost_smolvlm.sh && python3 -m py_compile scripts/ingest_training_photo.py scripts/build_release.py scripts/deploy_staging.py",
     "check:diff": "bash scripts/check_diff.sh",
     "start": "node server.mjs"
   },

server.mjs

@@ -39,6 +39,34 @@ function rejectCrossSite(req){const site=String(req.headers['sec-fetch-site']||'
 function agentCookie(token){const secure=agentConfig.publicOrigin.startsWith('https://')?'; Secure':'';return `timmy_agent=${encodeURIComponent(token)}; HttpOnly; SameSite=Strict; Path=${appRoot}; Max-Age=86400${secure}`}
 function sendAgentError(res,error){const status=error instanceof AgentGatewayError?error.status:503;const message=error instanceof AgentGatewayError?error.message:'Hermes is temporarily unavailable. Your local journal still works.';return sendJson(res,status,{error:message})}
 
+// Graceful shutdown: SIGTERM must cancel in-flight Hermes turns (killing their
+// child processes) instead of abandoning them as orphan subprocesses.
+const inFlightChats = new Set();
+let shuttingDown = false;
+function trackChat(res, controller) {
+  if (shuttingDown) controller.abort(new Error('Timmy is shutting down. Your journal still works — try again shortly.'));
+  const entry = { res, controller };
+  inFlightChats.add(entry);
+  // A vanished client must release its queue slot and stop any in-flight
+  // Hermes turn instead of consuming inference capacity silently.
+  res.on('close', () => {
+    inFlightChats.delete(entry);
+    if (!res.writableEnded) {
+      try { controller.abort(new Error('The browser closed the request. Your journal still works.')); } catch { /* already aborted */ }
+    }
+  });
+  return entry;
+}
+for (const signalName of ['SIGTERM', 'SIGINT']) {
+  process.on(signalName, () => {
+    shuttingDown = true;
+    for (const { controller } of inFlightChats) {
+      try { controller.abort(new Error('Server is shutting down.')); } catch { /* already aborted */ }
+    }
+    process.exit(0);
+  });
+}
+
 http.createServer(async(req,res)=>{
  try{
   const url=new URL(req.url,'http://localhost');
@@ -62,7 +90,7 @@ http.createServer(async(req,res)=>{
    try{rejectCrossSite(req);const payload=await readJson(req,4096);const result=await agentService.unlock({origin:requestOrigin(req),accessCode:String(payload?.accessCode||'')});res.setHeader('set-cookie',agentCookie(result.cookieToken));return sendJson(res,200,result.public)}catch(error){return sendAgentError(res,error)}
   }
   if(appPath==='/api/agent/chat'&&req.method==='POST'){
-   try{rejectCrossSite(req);const payload=await readJson(req,128*1024);return sendJson(res,200,await agentService.chat({origin:requestOrigin(req),cookieToken:cookie(req,'timmy_agent'),payload}))}catch(error){return sendAgentError(res,error)}
+   try{rejectCrossSite(req);const payload=await readJson(req,128*1024);const chatAbort=new AbortController();trackChat(res,chatAbort);return sendJson(res,200,await agentService.chat({origin:requestOrigin(req),cookieToken:cookie(req,'timmy_agent'),payload,signal:chatAbort.signal}))}catch(error){return sendAgentError(res,error)}
   }
   if(appPath.startsWith('/api/'))return sendJson(res,404,{error:'Not found'});
   if(req.method!=='GET'&&req.method!=='HEAD'){res.writeHead(405,{'allow':'GET, HEAD'});return res.end()}

src/hermes-agent-service.js

@@ -2,6 +2,7 @@ import { execFile } from 'node:child_process';
 import { randomBytes, timingSafeEqual } from 'node:crypto';
 import { isAbsolute } from 'node:path';
 import { detectUrgentText, hasUrgentLedgerContext, urgentChatMessage } from './domain.js';
+import { OverloadError, createInferenceQueue } from './inference-queue.js';
 
 const MAX_MESSAGE_CHARS = 4000;
 const MAX_LEDGER_ENTRIES = 20;
@@ -44,6 +45,8 @@ export function resolveHermesAgentConfig(env = process.env) {
     maxRequestsPerMinute: positiveInt(env.TIMMY_AGENT_RATE_PER_MINUTE, 12, 1, 60),
     maxUnlockAttemptsPerMinute: positiveInt(env.TIMMY_AGENT_UNLOCK_RATE_PER_MINUTE, 5, 1, 30),
     maxSessions: positiveInt(env.TIMMY_AGENT_MAX_SESSIONS, 64, 1, 512),
+    maxConcurrentTurns: positiveInt(env.TIMMY_AGENT_MAX_CONCURRENT_TURNS, 1, 1, 8),
+    maxQueueDepth: positiveInt(env.TIMMY_AGENT_MAX_QUEUE_DEPTH, 4, 0, 64),
     publicStatus(authenticated = false) {
       return {
         enabled,
@@ -86,12 +89,30 @@ export function parseHermesCliOutput(output) {
   return { sessionId: marker[1], reply };
 }
 
-function execFilePromise(command, args, options) {
+function execFilePromise(command, args, options, signal) {
   return new Promise((resolve, reject) => {
-    execFile(command, args, options, (error, stdout, stderr) => {
+    // Abort must actually terminate the spawned CLI, not merely abandon the
+    // promise — otherwise every cancelled turn leaks an orphan subprocess.
+    let killedByAbort = false;
+    const child = execFile(command, args, options, (error, stdout, stderr) => {
+      if (killedByAbort && signal) {
+        reject(normalizeAbortReason(signal.reason));
+        return;
+      }
       if (error) reject(error);
       else resolve({ stdout, stderr });
     });
+    if (!signal) return;
+    const killTree = () => {
+      killedByAbort = true;
+      try { child.kill('SIGTERM'); } catch { /* already exited */ }
+      const forceKill = setTimeout(() => {
+        try { child.kill('SIGKILL'); } catch { /* already exited */ }
+      }, 2_000);
+      forceKill.unref?.();
+    };
+    if (signal.aborted) killTree();
+    else signal.addEventListener('abort', killTree, { once: true });
   });
 }
 
@@ -103,7 +124,8 @@ export function buildHermesEnvironment(source = process.env) {
   return env;
 }
 
-export async function runHermesCliTurn({ prompt, hermesSessionId, config, execImpl = execFilePromise }) {
+export async function runHermesCliTurn({ prompt, hermesSessionId, config, signal, execImpl = execFilePromise }) {
+  if (signal?.aborted) throw normalizeAbortReason(signal.reason);
   const args = ['chat', '-q', prompt, '-Q', '--source', 'tool', '--max-turns', String(config.maxTurns), '--in', config.workdir];
   if (hermesSessionId) args.push('--resume', hermesSessionId, '--no-restore-cwd');
   const output = await execImpl(config.command, args, {
@@ -112,10 +134,15 @@ export async function runHermesCliTurn({ prompt, hermesSessionId, config, execIm
     maxBuffer: 1024 * 1024,
     windowsHide: true,
     env: buildHermesEnvironment(),
-  });
+  }, signal);
   return parseHermesCliOutput(`${output.stderr || ''}\n${output.stdout || ''}`);
 }
 
+function normalizeAbortReason(reason) {
+  if (reason instanceof Error) return reason;
+  return new Error(typeof reason === 'string' && reason.trim() ? reason : 'Request cancelled.');
+}
+
 function sanitizeLedger(entries) {
   if (!Array.isArray(entries)) throw new AgentGatewayError(400, 'Ledger context must be an array.');
   return entries.slice(-MAX_LEDGER_ENTRIES).map(entry => ({
@@ -139,6 +166,7 @@ export function createHermesAgentService({
   runTurn = input => runHermesCliTurn(input),
   randomToken = () => randomBytes(32).toString('base64url'),
   now = () => Date.now(),
+  queue = createInferenceQueue({ concurrency: config.maxConcurrentTurns, maxQueueDepth: config.maxQueueDepth, requestTimeoutMs: config.timeoutMs, now }),
 } = {}) {
   const sessions = new Map();
   let unlockFailures = [];
@@ -182,7 +210,7 @@ export function createHermesAgentService({
       return { cookieToken, public: config.publicStatus(true) };
     },
 
-    async chat({ origin, cookieToken, payload }) {
+    async chat({ origin, cookieToken, payload, signal }) {
       requireOrigin(origin);
       const session = lookup(cookieToken);
       if (!payload || typeof payload !== 'object' || Array.isArray(payload)) throw new AgentGatewayError(400, 'Invalid chat request.');
@@ -195,25 +223,48 @@ export function createHermesAgentService({
       if (urgent) return { reply: urgentChatMessage, connected: true, safetyOverride: true };
       const cutoff = now() - RATE_WINDOW_MS;
       session.requests = session.requests.filter(time => time > cutoff);
-      if (session.requests.length >= config.maxRequestsPerMinute || session.busy) throw new AgentGatewayError(429, 'Timmy is already thinking. Try again shortly.');
+      if (session.requests.length >= config.maxRequestsPerMinute) throw new AgentGatewayError(429, 'Too many messages in a row. Try again shortly.');
       session.requests.push(now());
-      session.busy = true;
+      if (signal?.aborted) throw new AgentGatewayError(503, 'The browser closed the request before Timmy could think. Your journal still works.');
+      // Forward browser-side aborts into the inference queue by handle so a
+      // disconnected client releases its slot immediately.
+      let handle = null;
+      const forwardAbort = () => {
+        if (!handle) return;
+        const reason = signal.reason instanceof Error ? signal.reason : normalizeReason(signal.reason);
+        queue.cancel(handle, reason);
+      };
+      if (signal) signal.addEventListener('abort', forwardAbort, { once: true });
       try {
-        const result = await runTurn({
-          prompt: buildPrompt(message, ledger, !session.hermesSessionId),
-          hermesSessionId: session.hermesSessionId,
-          config,
-        });
+        const result = await (handle = queue.run(async runSignal => {
+          if (runSignal.aborted) throw normalizeReason(runSignal.reason);
+          // Session continuity is read at execution time so a request that
+          // waited in the queue resumes the conversation state left by the
+          // turn that ran before it.
+          return runTurn({
+            prompt: buildPrompt(message, ledger, !session.hermesSessionId),
+            hermesSessionId: session.hermesSessionId,
+            config,
+            signal: runSignal,
+          });
+        }));
         const reply = stripUnsafeControls(result?.reply || '');
         if (!reply || reply.length > 12_000 || !/^[A-Za-z0-9_-]{8,128}$/.test(String(result?.sessionId || ''))) throw new Error('invalid agent result');
-        session.hermesSessionId = result.sessionId;
+        if (!session.hermesSessionId) session.hermesSessionId = result.sessionId;
         return { reply, connected: true };
       } catch (error) {
         if (error instanceof AgentGatewayError) throw error;
+        const messageText = String(error?.message || '');
+        if (error instanceof OverloadError || /timed out|busy right now|disconnected|closed before/i.test(messageText)) {
+          throw new AgentGatewayError(503, messageText);
+        }
         throw new AgentGatewayError(503, 'Hermes is temporarily unavailable. Your local journal still works.');
-      } finally {
-        session.busy = false;
       }
     },
   };
 }
+
+function normalizeReason(reason) {
+  if (reason instanceof Error) return reason;
+  return new Error(typeof reason === 'string' && reason.trim() ? reason : 'Request cancelled.');
+}

src/inference-queue.js

@@ -0,0 +1,130 @@
+export class OverloadError extends Error {
+  constructor(message = 'Timmy is busy right now. Your journal still works — try again shortly or continue manually.') {
+    super(message);
+    this.name = 'OverloadError';
+  }
+}
+
+function normalizeReason(reason) {
+  if (reason instanceof Error) return reason;
+  const text = String(reason ?? '').trim().slice(0, 200);
+  return new Error(text || 'Request cancelled.');
+}
+
+function expiredMessage() {
+  return 'Timmy is busy right now and your request timed out waiting. Your journal still works — try again shortly or continue manually.';
+}
+
+export function createInferenceQueue({
+  concurrency = 1,
+  maxQueueDepth = 0,
+  requestTimeoutMs = 60_000,
+  now = () => Date.now(),
+} = {}) {
+  let active = 0;
+  const waiting = [];
+  const handles = new Map();
+
+  function expireStale(nowMs) {
+    while (waiting.length) {
+      if (nowMs - waiting[0].enqueuedAt <= requestTimeoutMs) break;
+      const expired = waiting.shift();
+      expired.settled = true;
+      expired.abort(new OverloadError(expiredMessage()));
+    }
+  }
+
+  function pump(nowMs = now()) {
+    expireStale(nowMs);
+    while (active < concurrency && waiting.length) {
+      const entry = waiting.shift();
+      if (entry.settled) continue;
+      active += 1;
+      entry.grant();
+    }
+  }
+
+  function release() {
+    active -= 1;
+    pump();
+  }
+
+  async function admit(record) {
+    const { controller } = record;
+    if (controller.signal.aborted) throw normalizeReason(controller.signal.reason);
+    expireStale(now());
+    if (active < concurrency) {
+      active += 1;
+      return;
+    }
+    if (waiting.length >= maxQueueDepth) throw new OverloadError();
+    await new Promise((resolve, reject) => {
+      const entry = {
+        enqueuedAt: now(),
+        settled: false,
+        grant: resolve,
+        abort: reject,
+      };
+      record.entry = entry;
+      waiting.push(entry);
+    });
+    record.entry = null;
+  }
+
+  function run(task) {
+    const controller = new AbortController();
+    const record = { controller, entry: null };
+    let resolveOutcome;
+    let rejectOutcome;
+    // Deliberately not an async function: the caller must receive the very
+    // promise registered in `handles`, or cancellation lookups would target
+    // a different object than the one they hold.
+    const promise = new Promise((resolve, reject) => {
+      resolveOutcome = resolve;
+      rejectOutcome = reject;
+    });
+    handles.set(promise, record);
+    (async () => {
+      let holdsSlot = false;
+      try {
+        await admit(record);
+        holdsSlot = true;
+        // The task body may not have started even though a slot is held; an
+        // external cancel that raced ahead must still stop it here.
+        if (controller.signal.aborted) throw normalizeReason(controller.signal.reason);
+        resolveOutcome(await task(controller.signal));
+      } catch (error) {
+        rejectOutcome(error instanceof Error ? error : normalizeReason(error));
+      } finally {
+        if (holdsSlot) release();
+      }
+      // Drop the handle one microtask after the caller-visible promise
+      // settles, so a same-tick external cancel still finds it.
+      promise.then(
+        () => handles.delete(promise),
+        () => handles.delete(promise),
+      );
+    })();
+    promise.catch(() => {});
+    return promise;
+  }
+
+  function cancel(handle, reason) {
+    const record = handles.get(handle);
+    if (!record) return false;
+    handles.delete(handle);
+    const normalized = normalizeReason(reason);
+    const entry = record.entry;
+    if (entry && !entry.settled) {
+      entry.settled = true;
+      const index = waiting.indexOf(entry);
+      if (index >= 0) waiting.splice(index, 1);
+      entry.abort(normalized);
+    }
+    record.controller.abort(normalized);
+    return true;
+  }
+
+  pump();
+  return { run, cancel };
+}

src/vision-service.js

@@ -7,18 +7,22 @@ function providerEndpoint(baseUrl) {
   return `${url.toString().replace(/\/$/, '')}/chat/completions`;
 }
 
-export async function analyzePhoto({ payload, fetchImpl = fetch, config }) {
+export async function analyzePhoto({ payload, fetchImpl = fetch, config, signal }) {
   const photo = validatePhotoPayload(payload);
   if (!config?.model) throw new Error('AI analysis is not configured.');
   const endpoint = providerEndpoint(config.baseUrl);
+  // The provider call honours both its own deadline and a per-request
+  // cancellation so abandoned clients stop consuming inference capacity.
+  const timeoutSignal = AbortSignal.timeout(config.requestTimeoutMs || 60_000);
+  const combinedSignal = signal ? AbortSignal.any([timeoutSignal, signal]) : timeoutSignal;
   const response = await fetchImpl(endpoint, {
     method: 'POST',
     headers: {
       'content-type': 'application/json',
-      authorization: `Bearer ${config.apiKey || 'local-proxy'}`,
+      authorization: 'Bearer ' + String(config.apiKey || 'local-proxy'),
     },
     body: JSON.stringify(buildVisionRequest({ imageDataUrl: photo.imageDataUrl, model: config.model })),
-    signal: AbortSignal.timeout(config.requestTimeoutMs || 60_000),
+    signal: combinedSignal,
   }).catch(() => { throw new Error('AI analysis is temporarily unavailable. Continue manually.'); });
   if (!response.ok) throw new Error('AI analysis is temporarily unavailable. Continue manually.');
   let data;

tests/agent-queue.test.js

@@ -0,0 +1,92 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+import { AgentGatewayError, createHermesAgentService, resolveHermesAgentConfig } from '../src/hermes-agent-service.js';
+
+const origin = 'http://127.0.0.1:4173';
+const accessCode = 'test-agent-access-code-2026';
+const configured = () => resolveHermesAgentConfig({
+  TIMMY_AGENT_ENABLED: 'true',
+  TIMMY_AGENT_ACCESS_TOKEN: accessCode,
+  TIMMY_PUBLIC_ORIGIN: origin,
+  TIMMY_AGENT_WORKDIR: '/tmp/timmy-agent-workspace',
+});
+
+function deferred() {
+  let resolve;
+  let reject;
+  const promise = new Promise((res, rej) => { resolve = res; reject = rej; });
+  return { promise, resolve, reject };
+}
+
+async function unlock(service, token) {
+  await service.unlock({ origin, accessCode });
+  return token;
+}
+
+test('chat requests beyond the queue depth get a stable sanitized overload fallback', async () => {
+  let clock = 0;
+  const gate = deferred();
+  let started = 0;
+  const service = createHermesAgentService({
+    config: { ...configured(), maxQueueDepth: 1 },
+    randomToken: () => 'queue-cookie',
+    now: () => clock,
+    runTurn: async () => {
+      started += 1;
+      return gate.promise;
+    },
+  });
+  await unlock(service);
+
+  const first = service.chat({ origin, cookieToken: 'queue-cookie', payload: { message: 'first', ledger: [] } });
+  await new Promise(resolve => setImmediate(resolve));
+  assert.equal(started, 1);
+
+  const second = service.chat({ origin, cookieToken: 'queue-cookie', payload: { message: 'second', ledger: [] } });
+  await new Promise(resolve => setImmediate(resolve));
+  assert.equal(started, 1, 'second request waits in the bounded queue');
+
+  await assert.rejects(() => service.chat({ origin, cookieToken: 'queue-cookie', payload: { message: 'third', ledger: [] } }), error => {
+    assert.equal(error.status, 503);
+    assert.match(error.message, /busy|try again/i);
+    assert.doesNotMatch(error.message, /spawn|hermes|session|token|workdir/i);
+    return true;
+  }, 'request beyond max queue depth fails fast with sanitized copy');
+
+  gate.resolve({ reply: 'eventually', sessionId: 'private-session' });
+  assert.match((await first).reply, /eventually/);
+  assert.match((await second).reply, /eventually/);
+});
+
+test('a chat request abandoned by its browser is cancelled without running Hermes and frees its slot', async () => {
+  let clock = 0;
+  const gate = deferred();
+  const started = [];
+  const service = createHermesAgentService({
+    config: { ...configured(), maxQueueDepth: 2 },
+    randomToken: () => 'disconnect-cookie',
+    now: () => clock,
+    runTurn: async ({ signal }) => {
+      started.push('run');
+      return new Promise((resolve, reject) => {
+        signal.addEventListener('abort', () => reject(new Error('aborted')));
+        gate.promise.then(resolve, reject);
+      });
+    },
+  });
+  await unlock(service, 'disconnect-cookie');
+
+  const active = service.chat({ origin, cookieToken: 'disconnect-cookie', payload: { message: 'holding slot', ledger: [] } });
+  await new Promise(resolve => setImmediate(resolve));
+  const abandonSignal = new AbortController();
+  const abandoned = service.chat({ origin, cookieToken: 'disconnect-cookie', payload: { message: 'queued', ledger: [] }, signal: abandonSignal.signal });
+  await new Promise(resolve => setImmediate(resolve));
+
+  abandonSignal.abort(new Error('client disconnected'));
+  await assert.rejects(() => abandoned, error => error.status === 503 && /closed|disconnect/i.test(error.message));
+
+  gate.resolve({ reply: 'late answer', sessionId: 'private-session' });
+  assert.match((await active).reply, /late answer/);
+  assert.deepEqual(started, ['run'], 'the cancelled request never invoked a Hermes turn');
+});

tests/agent-turn-timeout.test.js

@@ -0,0 +1,109 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+import { chmod, mkdir, rm, writeFile } from 'node:fs/promises';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { constants as fsConstants } from 'node:fs';
+
+import { AgentGatewayError, createHermesAgentService, resolveHermesAgentConfig, runHermesCliTurn } from '../src/hermes-agent-service.js';
+
+const origin = 'http://127.0.0.1:4173';
+const accessCode = 'test-agent-access-code-2026';
+const configured = () => resolveHermesAgentConfig({
+  TIMMY_AGENT_ENABLED: 'true',
+  TIMMY_AGENT_ACCESS_TOKEN: accessCode,
+  TIMMY_PUBLIC_ORIGIN: origin,
+  TIMMY_AGENT_WORKDIR: '/tmp/timmy-agent-workspace',
+});
+
+async function waitUntil(check, timeoutMs = 3_000, stepMs = 20) {
+  const deadline = Date.now() + timeoutMs;
+  while (Date.now() < deadline) {
+    if (await check()) return true;
+    await new Promise(resolve => setTimeout(resolve, stepMs));
+  }
+  return false;
+}
+
+function isAlive(pid) {
+  try { process.kill(pid, 0); return true; } catch (error) { return error.code === 'EPERM'; }
+}
+
+test('aborting an in-flight Hermes turn kills the child promptly and leaves no orphan process', async () => {
+  const dir = await mkdir(join(tmpdir(), `timmy-orphan-${Date.now()}-`), { recursive: true });
+  const pidFile = join(dir, 'child.pid');
+  const fixture = join(dir, 'slow-hermes.sh');
+  // Prints valid Hermes output immediately, then idles long enough that only
+  // an explicit kill can end it. `exec` makes the recorded PID the sleeper.
+  await writeFile(fixture, `#!/usr/bin/env bash\nprintf '%s\\n' $$ > '${pidFile}'\necho "session_id: fixture_kill_001"\necho "thinking"\nexec sleep 60\n`, { mode: fsConstants.S_IRWXU });
+
+  const controller = new AbortController();
+  const turn = runHermesCliTurn({
+    prompt: 'hello',
+    hermesSessionId: null,
+    config: { ...configured(), command: fixture, workdir: dir },
+    signal: controller.signal,
+    env: { TMPDIR: dir },
+  });
+
+  const pidAppeared = await waitUntil(async () => {
+    try { Number(await (await import('node:fs/promises')).readFile(pidFile, 'utf8')); return true; } catch { return false; }
+  }, 4_000);
+  assert.ok(pidAppeared, 'fixture child started');
+  const pid = Number(await (await import('node:fs/promises')).readFile(pidFile, 'utf8'));
+  assert.ok(isAlive(pid), 'child is running before cancellation');
+
+  const startedAt = Date.now();
+  controller.abort(new Error('client disconnected'));
+  await assert.rejects(() => Promise.race([
+    turn,
+    new Promise((_, reject) => setTimeout(() => reject(new Error('turn was not cancelled')), 5_000)),
+  ]), error => /disconnect|abort|killed|terminated/i.test(String(error?.message || error?.code || '')));
+  const cancelMs = Date.now() - startedAt;
+  assert.ok(cancelMs < 4_000, `cancellation returned promptly (took ${cancelMs}ms)`);
+
+  const reaped = await waitUntil(() => !isAlive(pid), 4_000);
+  assert.ok(reaped, 'child process was actually killed — no orphan subprocess remains');
+  await rm(dir, { recursive: true, force: true });
+}, { timeout: 20_000 });
+
+test('a queued chat past the queue deadline is cleaned up with a sanitized timeout response and zero extra Hermes calls', async () => {
+  let clock = 0;
+  let started = 0;
+  let releaseActive;
+  const service = createHermesAgentService({
+    config: { ...configured(), timeoutMs: 50 },
+    randomToken: () => 'deadline-cookie',
+    now: () => clock,
+    runTurn: async () => {
+      started += 1;
+      return new Promise(resolve => { releaseActive = resolve; });
+    },
+  });
+  await service.unlock({ origin, accessCode });
+
+  const active = service.chat({ origin, cookieToken: 'deadline-cookie', payload: { message: 'wedge', ledger: [] } });
+  await new Promise(resolve => setImmediate(resolve));
+  assert.equal(started, 1);
+
+  const queued = service.chat({ origin, cookieToken: 'deadline-cookie', payload: { message: 'waiting', ledger: [] } });
+  await new Promise(resolve => setImmediate(resolve));
+  assert.equal(started, 1);
+
+  clock += 200; // past the 50ms queue deadline
+  // Attach the rejection expectation before the sweep runs so the expired
+  // request never counts as an unhandled rejection.
+  const expiryExpectation = assert.rejects(() => queued, error => {
+    assert.ok(error instanceof AgentGatewayError);
+    assert.equal(error.status, 503);
+    assert.match(error.message, /timed out|busy/i);
+    assert.doesNotMatch(error.message, /spawn|hermes|session|token|workdir|auth/i);
+    return true;
+  });
+  service.chat({ origin, cookieToken: 'deadline-cookie', payload: { message: 'sweep trigger', ledger: [] } }).catch(() => {});
+  await expiryExpectation;
+  assert.equal(started, 1, 'expired request never reached Hermes');
+  releaseActive({ reply: 'late but valid', sessionId: 'deadline_session_01' });
+  assert.match((await active).reply, /late but valid/);
+});

tests/fixtures/slow-hermes.mjs

@@ -0,0 +1,20 @@
+#!/usr/bin/env node
+// Slow Hermes fixture for bounded-queue acceptance tests. Normal prompts
+// answer immediately; prompts containing HOLD block until a release file
+// appears inside the agent workdir (the value after `--in`), so tests can
+// saturate the queue deterministically without touching shared /tmp state.
+import { existsSync } from 'node:fs';
+import { appendFileSync } from 'node:fs';
+
+const args = process.argv.slice(2);
+const promptIndex = args.indexOf('-q');
+const prompt = promptIndex >= 0 ? args[promptIndex + 1] : '';
+const workdirIndex = args.indexOf('--in');
+const workdir = workdirIndex >= 0 ? args[workdirIndex + 1] : '/tmp';
+appendFileSync(`${workdir}/fixture-pids.log`, `${process.pid}\n`);
+process.stdout.write('session_id: slow_fixture_2026\n');
+if (/HOLD/.test(prompt)) {
+  const releaseFile = `${workdir}/hermes-release`;
+  while (!existsSync(releaseFile)) await new Promise(resolve => setTimeout(resolve, 20));
+}
+process.stdout.write('Hermes fixture answered the bounded request.\n');

tests/hermes-agent-service.test.js

@@ -193,16 +193,21 @@ test('Hermes child receives only an explicit non-secret environment allowlist',
   });
 });
 
-test('upstream failures and concurrency fail closed without leaking process detail', async () => {
+test('saturated turns queue within bounds and excess load fails closed without leaking process detail', async () => {
   let release;
   const runTurn = () => new Promise(resolve => { release = resolve; });
-  const service = createHermesAgentService({ config: configured(), randomToken: () => 'browser-cookie', runTurn });
+  const service = createHermesAgentService({ config: { ...configured(), maxQueueDepth: 1 }, randomToken: () => 'browser-cookie', runTurn });
   await service.unlock({ origin, accessCode: 'test-agent-access-code-2026' });
   const active = service.chat({ origin, cookieToken: 'browser-cookie', payload: { message: 'first', ledger: [] } });
   await new Promise(resolve => setImmediate(resolve));
-  await rejectsStatus(() => service.chat({ origin, cookieToken: 'browser-cookie', payload: { message: 'second', ledger: [] } }), 429);
-  release({ reply: 'done', sessionId: 'private-session' });
-  await active;
+  const queued = service.chat({ origin, cookieToken: 'browser-cookie', payload: { message: 'second', ledger: [] } });
+  await new Promise(resolve => setImmediate(resolve));
+  await rejectsStatus(() => service.chat({ origin, cookieToken: 'browser-cookie', payload: { message: 'third', ledger: [] } }), 503);
+  release({ reply: 'first done', sessionId: 'private-session' });
+  assert.match((await active).reply, /first done/);
+  await new Promise(resolve => setImmediate(resolve));
+  release({ reply: 'queued done', sessionId: 'private-session' });
+  assert.match((await queued).reply, /queued done/);
 
   const broken = createHermesAgentService({ config: configured(), randomToken: () => 'broken-cookie', runTurn: async () => { throw new Error('spawn /root/.hermes/auth.json SECRET'); } });
   await broken.unlock({ origin, accessCode: 'test-agent-access-code-2026' });

tests/inference-queue-cancellation.test.js

@@ -0,0 +1,96 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+import { OverloadError, createInferenceQueue } from '../src/inference-queue.js';
+
+function deferred() {
+  let resolve;
+  let reject;
+  const promise = new Promise((res, rej) => { resolve = res; reject = rej; });
+  return { promise, resolve, reject };
+}
+
+test('a queued request past its deadline is expired deterministically and its place is reusable', async () => {
+  let clock = 0;
+  const queue = createInferenceQueue({ concurrency: 1, maxQueueDepth: 1, requestTimeoutMs: 100, now: () => clock });
+
+  const wedgeGate = deferred();
+  const active = queue.run(signal => {
+    return new Promise((resolve, reject) => {
+      signal.addEventListener('abort', () => reject(new Error('upstream aborted')));
+      wedgeGate.promise.then(resolve, reject);
+    });
+  });
+  active.catch(() => {});
+  await Promise.resolve();
+
+  let started = false;
+  const queued = queue.run(() => { started = true; return 'never'; });
+  await Promise.resolve();
+
+  clock += 150; // deadline (submitted at t=0, limit 100) has elapsed
+  const probe = queue.run(() => 'ok'); // queue activity triggers the deadline sweep
+
+  await assert.rejects(() => queued, error => {
+    assert.ok(error instanceof OverloadError);
+    assert.match(error.message, /timed out|busy/i);
+    return true;
+  });
+  assert.equal(started, false, 'the expired request must never reach the provider');
+
+  // Capacity is intact: cancelling the wedged holder lets the surviving request through.
+  queue.cancel(active, 'cleanup');
+  assert.equal(await Promise.race([probe, new Promise((_, reject) => setTimeout(() => reject(new Error('probe starved')), 200))]), 'ok');
+});
+
+test('client disconnect cancels a queued request before it reaches the provider', async () => {
+  let clock = 0;
+  const queue = createInferenceQueue({ concurrency: 1, maxQueueDepth: 2, requestTimeoutMs: 60_000, now: () => clock });
+
+  const wedgeGate = deferred();
+  const active = queue.run(() => wedgeGate.promise);
+  active.catch(() => {});
+  await Promise.resolve();
+
+  let sawStart = false;
+  const queued = queue.run(() => { sawStart = true; return 'too late'; });
+  await Promise.resolve();
+
+  queue.cancel(queued, 'client disconnected');
+  await assert.rejects(() => queued, error => error.message === 'client disconnected');
+  assert.equal(sawStart, false, 'cancelled request never invoked its task');
+
+  const admitted = queue.run(() => 'admitted-after-cancel');
+  wedgeGate.resolve('wedge-done');
+  assert.equal(await admitted, 'admitted-after-cancel');
+  assert.equal(await active, 'wedge-done');
+});
+
+test('cancelling an already-running task aborts its signal so buffers can be released', async () => {
+  let clock = 0;
+  const queue = createInferenceQueue({ concurrency: 1, maxQueueDepth: 1, requestTimeoutMs: 60_000, now: () => clock });
+
+  const observed = [];
+  const running = queue.run(async signal => {
+    observed.push(signal);
+    signal.addEventListener('abort', () => observed.push(`aborted:${signal.reason?.message ?? signal.reason}`));
+    await new Promise((resolve, reject) => {
+      signal.addEventListener('abort', () => reject(new Error('task stopped')));
+      setTimeout(resolve, 5_000);
+    });
+  });
+
+  // Let the wrapper admit the request and actually invoke the task body.
+  await Promise.resolve();
+  await Promise.resolve();
+  assert.equal(observed.length, 1, 'task is genuinely running before cancellation');
+
+  queue.cancel(running, 'stop');
+  await assert.rejects(() => running, /task stopped/);
+  assert.ok(observed[0] instanceof AbortSignal, 'task received a real AbortSignal');
+  assert.match(String(observed[1]), /stop/, 'running task observed the abort with its reason');
+
+  // The slot was released by the cancelled task, so the next request starts immediately.
+  const next = queue.run(() => 'next-runs');
+  assert.equal(await next, 'next-runs');
+});

tests/inference-queue.acceptance.mjs

@@ -0,0 +1,136 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { chmod, mkdtemp, rm, readFile, writeFile } from 'node:fs/promises';
+import { spawn } from 'node:child_process';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { constants as fsConstants } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+
+const root = fileURLToPath(new URL('..', import.meta.url));
+const fixture = fileURLToPath(new URL('./fixtures/slow-hermes.mjs', import.meta.url));
+const origin = 'http://127.0.0.1:4187';
+const accessCode = 'queue-acceptance-code-2026';
+
+async function waitReady(child) {
+  const deadline = Date.now() + 10_000;
+  while (Date.now() < deadline) {
+    if (child.exitCode !== null) throw new Error(`server exited ${child.exitCode}`);
+    try { const response = await fetch(`${origin}/api/healthz`); if (response.ok) return; } catch {}
+    await new Promise(resolve => setTimeout(resolve, 50));
+  }
+  throw new Error('server did not become ready');
+}
+
+function post(path, body, { cookie = '', signal } = {}) {
+  return fetch(`${origin}${path}`, {
+    method: 'POST',
+    headers: {
+      'content-type': 'application/json',
+      origin,
+      'sec-fetch-site': 'same-origin',
+      ...(cookie ? { cookie } : {}),
+    },
+    body: JSON.stringify(body),
+    signal,
+  }).then(
+    async response => ({ status: response.status, body: await response.json().catch(() => ({})) }),
+    error => ({ failed: true, reason: String(error?.name || error) }),
+  );
+}
+
+test('bounded queue end-to-end: disconnect cleanup, overload fallback, urgent bypass, recovery', async t => {
+  const workdir = await mkdtemp(join(tmpdir(), 'timmy-queue-accept-'));
+  const releaseFile = join(workdir, 'hermes-release');
+  const pidLog = join(workdir, 'fixture-pids.log');
+  await chmod(fixture, fsConstants.S_IRWXU);
+  const child = spawn(process.execPath, ['server.mjs'], {
+    cwd: root,
+    env: {
+      ...process.env,
+      PORT: '4187',
+      TIMMY_AGENT_ENABLED: 'true',
+      TIMMY_AGENT_ACCESS_TOKEN: accessCode,
+      TIMMY_PUBLIC_ORIGIN: origin,
+      TIMMY_AGENT_WORKDIR: workdir,
+      TIMMY_HERMES_COMMAND: fixture,
+      TIMMY_AGENT_MAX_CONCURRENT_TURNS: '1',
+      TIMMY_AGENT_MAX_QUEUE_DEPTH: '2',
+    },
+    stdio: ['ignore', 'pipe', 'pipe'],
+  });
+  let serverLogs = '';
+  child.stderr.on('data', chunk => { serverLogs += String(chunk); });
+  t.after(async () => {
+    child.kill('SIGTERM');
+    await rm(workdir, { recursive: true, force: true }).catch(() => {});
+    await rm(releaseFile, { force: true }).catch(() => {});
+  });
+  await waitReady(child);
+
+  // Authenticate once and keep the cookie for the whole scenario.
+  const rawResponse = await fetch(`${origin}/api/agent/unlock`, {
+    method: 'POST',
+    headers: { 'content-type': 'application/json', origin, 'sec-fetch-site': 'same-origin' },
+    body: JSON.stringify({ accessCode }),
+  });
+  assert.equal(rawResponse.status, 200);
+  const browserCookie = rawResponse.headers.get('set-cookie').split(';', 1)[0];
+
+  // Fill the slot: this HOLD runs until the fixture release file appears.
+  const heldOne = post('/api/agent/chat', { message: 'HOLD one', ledger: [] }, { cookie: browserCookie });
+  await new Promise(resolve => setTimeout(resolve, 400));
+
+  // 1. Disconnect cleanup: this request waits in the queue. When its client
+  //    vanishes, Timmy must drop it and free the queued place.
+  const disconnectController = new AbortController();
+  const doomed = post('/api/agent/chat', { message: 'HOLD doomed', ledger: [] }, { cookie: browserCookie, signal: disconnectController.signal });
+  await new Promise(resolve => setTimeout(resolve, 400));
+  disconnectController.abort();
+  const doomedOutcome = await doomed;
+  assert.equal(doomedOutcome.failed, true, 'abandoned client saw its request cancelled');
+
+  // If the freed place was reclaimed, both follow-up HOLDs are admitted
+  // (slot + queue of two minus one freed place). A sanitized 503 here would
+  // mean the cancelled client did not release capacity.
+  const heldTwo = post('/api/agent/chat', { message: 'HOLD two', ledger: [] }, { cookie: browserCookie });
+  const heldThree = post('/api/agent/chat', { message: 'HOLD three', ledger: [] }, { cookie: browserCookie });
+  await new Promise(resolve => setTimeout(resolve, 500));
+
+  // 2. Overload: everything is saturated now, so a further request must fail
+  //    fast with the stable manual-fallback copy and zero provider contact.
+  const overloadResult = await post('/api/agent/chat', { message: 'overload probe', ledger: [] }, { cookie: browserCookie });
+  assert.equal(overloadResult.status, 503);
+  assert.match(overloadResult.body.error, /busy|try again|manually/i);
+  assert.doesNotMatch(overloadResult.body.error, /spawn|hermes|session|token|workdir|fixture/i);
+
+  // 3. Urgent bypass while saturated: deterministic safety answer, zero calls.
+  const urgentResult = await post('/api/agent/chat', { message: 'I have rectal bleeding', ledger: [] }, { cookie: browserCookie });
+  assert.equal(urgentResult.status, 200);
+  assert.equal(urgentResult.body.safetyOverride, true);
+  assert.match(urgentResult.body.reply, /medical help/i);
+
+  // 4. Recovery: releasing the fixture lets every surviving request complete.
+  await writeFile(releaseFile, 'go');
+  for (const [label, promise] of [['one', heldOne], ['two', heldTwo], ['three', heldThree]]) {
+    const settled = await Promise.race([
+      promise,
+      new Promise(resolve => setTimeout(() => resolve({ status: 'TEST-TIMEOUT' }), 12_000)),
+    ]);
+    assert.equal(settled.status, 200, `${label} completed after recovery`);
+    assert.match(settled.body.reply, /bounded request/);
+    assert.doesNotMatch(JSON.stringify(settled.body), /slow_fixture_2026|session_id/);
+  }
+
+  // 5. No orphan subprocesses: every fixture PID recorded at spawn is gone.
+  await new Promise(resolve => setTimeout(resolve, 400));
+  const recordedPids = ((await readFile(pidLog, 'utf8').catch(() => '')) || '')
+    .split('\n').map(Number).filter(Boolean);
+  assert.equal(recordedPids.length, 3, `exactly three turns ran (saw ${recordedPids.length})`);
+  for (const pid of [...new Set(recordedPids)]) {
+    let alive = true;
+    try { process.kill(pid, 0); } catch { alive = false; }
+    assert.equal(alive, false, `fixture process ${pid} outlived its turn`);
+  }
+  assert.doesNotMatch(serverLogs, /access-code|secret|token/i);
+}, { timeout: 40_000 });

tests/inference-queue.test.js

@@ -0,0 +1,72 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+import { OverloadError, createInferenceQueue } from '../src/inference-queue.js';
+
+function deferred() {
+  let resolve;
+  let reject;
+  const promise = new Promise((res, rej) => { resolve = res; reject = rej; });
+  return { promise, resolve, reject };
+}
+
+test('vision analysis runs with bounded concurrency and strict FIFO ordering', async () => {
+  let clock = 0;
+  const gates = new Map();
+  const started = [];
+  const queue = createInferenceQueue({ concurrency: 2, maxQueueDepth: 4, requestTimeoutMs: 10_000, now: () => clock });
+
+  const run = label => queue.run(signal => {
+    started.push(label);
+    const gate = deferred();
+    gates.set(label, gate);
+    gate.promise.catch(() => {});
+    return gate.promise;
+  });
+
+  const first = run('first');
+  const second = run('second');
+  const third = run('third');
+  const fourth = run('fourth');
+  while (started.length < 2) await Promise.resolve();
+  assert.deepEqual(started, ['first', 'second'], 'only concurrency slots start immediately');
+
+  gates.get('first').resolve({ ok: 'first' });
+  assert.equal((await first).ok, 'first');
+
+  // The slot freed by `first` must be granted to `third` (head of the queue).
+  while (!gates.has('third')) await Promise.resolve();
+  gates.get('third').resolve({ ok: 'third' });
+
+  // Only after `third` releases its slot may `fourth` start.
+  while (!gates.has('fourth')) await Promise.resolve();
+  gates.get('second').resolve({ ok: 'second' });
+  gates.get('fourth').resolve({ ok: 'fourth' });
+
+  assert.deepEqual(await Promise.all([second, third, fourth]), [{ ok: 'second' }, { ok: 'third' }, { ok: 'fourth' }]);
+  assert.equal(started.length, 4, 'every admitted request eventually started');
+  assert.deepEqual([...new Set(started)], ['first', 'second', 'third', 'fourth'], 'each queued request started exactly once');
+  assert.equal(started.indexOf('third') < started.indexOf('fourth'), true, 'queued requests were granted in FIFO order');
+});
+
+test('overloaded submissions fail fast with a stable sanitized manual-fallback error', async () => {
+  let clock = 0;
+  const gates = [];
+  const queue = createInferenceQueue({ concurrency: 1, maxQueueDepth: 1, requestTimeoutMs: 10_000, now: () => clock });
+
+  const hold = queue.run(() => { const gate = deferred(); gates.push(gate); gate.promise.catch(() => {}); return gate.promise; });
+  await Promise.resolve();
+  const held = queue.run(() => new Promise(() => {}));
+  await Promise.resolve();
+
+  await assert.rejects(() => queue.run(() => new Promise(() => {})), error => {
+    assert.ok(error instanceof OverloadError);
+    assert.match(error.message, /busy|full/i);
+    assert.doesNotMatch(error.message, /stack|internal|fetch|127\.0\.0\.1/i);
+    return true;
+  }, 'excess request beyond maxQueueDepth is rejected immediately');
+
+  // Active and already-queued work continues unaffected by the rejected submission.
+  gates[0].resolve('done');
+  assert.equal(await hold, 'done');
+});

tests/inference-zero-call.test.js

@@ -0,0 +1,108 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+
+import { createHermesAgentService, resolveHermesAgentConfig } from '../src/hermes-agent-service.js';
+import { analyzePhoto } from '../src/vision-service.js';
+import { OverloadError, createInferenceQueue } from '../src/inference-queue.js';
+
+const origin = 'http://127.0.0.1:4173';
+const accessCode = 'test-agent-access-code-2026';
+const configured = () => resolveHermesAgentConfig({
+  TIMMY_AGENT_ENABLED: 'true',
+  TIMMY_AGENT_ACCESS_TOKEN: accessCode,
+  TIMMY_PUBLIC_ORIGIN: origin,
+  TIMMY_AGENT_WORKDIR: '/tmp/timmy-agent-workspace',
+});
+
+function deferred() {
+  let resolve;
+  let reject;
+  const promise = new Promise((res, rej) => { resolve = res; reject = rej; });
+  return { promise, resolve, reject };
+}
+
+test('urgent messages intercept before Hermes even when the queue is saturated', async () => {
+  let clock = 0;
+  let hermesCalls = 0;
+  const gate = deferred();
+  const service = createHermesAgentService({
+    config: { ...configured(), maxQueueDepth: 0 },
+    randomToken: () => 'urgent-cookie',
+    now: () => clock,
+    runTurn: async () => {
+      hermesCalls += 1;
+      return gate.promise;
+    },
+  });
+  await service.unlock({ origin, accessCode });
+
+  // Saturate the single slot so any queued path would be rejected as overloaded.
+  const active = service.chat({ origin, cookieToken: 'urgent-cookie', payload: { message: 'normal question', ledger: [] } });
+  await new Promise(resolve => setImmediate(resolve));
+  assert.equal(hermesCalls, 1);
+
+  const urgent = await service.chat({ origin, cookieToken: 'urgent-cookie', payload: { message: 'I have rectal bleeding', ledger: [] } });
+
+  assert.equal(hermesCalls, 1, 'Hermes was called exactly once — only for the non-urgent request');
+  assert.equal(urgent.safetyOverride, true);
+  assert.match(urgent.reply, /medical help/i);
+  gate.resolve({ reply: 'late', sessionId: 'urgent_session_01' });
+  assert.match((await active).reply, /late/);
+});
+
+test('vision analysis is bounded by the shared queue and overload returns the manual-fallback copy', async () => {
+  let clock = 0;
+  let providerCalls = 0;
+  const gates = [];
+  const queue = createInferenceQueue({ concurrency: 1, maxQueueDepth: 1, requestTimeoutMs: 60_000, now: () => clock });
+
+  const visionFetch = async (url, options) => {
+    providerCalls += 1;
+    const gate = deferred();
+    gates.push(gate);
+    options.signal?.addEventListener('abort', () => gate.reject(new Error('aborted')));
+    return new Promise((resolve, reject) => {
+      gate.promise.then(() => resolve({ ok: true, json: async () => ({ choices: [{ message: { content: JSON.stringify({ isStool: true, bristolType: 4, color: 'brown', confidence: 0.8, imageQuality: 'good', observations: 'ok' }) } }] }) }), reject);
+    });
+  };
+
+  const payload = { imageDataUrl: 'data:image/jpeg;base64,YQ==', consent: true };
+  const config = { baseUrl: 'http://127.0.0.1:8645/v1', apiKey: 'test-key-not-real', model: 'm' };
+
+  const first = queue.run(signal => analyzePhoto({ payload, fetchImpl: visionFetch, config, signal }));
+  while (providerCalls < 1) await Promise.resolve();
+  const second = queue.run(signal => analyzePhoto({ payload, fetchImpl: visionFetch, config, signal }));
+  await Promise.resolve();
+
+  await assert.rejects(() => queue.run((signal) => analyzePhoto({ payload, fetchImpl: visionFetch, config, signal })), error => {
+    assert.ok(error instanceof OverloadError);
+    assert.match(error.message, /continue manually/i);
+    assert.doesNotMatch(error.message, /api|key|token|127\.0\.0\.1|fetch/i);
+    return true;
+  }, 'third request beyond depth is overloaded with sanitized copy');
+
+  assert.equal(providerCalls, 1, 'overloaded request never reached the provider');
+  gates[0].resolve('go');
+  const result = await Promise.race([first, new Promise(resolve => setTimeout(() => resolve('pending'), 100))]);
+  assert.notEqual(result, 'pending');
+});
+
+test('vision requests carry cancellation so disconnected clients release provider work', async () => {
+  let capturedSignal;
+  const controller = new AbortController();
+  const gate = deferred();
+  const fetchImpl = async (url, options) => {
+    capturedSignal = options.signal;
+    options.signal?.addEventListener('abort', () => gate.reject(new Error('aborted')));
+    return gate.promise;
+  };
+  const payload = { imageDataUrl: 'data:image/jpeg;base64,YQ==', consent: true };
+  const config = { baseUrl: 'http://127.0.0.1:8645/v1', apiKey: 'test-key-not-real', model: 'm' };
+
+  const attempt = analyzePhoto({ payload, fetchImpl, config, signal: controller.signal });
+  while (!capturedSignal) await Promise.resolve();
+
+  controller.abort(new Error('client disconnected'));
+  await assert.rejects(() => attempt, /continue manually|unavailable/i);
+  assert.equal(capturedSignal.aborted, true, 'the provider fetch observed the cancellation');
+});

tests/server-shutdown.acceptance.mjs

@@ -0,0 +1,93 @@
+import test from 'node:test';
+import assert from 'node:assert/strict';
+import { chmod, mkdtemp, rm, readFile } from 'node:fs/promises';
+import { spawn } from 'node:child_process';
+import { tmpdir } from 'node:os';
+import { join } from 'node:path';
+import { constants as fsConstants } from 'node:fs';
+import { fileURLToPath } from 'node:url';
+
+const root = fileURLToPath(new URL('..', import.meta.url));
+const fixture = fileURLToPath(new URL('./fixtures/slow-hermes.mjs', import.meta.url));
+
+test('server shutdown cancels in-flight turns and leaves no orphan Hermes subprocesses', async () => {
+  const workdir = await mkdtemp(join(tmpdir(), 'timmy-shutdown-'));
+  await chmod(fixture, fsConstants.S_IRWXU);
+  const child = spawn(process.execPath, ['server.mjs'], {
+    cwd: root,
+    env: {
+      ...process.env,
+      HOST: '127.0.0.1',
+      PORT: '4188',
+      TIMMY_AGENT_ENABLED: 'true',
+      TIMMY_AGENT_ACCESS_TOKEN: 'shutdown-accept-code-2026',
+      TIMMY_PUBLIC_ORIGIN: 'http://127.0.0.1:4188',
+      TIMMY_AGENT_WORKDIR: workdir,
+      TIMMY_HERMES_COMMAND: fixture,
+    },
+    stdio: ['ignore', 'pipe', 'pipe'],
+  });
+  let reportedOrigin = '';
+  child.stdout.on('data', chunk => {
+    const match = String(chunk).match(/listening on http:\/\/([^:]+):(\d+)/);
+    if (match && !reportedOrigin) reportedOrigin = `http://${match[1]}:${match[2]}`;
+  });
+  const deadline = Date.now() + 10_000;
+  while (!reportedOrigin && Date.now() < deadline && child.exitCode === null) {
+    await new Promise(resolve => setTimeout(resolve, 50));
+  }
+  assert.ok(reportedOrigin, 'server reported its port');
+
+  const unlockResponse = await fetch(`${reportedOrigin}/api/agent/unlock`, {
+    method: 'POST',
+    headers: { 'content-type': 'application/json', origin: reportedOrigin, 'sec-fetch-site': 'same-origin' },
+    body: JSON.stringify({ accessCode: 'shutdown-accept-code-2026' }),
+  });
+  assert.equal(unlockResponse.status, 200);
+  const browserCookie = unlockResponse.headers.get('set-cookie').split(';', 1)[0];
+
+  // Start a turn wedged on the fixture's HOLD gate.
+  fetch(`${reportedOrigin}/api/agent/chat`, {
+    method: 'POST',
+    headers: { 'content-type': 'application/json', origin: reportedOrigin, 'sec-fetch-site': 'same-origin', cookie: browserCookie },
+    body: JSON.stringify({ message: 'HOLD shutdown probe', ledger: [] }),
+  }).catch(() => {});
+  await new Promise(resolve => setTimeout(resolve, 600));
+
+  const pidLog = join(workdir, 'fixture-pids.log');
+  // Wait until the fixture for this run's HOLD turn is actually alive.
+  let lastPid = null;
+  const spawnDeadline = Date.now() + 8_000;
+  while (Date.now() < spawnDeadline) {
+    const recorded = ((await readFile(pidLog, 'utf8').catch(() => '')) || '')
+      .split('\n').map(Number).filter(Boolean);
+    const candidate = recorded.at(-1);
+    if (candidate) {
+      let alive = true;
+      try { process.kill(candidate, 0); } catch { alive = false; }
+      if (alive) {
+        lastPid = candidate;
+        break;
+      }
+    }
+    await new Promise(resolve => setTimeout(resolve, 100));
+  }
+  assert.ok(lastPid, 'fixture subprocess was spawned for the in-flight turn');
+  let aliveBefore = true;
+  try { process.kill(lastPid, 0); } catch { aliveBefore = false; }
+  assert.ok(aliveBefore, `fixture ${lastPid} is running before shutdown`);
+
+  // Shut the server down while the turn is still running.
+  child.kill('SIGTERM');
+  await new Promise((resolve, reject) => {
+    const timeout = setTimeout(() => reject(new Error('server did not exit after SIGTERM')), 5_000);
+    child.on('exit', code => { clearTimeout(timeout); resolve(code); });
+  });
+
+  // The in-flight fixture must be gone — no orphan subprocess may survive.
+  await new Promise(resolve => setTimeout(resolve, 500));
+  let aliveAfter = true;
+  try { process.kill(lastPid, 0); } catch { aliveAfter = false; }
+  assert.equal(aliveAfter, false, `fixture process ${lastPid} survived server shutdown`);
+  await rm(workdir, { recursive: true, force: true }).catch(() => {});
+}, { timeout: 30_000 });