feat: add agent run queue controls

.env.example

@@ -26,6 +26,10 @@ H5_PUBLIC_BASE_URL=http://127.0.0.1:5173
 # VITE_AGENT_CODE_RUNS_ENABLED=0
 # VITE_AGENT_CODE_RUNS_AUTODETECT=0
 
+# Tool Gateway Queue v0（/agent/runs 后台任务队列，默认保守限流）。
+# MEMIND_AGENT_RUN_QUEUE_CONCURRENCY=1
+# MEMIND_AGENT_RUN_TIMEOUT_MS=900000
+
 # ===== memind_adm（独立后台服务）=====
 # 平台超管 /admin-api/* 与广场运营 /api/ops/v1/* 已从公网服务拆出，独立进程运行。
 # 登录仍在主域完成；后台凭 H5_COOKIE_DOMAIN=.tkmind.cn 共享的会话 Cookie 鉴权（生产务必设置）。

.runtime/portal/RUNBOOK.txt

@@ -46,3 +46,4 @@ Streaming runtime operations:
   node scripts/runtime-worker-drain.mjs undrain goosed-3
   node scripts/check-tool-runtime.mjs
   node scripts/check-agent-code-run-entry.mjs
+  curl -sk https://mm.tkmind.cn/api/runtime/status  # includes toolRuntime.queue

.runtime/portal/scripts/runtime-slo-report.mjs

@@ -148,6 +148,7 @@ function summarizeRuntime(runtimeJson) {
     routerEnabled: Boolean(runtimeJson?.router?.enabled),
     publicBaseUrl: runtimeJson?.publicBaseUrl ?? null,
     toolRuntime: runtimeJson?.toolRuntime ?? null,
+    toolQueue: runtimeJson?.toolRuntime?.queue ?? null,
     workers: workers.map((worker) => ({
       id: worker.id,
       healthy: runtimeJson?.targets?.find((target) => target.target === worker.target)?.healthy ?? null,
@@ -220,6 +221,10 @@ for (const worker of runtimeSummary?.workers ?? []) {
   if (worker.firstTokenCount > 0 && worker.ewmaFirstTokenMs <= 0) failures.push(`${worker.id}_first_token_ewma_missing`);
 }
 if (runtimeSummary?.toolRuntime?.chatInjectsCodeTools !== false) failures.push('chat_injects_code_tools');
+if (runtimeSummary?.toolQueue?.error) failures.push('tool_queue_status_error');
+if (runtimeSummary?.toolQueue?.inFlight > runtimeSummary?.toolQueue?.maxConcurrentRuns) {
+  failures.push('tool_queue_concurrency_exceeded');
+}
 
 const report = {
   ok: failures.length === 0,

.runtime/portal/server.mjs

@@ -6509,6 +6509,8 @@ var DEFAULT_RUN_RETRY_DELAYS_MS = [1500, 5e3, 15e3];
 var TERMINAL_STATUSES = /* @__PURE__ */ new Set(["succeeded", "failed"]);
 var CODE_TOOL_MODES = /* @__PURE__ */ new Set(["code", "code-task", "code_task", "code-tool", "code_tool", "code_tool_task"]);
 var RUN_METADATA_KEY = "memindRun";
+var DEFAULT_MAX_CONCURRENT_RUNS = 1;
+var DEFAULT_RUN_TIMEOUT_MS = 15 * 60 * 1e3;
 function nowMs() {
   return Date.now();
 }
@@ -6522,6 +6524,16 @@ function safeJsonParse(value, fallback = null) {
 function serializeMessage(message) {
   return JSON.stringify(message ?? {});
 }
+function positiveInteger(value, fallback) {
+  const n = Number(value);
+  if (!Number.isFinite(n) || n <= 0) return fallback;
+  return Math.floor(n);
+}
+function timeoutError(ms) {
+  const err = new Error(`agent run timed out after ${ms}ms`);
+  err.code = "AGENT_RUN_TIMEOUT";
+  return err;
+}
 function normalizeAgentRunToolMode(value) {
   const normalized = String(value ?? "chat").trim().toLowerCase();
   if (!normalized || normalized === "chat") return "chat";
@@ -6577,9 +6589,30 @@ function createAgentRunGateway({
   userAuth: userAuth2,
   tkmindProxy: tkmindProxy2,
   retryDelaysMs = DEFAULT_RUN_RETRY_DELAYS_MS,
-  autoDispatch = true
+  autoDispatch = true,
+  maxConcurrentRuns = positiveInteger(
+    process.env.MEMIND_AGENT_RUN_QUEUE_CONCURRENCY,
+    DEFAULT_MAX_CONCURRENT_RUNS
+  ),
+  runTimeoutMs = positiveInteger(
+    process.env.MEMIND_AGENT_RUN_TIMEOUT_MS,
+    DEFAULT_RUN_TIMEOUT_MS
+  )
 }) {
   const inFlight = /* @__PURE__ */ new Set();
+  const queuedDispatches = [];
+  const queuedDispatchSet = /* @__PURE__ */ new Set();
+  function enqueueRun(runId) {
+    if (!runId || inFlight.has(runId) || queuedDispatchSet.has(runId)) return false;
+    queuedDispatchSet.add(runId);
+    queuedDispatches.push(runId);
+    return true;
+  }
+  function nextQueuedRunId() {
+    const runId = queuedDispatches.shift();
+    if (runId) queuedDispatchSet.delete(runId);
+    return runId ?? null;
+  }
   async function appendEvent(runId, type, data = null) {
     await pool2.query(
       `INSERT INTO h5_agent_run_events (id, run_id, event_type, data_json, created_at)
@@ -6675,6 +6708,47 @@ function createAgentRunGateway({
     );
     await appendEvent(runId, status, fields);
   }
+  async function runWithTimeout(runId, task) {
+    let timer = null;
+    const timeout = new Promise((_, reject) => {
+      timer = setTimeout(() => reject(timeoutError(runTimeoutMs)), runTimeoutMs);
+    });
+    try {
+      return await Promise.race([task(), timeout]);
+    } finally {
+      if (timer) clearTimeout(timer);
+    }
+  }
+  async function executeRun(row, runId) {
+    const userMessage = safeJsonParse(row.user_message_json, {});
+    const runOptions = getRunOptionsFromMessage(userMessage);
+    let sessionId = row.agent_session_id ?? null;
+    if (!sessionId) {
+      const sessionOptions = {};
+      if (runOptions.toolMode === "code" && userAuth2?.getCodeAgentSessionPolicy) {
+        sessionOptions.sessionPolicy = await userAuth2.getCodeAgentSessionPolicy(row.user_id);
+      }
+      const session = await tkmindProxy2.startSessionForUser(row.user_id, sessionOptions);
+      sessionId = session.id;
+      await pool2.query(
+        `UPDATE h5_agent_runs SET agent_session_id = ?, updated_at = ? WHERE id = ?`,
+        [sessionId, nowMs(), runId]
+      );
+      await appendEvent(runId, "session_started", {
+        sessionId,
+        toolMode: runOptions.toolMode,
+        taskType: runOptions.taskType
+      });
+    }
+    await tkmindProxy2.submitSessionReplyForUser(
+      row.user_id,
+      sessionId,
+      row.request_id,
+      userMessage,
+      { toolMode: runOptions.toolMode }
+    );
+    return { sessionId };
+  }
   async function processRun(runId) {
     const row = await getRunById(runId);
     if (!row || TERMINAL_STATUSES.has(row.status)) return;
@@ -6688,40 +6762,18 @@ function createAgentRunGateway({
     if (Number(claim?.affectedRows ?? 0) === 0) return;
     await appendEvent(runId, "running", { attempt: nextAttempt });
     try {
-      const userMessage = safeJsonParse(row.user_message_json, {});
-      const runOptions = getRunOptionsFromMessage(userMessage);
-      let sessionId = row.agent_session_id ?? null;
-      if (!sessionId) {
-        const sessionOptions = {};
-        if (runOptions.toolMode === "code" && userAuth2?.getCodeAgentSessionPolicy) {
-          sessionOptions.sessionPolicy = await userAuth2.getCodeAgentSessionPolicy(row.user_id);
-        }
-        const session = await tkmindProxy2.startSessionForUser(row.user_id, sessionOptions);
-        sessionId = session.id;
-        await pool2.query(
-          `UPDATE h5_agent_runs SET agent_session_id = ?, updated_at = ? WHERE id = ?`,
-          [sessionId, nowMs(), runId]
-        );
-        await appendEvent(runId, "session_started", {
-          sessionId,
-          toolMode: runOptions.toolMode,
-          taskType: runOptions.taskType
-        });
-      }
-      await tkmindProxy2.submitSessionReplyForUser(
-        row.user_id,
-        sessionId,
-        row.request_id,
-        userMessage,
-        { toolMode: runOptions.toolMode }
-      );
+      const { sessionId } = await runWithTimeout(runId, () => executeRun(row, runId));
       await markRun(runId, "succeeded", {
         agent_session_id: sessionId,
         completed_at: nowMs()
       });
     } catch (err) {
       const message = err instanceof Error ? err.message : String(err);
-      const retryable = nextAttempt < retryDelaysMs.length;
+      const timedOut = err?.code === "AGENT_RUN_TIMEOUT";
+      if (timedOut) {
+        await appendEvent(runId, "timeout", { timeoutMs: runTimeoutMs });
+      }
+      const retryable = !timedOut && nextAttempt < retryDelaysMs.length;
       await markRun(runId, retryable ? "retryable" : "failed", {
         error_message: message,
         completed_at: retryable ? null : nowMs()
@@ -6731,15 +6783,46 @@ function createAgentRunGateway({
       }
     }
   }
+  function drainQueue() {
+    while (inFlight.size < maxConcurrentRuns) {
+      const runId = nextQueuedRunId();
+      if (!runId) break;
+      inFlight.add(runId);
+      void processRun(runId).finally(() => {
+        inFlight.delete(runId);
+        drainQueue();
+      });
+    }
+  }
   function dispatchRun(runId) {
-    if (!runId || inFlight.has(runId)) return;
-    inFlight.add(runId);
-    void processRun(runId).finally(() => inFlight.delete(runId));
+    if (!enqueueRun(runId)) return;
+    drainQueue();
+  }
+  async function getQueueStatus() {
+    const [rows] = await pool2.query(
+      `SELECT status, COUNT(*) AS count
+       FROM h5_agent_runs
+       WHERE status IN ('queued', 'running', 'retryable')
+       GROUP BY status`
+    );
+    const statusCounts = {};
+    for (const row of rows) {
+      statusCounts[row.status] = Number(row.count ?? 0);
+    }
+    return {
+      maxConcurrentRuns,
+      runTimeoutMs,
+      inFlight: inFlight.size,
+      pendingDispatches: queuedDispatches.length,
+      statusCounts,
+      terminalStatuses: [...TERMINAL_STATUSES]
+    };
   }
   return {
     createRun,
     getRunForUser,
-    dispatchRun
+    dispatchRun,
+    getQueueStatus
   };
 }
 
@@ -34957,7 +35040,9 @@ async function bootstrapUserAuth() {
     agentRunGateway = createAgentRunGateway({
       pool: pool2,
       userAuth,
-      tkmindProxy
+      tkmindProxy,
+      maxConcurrentRuns: Number(process.env.MEMIND_AGENT_RUN_QUEUE_CONCURRENCY ?? 1),
+      runTimeoutMs: Number(process.env.MEMIND_AGENT_RUN_TIMEOUT_MS ?? 15 * 60 * 1e3)
     });
     wechatMpService = createWechatMpService({
       config: WECHAT_MP_CONFIG,
@@ -36043,6 +36128,15 @@ api.get("/runtime/status", async (_req, res) => {
   }
   try {
     const status = await tkmindProxy.getRuntimeStatus();
+    const toolQueue = agentRunGateway?.getQueueStatus ? await agentRunGateway.getQueueStatus().catch((err) => ({
+      error: err instanceof Error ? err.message : String(err)
+    })) : null;
+    if (toolQueue) {
+      status.toolRuntime = {
+        ...status.toolRuntime ?? {},
+        queue: toolQueue
+      };
+    }
     return res.json({
       ok: true,
       timestamp: (/* @__PURE__ */ new Date()).toISOString(),

agent-run-gateway.mjs

@@ -4,6 +4,8 @@ const DEFAULT_RUN_RETRY_DELAYS_MS = [1500, 5000, 15000];
 const TERMINAL_STATUSES = new Set(['succeeded', 'failed']);
 const CODE_TOOL_MODES = new Set(['code', 'code-task', 'code_task', 'code-tool', 'code_tool', 'code_tool_task']);
 const RUN_METADATA_KEY = 'memindRun';
+const DEFAULT_MAX_CONCURRENT_RUNS = 1;
+const DEFAULT_RUN_TIMEOUT_MS = 15 * 60 * 1000;
 
 function nowMs() {
   return Date.now();
@@ -21,6 +23,18 @@ function serializeMessage(message) {
   return JSON.stringify(message ?? {});
 }
 
+function positiveInteger(value, fallback) {
+  const n = Number(value);
+  if (!Number.isFinite(n) || n <= 0) return fallback;
+  return Math.floor(n);
+}
+
+function timeoutError(ms) {
+  const err = new Error(`agent run timed out after ${ms}ms`);
+  err.code = 'AGENT_RUN_TIMEOUT';
+  return err;
+}
+
 export function normalizeAgentRunToolMode(value) {
   const normalized = String(value ?? 'chat').trim().toLowerCase();
   if (!normalized || normalized === 'chat') return 'chat';
@@ -88,8 +102,31 @@ export function createAgentRunGateway({
   tkmindProxy,
   retryDelaysMs = DEFAULT_RUN_RETRY_DELAYS_MS,
   autoDispatch = true,
+  maxConcurrentRuns = positiveInteger(
+    process.env.MEMIND_AGENT_RUN_QUEUE_CONCURRENCY,
+    DEFAULT_MAX_CONCURRENT_RUNS,
+  ),
+  runTimeoutMs = positiveInteger(
+    process.env.MEMIND_AGENT_RUN_TIMEOUT_MS,
+    DEFAULT_RUN_TIMEOUT_MS,
+  ),
 }) {
   const inFlight = new Set();
+  const queuedDispatches = [];
+  const queuedDispatchSet = new Set();
+
+  function enqueueRun(runId) {
+    if (!runId || inFlight.has(runId) || queuedDispatchSet.has(runId)) return false;
+    queuedDispatchSet.add(runId);
+    queuedDispatches.push(runId);
+    return true;
+  }
+
+  function nextQueuedRunId() {
+    const runId = queuedDispatches.shift();
+    if (runId) queuedDispatchSet.delete(runId);
+    return runId ?? null;
+  }
 
   async function appendEvent(runId, type, data = null) {
     await pool.query(
@@ -193,6 +230,50 @@ export function createAgentRunGateway({
     await appendEvent(runId, status, fields);
   }
 
+  async function runWithTimeout(runId, task) {
+    let timer = null;
+    const timeout = new Promise((_, reject) => {
+      timer = setTimeout(() => reject(timeoutError(runTimeoutMs)), runTimeoutMs);
+    });
+    try {
+      return await Promise.race([task(), timeout]);
+    } finally {
+      if (timer) clearTimeout(timer);
+    }
+  }
+
+  async function executeRun(row, runId) {
+    const userMessage = safeJsonParse(row.user_message_json, {});
+    const runOptions = getRunOptionsFromMessage(userMessage);
+    let sessionId = row.agent_session_id ?? null;
+    if (!sessionId) {
+      const sessionOptions = {};
+      if (runOptions.toolMode === 'code' && userAuth?.getCodeAgentSessionPolicy) {
+        sessionOptions.sessionPolicy = await userAuth.getCodeAgentSessionPolicy(row.user_id);
+      }
+      const session = await tkmindProxy.startSessionForUser(row.user_id, sessionOptions);
+      sessionId = session.id;
+      await pool.query(
+        `UPDATE h5_agent_runs SET agent_session_id = ?, updated_at = ? WHERE id = ?`,
+        [sessionId, nowMs(), runId],
+      );
+      await appendEvent(runId, 'session_started', {
+        sessionId,
+        toolMode: runOptions.toolMode,
+        taskType: runOptions.taskType,
+      });
+    }
+
+    await tkmindProxy.submitSessionReplyForUser(
+      row.user_id,
+      sessionId,
+      row.request_id,
+      userMessage,
+      { toolMode: runOptions.toolMode },
+    );
+    return { sessionId };
+  }
+
   async function processRun(runId) {
     const row = await getRunById(runId);
     if (!row || TERMINAL_STATUSES.has(row.status)) return;
@@ -207,41 +288,18 @@ export function createAgentRunGateway({
     await appendEvent(runId, 'running', { attempt: nextAttempt });
 
     try {
-      const userMessage = safeJsonParse(row.user_message_json, {});
-      const runOptions = getRunOptionsFromMessage(userMessage);
-      let sessionId = row.agent_session_id ?? null;
-      if (!sessionId) {
-        const sessionOptions = {};
-        if (runOptions.toolMode === 'code' && userAuth?.getCodeAgentSessionPolicy) {
-          sessionOptions.sessionPolicy = await userAuth.getCodeAgentSessionPolicy(row.user_id);
-        }
-        const session = await tkmindProxy.startSessionForUser(row.user_id, sessionOptions);
-        sessionId = session.id;
-        await pool.query(
-          `UPDATE h5_agent_runs SET agent_session_id = ?, updated_at = ? WHERE id = ?`,
-          [sessionId, nowMs(), runId],
-        );
-        await appendEvent(runId, 'session_started', {
-          sessionId,
-          toolMode: runOptions.toolMode,
-          taskType: runOptions.taskType,
-        });
-      }
-
-      await tkmindProxy.submitSessionReplyForUser(
-        row.user_id,
-        sessionId,
-        row.request_id,
-        userMessage,
-        { toolMode: runOptions.toolMode },
-      );
+      const { sessionId } = await runWithTimeout(runId, () => executeRun(row, runId));
       await markRun(runId, 'succeeded', {
         agent_session_id: sessionId,
         completed_at: nowMs(),
       });
     } catch (err) {
       const message = err instanceof Error ? err.message : String(err);
-      const retryable = nextAttempt < retryDelaysMs.length;
+      const timedOut = err?.code === 'AGENT_RUN_TIMEOUT';
+      if (timedOut) {
+        await appendEvent(runId, 'timeout', { timeoutMs: runTimeoutMs });
+      }
+      const retryable = !timedOut && nextAttempt < retryDelaysMs.length;
       await markRun(runId, retryable ? 'retryable' : 'failed', {
         error_message: message,
         completed_at: retryable ? null : nowMs(),
@@ -252,15 +310,49 @@ export function createAgentRunGateway({
     }
   }
 
+  function drainQueue() {
+    while (inFlight.size < maxConcurrentRuns) {
+      const runId = nextQueuedRunId();
+      if (!runId) break;
+      inFlight.add(runId);
+      void processRun(runId)
+        .finally(() => {
+          inFlight.delete(runId);
+          drainQueue();
+        });
+    }
+  }
+
   function dispatchRun(runId) {
-    if (!runId || inFlight.has(runId)) return;
-    inFlight.add(runId);
-    void processRun(runId).finally(() => inFlight.delete(runId));
+    if (!enqueueRun(runId)) return;
+    drainQueue();
+  }
+
+  async function getQueueStatus() {
+    const [rows] = await pool.query(
+      `SELECT status, COUNT(*) AS count
+       FROM h5_agent_runs
+       WHERE status IN ('queued', 'running', 'retryable')
+       GROUP BY status`,
+    );
+    const statusCounts = {};
+    for (const row of rows) {
+      statusCounts[row.status] = Number(row.count ?? 0);
+    }
+    return {
+      maxConcurrentRuns,
+      runTimeoutMs,
+      inFlight: inFlight.size,
+      pendingDispatches: queuedDispatches.length,
+      statusCounts,
+      terminalStatuses: [...TERMINAL_STATUSES],
+    };
   }
 
   return {
     createRun,
     getRunForUser,
     dispatchRun,
+    getQueueStatus,
   };
 }

agent-run-gateway.test.mjs

@@ -22,6 +22,14 @@ function createFakePool() {
       if (sql.includes('SELECT * FROM h5_agent_runs WHERE id = ? LIMIT 1')) {
         return [[runs.get(params[0])].filter(Boolean)];
       }
+      if (sql.includes('SELECT status, COUNT(*) AS count')) {
+        const counts = new Map();
+        for (const row of runs.values()) {
+          if (!['queued', 'running', 'retryable'].includes(row.status)) continue;
+          counts.set(row.status, (counts.get(row.status) ?? 0) + 1);
+        }
+        return [[...counts].map(([status, count]) => ({ status, count }))];
+      }
       if (sql.includes('INSERT INTO h5_agent_runs')) {
         const [
           id,
@@ -232,3 +240,101 @@ test('agent run retries transient failures and then becomes terminal', async ()
   assert.equal(pool.runs.get(run.id).attempts, 2);
   assert.match(pool.runs.get(run.id).error_message, /upstream unavailable/);
 });
+
+test('agent run queue limits concurrent execution', async () => {
+  const pool = createFakePool();
+  let active = 0;
+  let maxActive = 0;
+  const release = [];
+  const gateway = createAgentRunGateway({
+    pool,
+    userAuth: {},
+    tkmindProxy: {
+      async startSessionForUser(_userId) {
+        return { id: `session-${release.length + 1}` };
+      },
+      async submitSessionReplyForUser() {
+        active += 1;
+        maxActive = Math.max(maxActive, active);
+        await new Promise((resolve) => release.push(resolve));
+        active -= 1;
+      },
+    },
+    retryDelaysMs: [],
+    maxConcurrentRuns: 1,
+  });
+
+  const run1 = await gateway.createRun('user-1', {
+    requestId: 'req-1',
+    userMessage: { role: 'user', content: [] },
+  });
+  const run2 = await gateway.createRun('user-1', {
+    requestId: 'req-2',
+    userMessage: { role: 'user', content: [] },
+  });
+
+  await waitFor(() => active === 1 && pool.runs.get(run2.id)?.status === 'queued');
+  assert.equal(maxActive, 1);
+  assert.equal((await gateway.getQueueStatus()).pendingDispatches, 1);
+  release.shift()();
+  await waitFor(() => pool.runs.get(run1.id)?.status === 'succeeded' && active === 1);
+  release.shift()();
+  await waitFor(() => pool.runs.get(run2.id)?.status === 'succeeded');
+  assert.equal(maxActive, 1);
+});
+
+test('agent run timeout fails without retrying', async () => {
+  const pool = createFakePool();
+  const gateway = createAgentRunGateway({
+    pool,
+    userAuth: {},
+    tkmindProxy: {
+      async startSessionForUser() {
+        return { id: 'session-timeout' };
+      },
+      async submitSessionReplyForUser() {
+        await new Promise(() => {});
+      },
+    },
+    retryDelaysMs: [0, 0],
+    maxConcurrentRuns: 1,
+    runTimeoutMs: 5,
+  });
+
+  const run = await gateway.createRun('user-1', {
+    requestId: 'req-timeout',
+    userMessage: { role: 'user', content: [] },
+  });
+
+  await waitFor(() => pool.runs.get(run.id)?.status === 'failed');
+  assert.equal(pool.runs.get(run.id).attempts, 1);
+  assert.match(pool.runs.get(run.id).error_message, /timed out/);
+  assert.equal(
+    pool.events.some((event) => event.runId === run.id && event.eventType === 'timeout'),
+    true,
+  );
+});
+
+test('agent run queue status reports active database and local queue state', async () => {
+  const pool = createFakePool();
+  const gateway = createAgentRunGateway({
+    pool,
+    userAuth: {},
+    tkmindProxy: {},
+    autoDispatch: false,
+    maxConcurrentRuns: 2,
+    runTimeoutMs: 1234,
+  });
+
+  await gateway.createRun('user-1', {
+    requestId: 'req-status',
+    userMessage: { role: 'user', content: [] },
+  });
+
+  const status = await gateway.getQueueStatus();
+  assert.equal(status.maxConcurrentRuns, 2);
+  assert.equal(status.runTimeoutMs, 1234);
+  assert.equal(status.inFlight, 0);
+  assert.equal(status.pendingDispatches, 0);
+  assert.equal(status.statusCounts.queued, 1);
+});

scripts/build-portal-runtime.mjs

@@ -346,6 +346,7 @@ async function writeMetadata() {
       '  node scripts/runtime-worker-drain.mjs undrain goosed-3',
       '  node scripts/check-tool-runtime.mjs',
       '  node scripts/check-agent-code-run-entry.mjs',
+      '  curl -sk https://mm.tkmind.cn/api/runtime/status  # includes toolRuntime.queue',
       '',
     ].join('\n'),
   );

scripts/runtime-slo-report.mjs

@@ -148,6 +148,7 @@ function summarizeRuntime(runtimeJson) {
     routerEnabled: Boolean(runtimeJson?.router?.enabled),
     publicBaseUrl: runtimeJson?.publicBaseUrl ?? null,
     toolRuntime: runtimeJson?.toolRuntime ?? null,
+    toolQueue: runtimeJson?.toolRuntime?.queue ?? null,
     workers: workers.map((worker) => ({
       id: worker.id,
       healthy: runtimeJson?.targets?.find((target) => target.target === worker.target)?.healthy ?? null,
@@ -220,6 +221,10 @@ for (const worker of runtimeSummary?.workers ?? []) {
   if (worker.firstTokenCount > 0 && worker.ewmaFirstTokenMs <= 0) failures.push(`${worker.id}_first_token_ewma_missing`);
 }
 if (runtimeSummary?.toolRuntime?.chatInjectsCodeTools !== false) failures.push('chat_injects_code_tools');
+if (runtimeSummary?.toolQueue?.error) failures.push('tool_queue_status_error');
+if (runtimeSummary?.toolQueue?.inFlight > runtimeSummary?.toolQueue?.maxConcurrentRuns) {
+  failures.push('tool_queue_concurrency_exceeded');
+}
 
 const report = {
   ok: failures.length === 0,

server.mjs

@@ -572,6 +572,8 @@ async function bootstrapUserAuth() {
       pool,
       userAuth,
       tkmindProxy,
+      maxConcurrentRuns: Number(process.env.MEMIND_AGENT_RUN_QUEUE_CONCURRENCY ?? 1),
+      runTimeoutMs: Number(process.env.MEMIND_AGENT_RUN_TIMEOUT_MS ?? 15 * 60 * 1000),
     });
     wechatMpService = createWechatMpService({
       config: WECHAT_MP_CONFIG,
@@ -1781,6 +1783,17 @@ api.get('/runtime/status', async (_req, res) => {
   }
   try {
     const status = await tkmindProxy.getRuntimeStatus();
+    const toolQueue = agentRunGateway?.getQueueStatus
+      ? await agentRunGateway.getQueueStatus().catch((err) => ({
+          error: err instanceof Error ? err.message : String(err),
+        }))
+      : null;
+    if (toolQueue) {
+      status.toolRuntime = {
+        ...(status.toolRuntime ?? {}),
+        queue: toolQueue,
+      };
+    }
     return res.json({
       ok: true,
       timestamp: new Date().toISOString(),