@dudousxd/nestjs-agent 1.16.1 → 1.17.1

package/dist/durable/index.js

@@ -2,10 +2,10 @@ var __defProp = Object.defineProperty;
 var __name = (target, value) => __defProp(target, "name", { value, configurable: true });
 
 // src/durable/agent-run.workflow.ts
-import { AGENT_DEPS_FACTORY as AGENT_DEPS_FACTORY2, AGENT_STORE as AGENT_STORE2, RunCancelledError, agentFailureCode, encodeStreamEvent as encodeStreamEvent3, isReplayIntegrityError, publishAgentRunFailed, releaseThreadRun, runAgentLoop, settleAll, settleUnsettledDelegation } from "@dudousxd/nestjs-agent-core";
+import { AGENT_DEPS_FACTORY as AGENT_DEPS_FACTORY2, AGENT_STORE, RUN_ENDED_BEFORE_TOOL_CALL, RunCancelledError, agentFailureCode, encodeStreamEvent as encodeStreamEvent2, isReplayIntegrityError, publishAgentRunFailed, releaseThreadRun, runAgentLoop, settleAll, settleDeadRun, settleUnsettledDelegation, streamFailure } from "@dudousxd/nestjs-agent-core";
 import { RUN_GATEWAY, Workflow } from "@dudousxd/nestjs-durable";
 import { isWorkflowControlFlowSignal } from "@dudousxd/nestjs-durable-core";
-import { Inject as Inject3, Injectable as Injectable3, Optional as Optional2 } from "@nestjs/common";
+import { Inject as Inject2, Injectable as Injectable2, Logger, Optional } from "@nestjs/common";
 
 // src/agent-deps.ts
 function utcDay(date = /* @__PURE__ */ new Date()) {
@@ -23,9 +23,26 @@ function childSinkWriter(writer) {
 }
 __name(childSinkWriter, "childSinkWriter");
 
-// src/queue/chat-queue.service.ts
-import { AGENT_ATTACHMENT_STAGING, AGENT_OPTIONS, AGENT_QUOTA_PROVIDER, AGENT_SINK, AGENT_STORE, encodeStreamEvent, isChatQueueStore, queuedMessageView } from "@dudousxd/nestjs-agent-core";
-import { Inject, Injectable, Logger, NotImplementedException, Optional } from "@nestjs/common";
+// src/queue/chat-queue.token.ts
+var AGENT_CHAT_QUEUE = /* @__PURE__ */ Symbol.for("@dudousxd/nestjs-agent:chat-queue");
+
+// src/durable/agent-run.steps.ts
+import { AGENT_DEPS_FACTORY, createFrameBuffer, createUiCollector, encodeStreamEvent, invokeWithTransientRetry, observeTurnFrames, publishAgentToolRetry, stampToolKinds, toolCallContext, traceLlmTurn, traceToolExecution, withAskTool, withMemoryTool, withSkillTool, withToolTimeout, withTurnFrames, wrapToolStepOutput } from "@dudousxd/nestjs-agent-core";
+import { Step } from "@dudousxd/nestjs-durable";
+import { Inject, Injectable } from "@nestjs/common";
+
+// src/durable/outside-workflow-ctx.ts
+import { runInWorkflowCtx } from "@dudousxd/nestjs-durable-core";
+function outsideWorkflowCtx(fn) {
+  return runInWorkflowCtx(void 0, fn);
+}
+__name(outsideWorkflowCtx, "outsideWorkflowCtx");
+function stepOf(ctx) {
+  return (name, fn, ...rest) => ctx.localStep(name, (...args) => outsideWorkflowCtx(() => fn(...args)), ...rest);
+}
+__name(stepOf, "stepOf");
+
+// src/durable/agent-run.steps.ts
 function _ts_decorate(decorators, target, key, desc) {
   var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
   if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
@@ -52,729 +69,289 @@ function _ts_param(paramIndex, decorator) {
   };
 }
 __name(_ts_param, "_ts_param");
-var MAX_DRAIN_ATTEMPTS = 8;
-var ChatQueueService = class _ChatQueueService {
+var AgentRunSteps = class {
   static {
-    __name(this, "ChatQueueService");
+    __name(this, "AgentRunSteps");
+  }
+  factory;
+  constructor(factory) {
+    this.factory = factory;
   }
-  store;
-  sink;
-  quotaProvider;
-  options;
-  staging;
-  logger = new Logger(_ChatQueueService.name);
   /**
-  * Runs this process claimed a thread for and has not handed to a runner yet. A holder in here is
-  * alive whatever the runner says — it simply has not registered the run yet.
+  * Retryable: every side effect (persist/quota/stream markers) happens loop-side under
+  * `hooks.step`, so a durable retry only re-runs the model call and re-streams its tokens — never
+  * double-writes. NEVER call `writer.end()`/`fail()` here: the workflow owns the stream's lifecycle,
+  * across however many turns/retries it takes.
   */
-  starting = /* @__PURE__ */ new Set();
-  constructor(store, sink, quotaProvider, options, staging) {
-    this.store = store;
-    this.sink = sink;
-    this.quotaProvider = quotaProvider;
-    this.options = options;
-    this.staging = staging;
-  }
-  /** Whether the bound store can hold a queue. Without one, a busy thread is not queued. */
-  get supported() {
-    return isChatQueueStore(this.store);
+  llm(input) {
+    return outsideWorkflowCtx(() => this.runLlm(input));
   }
-  /** The bound store as a queue store, or a 501 naming what is missing. */
-  queueStore() {
-    if (!isChatQueueStore(this.store)) {
-      throw new NotImplementedException("Queueing messages requires an AgentStore that implements ChatQueueStore (enqueueMessage, listQueue, \u2026, claimActiveStream, releaseActiveStream); the bound store does not.");
+  async runLlm(input) {
+    const deps = this.factory.forAgent(input.agentName);
+    const tools = withMemoryTool({
+      tools: withSkillTool({
+        tools: withAskTool({
+          tools: await deps.registry.definitionsFor(input.actor, deps.rolesPolicy, deps.toolAllowList, {
+            ...input.threadId !== void 0 ? {
+              threadId: input.threadId
+            } : {},
+            ...input.agentName !== void 0 ? {
+              agentName: input.agentName
+            } : {}
+          }),
+          ask: deps.ask
+        }),
+        enabled: deps.skills !== void 0
+      }),
+      enabled: deps.memory?.provider.write !== void 0
+    });
+    const buffer = input.bufferOutput === true ? createFrameBuffer() : void 0;
+    let writer;
+    if (buffer !== void 0) {
+      writer = buffer.writer;
+    } else {
+      const opened = await deps.sink.open(input.sinkRunId);
+      writer = input.childSink ? childSinkWriter(opened) : opened;
     }
-    return this.store;
-  }
-  async state(threadId) {
-    const store = this.queueStore();
-    const [items, paused] = await Promise.all([
-      store.listQueue(threadId),
-      store.queuePause(threadId)
-    ]);
-    return {
-      items: items.map(queuedMessageView),
-      paused
+    const frames = observeTurnFrames(writer);
+    const turn = withTurnFrames(await traceLlmTurn(input.runId, input.step, () => deps.model.runTurn({
+      system: input.system,
+      messages: input.messages,
+      tools,
+      sink: frames.writer,
+      ...input.model !== void 0 ? {
+        model: input.model
+      } : {}
+    })), frames.summary());
+    const stamped = stampToolKinds(turn, deps);
+    return buffer === void 0 ? stamped : {
+      ...stamped,
+      bufferedFrames: buffer.frames()
     };
   }
   /**
-  * Whether `holder` is still running. A run this process is about to start is; otherwise the
-  * runner decides, and a runner that cannot tell says yes.
+  * NO DURABLE retries — a tool may not be idempotent (a bare `@Step()` carries none, unlike `llm`
+  * above). What it does retry, in place, is a CLASSIFIED-TRANSIENT error from the tool's own
+  * invocation (a DB deadlock, a lock-wait timeout — see `isTransientToolError`): the server rolled
+  * that work back, so re-invoking it is safe, and doing it here (inside this one step, never a new
+  * checkpoint) is exactly how the non-dispatched loop path handles it too. The timeout is applied
+  * HERE, handler-side, via `withToolTimeout` — never as a durable dispatch `timeoutMs`, which
+  * bounds engine liveness (retryable) rather than the tool's own business outcome — and per
+  * ATTEMPT, so a retried invocation gets its own fresh timeout window.
   */
-  async isAlive(holder, runner) {
-    if (this.starting.has(holder)) {
-      return true;
-    }
-    if (typeof runner.isRunActive !== "function") {
-      return true;
-    }
-    try {
-      return await runner.isRunActive(holder);
-    } catch {
-      return true;
-    }
+  tool(input) {
+    return outsideWorkflowCtx(() => this.runTool(input));
   }
-  /**
-  * The run holding `threadId`, or `null` when it is free — a holder the runner reports dead counts
-  * as free, and is returned as `stale` so the caller can claim over it.
-  */
-  async holder(threadId, runner) {
-    const holder = await this.queueStore().activeRunForThread(threadId);
-    if (holder === null) {
-      return {
-        live: null,
-        stale: null
-      };
+  async runTool(input) {
+    const deps = this.factory.forAgent(input.ctx.agentName);
+    if (deps.registry.spec(input.toolName)?.kind === "action" && input.toolType !== "action") {
+      throw new Error(`tool "${input.toolName}" is an action tool but was dispatched without approval: the dispatching process resolved its kind against a registry that does not have it`);
     }
-    return await this.isAlive(holder, runner) ? {
-      live: holder,
-      stale: null
-    } : {
-      live: null,
-      stale: holder
+    const sinkRunId = input.sinkRunId;
+    let opened;
+    const ui = createUiCollector(input.toolCallId, sinkRunId === void 0 ? void 0 : async (event) => {
+      opened ??= Promise.resolve(deps.sink.open(sinkRunId)).then((writer) => input.childSink === true ? childSinkWriter(writer) : writer);
+      await (await opened).write(encodeStreamEvent(event));
+    });
+    const ctx = {
+      ...toolCallContext(input.ctx, input.toolCallId),
+      emitUi: ui.emit
+    };
+    const localClassify = deps.toolTransientRetry !== false ? deps.toolTransientRetry?.classify : void 0;
+    const transientRetry = input.transientRetry === false ? false : {
+      attempts: input.transientRetry.attempts,
+      backoffMs: input.transientRetry.backoffMs,
+      ...localClassify !== void 0 ? {
+        classify: localClassify
+      } : {}
     };
+    const output = await traceToolExecution(input.ctx.runId, {
+      toolCallId: input.toolCallId,
+      toolName: input.toolName,
+      toolType: input.toolType
+    }, () => invokeWithTransientRetry(() => {
+      ui.restart();
+      const invocation = deps.registry.invoke(input.toolName, input.input, ctx, deps.rolesPolicy);
+      return input.timeoutMs !== void 0 ? withToolTimeout(invocation, input.timeoutMs, input.toolName) : invocation;
+    }, transientRetry, {
+      onRetry: /* @__PURE__ */ __name((attempt, error) => {
+        publishAgentToolRetry({
+          toolName: input.toolName,
+          toolCallId: input.toolCallId,
+          attempt,
+          message: error instanceof Error ? error.message : String(error)
+        });
+      }, "onRetry")
+    }));
+    return input.collectUi === true ? wrapToolStepOutput(output, ui.components()) : output;
+  }
+};
+_ts_decorate([
+  Step({
+    retries: 3
+  }),
+  _ts_metadata("design:type", Function),
+  _ts_metadata("design:paramtypes", [
+    typeof DispatchedLlmInput === "undefined" ? Object : DispatchedLlmInput
+  ]),
+  _ts_metadata("design:returntype", typeof Promise === "undefined" ? Object : Promise)
+], AgentRunSteps.prototype, "llm", null);
+_ts_decorate([
+  Step(),
+  _ts_metadata("design:type", Function),
+  _ts_metadata("design:paramtypes", [
+    typeof DispatchedToolInput === "undefined" ? Object : DispatchedToolInput
+  ]),
+  _ts_metadata("design:returntype", typeof Promise === "undefined" ? Object : Promise)
+], AgentRunSteps.prototype, "tool", null);
+AgentRunSteps = _ts_decorate([
+  Injectable(),
+  _ts_param(0, Inject(AGENT_DEPS_FACTORY)),
+  _ts_metadata("design:type", Function),
+  _ts_metadata("design:paramtypes", [
+    typeof AgentDepsFactory === "undefined" ? Object : AgentDepsFactory
+  ])
+], AgentRunSteps);
+
+// src/durable/agent-run.workflow.ts
+function _ts_decorate2(decorators, target, key, desc) {
+  var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
+  if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
+    r = Reflect.decorate(decorators, target, key, desc);
+  } else {
+    for (var i = decorators.length - 1; i >= 0; i--) {
+      if (d = decorators[i]) {
+        r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
+      }
+    }
+  }
+  return c > 3 && r && Object.defineProperty(target, key, r), r;
+}
+__name(_ts_decorate2, "_ts_decorate");
+function _ts_metadata2(metadataKey, metadataValue) {
+  if (typeof Reflect === "object" && typeof Reflect.metadata === "function") {
+    return Reflect.metadata(metadataKey, metadataValue);
+  }
+}
+__name(_ts_metadata2, "_ts_metadata");
+function _ts_param2(paramIndex, decorator) {
+  return function(target, key) {
+    decorator(target, key, paramIndex);
+  };
+}
+__name(_ts_param2, "_ts_param");
+var AgentRunWorkflow = class _AgentRunWorkflow {
+  static {
+    __name(this, "AgentRunWorkflow");
+  }
+  factory;
+  store;
+  runs;
+  queue;
+  logger = new Logger(_AgentRunWorkflow.name);
+  constructor(factory, store, runs, queue) {
+    this.factory = factory;
+    this.store = store;
+    this.runs = runs;
+    this.queue = queue;
   }
   /**
-  * Decide what follows a TOP-LEVEL run that is settling on `threadId`, and move the thread there:
-  * hand it to the next queued message (claimed and popped here, started by the caller), or pause
-  * the queue, or release the thread. Call it BEFORE the run's terminal frame, so a client reading
-  * the stream learns what comes next from the `queue` frame, and so the thread is already free
-  * when the client sees the run end.
+  * Move the thread past this settling turn — to the next queued message, started here as a
+  * fire-and-forget `ctx.startChild` under the message's own id, or to a paused/empty queue — and
+  * answer the `queue` frame to write before the turn's terminal.
   *
-  * A no-op returning `{}` on a store without a queue.
-  */
-  async plan(args) {
-    if (!isChatQueueStore(this.store)) {
-      return {};
-    }
-    const store = this.store;
-    const { threadId, outcome } = args;
-    if (await store.activeRunForThread(threadId) !== args.runId) {
-      return {};
-    }
-    let holding = args.runId;
-    for (let attempt = 0; attempt < MAX_DRAIN_ATTEMPTS; attempt += 1) {
-      const items = await store.listQueue(threadId);
-      const head = items[0];
-      if (head === void 0) {
-        if (holding !== null) {
-          await store.releaseActiveStream(threadId, holding);
-          holding = null;
-          continue;
-        }
-        return {};
-      }
-      let pause = await store.queuePause(threadId);
-      if (pause === null) {
-        const reason = pauseReasonFor(outcome, head);
-        if (reason !== null) {
-          pause = {
-            reason,
-            ...outcome === "failed" && args.error !== void 0 ? {
-              message: args.error
-            } : {},
-            at: (/* @__PURE__ */ new Date()).toISOString()
-          };
-          await store.setQueuePause(threadId, pause);
-        }
-      }
-      if (pause === null) {
-        const blocked = await this.quotaBlock(head.actor);
-        if (blocked !== null) {
-          pause = {
-            reason: "quota_exceeded",
-            message: blocked,
-            at: (/* @__PURE__ */ new Date()).toISOString()
-          };
-          await store.setQueuePause(threadId, pause);
-        }
-      }
-      if (pause !== null) {
-        if (holding !== null) {
-          await store.releaseActiveStream(threadId, holding);
-        }
-        return {
-          frame: queueFrame(items, pause)
-        };
-      }
-      const claimed = await store.claimActiveStream(threadId, head.id, holding !== null ? {
-        replacing: holding
-      } : {});
-      if (!claimed) {
-        return {};
-      }
-      holding = head.id;
-      if (!await store.removeQueuedMessage(head.id)) {
-        await store.releaseActiveStream(threadId, head.id);
-        holding = null;
-        continue;
-      }
-      const next = {
-        runId: head.id,
-        messageId: head.id,
-        input: await this.inputFor(head)
-      };
-      return {
-        next,
-        frame: queueFrame(items.slice(1), null, {
-          messageId: head.id,
-          runId: head.id
-        })
-      };
-    }
-    if (holding !== null) {
-      await store.releaseActiveStream(threadId, holding);
-    }
-    return {};
-  }
-  /**
-  * {@link plan}, then start what it picked with `start`. What an in-process settle (the inline
-  * runner, a durable cancel issued outside a workflow) calls; answers the frame to write.
-  */
-  async handoff(args, start) {
-    const plan = await this.plan(args);
-    if (plan.next === void 0) {
-      return plan.frame;
-    }
-    return await this.launch(args.threadId, plan.next, start) ? plan.frame : this.pausedFrame(args.threadId);
-  }
-  /**
-  * Start the head of an IDLE thread's queue — after a send queued behind a run that turned out to
-  * be gone, after a resume, after a stale holder. Does nothing while a live run holds the thread
-  * (its settle drains) or while the queue is paused. Answers the started run's id, if any.
+  * Journaled: the decision (which message, claimed and popped) is one `localStep`, so a replay
+  * reads it back instead of popping a second message, and the start is the runtime's own
+  * replay-safe spawn. Gated by `ctx.patched`, so a run recorded before the queue existed replays
+  * against the history it has. A no-op for a sub-agent's run and on a store without a queue.
   */
-  async kick(threadId, runner) {
-    if (!isChatQueueStore(this.store)) {
+  async advanceQueue(ctx, input, outcome, error) {
+    const queue = this.queue;
+    if (queue === void 0 || !queue.supported || input.sinkRunId !== void 0 || input.deliverTo !== void 0) {
       return void 0;
     }
-    const { live, stale } = await this.holder(threadId, runner);
-    if (live !== null) {
+    if (!await ctx.patched("agent:chat-queue")) {
       return void 0;
     }
-    if (stale !== null) {
-      this.logger.warn(`thread ${threadId} was held by run ${stale}, which is no longer running`);
-    }
-    const plan = await this.plan({
-      threadId,
-      runId: stale,
-      outcome: "completed"
+    const plan = await stepOf(ctx)(`queue:${outcome}`, async () => {
+      try {
+        const settled = outcome === "completed" && await this.isCancelled(ctx.runId) ? "cancelled" : outcome;
+        return await queue.plan({
+          threadId: input.threadId,
+          runId: ctx.runId,
+          outcome: settled,
+          ...error !== void 0 ? {
+            error
+          } : {}
+        });
+      } catch {
+        return {};
+      }
     });
-    if (plan.next === void 0) {
-      return void 0;
+    const next = plan.next;
+    if (next === void 0) {
+      return plan.frame;
     }
-    const started = await this.launch(threadId, plan.next, (input, runId) => runner.start(input, {
-      runId
-    }));
-    if (started && plan.frame !== void 0) {
-      await this.write(plan.next.runId, plan.frame);
+    try {
+      await ctx.startChild(_AgentRunWorkflow, next.input, next.runId);
+      return plan.frame;
+    } catch (failure) {
+      if (isWorkflowControlFlowSignal(failure)) {
+        throw failure;
+      }
+      const reason = failure instanceof Error ? failure.message : String(failure);
+      return stepOf(ctx)("queue:restore", async () => {
+        await queue.restore(input.threadId, next, "start_failed", reason);
+        return queue.pausedFrame(input.threadId);
+      });
     }
-    return started ? plan.next.runId : void 0;
   }
   /**
-  * Start a turn the thread is already claimed for. On failure the message goes back to the head of
-  * the queue, the queue pauses (`start_failed`), and the claim is released — nothing is lost and
-  * nothing is left holding the thread.
+  * Has someone asked this run to stop? Read from the runtime's own run status, which is where
+  * `DurableAgentRunner.cancel` puts it — `cancelling` while a compensating cancel is in flight,
+  * `cancelled` once it has settled.
+  *
+  * Costs a run read at each of the loop's observation points (two per step). That is a row the
+  * control plane already has, against a turn that is otherwise making model calls, and it buys the
+  * property that matters: the answer comes from a place BOTH the canceller and every pod that might
+  * replay this run can see.
+  *
+  * Failing to ask is not a cancel. A tenant with no gateway wired, or an operator briefly
+  * unreachable, answers "no" and the run carries on — the alternative is failing healthy runs
+  * because a status read timed out. The loop journals that answer like any other, so two processes
+  * can never disagree about what was seen here.
   */
-  async launch(threadId, next, start) {
-    this.starting.add(next.runId);
+  async isCancelled(runId) {
     try {
-      await start(next.input, next.runId);
-      return true;
-    } catch (error) {
-      const message = error instanceof Error ? error.message : String(error);
-      this.logger.error(`queued message ${next.messageId} could not start: ${message}`);
-      await this.restore(threadId, next, "start_failed", message);
+      const status = (await this.runs?.getRunDetail(runId))?.run.status;
+      return status === "cancelled" || status === "cancelling";
+    } catch {
       return false;
-    } finally {
-      this.starting.delete(next.runId);
     }
   }
-  /** Put a turn that never started back at the head of its queue, paused, and free the thread. */
-  async restore(threadId, next, reason, message) {
-    const store = this.queueStore();
-    const { input } = next;
-    await store.enqueueMessage({
-      threadId,
-      actor: input.actor,
-      content: input.userText,
-      at: "head",
-      ...input.attachments !== void 0 ? {
-        attachments: input.attachments
-      } : {},
-      ...input.agentName !== void 0 ? {
-        agentName: input.agentName
-      } : {},
-      ...input.model !== void 0 ? {
-        model: input.model
-      } : {},
-      ...input.pageContext !== void 0 ? {
-        pageContext: input.pageContext
-      } : {}
-    });
-    await store.setQueuePause(threadId, {
-      reason,
-      message,
-      at: (/* @__PURE__ */ new Date()).toISOString()
-    });
-    await store.releaseActiveStream(threadId, next.runId);
-  }
-  /** A `queue` frame with the thread's current state. */
-  async pausedFrame(threadId) {
-    return {
-      kind: "queue",
-      queue: await this.state(threadId)
-    };
-  }
   /**
-  * Tell whoever is watching the thread that its queue changed: a `queue` frame into the stream of
-  * the run holding it. Best-effort — a thread nobody is streaming has no one to tell, and the next
-  * read of the thread carries the queue anyway.
+  * Settle the delegation a DETACHED run was started for, when that run ends without an answer.
+  *
+  * The loop delivers its own success (`deliver:detached`) but cannot catch its own crash, and the
+  * runtime settles the run row without knowing anything delegated it. Left alone, the card in the
+  * calling conversation says "started" for ever — the one state a reader can neither wait on nor
+  * act on. A no-op for every other run: only a detached one carries a delivery address.
   */
-  async publish(threadId) {
-    const state = await this.state(threadId);
-    const holder = await this.queueStore().activeRunForThread(threadId);
-    if (holder !== null) {
-      await this.write(holder, {
-        kind: "queue",
-        queue: state
-      }).catch((error) => {
-        this.logger.warn(`could not publish the queue of thread ${threadId}: ${error instanceof Error ? error.message : String(error)}`);
-      });
+  async settleDetachedParent(ctx, input, outcome) {
+    const delivery = input.deliverTo;
+    if (delivery === void 0) {
+      return;
     }
-    return state;
-  }
-  async write(runId, frame) {
-    const writer = await this.sink.open(runId);
-    await writer.write(encodeStreamEvent(frame));
-  }
-  /** Why `actor` may not start a turn now (the quota message), or `null`. */
-  async quotaBlock(actor) {
-    if (this.options?.quota === void 0 || this.quotaProvider === void 0) {
-      return null;
-    }
-    try {
-      const { blocked } = await this.quotaProvider.report({
-        actor
-      });
-      if (blocked === void 0) {
-        return null;
-      }
-      return blocked.reason ?? `The ${blocked.period === "day" ? "daily" : "monthly"} quota is used up`;
-    } catch {
-      return null;
-    }
-  }
-  /**
-  * The turn a queued message starts: everything it was queued with, today's date, and its
-  * attachments' urls minted afresh — the ones stored at enqueue time may have expired while it
-  * waited. An attachment the staging store no longer resolves keeps its stored url.
-  */
-  async inputFor(message) {
-    const attachments = await this.freshAttachments(message.actor, message.attachments ?? []);
-    return {
-      threadId: message.threadId,
-      actor: message.actor,
-      userText: message.content,
-      day: utcDay(),
-      ...message.agentName !== void 0 ? {
-        agentName: message.agentName
-      } : {},
-      ...attachments.length > 0 ? {
-        attachments
-      } : {},
-      ...message.pageContext !== void 0 ? {
-        pageContext: message.pageContext
-      } : {},
-      ...message.model !== void 0 ? {
-        model: message.model
-      } : {}
-    };
-  }
-  async freshAttachments(actor, attachments) {
-    const staging = this.staging;
-    if (attachments.length === 0 || staging === void 0 || typeof staging.resolve !== "function") {
-      return attachments;
-    }
-    return Promise.all(attachments.map(async (attachment) => {
-      try {
-        return await staging.resolve({
-          mediaId: attachment.mediaId,
-          actor
-        }) ?? attachment;
-      } catch {
-        return attachment;
-      }
-    }));
-  }
-};
-ChatQueueService = _ts_decorate([
-  Injectable(),
-  _ts_param(0, Inject(AGENT_STORE)),
-  _ts_param(1, Inject(AGENT_SINK)),
-  _ts_param(2, Optional()),
-  _ts_param(2, Inject(AGENT_QUOTA_PROVIDER)),
-  _ts_param(3, Optional()),
-  _ts_param(3, Inject(AGENT_OPTIONS)),
-  _ts_param(4, Optional()),
-  _ts_param(4, Inject(AGENT_ATTACHMENT_STAGING)),
-  _ts_metadata("design:type", Function),
-  _ts_metadata("design:paramtypes", [
-    typeof AgentStore === "undefined" ? Object : AgentStore,
-    typeof TokenStreamSink === "undefined" ? Object : TokenStreamSink,
-    typeof QuotaProvider === "undefined" ? Object : QuotaProvider,
-    typeof Pick === "undefined" ? Object : Pick,
-    typeof AttachmentStagingStore === "undefined" ? Object : AttachmentStagingStore
-  ])
-], ChatQueueService);
-function pauseReasonFor(outcome, head) {
-  if (outcome === "failed") {
-    return "run_failed";
-  }
-  if (outcome === "cancelled" && head.interrupt !== true) {
-    return "cancelled";
-  }
-  return null;
-}
-__name(pauseReasonFor, "pauseReasonFor");
-function queueFrame(items, paused, started) {
-  return {
-    kind: "queue",
-    queue: {
-      items: items.map(queuedMessageView),
-      paused
-    },
-    ...started !== void 0 ? {
-      started
-    } : {}
-  };
-}
-__name(queueFrame, "queueFrame");
-
-// src/durable/agent-run.steps.ts
-import { AGENT_DEPS_FACTORY, createFrameBuffer, createUiCollector, encodeStreamEvent as encodeStreamEvent2, invokeWithTransientRetry, observeTurnFrames, publishAgentToolRetry, stampToolKinds, traceLlmTurn, traceToolExecution, withAskTool, withMemoryTool, withSkillTool, withToolTimeout, withTurnFrames, wrapToolStepOutput } from "@dudousxd/nestjs-agent-core";
-import { Step } from "@dudousxd/nestjs-durable";
-import { Inject as Inject2, Injectable as Injectable2 } from "@nestjs/common";
-
-// src/durable/outside-workflow-ctx.ts
-import { runInWorkflowCtx } from "@dudousxd/nestjs-durable-core";
-function outsideWorkflowCtx(fn) {
-  return runInWorkflowCtx(void 0, fn);
-}
-__name(outsideWorkflowCtx, "outsideWorkflowCtx");
-function stepOf(ctx) {
-  return (name, fn, ...rest) => ctx.localStep(name, (...args) => outsideWorkflowCtx(() => fn(...args)), ...rest);
-}
-__name(stepOf, "stepOf");
-
-// src/durable/agent-run.steps.ts
-function _ts_decorate2(decorators, target, key, desc) {
-  var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
-  if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
-    r = Reflect.decorate(decorators, target, key, desc);
-  } else {
-    for (var i = decorators.length - 1; i >= 0; i--) {
-      if (d = decorators[i]) {
-        r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
-      }
-    }
-  }
-  return c > 3 && r && Object.defineProperty(target, key, r), r;
-}
-__name(_ts_decorate2, "_ts_decorate");
-function _ts_metadata2(metadataKey, metadataValue) {
-  if (typeof Reflect === "object" && typeof Reflect.metadata === "function") {
-    return Reflect.metadata(metadataKey, metadataValue);
-  }
-}
-__name(_ts_metadata2, "_ts_metadata");
-function _ts_param2(paramIndex, decorator) {
-  return function(target, key) {
-    decorator(target, key, paramIndex);
-  };
-}
-__name(_ts_param2, "_ts_param");
-var AgentRunSteps = class {
-  static {
-    __name(this, "AgentRunSteps");
-  }
-  factory;
-  constructor(factory) {
-    this.factory = factory;
-  }
-  /**
-  * Retryable: every side effect (persist/quota/stream markers) happens loop-side under
-  * `hooks.step`, so a durable retry only re-runs the model call and re-streams its tokens — never
-  * double-writes. NEVER call `writer.end()`/`fail()` here: the workflow owns the stream's lifecycle,
-  * across however many turns/retries it takes.
-  */
-  llm(input) {
-    return outsideWorkflowCtx(() => this.runLlm(input));
-  }
-  async runLlm(input) {
-    const deps = this.factory.forAgent(input.agentName);
-    const tools = withMemoryTool({
-      tools: withSkillTool({
-        tools: withAskTool({
-          tools: await deps.registry.definitionsFor(input.actor, deps.rolesPolicy, deps.toolAllowList, {
-            ...input.threadId !== void 0 ? {
-              threadId: input.threadId
-            } : {},
-            ...input.agentName !== void 0 ? {
-              agentName: input.agentName
-            } : {}
-          }),
-          ask: deps.ask
-        }),
-        enabled: deps.skills !== void 0
-      }),
-      enabled: deps.memory?.provider.write !== void 0
-    });
-    const buffer = input.bufferOutput === true ? createFrameBuffer() : void 0;
-    let writer;
-    if (buffer !== void 0) {
-      writer = buffer.writer;
-    } else {
-      const opened = await deps.sink.open(input.sinkRunId);
-      writer = input.childSink ? childSinkWriter(opened) : opened;
-    }
-    const frames = observeTurnFrames(writer);
-    const turn = withTurnFrames(await traceLlmTurn(input.runId, input.step, () => deps.model.runTurn({
-      system: input.system,
-      messages: input.messages,
-      tools,
-      sink: frames.writer,
-      ...input.model !== void 0 ? {
-        model: input.model
-      } : {}
-    })), frames.summary());
-    const stamped = stampToolKinds(turn, deps);
-    return buffer === void 0 ? stamped : {
-      ...stamped,
-      bufferedFrames: buffer.frames()
-    };
-  }
-  /**
-  * NO DURABLE retries — a tool may not be idempotent (a bare `@Step()` carries none, unlike `llm`
-  * above). What it does retry, in place, is a CLASSIFIED-TRANSIENT error from the tool's own
-  * invocation (a DB deadlock, a lock-wait timeout — see `isTransientToolError`): the server rolled
-  * that work back, so re-invoking it is safe, and doing it here (inside this one step, never a new
-  * checkpoint) is exactly how the non-dispatched loop path handles it too. The timeout is applied
-  * HERE, handler-side, via `withToolTimeout` — never as a durable dispatch `timeoutMs`, which
-  * bounds engine liveness (retryable) rather than the tool's own business outcome — and per
-  * ATTEMPT, so a retried invocation gets its own fresh timeout window.
-  */
-  tool(input) {
-    return outsideWorkflowCtx(() => this.runTool(input));
-  }
-  async runTool(input) {
-    const deps = this.factory.forAgent(input.ctx.agentName);
-    if (deps.registry.spec(input.toolName)?.kind === "action" && input.toolType !== "action") {
-      throw new Error(`tool "${input.toolName}" is an action tool but was dispatched without approval: the dispatching process resolved its kind against a registry that does not have it`);
-    }
-    const sinkRunId = input.sinkRunId;
-    let opened;
-    const ui = createUiCollector(input.toolCallId, sinkRunId === void 0 ? void 0 : async (event) => {
-      opened ??= Promise.resolve(deps.sink.open(sinkRunId)).then((writer) => input.childSink === true ? childSinkWriter(writer) : writer);
-      await (await opened).write(encodeStreamEvent2(event));
-    });
-    const ctx = {
-      ...input.ctx,
-      emitUi: ui.emit
-    };
-    const localClassify = deps.toolTransientRetry !== false ? deps.toolTransientRetry?.classify : void 0;
-    const transientRetry = input.transientRetry === false ? false : {
-      attempts: input.transientRetry.attempts,
-      backoffMs: input.transientRetry.backoffMs,
-      ...localClassify !== void 0 ? {
-        classify: localClassify
-      } : {}
-    };
-    const output = await traceToolExecution(input.ctx.runId, {
-      toolCallId: input.toolCallId,
-      toolName: input.toolName,
-      toolType: input.toolType
-    }, () => invokeWithTransientRetry(() => {
-      ui.restart();
-      const invocation = deps.registry.invoke(input.toolName, input.input, ctx, deps.rolesPolicy);
-      return input.timeoutMs !== void 0 ? withToolTimeout(invocation, input.timeoutMs, input.toolName) : invocation;
-    }, transientRetry, {
-      onRetry: /* @__PURE__ */ __name((attempt, error) => {
-        publishAgentToolRetry({
-          toolName: input.toolName,
-          toolCallId: input.toolCallId,
-          attempt,
-          message: error instanceof Error ? error.message : String(error)
-        });
-      }, "onRetry")
-    }));
-    return input.collectUi === true ? wrapToolStepOutput(output, ui.components()) : output;
-  }
-};
-_ts_decorate2([
-  Step({
-    retries: 3
-  }),
-  _ts_metadata2("design:type", Function),
-  _ts_metadata2("design:paramtypes", [
-    typeof DispatchedLlmInput === "undefined" ? Object : DispatchedLlmInput
-  ]),
-  _ts_metadata2("design:returntype", typeof Promise === "undefined" ? Object : Promise)
-], AgentRunSteps.prototype, "llm", null);
-_ts_decorate2([
-  Step(),
-  _ts_metadata2("design:type", Function),
-  _ts_metadata2("design:paramtypes", [
-    typeof DispatchedToolInput === "undefined" ? Object : DispatchedToolInput
-  ]),
-  _ts_metadata2("design:returntype", typeof Promise === "undefined" ? Object : Promise)
-], AgentRunSteps.prototype, "tool", null);
-AgentRunSteps = _ts_decorate2([
-  Injectable2(),
-  _ts_param2(0, Inject2(AGENT_DEPS_FACTORY)),
-  _ts_metadata2("design:type", Function),
-  _ts_metadata2("design:paramtypes", [
-    typeof AgentDepsFactory === "undefined" ? Object : AgentDepsFactory
-  ])
-], AgentRunSteps);
-
-// src/durable/agent-run.workflow.ts
-function _ts_decorate3(decorators, target, key, desc) {
-  var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
-  if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
-    r = Reflect.decorate(decorators, target, key, desc);
-  } else {
-    for (var i = decorators.length - 1; i >= 0; i--) {
-      if (d = decorators[i]) {
-        r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
-      }
-    }
-  }
-  return c > 3 && r && Object.defineProperty(target, key, r), r;
-}
-__name(_ts_decorate3, "_ts_decorate");
-function _ts_metadata3(metadataKey, metadataValue) {
-  if (typeof Reflect === "object" && typeof Reflect.metadata === "function") {
-    return Reflect.metadata(metadataKey, metadataValue);
-  }
-}
-__name(_ts_metadata3, "_ts_metadata");
-function _ts_param3(paramIndex, decorator) {
-  return function(target, key) {
-    decorator(target, key, paramIndex);
-  };
-}
-__name(_ts_param3, "_ts_param");
-var AgentRunWorkflow = class _AgentRunWorkflow {
-  static {
-    __name(this, "AgentRunWorkflow");
-  }
-  factory;
-  store;
-  runs;
-  queue;
-  constructor(factory, store, runs, queue) {
-    this.factory = factory;
-    this.store = store;
-    this.runs = runs;
-    this.queue = queue;
-  }
-  /**
-  * Move the thread past this settling turn — to the next queued message, started here as a
-  * fire-and-forget `ctx.startChild` under the message's own id, or to a paused/empty queue — and
-  * answer the `queue` frame to write before the turn's terminal.
-  *
-  * Journaled: the decision (which message, claimed and popped) is one `localStep`, so a replay
-  * reads it back instead of popping a second message, and the start is the runtime's own
-  * replay-safe spawn. Gated by `ctx.patched`, so a run recorded before the queue existed replays
-  * against the history it has. A no-op for a sub-agent's run and on a store without a queue.
-  */
-  async advanceQueue(ctx, input, outcome, error) {
-    const queue = this.queue;
-    if (queue === void 0 || !queue.supported || input.sinkRunId !== void 0 || input.deliverTo !== void 0) {
-      return void 0;
-    }
-    if (!await ctx.patched("agent:chat-queue")) {
-      return void 0;
-    }
-    const plan = await stepOf(ctx)(`queue:${outcome}`, async () => {
-      try {
-        const settled = outcome === "completed" && await this.isCancelled(ctx.runId) ? "cancelled" : outcome;
-        return await queue.plan({
-          threadId: input.threadId,
-          runId: ctx.runId,
-          outcome: settled,
-          ...error !== void 0 ? {
-            error
-          } : {}
-        });
-      } catch {
-        return {};
-      }
-    });
-    const next = plan.next;
-    if (next === void 0) {
-      return plan.frame;
-    }
-    try {
-      await ctx.startChild(_AgentRunWorkflow, next.input, next.runId);
-      return plan.frame;
-    } catch (failure) {
-      if (isWorkflowControlFlowSignal(failure)) {
-        throw failure;
-      }
-      const reason = failure instanceof Error ? failure.message : String(failure);
-      return stepOf(ctx)("queue:restore", async () => {
-        await queue.restore(input.threadId, next, "start_failed", reason);
-        return queue.pausedFrame(input.threadId);
-      });
-    }
-  }
-  /**
-  * Has someone asked this run to stop? Read from the runtime's own run status, which is where
-  * `DurableAgentRunner.cancel` puts it — `cancelling` while a compensating cancel is in flight,
-  * `cancelled` once it has settled.
-  *
-  * Costs a run read at each of the loop's observation points (two per step). That is a row the
-  * control plane already has, against a turn that is otherwise making model calls, and it buys the
-  * property that matters: the answer comes from a place BOTH the canceller and every pod that might
-  * replay this run can see.
-  *
-  * Failing to ask is not a cancel. A tenant with no gateway wired, or an operator briefly
-  * unreachable, answers "no" and the run carries on — the alternative is failing healthy runs
-  * because a status read timed out. The loop journals that answer like any other, so two processes
-  * can never disagree about what was seen here.
-  */
-  async isCancelled(runId) {
-    try {
-      const status = (await this.runs?.getRunDetail(runId))?.run.status;
-      return status === "cancelled" || status === "cancelling";
-    } catch {
-      return false;
-    }
-  }
-  /**
-  * Settle the delegation a DETACHED run was started for, when that run ends without an answer.
-  *
-  * The loop delivers its own success (`deliver:detached`) but cannot catch its own crash, and the
-  * runtime settles the run row without knowing anything delegated it. Left alone, the card in the
-  * calling conversation says "started" for ever — the one state a reader can neither wait on nor
-  * act on. A no-op for every other run: only a detached one carries a delivery address.
-  */
-  async settleDetachedParent(ctx, input, outcome) {
-    const delivery = input.deliverTo;
-    if (delivery === void 0) {
-      return;
-    }
-    await stepOf(ctx)("deliver:detached:unsettled", () => settleUnsettledDelegation({
-      store: this.store,
-      delivery,
-      agent: input.agentName ?? "default",
-      runId: ctx.runId,
-      status: outcome.status,
-      ...outcome.error !== void 0 ? {
-        error: outcome.error
-      } : {}
-    }));
+    await stepOf(ctx)("deliver:detached:unsettled", () => settleUnsettledDelegation({
+      store: this.store,
+      delivery,
+      agent: input.agentName ?? "default",
+      runId: ctx.runId,
+      status: outcome.status,
+      ...outcome.error !== void 0 ? {
+        error: outcome.error
+      } : {}
+    }));
   }
   async run(ctx, input) {
     const day = input.day ?? utcDay();
@@ -806,7 +383,7 @@ var AgentRunWorkflow = class _AgentRunWorkflow {
           end: /* @__PURE__ */ __name(async () => {
             const frame = await this.advanceQueue(ctx, input, "completed");
             if (frame !== void 0) {
-              await writer.write(encodeStreamEvent3(frame));
+              await writer.write(encodeStreamEvent2(frame));
             }
             await writer.end();
           }, "end")
@@ -950,30 +527,42 @@ var AgentRunWorkflow = class _AgentRunWorkflow {
         throw error;
       }
       if (isReplayIntegrityError(error)) {
-        const failure = {
-          code: "run_failed",
-          message: error.message
-        };
+        const detail = error.message;
+        const code2 = agentFailureCode(error);
+        this.logger.error(`agent run ${ctx.runId} failed (${code2}): ${detail}`);
         publishAgentRunFailed({
           runId: ctx.runId,
-          ...failure
+          code: code2,
+          message: detail
         });
-        await (await hooks.openSink()).fail(failure);
+        await settleDeadRun(this.store, {
+          runId: ctx.runId,
+          threadId: input.threadId,
+          failure: {
+            code: code2,
+            message: detail
+          }
+        });
+        await (await hooks.openSink()).fail(streamFailure(error));
         throw error;
       }
       const message = error instanceof Error ? error.message : String(error);
       const code = agentFailureCode(error);
+      this.logger.error(`agent run ${ctx.runId} failed (${code}): ${message}`);
       publishAgentRunFailed({
         runId: ctx.runId,
         code,
         message
       });
-      await stepOf(ctx)("persist:run:fail", () => this.store.recordRunEnd?.({
-        runId: ctx.runId,
-        status: "failed",
-        errorCode: code,
-        errorMessage: message
-      }) ?? Promise.resolve());
+      await stepOf(ctx)("persist:run:fail", async () => {
+        await this.store.recordRunEnd?.({
+          runId: ctx.runId,
+          status: "failed",
+          errorCode: code,
+          errorMessage: message
+        });
+        await Promise.resolve(this.store.failUnsettledToolCalls?.(ctx.runId, RUN_ENDED_BEFORE_TOOL_CALL)).catch(() => 0);
+      });
       const queueFrame2 = await this.advanceQueue(ctx, input, "failed", message);
       await stepOf(ctx)("deactivate", () => releaseThreadRun(this.store, input.threadId, ctx.runId));
       await this.settleDetachedParent(ctx, input, {
@@ -982,29 +571,27 @@ var AgentRunWorkflow = class _AgentRunWorkflow {
       });
       const writer = await hooks.openSink();
       if (queueFrame2 !== void 0) {
-        await writer.write(encodeStreamEvent3(queueFrame2));
+        await writer.write(encodeStreamEvent2(queueFrame2));
       }
-      await writer.fail({
-        code,
-        message
-      });
+      await writer.fail(streamFailure(error));
       throw error;
     }
   }
 };
-AgentRunWorkflow = _ts_decorate3([
-  Injectable3(),
+AgentRunWorkflow = _ts_decorate2([
+  Injectable2(),
   Workflow({
     name: "agent.run",
     version: "1"
   }),
-  _ts_param3(0, Inject3(AGENT_DEPS_FACTORY2)),
-  _ts_param3(1, Inject3(AGENT_STORE2)),
-  _ts_param3(2, Optional2()),
-  _ts_param3(2, Inject3(RUN_GATEWAY)),
-  _ts_param3(3, Optional2()),
-  _ts_metadata3("design:type", Function),
-  _ts_metadata3("design:paramtypes", [
+  _ts_param2(0, Inject2(AGENT_DEPS_FACTORY2)),
+  _ts_param2(1, Inject2(AGENT_STORE)),
+  _ts_param2(2, Optional()),
+  _ts_param2(2, Inject2(RUN_GATEWAY)),
+  _ts_param2(3, Optional()),
+  _ts_param2(3, Inject2(AGENT_CHAT_QUEUE)),
+  _ts_metadata2("design:type", Function),
+  _ts_metadata2("design:paramtypes", [
     typeof AgentDepsFactory === "undefined" ? Object : AgentDepsFactory,
     typeof AgentStore === "undefined" ? Object : AgentStore,
     typeof RunGateway === "undefined" ? Object : RunGateway,
@@ -1018,10 +605,10 @@ __name(isSignalTimeout, "isSignalTimeout");
 
 // src/durable/durable-agent-runner.ts
 import { randomUUID } from "crypto";
-import { AGENT_SINK as AGENT_SINK2, AGENT_STORE as AGENT_STORE3, encodeStreamEvent as encodeStreamEvent4, releaseThreadRun as releaseThreadRun2 } from "@dudousxd/nestjs-agent-core";
+import { AGENT_SINK, AGENT_STORE as AGENT_STORE2, encodeStreamEvent as encodeStreamEvent3, releaseThreadRun as releaseThreadRun2 } from "@dudousxd/nestjs-agent-core";
 import { RUN_GATEWAY as RUN_GATEWAY2, WorkflowService } from "@dudousxd/nestjs-durable";
 import { isWorkflowControlFlowSignal as isWorkflowControlFlowSignal2 } from "@dudousxd/nestjs-durable-core";
-import { Inject as Inject4, Injectable as Injectable4, Logger as Logger2, Optional as Optional3 } from "@nestjs/common";
+import { Inject as Inject3, Injectable as Injectable3, Logger as Logger2, Optional as Optional2 } from "@nestjs/common";
 
 // src/in-process-sink.ts
 import { AgentStreamError } from "@dudousxd/nestjs-agent-core";
@@ -1092,7 +679,7 @@ var InProcessTokenStreamSink = class {
 };
 
 // src/durable/durable-agent-runner.ts
-function _ts_decorate4(decorators, target, key, desc) {
+function _ts_decorate3(decorators, target, key, desc) {
   var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
   if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
     r = Reflect.decorate(decorators, target, key, desc);
@@ -1105,19 +692,19 @@ function _ts_decorate4(decorators, target, key, desc) {
   }
   return c > 3 && r && Object.defineProperty(target, key, r), r;
 }
-__name(_ts_decorate4, "_ts_decorate");
-function _ts_metadata4(metadataKey, metadataValue) {
+__name(_ts_decorate3, "_ts_decorate");
+function _ts_metadata3(metadataKey, metadataValue) {
   if (typeof Reflect === "object" && typeof Reflect.metadata === "function") {
     return Reflect.metadata(metadataKey, metadataValue);
   }
 }
-__name(_ts_metadata4, "_ts_metadata");
-function _ts_param4(paramIndex, decorator) {
+__name(_ts_metadata3, "_ts_metadata");
+function _ts_param3(paramIndex, decorator) {
   return function(target, key) {
     decorator(target, key, paramIndex);
   };
 }
-__name(_ts_param4, "_ts_param");
+__name(_ts_param3, "_ts_param");
 function threadOfRun(input) {
   if (typeof input !== "object" || input === null || !("threadId" in input)) {
     return void 0;
@@ -1225,7 +812,7 @@ var DurableAgentRunner = class _DurableAgentRunner {
             runId: nextRunId
           }));
           if (frame !== void 0) {
-            await writer.write(encodeStreamEvent4(frame));
+            await writer.write(encodeStreamEvent3(frame));
           }
         } catch (error) {
           this.logger.error(`could not advance the queue of thread ${threadId} after cancelling ${runId}: ${error instanceof Error ? error.message : String(error)}`);
@@ -1233,7 +820,7 @@ var DurableAgentRunner = class _DurableAgentRunner {
       }
       await releaseThreadRun2(this.store, threadId, runId);
     }
-    await writer.write(encodeStreamEvent4({
+    await writer.write(encodeStreamEvent3({
       kind: "cancelled"
     }));
     await writer.end();
@@ -1260,14 +847,15 @@ var DurableAgentRunner = class _DurableAgentRunner {
     }
   }
 };
-DurableAgentRunner = _ts_decorate4([
-  Injectable4(),
-  _ts_param4(1, Inject4(RUN_GATEWAY2)),
-  _ts_param4(2, Inject4(AGENT_STORE3)),
-  _ts_param4(3, Inject4(AGENT_SINK2)),
-  _ts_param4(4, Optional3()),
-  _ts_metadata4("design:type", Function),
-  _ts_metadata4("design:paramtypes", [
+DurableAgentRunner = _ts_decorate3([
+  Injectable3(),
+  _ts_param3(1, Inject3(RUN_GATEWAY2)),
+  _ts_param3(2, Inject3(AGENT_STORE2)),
+  _ts_param3(3, Inject3(AGENT_SINK)),
+  _ts_param3(4, Optional2()),
+  _ts_param3(4, Inject3(AGENT_CHAT_QUEUE)),
+  _ts_metadata3("design:type", Function),
+  _ts_metadata3("design:paramtypes", [
     typeof WorkflowService === "undefined" ? Object : WorkflowService,
     typeof RunGateway === "undefined" ? Object : RunGateway,
     typeof AgentStore === "undefined" ? Object : AgentStore,
@@ -1279,7 +867,7 @@ DurableAgentRunner = _ts_decorate4([
 // src/durable/agent-durable.module.ts
 import { AGENT_DURABLE_RUNNER } from "@dudousxd/nestjs-agent-core";
 import { Global, Module } from "@nestjs/common";
-function _ts_decorate5(decorators, target, key, desc) {
+function _ts_decorate4(decorators, target, key, desc) {
   var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
   if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
     r = Reflect.decorate(decorators, target, key, desc);
@@ -1292,7 +880,7 @@ function _ts_decorate5(decorators, target, key, desc) {
   }
   return c > 3 && r && Object.defineProperty(target, key, r), r;
 }
-__name(_ts_decorate5, "_ts_decorate");
+__name(_ts_decorate4, "_ts_decorate");
 function runnerProviders() {
   return [
     DurableAgentRunner,
@@ -1308,7 +896,7 @@ var AgentDurableHttpModule = class AgentDurableHttpModule2 {
     __name(this, "AgentDurableHttpModule");
   }
 };
-AgentDurableHttpModule = _ts_decorate5([
+AgentDurableHttpModule = _ts_decorate4([
   Global(),
   Module({
     providers: [
@@ -1346,7 +934,7 @@ var AgentDurableModule = class _AgentDurableModule {
     };
   }
 };
-AgentDurableModule = _ts_decorate5([
+AgentDurableModule = _ts_decorate4([
   Global(),
   Module({
     providers: [
@@ -1364,14 +952,14 @@ AgentDurableModule = _ts_decorate5([
 
 // src/agent.module.ts
 import { AGENT_ACTOR_RESOLVER as AGENT_ACTOR_RESOLVER13, AGENT_APPROVAL_PORT, AGENT_DEPS_FACTORY as AGENT_DEPS_FACTORY7, AGENT_DURABLE_RUNNER as AGENT_DURABLE_RUNNER2, AGENT_MEMORY as AGENT_MEMORY3, AGENT_MODEL as AGENT_MODEL2, AGENT_MODEL_CATALOG as AGENT_MODEL_CATALOG3, AGENT_OPTIONS as AGENT_OPTIONS7, AGENT_PROMPT_CONTRIBUTORS as AGENT_PROMPT_CONTRIBUTORS3, AGENT_QUOTA_PROVIDER as AGENT_QUOTA_PROVIDER3, AGENT_REGISTRY as AGENT_REGISTRY6, AGENT_ROLES_POLICY as AGENT_ROLES_POLICY2, AGENT_RUNNER as AGENT_RUNNER2, AGENT_SINK as AGENT_SINK4, AGENT_SKILLS as AGENT_SKILLS3, AGENT_SKILL_SOURCES as AGENT_SKILL_SOURCES2, AGENT_STORE as AGENT_STORE7, AGENT_TOOL_REGISTRY as AGENT_TOOL_REGISTRY3, AgentRegistry as AgentRegistry6, DefaultRolesPolicy, InMemoryAgentStore, ToolRegistry as ToolRegistry3 } from "@dudousxd/nestjs-agent-core";
-import { Global as Global2, Logger as Logger7, Module as Module2 } from "@nestjs/common";
+import { Global as Global2, Logger as Logger8, Module as Module2 } from "@nestjs/common";
 import { DiscoveryModule, ModulesContainer, RouterModule } from "@nestjs/core";
 
 // src/agent-deps.factory.ts
-import { AGENT_MEMORY, AGENT_MODEL, AGENT_PRICING_STORE, AGENT_PROMPT_CONTRIBUTORS, AGENT_REGISTRY, AGENT_ROLES_POLICY, AGENT_SINK as AGENT_SINK3, AGENT_SKILLS, AGENT_STORE as AGENT_STORE4, AGENT_TOOL_REGISTRY, AgentRegistry as AgentRegistry2, ToolRegistry as ToolRegistry2, normalizeDelegation, summarizeWithModel, windowHistory } from "@dudousxd/nestjs-agent-core";
-import { AGENT_OPTIONS as AGENT_OPTIONS2 } from "@dudousxd/nestjs-agent-core";
-import { Inject as Inject5, Injectable as Injectable5, Optional as Optional4 } from "@nestjs/common";
-function _ts_decorate6(decorators, target, key, desc) {
+import { AGENT_MEMORY, AGENT_MODEL, AGENT_PRICING_STORE, AGENT_PROMPT_CONTRIBUTORS, AGENT_REGISTRY, AGENT_ROLES_POLICY, AGENT_SINK as AGENT_SINK2, AGENT_SKILLS, AGENT_STORE as AGENT_STORE3, AGENT_TOOL_REGISTRY, AgentRegistry as AgentRegistry2, ToolRegistry as ToolRegistry2, normalizeDelegation, summarizeWithModel, windowHistory } from "@dudousxd/nestjs-agent-core";
+import { AGENT_OPTIONS } from "@dudousxd/nestjs-agent-core";
+import { Inject as Inject4, Injectable as Injectable4, Optional as Optional3 } from "@nestjs/common";
+function _ts_decorate5(decorators, target, key, desc) {
   var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
   if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
     r = Reflect.decorate(decorators, target, key, desc);
@@ -1384,19 +972,19 @@ function _ts_decorate6(decorators, target, key, desc) {
   }
   return c > 3 && r && Object.defineProperty(target, key, r), r;
 }
-__name(_ts_decorate6, "_ts_decorate");
-function _ts_metadata5(metadataKey, metadataValue) {
+__name(_ts_decorate5, "_ts_decorate");
+function _ts_metadata4(metadataKey, metadataValue) {
   if (typeof Reflect === "object" && typeof Reflect.metadata === "function") {
     return Reflect.metadata(metadataKey, metadataValue);
   }
 }
-__name(_ts_metadata5, "_ts_metadata");
-function _ts_param5(paramIndex, decorator) {
+__name(_ts_metadata4, "_ts_metadata");
+function _ts_param4(paramIndex, decorator) {
   return function(target, key) {
     decorator(target, key, paramIndex);
   };
 }
-__name(_ts_param5, "_ts_param");
+__name(_ts_param4, "_ts_param");
 function delegateToolName(args) {
   const slug = args.target.replace(/[^a-zA-Z0-9]+/g, "_");
   return args.detached === true ? `start_${slug}` : `ask_${slug}`;
@@ -1430,204 +1018,651 @@ var AgentDepsFactory2 = class {
     this.skills = skills;
     this.memory = memory;
   }
-  /** The turn's memory seam, shared with the read-back endpoint. Undefined → memory is not configured. */
-  memoryConfig() {
-    return this.memory;
+  /** The turn's memory seam, shared with the read-back endpoint. Undefined → memory is not configured. */
+  memoryConfig() {
+    return this.memory;
+  }
+  /** The turn's skills seam, shared with the listing endpoint. Undefined → skills are not configured. */
+  skillsConfig() {
+    return this.skills;
+  }
+  /**
+  * The agent a turn uses when the caller names none: the explicit `defaultAgent` option, else the
+  * sole registered `@Agent` when there is exactly one, else `'default'` (a bare assistant).
+  */
+  defaultAgentName() {
+    if (this.options.defaultAgent !== void 0) {
+      return this.options.defaultAgent;
+    }
+    const registered = this.agents.list();
+    return registered.length === 1 ? registered[0]?.name ?? "default" : "default";
+  }
+  effectiveTools(definition) {
+    if (definition === void 0) {
+      return void 0;
+    }
+    const delegated = (definition.delegatesTo ?? []).map((edge) => {
+      const { agent, detached } = normalizeDelegation(edge);
+      return delegateToolName({
+        target: agent,
+        detached
+      });
+    });
+    if (definition.tools === void 0 && delegated.length === 0) {
+      return void 0;
+    }
+    return [
+      ...definition.tools ?? [],
+      ...delegated
+    ];
+  }
+  forAgent(agentName) {
+    const name = agentName ?? this.defaultAgentName();
+    const definition = this.agents.get(name);
+    const toolAllowList = this.effectiveTools(definition);
+    const followUpsCount = this.followUpsCount();
+    const retrieval = this.options.retrieval;
+    const historyPolicy = this.historyPolicy(definition);
+    const ask = definition?.ask ?? this.options.ask;
+    return {
+      model: this.model,
+      store: this.store,
+      sink: this.sink,
+      rolesPolicy: this.rolesPolicy,
+      ...this.options.approvalPolicy !== void 0 ? {
+        approvalPolicy: this.options.approvalPolicy
+      } : {},
+      registry: this.registry,
+      promptContributors: this.promptContributors,
+      systemPrompt: definition?.systemPrompt ?? this.options.systemPrompt ?? "You are a helpful assistant.",
+      maxSteps: definition?.maxSteps ?? 8,
+      ...definition?.maxDelegationDepth !== void 0 ? {
+        maxDelegationDepth: definition.maxDelegationDepth
+      } : {},
+      ...definition?.maxAgentAppearances !== void 0 ? {
+        maxAgentAppearances: definition.maxAgentAppearances
+      } : {},
+      ...definition?.modelId !== void 0 ? {
+        modelId: definition.modelId
+      } : {},
+      ...this.pricingStore !== void 0 ? {
+        pricingStore: this.pricingStore
+      } : {},
+      ...historyPolicy !== void 0 ? {
+        historyPolicy
+      } : {},
+      inputProcessors: this.options.inputProcessors ?? [],
+      outputProcessors: this.options.outputProcessors ?? [],
+      ...definition?.outputSchema !== void 0 ? {
+        outputSchema: definition.outputSchema
+      } : {},
+      ...definition?.outputRepairAttempts !== void 0 ? {
+        outputRepairAttempts: definition.outputRepairAttempts
+      } : {},
+      ...definition?.intake !== void 0 ? {
+        intake: definition.intake
+      } : {},
+      // Most specific wins, exactly as `historyPolicy` resolves: the agent's own `@Agent({ ask })`,
+      // else the module-wide flag. Both are module config, so every process of a deployment agrees.
+      ...ask === true ? {
+        ask: true
+      } : {},
+      // Module-wide, never per-agent: a skill is a procedure, not a persona, and an agent that could
+      // opt out of one would make the catalog depend on which agent a turn happened to select — a
+      // per-turn fact the worker re-deriving the tool list cannot see.
+      ...this.skills !== void 0 ? {
+        skills: this.skills
+      } : {},
+      // Module-wide, never per-agent, for the same reason skills are: which memories a turn carries
+      // cannot depend on which persona answered it, or the worker re-deriving a dispatched turn's
+      // tool list would disagree with the loop about whether `remember` was offered.
+      ...this.memory !== void 0 ? {
+        memory: this.memory
+      } : {},
+      ...toolAllowList !== void 0 ? {
+        toolAllowList
+      } : {},
+      ...this.options.toolTimeoutMs !== void 0 ? {
+        toolTimeoutMs: this.options.toolTimeoutMs
+      } : {},
+      ...this.options.toolTransientRetry !== void 0 ? {
+        toolTransientRetry: this.options.toolTransientRetry
+      } : {},
+      ...followUpsCount !== void 0 ? {
+        followUpsCount
+      } : {},
+      ...retrieval?.mode === "inject" ? {
+        retriever: retrieval.retriever,
+        ...retrieval.topK !== void 0 ? {
+          retrievalTopK: retrieval.topK
+        } : {}
+      } : {}
+    };
+  }
+  /**
+  * The ceiling on how much of a thread this agent's turns carry, most specific wins: the agent's
+  * own `@Agent({ history })`, else a module-wide custom `historyPolicy`, else the module-wide
+  * `history` window. Undefined → unbounded, the behaviour of a module that configures none.
+  */
+  historyPolicy(definition) {
+    if (definition?.history !== void 0) {
+      return this.window(definition.history);
+    }
+    if (this.options.historyPolicy !== void 0) {
+      return this.options.historyPolicy;
+    }
+    return this.options.history === void 0 ? void 0 : this.window(this.options.history);
+  }
+  /** Build the built-in window policy from the declarative `{ maxMessages, maxTokens, summarize }`. */
+  window(config) {
+    return windowHistory({
+      ...config.maxMessages !== void 0 ? {
+        maxMessages: config.maxMessages
+      } : {},
+      ...config.maxTokens !== void 0 ? {
+        maxTokens: config.maxTokens
+      } : {},
+      // The shared module model, not a per-agent one: an agent's `modelId` is an accounting label,
+      // and only one provider is ever bound.
+      ...config.summarize === true ? {
+        summarize: summarizeWithModel(this.model)
+      } : {}
+    });
+  }
+  /** Normalize the `followUps` option (`true` → 3, `{ count }` → count) to a number, or undefined. */
+  followUpsCount() {
+    const followUps = this.options.followUps;
+    if (followUps === true) {
+      return 3;
+    }
+    if (typeof followUps === "object") {
+      return followUps.count;
+    }
+    return void 0;
+  }
+};
+AgentDepsFactory2 = _ts_decorate5([
+  Injectable4(),
+  _ts_param4(0, Inject4(AGENT_OPTIONS)),
+  _ts_param4(1, Inject4(AGENT_MODEL)),
+  _ts_param4(2, Inject4(AGENT_STORE3)),
+  _ts_param4(3, Inject4(AGENT_SINK2)),
+  _ts_param4(4, Inject4(AGENT_ROLES_POLICY)),
+  _ts_param4(5, Inject4(AGENT_TOOL_REGISTRY)),
+  _ts_param4(6, Inject4(AGENT_REGISTRY)),
+  _ts_param4(7, Inject4(AGENT_PROMPT_CONTRIBUTORS)),
+  _ts_param4(8, Optional3()),
+  _ts_param4(8, Inject4(AGENT_PRICING_STORE)),
+  _ts_param4(9, Optional3()),
+  _ts_param4(9, Inject4(AGENT_SKILLS)),
+  _ts_param4(10, Optional3()),
+  _ts_param4(10, Inject4(AGENT_MEMORY)),
+  _ts_metadata4("design:type", Function),
+  _ts_metadata4("design:paramtypes", [
+    typeof AgentModuleOptions === "undefined" ? Object : AgentModuleOptions,
+    typeof ModelProvider === "undefined" ? Object : ModelProvider,
+    typeof AgentStore === "undefined" ? Object : AgentStore,
+    typeof TokenStreamSink === "undefined" ? Object : TokenStreamSink,
+    typeof RolesPolicy === "undefined" ? Object : RolesPolicy,
+    typeof ToolRegistry2 === "undefined" ? Object : ToolRegistry2,
+    typeof AgentRegistry2 === "undefined" ? Object : AgentRegistry2,
+    Array,
+    Object,
+    Object,
+    Object
+  ])
+], AgentDepsFactory2);
+
+// src/agent.service.ts
+import { AGENT_ATTACHMENT_STAGING as AGENT_ATTACHMENT_STAGING2, AGENT_DEPS_FACTORY as AGENT_DEPS_FACTORY3, AGENT_MODEL_CATALOG, AGENT_OPTIONS as AGENT_OPTIONS3, AGENT_QUOTA_PROVIDER as AGENT_QUOTA_PROVIDER2, AGENT_RUNNER, AGENT_STORE as AGENT_STORE5, REQUESTER_APPROVER, findCatalogModel, mayDecideApproval, readElicitationQuestions, settleDeadRun as settleDeadRun2, validateElicitationAnswer } from "@dudousxd/nestjs-agent-core";
+import { BadRequestException, ConflictException as ConflictException2, ForbiddenException, GoneException, HttpException, HttpStatus, Inject as Inject6, Injectable as Injectable6, NotFoundException, NotImplementedException as NotImplementedException2, Optional as Optional5 } from "@nestjs/common";
+
+// src/queue/chat-queue.service.ts
+import { AGENT_ATTACHMENT_STAGING, AGENT_OPTIONS as AGENT_OPTIONS2, AGENT_QUOTA_PROVIDER, AGENT_SINK as AGENT_SINK3, AGENT_STORE as AGENT_STORE4, encodeStreamEvent as encodeStreamEvent4, isChatQueueStore, queuedMessageView } from "@dudousxd/nestjs-agent-core";
+import { Inject as Inject5, Injectable as Injectable5, Logger as Logger3, NotImplementedException, Optional as Optional4 } from "@nestjs/common";
+function _ts_decorate6(decorators, target, key, desc) {
+  var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
+  if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
+    r = Reflect.decorate(decorators, target, key, desc);
+  } else {
+    for (var i = decorators.length - 1; i >= 0; i--) {
+      if (d = decorators[i]) {
+        r = (c < 3 ? d(r) : c > 3 ? d(target, key, r) : d(target, key)) || r;
+      }
+    }
+  }
+  return c > 3 && r && Object.defineProperty(target, key, r), r;
+}
+__name(_ts_decorate6, "_ts_decorate");
+function _ts_metadata5(metadataKey, metadataValue) {
+  if (typeof Reflect === "object" && typeof Reflect.metadata === "function") {
+    return Reflect.metadata(metadataKey, metadataValue);
+  }
+}
+__name(_ts_metadata5, "_ts_metadata");
+function _ts_param5(paramIndex, decorator) {
+  return function(target, key) {
+    decorator(target, key, paramIndex);
+  };
+}
+__name(_ts_param5, "_ts_param");
+var MAX_DRAIN_ATTEMPTS = 8;
+var ChatQueueService2 = class _ChatQueueService {
+  static {
+    __name(this, "ChatQueueService");
+  }
+  store;
+  sink;
+  quotaProvider;
+  options;
+  staging;
+  logger = new Logger3(_ChatQueueService.name);
+  /**
+  * Runs this process claimed a thread for and has not handed to a runner yet. A holder in here is
+  * alive whatever the runner says — it simply has not registered the run yet.
+  */
+  starting = /* @__PURE__ */ new Set();
+  constructor(store, sink, quotaProvider, options, staging) {
+    this.store = store;
+    this.sink = sink;
+    this.quotaProvider = quotaProvider;
+    this.options = options;
+    this.staging = staging;
+  }
+  /** Whether the bound store can hold a queue. Without one, a busy thread is not queued. */
+  get supported() {
+    return isChatQueueStore(this.store);
   }
-  /** The turn's skills seam, shared with the listing endpoint. Undefined → skills are not configured. */
-  skillsConfig() {
-    return this.skills;
+  /** The bound store as a queue store, or a 501 naming what is missing. */
+  queueStore() {
+    if (!isChatQueueStore(this.store)) {
+      throw new NotImplementedException("Queueing messages requires an AgentStore that implements ChatQueueStore (enqueueMessage, listQueue, \u2026, claimActiveStream, releaseActiveStream); the bound store does not.");
+    }
+    return this.store;
+  }
+  async state(threadId) {
+    const store = this.queueStore();
+    const [items, paused] = await Promise.all([
+      store.listQueue(threadId),
+      store.queuePause(threadId)
+    ]);
+    return {
+      items: items.map(queuedMessageView),
+      paused
+    };
   }
   /**
-  * The agent a turn uses when the caller names none: the explicit `defaultAgent` option, else the
-  * sole registered `@Agent` when there is exactly one, else `'default'` (a bare assistant).
+  * Whether `holder` is still running. A run this process is about to start is; otherwise the
+  * runner decides, and a runner that cannot tell says yes.
   */
-  defaultAgentName() {
-    if (this.options.defaultAgent !== void 0) {
-      return this.options.defaultAgent;
+  async isAlive(holder, runner) {
+    if (this.starting.has(holder)) {
+      return true;
+    }
+    if (typeof runner.isRunActive !== "function") {
+      return true;
+    }
+    try {
+      return await runner.isRunActive(holder);
+    } catch {
+      return true;
     }
-    const registered = this.agents.list();
-    return registered.length === 1 ? registered[0]?.name ?? "default" : "default";
   }
-  effectiveTools(definition) {
-    if (definition === void 0) {
-      return void 0;
+  /**
+  * The run holding `threadId`, or `null` when it is free — a holder the runner reports dead counts
+  * as free, and is returned as `stale` so the caller can claim over it.
+  */
+  async holder(threadId, runner) {
+    const holder = await this.queueStore().activeRunForThread(threadId);
+    if (holder === null) {
+      return {
+        live: null,
+        stale: null
+      };
     }
-    const delegated = (definition.delegatesTo ?? []).map((edge) => {
-      const { agent, detached } = normalizeDelegation(edge);
-      return delegateToolName({
-        target: agent,
-        detached
-      });
-    });
-    if (definition.tools === void 0 && delegated.length === 0) {
-      return void 0;
+    return await this.isAlive(holder, runner) ? {
+      live: holder,
+      stale: null
+    } : {
+      live: null,
+      stale: holder
+    };
+  }
+  /**
+  * Decide what follows a TOP-LEVEL run that is settling on `threadId`, and move the thread there:
+  * hand it to the next queued message (claimed and popped here, started by the caller), or pause
+  * the queue, or release the thread. Call it BEFORE the run's terminal frame, so a client reading
+  * the stream learns what comes next from the `queue` frame, and so the thread is already free
+  * when the client sees the run end.
+  *
+  * A no-op returning `{}` on a store without a queue.
+  */
+  async plan(args) {
+    if (!isChatQueueStore(this.store)) {
+      return {};
     }
-    return [
-      ...definition.tools ?? [],
-      ...delegated
-    ];
+    const store = this.store;
+    const { threadId, outcome } = args;
+    if (await store.activeRunForThread(threadId) !== args.runId) {
+      return {};
+    }
+    let holding = args.runId;
+    for (let attempt = 0; attempt < MAX_DRAIN_ATTEMPTS; attempt += 1) {
+      const items = await store.listQueue(threadId);
+      const head = items[0];
+      if (head === void 0) {
+        if (holding !== null) {
+          await store.releaseActiveStream(threadId, holding);
+          holding = null;
+          continue;
+        }
+        return {};
+      }
+      let pause = await store.queuePause(threadId);
+      if (pause === null) {
+        const reason = pauseReasonFor(outcome, head);
+        if (reason !== null) {
+          pause = {
+            reason,
+            ...outcome === "failed" && args.error !== void 0 ? {
+              message: args.error
+            } : {},
+            at: (/* @__PURE__ */ new Date()).toISOString()
+          };
+          await store.setQueuePause(threadId, pause);
+        }
+      }
+      if (pause === null) {
+        const blocked = await this.quotaBlock(head.actor);
+        if (blocked !== null) {
+          pause = {
+            reason: "quota_exceeded",
+            message: blocked,
+            at: (/* @__PURE__ */ new Date()).toISOString()
+          };
+          await store.setQueuePause(threadId, pause);
+        }
+      }
+      if (pause !== null) {
+        if (holding !== null) {
+          await store.releaseActiveStream(threadId, holding);
+        }
+        return {
+          frame: queueFrame(items, pause)
+        };
+      }
+      const claimed = await store.claimActiveStream(threadId, head.id, holding !== null ? {
+        replacing: holding
+      } : {});
+      if (!claimed) {
+        return {};
+      }
+      holding = head.id;
+      if (!await store.removeQueuedMessage(head.id)) {
+        await store.releaseActiveStream(threadId, head.id);
+        holding = null;
+        continue;
+      }
+      const next = {
+        runId: head.id,
+        messageId: head.id,
+        input: await this.inputFor(head)
+      };
+      return {
+        next,
+        frame: queueFrame(items.slice(1), null, {
+          messageId: head.id,
+          runId: head.id
+        })
+      };
+    }
+    if (holding !== null) {
+      await store.releaseActiveStream(threadId, holding);
+    }
+    return {};
   }
-  forAgent(agentName) {
-    const name = agentName ?? this.defaultAgentName();
-    const definition = this.agents.get(name);
-    const toolAllowList = this.effectiveTools(definition);
-    const followUpsCount = this.followUpsCount();
-    const retrieval = this.options.retrieval;
-    const historyPolicy = this.historyPolicy(definition);
-    const ask = definition?.ask ?? this.options.ask;
-    return {
-      model: this.model,
-      store: this.store,
-      sink: this.sink,
-      rolesPolicy: this.rolesPolicy,
-      ...this.options.approvalPolicy !== void 0 ? {
-        approvalPolicy: this.options.approvalPolicy
-      } : {},
-      registry: this.registry,
-      promptContributors: this.promptContributors,
-      systemPrompt: definition?.systemPrompt ?? this.options.systemPrompt ?? "You are a helpful assistant.",
-      maxSteps: definition?.maxSteps ?? 8,
-      ...definition?.maxDelegationDepth !== void 0 ? {
-        maxDelegationDepth: definition.maxDelegationDepth
-      } : {},
-      ...definition?.maxAgentAppearances !== void 0 ? {
-        maxAgentAppearances: definition.maxAgentAppearances
-      } : {},
-      ...definition?.modelId !== void 0 ? {
-        modelId: definition.modelId
-      } : {},
-      ...this.pricingStore !== void 0 ? {
-        pricingStore: this.pricingStore
-      } : {},
-      ...historyPolicy !== void 0 ? {
-        historyPolicy
-      } : {},
-      inputProcessors: this.options.inputProcessors ?? [],
-      outputProcessors: this.options.outputProcessors ?? [],
-      ...definition?.outputSchema !== void 0 ? {
-        outputSchema: definition.outputSchema
-      } : {},
-      ...definition?.outputRepairAttempts !== void 0 ? {
-        outputRepairAttempts: definition.outputRepairAttempts
-      } : {},
-      ...definition?.intake !== void 0 ? {
-        intake: definition.intake
-      } : {},
-      // Most specific wins, exactly as `historyPolicy` resolves: the agent's own `@Agent({ ask })`,
-      // else the module-wide flag. Both are module config, so every process of a deployment agrees.
-      ...ask === true ? {
-        ask: true
-      } : {},
-      // Module-wide, never per-agent: a skill is a procedure, not a persona, and an agent that could
-      // opt out of one would make the catalog depend on which agent a turn happened to select — a
-      // per-turn fact the worker re-deriving the tool list cannot see.
-      ...this.skills !== void 0 ? {
-        skills: this.skills
-      } : {},
-      // Module-wide, never per-agent, for the same reason skills are: which memories a turn carries
-      // cannot depend on which persona answered it, or the worker re-deriving a dispatched turn's
-      // tool list would disagree with the loop about whether `remember` was offered.
-      ...this.memory !== void 0 ? {
-        memory: this.memory
-      } : {},
-      ...toolAllowList !== void 0 ? {
-        toolAllowList
-      } : {},
-      ...this.options.toolTimeoutMs !== void 0 ? {
-        toolTimeoutMs: this.options.toolTimeoutMs
+  /**
+  * {@link plan}, then start what it picked with `start`. What an in-process settle (the inline
+  * runner, a durable cancel issued outside a workflow) calls; answers the frame to write.
+  */
+  async handoff(args, start) {
+    const plan = await this.plan(args);
+    if (plan.next === void 0) {
+      return plan.frame;
+    }
+    return await this.launch(args.threadId, plan.next, start) ? plan.frame : this.pausedFrame(args.threadId);
+  }
+  /**
+  * Start the head of an IDLE thread's queue — after a send queued behind a run that turned out to
+  * be gone, after a resume, after a stale holder. Does nothing while a live run holds the thread
+  * (its settle drains) or while the queue is paused. Answers the started run's id, if any.
+  */
+  async kick(threadId, runner) {
+    if (!isChatQueueStore(this.store)) {
+      return void 0;
+    }
+    const { live, stale } = await this.holder(threadId, runner);
+    if (live !== null) {
+      return void 0;
+    }
+    if (stale !== null) {
+      this.logger.warn(`thread ${threadId} was held by run ${stale}, which is no longer running`);
+    }
+    const plan = await this.plan({
+      threadId,
+      runId: stale,
+      outcome: "completed"
+    });
+    if (plan.next === void 0) {
+      return void 0;
+    }
+    const started = await this.launch(threadId, plan.next, (input, runId) => runner.start(input, {
+      runId
+    }));
+    if (started && plan.frame !== void 0) {
+      await this.write(plan.next.runId, plan.frame);
+    }
+    return started ? plan.next.runId : void 0;
+  }
+  /**
+  * Start a turn the thread is already claimed for. On failure the message goes back to the head of
+  * the queue, the queue pauses (`start_failed`), and the claim is released — nothing is lost and
+  * nothing is left holding the thread.
+  */
+  async launch(threadId, next, start) {
+    this.starting.add(next.runId);
+    try {
+      await start(next.input, next.runId);
+      return true;
+    } catch (error) {
+      const message = error instanceof Error ? error.message : String(error);
+      this.logger.error(`queued message ${next.messageId} could not start: ${message}`);
+      await this.restore(threadId, next, "start_failed", message);
+      return false;
+    } finally {
+      this.starting.delete(next.runId);
+    }
+  }
+  /** Put a turn that never started back at the head of its queue, paused, and free the thread. */
+  async restore(threadId, next, reason, message) {
+    const store = this.queueStore();
+    const { input } = next;
+    await store.enqueueMessage({
+      threadId,
+      actor: input.actor,
+      content: input.userText,
+      at: "head",
+      ...input.attachments !== void 0 ? {
+        attachments: input.attachments
       } : {},
-      ...this.options.toolTransientRetry !== void 0 ? {
-        toolTransientRetry: this.options.toolTransientRetry
+      ...input.agentName !== void 0 ? {
+        agentName: input.agentName
       } : {},
-      ...followUpsCount !== void 0 ? {
-        followUpsCount
+      ...input.model !== void 0 ? {
+        model: input.model
       } : {},
-      ...retrieval?.mode === "inject" ? {
-        retriever: retrieval.retriever,
-        ...retrieval.topK !== void 0 ? {
-          retrievalTopK: retrieval.topK
-        } : {}
+      ...input.pageContext !== void 0 ? {
+        pageContext: input.pageContext
       } : {}
+    });
+    await store.setQueuePause(threadId, {
+      reason,
+      message,
+      at: (/* @__PURE__ */ new Date()).toISOString()
+    });
+    await store.releaseActiveStream(threadId, next.runId);
+  }
+  /** A `queue` frame with the thread's current state. */
+  async pausedFrame(threadId) {
+    return {
+      kind: "queue",
+      queue: await this.state(threadId)
     };
   }
   /**
-  * The ceiling on how much of a thread this agent's turns carry, most specific wins: the agent's
-  * own `@Agent({ history })`, else a module-wide custom `historyPolicy`, else the module-wide
-  * `history` window. Undefined → unbounded, the behaviour of a module that configures none.
+  * Tell whoever is watching the thread that its queue changed: a `queue` frame into the stream of
+  * the run holding it. Best-effort — a thread nobody is streaming has no one to tell, and the next
+  * read of the thread carries the queue anyway.
   */
-  historyPolicy(definition) {
-    if (definition?.history !== void 0) {
-      return this.window(definition.history);
+  async publish(threadId) {
+    const state = await this.state(threadId);
+    const holder = await this.queueStore().activeRunForThread(threadId);
+    if (holder !== null) {
+      await this.write(holder, {
+        kind: "queue",
+        queue: state
+      }).catch((error) => {
+        this.logger.warn(`could not publish the queue of thread ${threadId}: ${error instanceof Error ? error.message : String(error)}`);
+      });
     }
-    if (this.options.historyPolicy !== void 0) {
-      return this.options.historyPolicy;
+    return state;
+  }
+  async write(runId, frame) {
+    const writer = await this.sink.open(runId);
+    await writer.write(encodeStreamEvent4(frame));
+  }
+  /** Why `actor` may not start a turn now (the quota message), or `null`. */
+  async quotaBlock(actor) {
+    if (this.options?.quota === void 0 || this.quotaProvider === void 0) {
+      return null;
+    }
+    try {
+      const { blocked } = await this.quotaProvider.report({
+        actor
+      });
+      if (blocked === void 0) {
+        return null;
+      }
+      return blocked.reason ?? `The ${blocked.period === "day" ? "daily" : "monthly"} quota is used up`;
+    } catch {
+      return null;
     }
-    return this.options.history === void 0 ? void 0 : this.window(this.options.history);
   }
-  /** Build the built-in window policy from the declarative `{ maxMessages, maxTokens, summarize }`. */
-  window(config) {
-    return windowHistory({
-      ...config.maxMessages !== void 0 ? {
-        maxMessages: config.maxMessages
+  /**
+  * The turn a queued message starts: everything it was queued with, today's date, and its
+  * attachments' urls minted afresh — the ones stored at enqueue time may have expired while it
+  * waited. An attachment the staging store no longer resolves keeps its stored url.
+  */
+  async inputFor(message) {
+    const attachments = await this.freshAttachments(message.actor, message.attachments ?? []);
+    return {
+      threadId: message.threadId,
+      actor: message.actor,
+      userText: message.content,
+      day: utcDay(),
+      ...message.agentName !== void 0 ? {
+        agentName: message.agentName
       } : {},
-      ...config.maxTokens !== void 0 ? {
-        maxTokens: config.maxTokens
+      ...attachments.length > 0 ? {
+        attachments
       } : {},
-      // The shared module model, not a per-agent one: an agent's `modelId` is an accounting label,
-      // and only one provider is ever bound.
-      ...config.summarize === true ? {
-        summarize: summarizeWithModel(this.model)
+      ...message.pageContext !== void 0 ? {
+        pageContext: message.pageContext
+      } : {},
+      ...message.model !== void 0 ? {
+        model: message.model
       } : {}
-    });
+    };
   }
-  /** Normalize the `followUps` option (`true` → 3, `{ count }` → count) to a number, or undefined. */
-  followUpsCount() {
-    const followUps = this.options.followUps;
-    if (followUps === true) {
-      return 3;
-    }
-    if (typeof followUps === "object") {
-      return followUps.count;
+  async freshAttachments(actor, attachments) {
+    const staging = this.staging;
+    if (attachments.length === 0 || staging === void 0 || typeof staging.resolve !== "function") {
+      return attachments;
     }
-    return void 0;
+    return Promise.all(attachments.map(async (attachment) => {
+      try {
+        return await staging.resolve({
+          mediaId: attachment.mediaId,
+          actor
+        }) ?? attachment;
+      } catch {
+        return attachment;
+      }
+    }));
   }
 };
-AgentDepsFactory2 = _ts_decorate6([
+ChatQueueService2 = _ts_decorate6([
   Injectable5(),
-  _ts_param5(0, Inject5(AGENT_OPTIONS2)),
-  _ts_param5(1, Inject5(AGENT_MODEL)),
-  _ts_param5(2, Inject5(AGENT_STORE4)),
-  _ts_param5(3, Inject5(AGENT_SINK3)),
-  _ts_param5(4, Inject5(AGENT_ROLES_POLICY)),
-  _ts_param5(5, Inject5(AGENT_TOOL_REGISTRY)),
-  _ts_param5(6, Inject5(AGENT_REGISTRY)),
-  _ts_param5(7, Inject5(AGENT_PROMPT_CONTRIBUTORS)),
-  _ts_param5(8, Optional4()),
-  _ts_param5(8, Inject5(AGENT_PRICING_STORE)),
-  _ts_param5(9, Optional4()),
-  _ts_param5(9, Inject5(AGENT_SKILLS)),
-  _ts_param5(10, Optional4()),
-  _ts_param5(10, Inject5(AGENT_MEMORY)),
+  _ts_param5(0, Inject5(AGENT_STORE4)),
+  _ts_param5(1, Inject5(AGENT_SINK3)),
+  _ts_param5(2, Optional4()),
+  _ts_param5(2, Inject5(AGENT_QUOTA_PROVIDER)),
+  _ts_param5(3, Optional4()),
+  _ts_param5(3, Inject5(AGENT_OPTIONS2)),
+  _ts_param5(4, Optional4()),
+  _ts_param5(4, Inject5(AGENT_ATTACHMENT_STAGING)),
   _ts_metadata5("design:type", Function),
   _ts_metadata5("design:paramtypes", [
-    typeof AgentModuleOptions === "undefined" ? Object : AgentModuleOptions,
-    typeof ModelProvider === "undefined" ? Object : ModelProvider,
     typeof AgentStore === "undefined" ? Object : AgentStore,
     typeof TokenStreamSink === "undefined" ? Object : TokenStreamSink,
-    typeof RolesPolicy === "undefined" ? Object : RolesPolicy,
-    typeof ToolRegistry2 === "undefined" ? Object : ToolRegistry2,
-    typeof AgentRegistry2 === "undefined" ? Object : AgentRegistry2,
-    Array,
-    Object,
-    Object,
-    Object
+    typeof QuotaProvider === "undefined" ? Object : QuotaProvider,
+    typeof Pick === "undefined" ? Object : Pick,
+    typeof AttachmentStagingStore === "undefined" ? Object : AttachmentStagingStore
   ])
-], AgentDepsFactory2);
+], ChatQueueService2);
+function pauseReasonFor(outcome, head) {
+  if (outcome === "failed") {
+    return "run_failed";
+  }
+  if (outcome === "cancelled" && head.interrupt !== true) {
+    return "cancelled";
+  }
+  return null;
+}
+__name(pauseReasonFor, "pauseReasonFor");
+function queueFrame(items, paused, started) {
+  return {
+    kind: "queue",
+    queue: {
+      items: items.map(queuedMessageView),
+      paused
+    },
+    ...started !== void 0 ? {
+      started
+    } : {}
+  };
+}
+__name(queueFrame, "queueFrame");
+
+// src/run-not-active.exception.ts
+import { RUN_NOT_ACTIVE_CODE, RUN_NOT_ACTIVE_MESSAGE } from "@dudousxd/nestjs-agent-core";
+import { ConflictException } from "@nestjs/common";
+var RunNotActiveException = class extends ConflictException {
+  static {
+    __name(this, "RunNotActiveException");
+  }
+  runId;
+  code = RUN_NOT_ACTIVE_CODE;
+  constructor(runId) {
+    super({
+      statusCode: 409,
+      code: RUN_NOT_ACTIVE_CODE,
+      message: RUN_NOT_ACTIVE_MESSAGE
+    }), this.runId = runId;
+  }
+};
 
 // src/agent.service.ts
-import { AGENT_ATTACHMENT_STAGING as AGENT_ATTACHMENT_STAGING2, AGENT_DEPS_FACTORY as AGENT_DEPS_FACTORY3, AGENT_MODEL_CATALOG, AGENT_OPTIONS as AGENT_OPTIONS3, AGENT_QUOTA_PROVIDER as AGENT_QUOTA_PROVIDER2, AGENT_RUNNER, AGENT_STORE as AGENT_STORE5, REQUESTER_APPROVER, findCatalogModel, mayDecideApproval, readElicitationQuestions, validateElicitationAnswer } from "@dudousxd/nestjs-agent-core";
-import { BadRequestException, ConflictException, ForbiddenException, GoneException, HttpException, HttpStatus, Inject as Inject6, Injectable as Injectable6, NotFoundException, NotImplementedException as NotImplementedException2, Optional as Optional5 } from "@nestjs/common";
 function _ts_decorate7(decorators, target, key, desc) {
   var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
   if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
@@ -1824,7 +1859,7 @@ var AgentService = class {
       };
     }
     await this.requireQueue().publish(result.threadId);
-    throw new ConflictException({
+    throw new ConflictException2({
       statusCode: 409,
       code: "run_active",
       message: `thread ${result.threadId} already has a turn running`
@@ -1884,6 +1919,11 @@ var AgentService = class {
     }
     const mode = params.mode ?? "auto";
     const { live, stale } = await queue.holder(threadId, this.runner);
+    if (stale !== null) {
+      await settleDeadRun2(this.store, {
+        runId: stale
+      });
+    }
     if (live === null && mode !== "queue") {
       const runId = crypto.randomUUID();
       if (await queue.queueStore().claimActiveStream(threadId, runId, stale !== null ? {
@@ -1896,7 +1936,7 @@ var AgentService = class {
       }
     }
     if (params.regenerate === true) {
-      throw new ConflictException({
+      throw new ConflictException2({
         statusCode: 409,
         code: "run_active",
         message: "cannot regenerate while a turn is running on this thread"
@@ -2244,8 +2284,29 @@ var AgentService = class {
   async signalToolCall(toolCallId2, reply) {
     await this.assertNotExpired(toolCallId2);
     const runId = await this.resolveRunForToolCall(toolCallId2);
+    await this.assertRunWaiting(runId);
     return this.runner.signal(runId, toolCallId2, reply);
   }
+  /**
+  * Refuse a decision addressed at a run that is over (`409 run_not_active`). The runtime would take
+  * the signal and buffer it for a run that never comes back — the person's "yes" is accepted, the
+  * card says so, and nothing runs. Settles the calls the dead run left awaiting a decision on the
+  * way out, so the card stops asking. A runner that cannot say whether a run is alive is taken at
+  * its word that it is, and so is one that fails to answer.
+  */
+  async assertRunWaiting(runId) {
+    if (typeof this.runner.isRunActive !== "function") {
+      return;
+    }
+    const active = await this.runner.isRunActive(runId).catch(() => true);
+    if (active) {
+      return;
+    }
+    await settleDeadRun2(this.store, {
+      runId
+    });
+    throw new RunNotActiveException(runId);
+  }
   async cancel(actor, runId) {
     await this.assertOwnsActiveStream(actor, runId);
     return this.runner.cancel(runId);
@@ -2638,7 +2699,7 @@ AgentService = _ts_decorate7([
     typeof ModelCatalog === "undefined" ? Object : ModelCatalog,
     typeof QuotaProvider === "undefined" ? Object : QuotaProvider,
     typeof Pick === "undefined" ? Object : Pick,
-    typeof ChatQueueService === "undefined" ? Object : ChatQueueService
+    typeof ChatQueueService2 === "undefined" ? Object : ChatQueueService2
   ])
 ], AgentService);
 
@@ -2929,7 +2990,7 @@ AttachmentsController = _ts_decorate10([
 ], AttachmentsController);
 
 // src/controller/chat.controller.ts
-import { AGENT_ACTOR_RESOLVER as AGENT_ACTOR_RESOLVER2, AgentStreamError as AgentStreamError2 } from "@dudousxd/nestjs-agent-core";
+import { AGENT_ACTOR_RESOLVER as AGENT_ACTOR_RESOLVER2, AgentStreamError as AgentStreamError2, streamFailure as streamFailure2 } from "@dudousxd/nestjs-agent-core";
 import { BadRequestException as BadRequestException3, Body, Controller as Controller3, Get as Get3, Inject as Inject9, Param, Post as Post2, Query, Req as Req2, Res } from "@nestjs/common";
 function _ts_decorate11(decorators, target, key, desc) {
   var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
@@ -3125,10 +3186,7 @@ data: ${line}
       const payload = error instanceof AgentStreamError2 ? {
         code: error.code,
         message: error.message
-      } : {
-        code: "run_failed",
-        message: error instanceof Error ? error.message : "stream error"
-      };
+      } : streamFailure2(error);
       res.write(`event: error
 data: ${JSON.stringify(payload)}
 
@@ -4453,7 +4511,7 @@ ToolsController = _ts_decorate21([
 
 // src/discovery/agent-discovery.service.ts
 import { AGENT_PROMPT_CONTRIBUTORS as AGENT_PROMPT_CONTRIBUTORS2, AGENT_REGISTRY as AGENT_REGISTRY4, AgentRegistry as AgentRegistry5 } from "@dudousxd/nestjs-agent-core";
-import { Inject as Inject20, Injectable as Injectable8, Logger as Logger3 } from "@nestjs/common";
+import { Inject as Inject20, Injectable as Injectable8, Logger as Logger4 } from "@nestjs/common";
 import { DiscoveryService, MetadataScanner } from "@nestjs/core";
 
 // src/decorator/agent.decorator.ts
@@ -4504,7 +4562,7 @@ var AgentDiscoveryService = class _AgentDiscoveryService {
   scanner;
   registry;
   contributors;
-  logger = new Logger3(_AgentDiscoveryService.name);
+  logger = new Logger4(_AgentDiscoveryService.name);
   constructor(discovery, scanner, registry, contributors) {
     this.discovery = discovery;
     this.scanner = scanner;
@@ -4637,7 +4695,7 @@ AgentDiscoveryService = _ts_decorate22([
 
 // src/discovery/ai-tool-discovery.service.ts
 import { AGENT_OPTIONS as AGENT_OPTIONS6, AGENT_REGISTRY as AGENT_REGISTRY5, AGENT_TOOL_REGISTRY as AGENT_TOOL_REGISTRY2, normalizeDelegation as normalizeDelegation2 } from "@dudousxd/nestjs-agent-core";
-import { Inject as Inject21, Injectable as Injectable9, Logger as Logger4 } from "@nestjs/common";
+import { Inject as Inject21, Injectable as Injectable9, Logger as Logger5 } from "@nestjs/common";
 import { DiscoveryService as DiscoveryService2 } from "@nestjs/core";
 import { z } from "zod";
 
@@ -4691,7 +4749,7 @@ var AiToolDiscoveryService = class _AiToolDiscoveryService {
   registry;
   agents;
   options;
-  logger = new Logger4(_AiToolDiscoveryService.name);
+  logger = new Logger5(_AiToolDiscoveryService.name);
   constructor(discovery, registry, agents, options) {
     this.discovery = discovery;
     this.registry = registry;
@@ -4842,7 +4900,7 @@ AiToolDiscoveryService = _ts_decorate23([
 
 // src/discovery/skill-discovery.service.ts
 import { AGENT_SKILL_SOURCES } from "@dudousxd/nestjs-agent-core";
-import { Inject as Inject22, Injectable as Injectable10, Logger as Logger5 } from "@nestjs/common";
+import { Inject as Inject22, Injectable as Injectable10, Logger as Logger6 } from "@nestjs/common";
 import { DiscoveryService as DiscoveryService3 } from "@nestjs/core";
 
 // src/decorator/skill.decorator.ts
@@ -4891,7 +4949,7 @@ var SkillDiscoveryService = class _SkillDiscoveryService {
   }
   discovery;
   declared;
-  logger = new Logger5(_SkillDiscoveryService.name);
+  logger = new Logger6(_SkillDiscoveryService.name);
   constructor(discovery, declared) {
     this.discovery = discovery;
     this.declared = declared;
@@ -5001,8 +5059,8 @@ var LedgerQuotaProvider = class {
 };
 
 // src/runner/inline-agent-runner.ts
-import { AGENT_DEPS_FACTORY as AGENT_DEPS_FACTORY6, AGENT_STORE as AGENT_STORE6, RunCancelledError as RunCancelledError2, agentFailureCode as agentFailureCode2, encodeStreamEvent as encodeStreamEvent5, publishAgentRunFailed as publishAgentRunFailed2, releaseThreadRun as releaseThreadRun3, runAgentLoop as runAgentLoop2, settleAll as settleAll2, settleUnsettledDelegation as settleUnsettledDelegation2 } from "@dudousxd/nestjs-agent-core";
-import { Inject as Inject23, Injectable as Injectable11, Logger as Logger6, Optional as Optional10 } from "@nestjs/common";
+import { AGENT_DEPS_FACTORY as AGENT_DEPS_FACTORY6, AGENT_STORE as AGENT_STORE6, RUN_ENDED_BEFORE_TOOL_CALL as RUN_ENDED_BEFORE_TOOL_CALL2, RunCancelledError as RunCancelledError2, agentFailureCode as agentFailureCode2, encodeStreamEvent as encodeStreamEvent5, publishAgentRunFailed as publishAgentRunFailed2, releaseThreadRun as releaseThreadRun3, runAgentLoop as runAgentLoop2, settleAll as settleAll2, settleUnsettledDelegation as settleUnsettledDelegation2, streamFailure as streamFailure3 } from "@dudousxd/nestjs-agent-core";
+import { Inject as Inject23, Injectable as Injectable11, Logger as Logger7, Optional as Optional10 } from "@nestjs/common";
 function _ts_decorate25(decorators, target, key, desc) {
   var c = arguments.length, r = c < 3 ? target : desc === null ? desc = Object.getOwnPropertyDescriptor(target, key) : desc, d;
   if (typeof Reflect === "object" && typeof Reflect.decorate === "function") {
@@ -5036,7 +5094,7 @@ var InlineAgentRunner = class _InlineAgentRunner {
   factory;
   store;
   queue;
-  logger = new Logger6(_InlineAgentRunner.name);
+  logger = new Logger7(_InlineAgentRunner.name);
   pending = /* @__PURE__ */ new Map();
   /**
   * Runs someone has asked to stop. In-process, like `pending`, and for the same reason: this runner
@@ -5056,7 +5114,15 @@ var InlineAgentRunner = class _InlineAgentRunner {
   * that was running it is gone (a restart), and a thread it still holds is a stale claim.
   */
   async isRunActive(runId) {
-    return this.live.has(runId);
+    if (this.live.has(runId)) {
+      return true;
+    }
+    for (const key of this.pending.keys()) {
+      if (key.startsWith(`${runId}:`)) {
+        return true;
+      }
+    }
+    return false;
   }
   async start(input, options = {}) {
     const runId = options.runId ?? crypto.randomUUID();
@@ -5099,13 +5165,11 @@ var InlineAgentRunner = class _InlineAgentRunner {
         errorCode: code,
         errorMessage: message
       });
+      await Promise.resolve(this.store.failUnsettledToolCalls?.(runId, RUN_ENDED_BEFORE_TOOL_CALL2)).catch(() => 0);
       const writer = await deps.sink.open(runId);
       await this.settleQueue(writer, input, runId, "failed", message);
       await releaseThreadRun3(this.store, input.threadId, runId);
-      await writer.fail({
-        code,
-        message
-      });
+      await writer.fail(streamFailure3(error));
     }).finally(() => {
       this.cancelled.delete(runId);
       this.live.delete(runId);
@@ -5379,6 +5443,7 @@ var InlineAgentRunner = class _InlineAgentRunner {
           errorMessage: message
         }
       });
+      await Promise.resolve(this.store.failUnsettledToolCalls?.(runId, RUN_ENDED_BEFORE_TOOL_CALL2)).catch(() => 0);
       await settleUnsettledDelegation2({
         store: this.store,
         delivery: deliverTo,
@@ -5481,7 +5546,7 @@ InlineAgentRunner = _ts_decorate25([
   _ts_metadata24("design:paramtypes", [
     typeof AgentDepsFactory === "undefined" ? Object : AgentDepsFactory,
     typeof AgentStore === "undefined" ? Object : AgentStore,
-    typeof ChatQueueService === "undefined" ? Object : ChatQueueService
+    typeof ChatQueueService2 === "undefined" ? Object : ChatQueueService2
   ])
 ], InlineAgentRunner);
 function isThreadTurn2(input) {
@@ -5543,7 +5608,7 @@ function _ts_decorate26(decorators, target, key, desc) {
 }
 __name(_ts_decorate26, "_ts_decorate");
 var DEFAULT_PATH = "agent";
-var logger = new Logger7("AgentModule");
+var logger = new Logger8("AgentModule");
 function externalStoreWrapper(modules) {
   for (const module of modules.values()) {
     if (module.metatype === AgentModule) continue;
@@ -5696,7 +5761,11 @@ function sharedProviders(durable) {
     SkillDiscoveryService,
     // The thread message queue — every surface, since both runners and the durable workflow settle
     // into it, and the chat routes enqueue into it.
-    ChatQueueService,
+    ChatQueueService2,
+    {
+      provide: AGENT_CHAT_QUEUE,
+      useExisting: ChatQueueService2
+    },
     InlineAgentRunner,
     AgentService,
     // Bound ALWAYS (durable or inline) — the console's cross-thread approvals inbox routes decisions
@@ -5758,7 +5827,8 @@ function exportsFor() {
     AGENT_DEPS_FACTORY7,
     AgentDepsFactory2,
     AgentService,
-    ChatQueueService,
+    ChatQueueService2,
+    AGENT_CHAT_QUEUE,
     InlineAgentRunner,
     AGENT_APPROVAL_PORT
   ];