fix(agent-loop): quiesce scheduler failures

packages/core/agent-loop/README.i18n.yaml

@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/core/agent-loop/README.md
-README.md: 6bb8b12af69f54c2a75cd672e4d3802887808c76
-README.zh.md: 80a01f3e3fdddba8c243cad28c43072148af1dd9
+README.md: 16d70cc06498fec1221b7872f988a0126f69f39f
+README.zh.md: ce68595072766ebbf1e4cbd9f7c262cee36c5eff

packages/core/agent-loop/README.md

@@ -67,7 +67,7 @@ After `agent/request` returns a provider/model call config, the loop asks `ctx.l
 
 Plugin failure ends the current turn, not the loop. Only final adapter dispatch/iteration failures and terminal in-band error or aborted finishes enter `agent/request-error`; middleware, result processing, tools, and other extension failures close directly. Recovery receives the exact live error, immutable provider facts, immutable prior failures, the immutable retry policy of the adapter registration that served the request, and the turn signal after the failed step closes; the policy is absent if no final adapter served it. A handling listener returns `{ kind: 'retry' }`; the loop closes the failed turn with its error and opens one numbered retry turn without an intervening idle notification. Success clears the consecutive history, and an unhandled failure is terminal. AgentLoop owns one cancellation signal for the current admission or turn. An effective `cancel(cause)` clears pending work unless `keepInbox` is set and cooperatively aborts that signal; idle cancellation is a no-op. Durable `turn/end` records `aborted` for `user` and `parent`, while disposal records `disposed`; undispatched model tool calls receive synthetic `tool/call` and `ABORTED_BEFORE_DISPATCH` result pairs. The cancellation cause changes reporting, not how result context finalized after cancellation is handled. Disposal waits for signal-ignoring work before registry removal. The [explicit-cancellation decision](../../../.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.md) owns the lifecycle and race contract.
 
-Within a step, exclusive calls form barriers; parallel-safe calls use a bounded rolling pool and are reclassified before start. Only dispatch/body overlaps. Policy, durable results, and result context remain model-ordered. Abort stops new calls, drains started results, and retains their finalized result context without distinguishing the cancellation cause.
+Within a step, exclusive calls form barriers; parallel-safe calls use a bounded rolling pool and are reclassified before start. Only dispatch/body overlaps. Policy, durable results, and result context remain model-ordered. Abort stops new calls, drains started results, and retains their finalized result context without distinguishing the cancellation cause. An internal scheduler failure stops new dispatches, waits for already-started dispatches, and reaches the turn error boundary without fabricating tool results.
 
 ### What belongs to plugins
 

packages/core/agent-loop/README.zh.md

@@ -67,7 +67,7 @@ interface Config {
 
 插件失败会结束当前轮次，而不是结束循环。只有最终适配器分发／迭代失败以及带内的终止错误或中止结束才进入 `agent/request-error`；中间件、结果处理、工具及其他扩展失败会直接关闭轮次。失败步骤关闭后，恢复逻辑会接收确切的实时错误、不可变的提供方事实、不可变的先前失败、为请求提供服务的适配器注册所对应的不可变重试策略，以及轮次信号；如果没有最终适配器为其提供服务，则该策略缺失。处理失败的监听器返回 `{ kind: 'retry' }`；循环用其错误关闭失败轮次，并在不插入空闲通知的情况下开启一个编号重试轮次。成功会清除连续失败历史；未被处理的失败是终态。AgentLoop 为当前接纳或轮次拥有一个取消信号。有效的 `cancel(cause)` 在未设置 `keepInbox` 时清除待处理工作，并以协作方式中止该信号；空闲取消是空操作。持久 `turn/end` 为 `user` 和 `parent` 记录 `aborted`，dispose（资源释放）则记录 `disposed`；未分发的模型工具调用会收到合成的 `tool/call` 与 `ABORTED_BEFORE_DISPATCH` 结果对。取消原因只改变报告方式，不改变对取消后已定案结果上下文的处理。dispose 会等待忽略信号的工作完成，然后才从注册表移除。[显式取消决策](../../../.agents/notes/implemented/architecture/2026-07-16-explicit-turn-cancellation.md)规定生命周期与竞态契约。
 
-在步骤内，独占调用形成屏障；并行安全调用使用有界滚动池，并在启动前重新分类。只有分发／主体会重叠。策略、持久结果和结果上下文仍保持模型顺序。中止会停止新调用，drain 已启动的结果，并保留其已定案的结果上下文，不区分取消原因。
+在步骤内，独占调用形成屏障；并行安全调用使用有界滚动池，并在启动前重新分类。只有分发／主体会重叠。策略、持久结果和结果上下文仍保持模型顺序。中止会停止新调用，drain 已启动的结果，并保留其已定案的结果上下文，不区分取消原因。内部调度器故障会停止新的分发，等待已启动的分发，然后在不虚构工具结果的情况下到达轮次错误边界。
 
 ### 插件负责的内容
 

packages/core/agent-loop/src/tool-calls.ts

@@ -2,10 +2,12 @@
  * Schedules one assistant step's tool calls. Exclusive calls form barriers;
  * parallel calls use a bounded rolling pool and are reclassified before start.
  * Dispatch may overlap, while policy, results, and result context remain
- * model-ordered. Abort stops replenishment and drains started calls.
+ * model-ordered. Abort or an internal scheduler failure stops replenishment
+ * and drains started calls.
  *
- * Each advertised call records a balanced `tool/call`/`tool/result` pair. Calls
- * skipped after abort receive synthetic error results so replay stays valid.
+ * Abort records synthetic error results for skipped calls so replay stays
+ * valid. A terminal scheduler failure preserves already-recorded `tool/call`
+ * events without fabricating results.
  * @module dsh-agent-loop/tool-calls
  */
 
@@ -37,10 +39,13 @@ interface GroupOutcome {
 
 /**
  * Schedule one assistant step's tool calls by their live concurrency mode.
- * Started calls receive ordered results. Abort drains them, records synthetic
- * results for unstarted calls, and returns with the signal still aborted after
- * accepting started-call context through the caller-supplied acceptor (the
- * machine stages it on its outbox for the next step boundary).
+ * Ordinary completion and abort commit started-call results in order. Abort
+ * drains them, records synthetic results for unstarted calls, and returns with
+ * the signal still aborted after accepting started-call context through the
+ * caller-supplied acceptor (the machine stages it on its outbox for the next
+ * step boundary). An internal scheduler failure stops new dispatches, drains
+ * already-started dispatches, and rejects with the first failure without
+ * fabricating tool results.
  * The committed step's AgentLoop driver boundary supplies the initiating Agent
  * that becomes each explicit {@link ToolExecutionInput.agent}.
  *
@@ -110,7 +115,8 @@ function parseArguments(raw: string): unknown {
  * drain and remains for the caller's next barrier. Results and contexts commit
  * in model order. Abort stops starts, drains and commits started calls, accepts
  * their contexts into the owning batch, records results for skipped calls, and
- * returns an aborted outcome.
+ * returns an aborted outcome. Scheduler failure drains dispatches without
+ * committing synthetic recovery results.
  */
 async function runGroup(
   ctx: Context,
@@ -131,6 +137,10 @@ async function runGroup(
   let started = 0
   let aborted: boolean = signal.aborted
   let concluded = false
+  let schedulerFailure: { error: unknown } | undefined
+  const throwSchedulerFailure = (): void => {
+    if (schedulerFailure !== undefined) throw schedulerFailure.error
+  }
 
   // `committed` advances only across contiguous model-order slots.
   const commitReady = async (): Promise<void> => {
@@ -157,12 +167,19 @@ async function runGroup(
     callSeqs[index] = appendToolCall(session, turn, step, call.block)
     started++
     const prepared = await ctx.tools[TOOL_REGISTRY_SCHEDULER].prepare(call.exec)
+    throwSchedulerFailure()
     switch (prepared.kind) {
       case 'dispatch': {
-        const promise = ctx.tools[TOOL_REGISTRY_SCHEDULER].dispatch(prepared.exec).then((outcome) => {
-          slots[index] = { exec: prepared.exec, result: outcome.result, needsPost: outcome.kind === 'post-result' }
-          return index
-        })
+        const promise = ctx.tools[TOOL_REGISTRY_SCHEDULER].dispatch(prepared.exec).then(
+          (outcome) => {
+            slots[index] = { exec: prepared.exec, result: outcome.result, needsPost: outcome.kind === 'post-result' }
+            return index
+          },
+          (error: unknown) => {
+            schedulerFailure ??= { error }
+            return index
+          },
+        )
         inFlight.set(index, promise)
         break
       }
@@ -187,24 +204,34 @@ async function runGroup(
         && ctx.tools.executionMode(nextCall.exec).kind !== 'parallel') break
       await startCall(nextToStart)
       nextToStart++
+      throwSchedulerFailure()
       await commitReady()
+      throwSchedulerFailure()
       // Abort may arrive while pre-execute awaits.
       if (signal.aborted) aborted = true
     }
   }
 
-  // Ordered pre-execute may await; only dispatch/body overlaps.
-  // TODO: Drain every started call before rethrowing a scheduler error; tool
-  // bodies must not outlive the failed turn.
-  await fillPool()
-  while (inFlight.size > 0) {
-    const settledIndex = await Promise.race(inFlight.values())
-    inFlight.delete(settledIndex)
-    await commitReady()
-    // Abort may arrive while a tool or ordered commit awaits.
-
-    if (signal.aborted) aborted = true
+  // Ordered pre-execute may await; only dispatch/body overlaps. A scheduler
+  // failure stops new dispatches and reaches the turn boundary after every
+  // already-started dispatch settles.
+  try {
     await fillPool()
+    while (inFlight.size > 0) {
+      const settledIndex = await Promise.race(inFlight.values())
+      inFlight.delete(settledIndex)
+      throwSchedulerFailure()
+      await commitReady()
+      throwSchedulerFailure()
+      // Abort may arrive while a tool or ordered commit awaits.
+
+      if (signal.aborted) aborted = true
+      await fillPool()
+    }
+  } catch (error: unknown) {
+    schedulerFailure ??= { error }
+    await Promise.allSettled(inFlight.values())
+    throw schedulerFailure.error
   }
 
   if (aborted) {

packages/core/agent-loop/tests/tool-calls.spec.ts

@@ -9,7 +9,7 @@ import { createUserMessage, CallId, StreamChunk  } from '@deepseek-ai/dsh-llm'
 import SessionStore, { SessionEvent, SessionId } from '@deepseek-ai/dsh-session'
 import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
 import LlmService from '@deepseek-ai/dsh-llm'
-import ToolRegistry, { defineContentToolFixture, TOOL_ABORTED_BEFORE_DISPATCH, type PostToolDecision, type PreToolDecision } from '@deepseek-ai/dsh-tools'
+import ToolRegistry, { defineContentToolFixture, TOOL_ABORTED_BEFORE_DISPATCH, TOOL_REGISTRY_SCHEDULER, type PostToolDecision, type PreToolDecision } from '@deepseek-ai/dsh-tools'
 import AgentRegistry, { type Agent } from '@deepseek-ai/dsh-agent'
 import AgentLoop, { DEFAULT_MAX_PARALLEL_TOOL_CALLS } from '@deepseek-ai/dsh-agent-loop'
 import { MockAdapter, textResponse } from './mock-adapter.ts'
@@ -613,3 +613,66 @@ describe('tool-call scheduler: abort handling', () => {
       })
   })
 })
+
+describe('tool-call scheduler: failure quiescence', () => {
+  it('stops new dispatches and drains started bodies before surfacing the first failure', async () => {
+    const adapter = new MockAdapter([
+      multiCall([
+        { id: 'c1', name: 'p', args: { id: '1' } },
+        { id: 'c2', name: 'p', args: { id: '2' } },
+        { id: 'c3', name: 'p', args: { id: '3' } },
+      ]),
+    ])
+    const ctx = await harness(adapter, 3)
+    const gated = gatedParallelTool('p')
+    ctx.tools.register(gated.tool)
+    // The registry contains expected failures as results; replace its internal
+    // view only to inject the invariant violation this boundary must contain.
+    const scheduler = ctx.tools[TOOL_REGISTRY_SCHEDULER]
+    const prepare = scheduler.prepare.bind(scheduler)
+    const dispatch = scheduler.dispatch.bind(scheduler)
+    const prepareGate = Promise.withResolvers<undefined>()
+    let thirdPrepareEntered = false
+    scheduler.prepare = async (exec) => {
+      const prepared = await prepare(exec)
+      if (exec.callId === CallId('c3')) {
+        thirdPrepareEntered = true
+        await prepareGate.promise
+      }
+      return prepared
+    }
+    const schedulerError = new Error('scheduler exploded')
+    const drainedError = new Error('sibling failed while draining')
+    let rejectFirst: ((error: Error) => void) | undefined
+    scheduler.dispatch = exec => exec.callId === CallId('c1')
+      ? new Promise((_resolve, reject) => { rejectFirst = reject })
+      : dispatch(exec).then(() => { throw drainedError })
+    const agent = ctx.agentLoop.create(SessionId('scheduler-failure'), { provider: 'mock', model: 'mock' })
+    const errors: unknown[] = []
+    ctx.on('agent/error', (subject, _turn, _step, error) => {
+      if (subject === agent) errors.push(error)
+    })
+    let idle = false
+    const idlePromise = waitForIdle(ctx, agent).then(() => { idle = true })
+
+    agent.followup(createUserMessage({ content: [{ type: 'text', text: 'go' }], source: { kind: 'user' } }))
+    await until(() => gated.started.includes('2') && thirdPrepareEntered && rejectFirst !== undefined)
+    rejectFirst?.(schedulerError)
+    await new Promise<void>(resolve => setImmediate(resolve))
+    prepareGate.resolve(undefined)
+    await new Promise<void>(resolve => setImmediate(resolve))
+
+    const startedBeforeDrain = [...gated.started]
+    const idleBeforeDrain = idle
+    const errorsBeforeDrain = [...errors]
+    for (const id of gated.pending()) gated.release(id)
+    await idlePromise
+
+    expect(startedBeforeDrain).toEqual(['2'])
+    expect(idleBeforeDrain).toBe(false)
+    expect(errorsBeforeDrain).toEqual([])
+    expect(gated.pending()).toEqual([])
+    expect(errors).toEqual([schedulerError])
+    expect(errors[0]).toBe(schedulerError)
+  })
+})