feat(ai, ai-harness): signal a run stopped by the agent loop strategy

.changeset/loop-stopped-signal.md

@@ -0,0 +1,8 @@
+---
+'@tanstack/ai': minor
+'@tanstack/ai-harness': minor
+---
+
+A run that the agent loop stops before the model finished now says so. When `agentLoopStrategy` (for example `maxIterations`) or a middleware `onShouldContinue` returns `false` while tool results wait for the next model call, `chat()` ends the stream with a `RUN_FINISHED` that has `metadata.tanstack.loopStopped: true`, and the last stored assistant message gets the same field. `finishReason` stays the reason of the last model call (`tool_calls`, or `length` for a cut answer). A run that the model ends has no `loopStopped`.
+
+In `@tanstack/ai-harness`, a turn that the loop stops settles with `{ outcome: 'completed', loopStopped: true }`. `session.settled()`, the `harness.input.settled` event of `createHarnessClient().events()`, and the transcript all carry it.

docs/api/ai.md

@@ -259,6 +259,8 @@ An iteration is one model turn, not one tool call. One turn can call many tools
 
 A strategy that goes on after a `length` answer makes the model send the [cut tool call](../chat/agentic-cycle#when-the-output-limit-cuts-a-tool-call) again.
 
+When a strategy stops the loop while tool results wait for the model, the last `RUN_FINISHED` has `metadata.tanstack.loopStopped: true`. See [Know when the loop stopped the run](../chat/agentic-cycle#know-when-the-loop-stopped-the-run).
+
 ## Response helpers
 
 Each helper turns a `chat()` stream into a response for one transport. Pick the one that matches the client connection adapter. See [Transports](../transports/overview).

docs/chat/agentic-cycle.md

@@ -213,6 +213,66 @@ A strategy is a function. It gets the loop state and returns `true` to allow one
 - `toolCallCount`: the tool calls so far in this run.
 - `lastTurnToolCallCount`: the tool calls in the last batch.
 
+### Know when the loop stopped the run
+
+Read `metadata.tanstack.loopStopped` on the last `RUN_FINISHED` event. It is `true` when the strategy, or a middleware `onShouldContinue`, stopped the loop while tool results waited for the model. In that case, the model did not finish its task.
+
+A run that the model ends has no `loopStopped`. The `finishReason` is always the reason of the last model call, often `tool_calls` for a stopped run.
+
+On the server:
+
+```typescript
+// api/chat.ts
+import { chat, maxIterations } from "@tanstack/ai";
+import { openaiText } from "@tanstack/ai-openai";
+import { getWeather } from "./tools";
+
+const stream = chat({
+  adapter: openaiText("gpt-6.1-sol"),
+  messages: [{ role: "user", content: "Plan my week of outfits." }],
+  threadId: "thread-1",
+  tools: [getWeather],
+  agentLoopStrategy: maxIterations(3),
+});
+
+for await (const chunk of stream) {
+  if (chunk.type === "RUN_FINISHED" && chunk.metadata?.tanstack?.loopStopped) {
+    console.log("The loop stopped before the model finished.");
+  }
+}
+```
+
+On the client, read the same event in `onChunk`:
+
+```tsx
+import { useState } from "react";
+import { useChat, fetchServerSentEvents } from "@tanstack/ai-react";
+
+export function Chat() {
+  const [loopStopped, setLoopStopped] = useState(false);
+  const { sendMessage } = useChat({
+    connection: fetchServerSentEvents("/api/chat"),
+    threadId: "support-chat",
+    onChunk: (chunk) => {
+      if (chunk.type === "RUN_FINISHED" && chunk.metadata?.tanstack?.loopStopped) {
+        setLoopStopped(true);
+      }
+    },
+  });
+
+  return (
+    <>
+      <button type="button" onClick={() => sendMessage("Plan my week of outfits.")}>
+        Ask
+      </button>
+      {loopStopped ? <p>The agent stopped at its step limit.</p> : null}
+    </>
+  );
+}
+```
+
+The last assistant message that `chat()` stores on the server also gets `metadata.tanstack.loopStopped: true`. So a saved transcript keeps the signal.
+
 ## Advanced
 
 ### When the output limit cuts a tool call
@@ -225,7 +285,7 @@ A model can run out of output tokens in the middle of a tool call, for example w
 - It starts no client tool.
 - Each cut call gets an error tool result.
 
-By default, the run then ends with the finish reason `length`. To let the model send the cut call again, use a strategy that does not stop on `length`:
+By default, the run then ends with the finish reason `length` and [`loopStopped`](#know-when-the-loop-stopped-the-run). To let the model send the cut call again, use a strategy that does not stop on `length`:
 
 ```typescript
 // api/chat.ts

docs/config.json

@@ -216,7 +216,7 @@
           "label": "Agentic Cycle",
           "to": "chat/agentic-cycle",
           "addedAt": "2026-04-15",
-          "updatedAt": "2026-10-22"
+          "updatedAt": "2026-11-05"
         },
         {
           "label": "Thinking & Reasoning",
@@ -1133,7 +1133,7 @@
           "label": "Send inputs safely",
           "to": "harness/inputs",
           "addedAt": "2026-10-13",
-          "updatedAt": "2026-10-22"
+          "updatedAt": "2026-11-05"
         },
         {
           "label": "Run a tool without a model",
@@ -2038,7 +2038,7 @@
           "label": "@tanstack/ai",
           "to": "api/ai",
           "addedAt": "2026-04-15",
-          "updatedAt": "2026-11-02"
+          "updatedAt": "2026-11-05"
         },
         {
           "label": "@tanstack/ai-client",

docs/harness/inputs.md

@@ -78,13 +78,16 @@ const settlement = await session.settled({
 if (settlement.outcome === 'failed') {
   console.log(settlement.error?.code) // 'attempts_exhausted', 'timeout', or undefined
 }
+if (settlement.outcome === 'completed' && settlement.loopStopped) {
+  console.log('The agent loop stopped the turn before the model finished.')
+}
 ```
 
 `signal` is optional. When it aborts first, `settled` rejects with its reason. The input keeps running.
 
 On a host that does not own the thread, the input waits for the owner. If the owner refuses it, `settled` rejects with `InputRejectedError`. The owner refuses an input when the thread is busy and the input has `busy: 'reject'`, or when the log has its id with another message. The log then has a `harness.offer.rejected` record.
 
-- `completed`: the turn has its answer.
+- `completed`: the turn has its answer. With `loopStopped: true`, the agent loop stopped the turn while tool results waited for the model, so the model did not finish. The default `agentLoopStrategy` of a harness is `maxIterations(50)`. See [Know when the loop stopped the run](../chat/agentic-cycle#know-when-the-loop-stopped-the-run).
 - `failed`: the turn failed. `error.code` is set for a limit.
 - `aborted`: someone cancelled the turn. The text that the model streamed before the cancel stays in the transcript as its answer.
 - `interrupted`: the turn waits for a person. The `resolve` that answers it is a new input. See [Wait for a person](./waits).

packages/ai-client/src/chat-client.ts

@@ -220,6 +220,8 @@ function connectionDrainsOnSend(connection: ConnectionAdapter): boolean {
 function isIntermediateToolTurn(chunk: StreamChunk): boolean {
   if (chunk.type !== 'RUN_FINISHED') return false
   if (chunk.outcome?.type === 'interrupt') return false
+  // A loop stop keeps finishReason tool_calls. That event ends the stream.
+  if (tanstackMetadata(chunk)?.loopStopped === true) return false
   const extra = chunk as StreamChunk & { finishReason?: unknown }
   if (extra.finishReason !== undefined) {
     return extra.finishReason === 'tool_calls'

packages/ai-client/tests/chat-client-loop-stopped.test.ts

@@ -0,0 +1,48 @@
+import { describe, expect, it } from 'vitest'
+import { EventType, type StreamChunk } from '@tanstack/ai/client'
+import { ChatClient } from '../src/chat-client'
+import type { SubscribeConnectionAdapter } from '../src/connection-adapters'
+
+function finished(loopStopped: boolean): StreamChunk {
+  return {
+    type: EventType.RUN_FINISHED,
+    runId: 'run-1',
+    threadId: 'thread-1',
+    timestamp: 1,
+    metadata: {
+      tanstack: {
+        finishReason: 'tool_calls',
+        ...(loopStopped ? { loopStopped: true as const } : {}),
+      },
+    },
+  }
+}
+
+function subscribeConnection(chunk: StreamChunk): SubscribeConnectionAdapter {
+  return {
+    async *subscribe() {
+      yield {
+        type: EventType.RUN_STARTED,
+        runId: 'run-1',
+        threadId: 'thread-1',
+        timestamp: 0,
+      }
+      yield chunk
+    },
+    send: async () => {},
+  }
+}
+
+describe('loopStopped on a subscribe connection', () => {
+  it('resolves sendMessage when the loop stops on tool_calls', async () => {
+    const client = new ChatClient({
+      threadId: 'thread-1',
+      connection: subscribeConnection(finished(true)),
+    })
+
+    await client.sendMessage('hi')
+
+    expect(client.getIsLoading()).toBe(false)
+    expect(client.getSnapshot().status).toBe('ready')
+  })
+})

packages/ai-harness/src/log.ts

@@ -944,7 +944,12 @@ export function foldEntry(
           ...(record.operationId ? { operationId: record.operationId } : {}),
           ...(record.outcome === 'failed'
             ? { outcome: 'failed', error: record.error }
-            : { outcome: record.outcome }),
+            : {
+                outcome: record.outcome,
+                ...(record.outcome === 'completed' && record.loopStopped
+                  ? { loopStopped: true as const }
+                  : {}),
+              }),
         },
       }))
       state.thread = settleRunNode(state.thread, record)

packages/ai-harness/src/session/turn.ts

@@ -678,6 +678,8 @@ export abstract class SessionTurn<
     let openMessageId: string | undefined
     let interrupts: Array<Interrupt> | undefined
     let failure: string | undefined
+    /** The last RUN_FINISHED says the agent loop stopped the model. */
+    let loopStopped = false
     /** With `outputSchema`: the value of the last call, or why its check failed. */
     let structured: unknown
     let structuredError: { code: string; rawText?: string } | undefined
@@ -1364,6 +1366,12 @@ export abstract class SessionTurn<
               ) {
                 isRunOpen = false
               }
+              if (
+                chunk.type === EventType.RUN_FINISHED &&
+                chunk.metadata?.tanstack?.loopStopped === true
+              ) {
+                loopStopped = true
+              }
               if (
                 chunk.type === EventType.TEXT_MESSAGE_CONTENT ||
                 chunk.type === EventType.REASONING_MESSAGE_CONTENT ||
@@ -1853,7 +1861,10 @@ export abstract class SessionTurn<
       await settleTurn({ outcome: 'interrupted' })
       operation.finish('interrupted', { text, interrupts: [], media: captured })
     } else {
-      await settleTurn({ outcome: 'completed' })
+      await settleTurn({
+        outcome: 'completed',
+        ...(loopStopped ? { loopStopped: true } : {}),
+      })
       operation.finish('completed', {
         text,
         media: captured,

packages/ai-harness/src/types.ts

@@ -481,9 +481,19 @@ export type InputSettlement = {
        * `interrupted`: the turn stopped for human input. The `resolve` that
        * answers it is a new input with its own settlement.
        */
-      outcome: 'completed' | 'aborted' | 'interrupted'
+      outcome: 'aborted' | 'interrupted'
       error?: never
     }
+  | {
+      outcome: 'completed'
+      error?: never
+      /**
+       * The agent loop stopped before the model was done:
+       * `agentLoopStrategy` (default `maxIterations(50)`) or a middleware
+       * `onShouldContinue` said stop while tool results waited.
+       */
+      loopStopped?: true
+    }
   | {
       outcome: 'failed'
       /**

packages/ai-harness/tests/loop-stopped.test.ts

@@ -0,0 +1,117 @@
+import { describe, expect, it } from 'vitest'
+import { z } from 'zod'
+import { EventType, maxIterations, toolDefinition } from '@tanstack/ai'
+import { memoryPersistence } from '@tanstack/ai-persistence'
+import { HARNESS_EVENTS, createHarnessHost, defineHarness } from '../src'
+import { createHarnessHandler } from '../src/server'
+import { createHarnessClient } from '../src/client'
+import { mockAdapter, text, toolCall } from './helpers'
+import type { ModelMessage } from '@tanstack/ai'
+import type { InputSettlement } from '../src'
+import type { Reply } from './helpers'
+
+const lookup = toolDefinition({
+  name: 'lookup',
+  description: 'Look something up',
+  inputSchema: z.object({ q: z.string() }),
+}).server(async () => ({ found: true }))
+
+/** A harness with `maxIterations(2)`, its host session, and a client. */
+async function open(replies: Array<Reply>) {
+  const { adapter, calls } = mockAdapter(replies)
+  const harness = defineHarness({
+    name: 'test/loop-stopped',
+    adapter,
+    tools: [lookup],
+    agentLoopStrategy: maxIterations(2),
+  })
+  const host = createHarnessHost({
+    persistence: memoryPersistence(),
+    access: {
+      getUser: () => ({ id: 'u' }),
+      rules: { isAllowed: () => true },
+    },
+  })
+  const handler = createHarnessHandler({ host, harness })
+  const client = createHarnessClient<typeof harness>({
+    url: 'http://local/api/harness/',
+    threadId: 't1',
+    fetch: (input, init) => handler(new Request(input, init)),
+  })
+  const session = await host.open(harness, {
+    threadId: 't1',
+    user: { id: 'u' },
+  })
+  return { host, session, client, calls }
+}
+
+/** The first `harness.input.settled` value that the client reads. */
+async function clientSettlement(
+  client: Awaited<ReturnType<typeof open>>['client'],
+): Promise<InputSettlement | undefined> {
+  for await (const { event } of client.events({
+    from: '0',
+    signal: AbortSignal.timeout(2_000),
+  })) {
+    if (
+      event.type === EventType.CUSTOM &&
+      event.name === HARNESS_EVENTS.inputSettled
+    ) {
+      return event.value
+    }
+  }
+  return undefined
+}
+
+const loopStoppedOf = (message: ModelMessage | undefined) => {
+  const metadata = message?.metadata
+  if (typeof metadata !== 'object' || metadata === null) return undefined
+  const tanstack: unknown = 'tanstack' in metadata ? metadata.tanstack : null
+  if (typeof tanstack !== 'object' || tanstack === null) return undefined
+  return 'loopStopped' in tanstack ? tanstack.loopStopped : undefined
+}
+
+describe('a turn that the agent loop stops', () => {
+  it('settles completed with loopStopped, for the session and the client', async () => {
+    const { host, session, client, calls } = await open([
+      () => toolCall('lookup', { q: 'a' }, 'call-1'),
+      () => toolCall('lookup', { q: 'b' }, 'call-2'),
+      () => text('never'),
+    ])
+
+    await session.send('look it up', { inputId: 'in-1' })
+
+    expect(calls).toHaveLength(2)
+    expect(await session.settled({ inputId: 'in-1' })).toMatchObject({
+      outcome: 'completed',
+      loopStopped: true,
+    })
+    expect(await clientSettlement(client)).toMatchObject({
+      inputId: 'in-1',
+      outcome: 'completed',
+      loopStopped: true,
+    })
+    const assistant = (await client.transcript()).findLast(
+      (message) => message.role === 'assistant',
+    )
+    expect(loopStoppedOf(assistant)).toBe(true)
+    await host.close()
+  })
+
+  it('has no loopStopped when the model ends the turn', async () => {
+    const { host, session, client } = await open([
+      () => toolCall('lookup', { q: 'a' }, 'call-1'),
+      () => text('Found it.'),
+    ])
+
+    await session.send('look it up', { inputId: 'in-1' })
+
+    const settlement = await session.settled({ inputId: 'in-1' })
+    expect(settlement).toMatchObject({ outcome: 'completed' })
+    expect(settlement).not.toHaveProperty('loopStopped')
+    expect(await clientSettlement(client)).not.toHaveProperty('loopStopped')
+    const transcript = await client.transcript()
+    expect(transcript.some((message) => loopStoppedOf(message))).toBe(false)
+    await host.close()
+  })
+})