feat(agent-loop): surface max-tokens as a distinct turn-end reason

Add a `max-tokens` variant to `TurnEndReasonMap` and carry the model
finish reason up from `runStep` to `runTurn`, applying the rule "any
max-tokens step in the turn surfaces as max-tokens" (disposed/aborted/
error still take precedence). This lets consumers distinguish a clean
stop from a truncated one — the contract RFC 010's ACP bridge maps to
the `max_tokens` stop reason.

Also add an AGENTS.md rule: write an ADR when (and only when) a PR makes
a durable, contested, surprising decision.

packages/agent-loop/README.md

@@ -64,7 +64,7 @@ forever:
   idle unless more queued
 ```
 
-Error containment: a throwing plugin ends the **turn**, never the loop. Dispose mid-turn emits `agent/status('disposed')` and ends with reason `disposed`.
+Error containment: a throwing plugin ends the **turn**, never the loop. Dispose mid-turn emits `agent/status('disposed')` and ends with reason `disposed`. A step that hits the model's output-token ceiling makes the turn end `max-tokens` (the rule: any `max-tokens` step in the turn surfaces as `max-tokens`; `disposed`/`aborted`/`error` still take precedence) — distinct from a clean `completed` stop.
 
 ### What is NOT here
 

packages/agent-loop/src/loop.ts

@@ -71,6 +71,31 @@ function errorData(err: CodedError): { message: string; code?: string } {
   return { message: err.message, ...typeof err.code === 'string' ? { code: err.code } : {} }
 }
 
+/**
+ * The turn-end contribution of a step's *successful* finish, or `undefined`
+ * when the step finished ordinarily (a plain `completed`).
+ *
+ * {@link finishError} has already converted `error`/`aborted` finishes into
+ * thrown step errors, so the finishes that reach here are `stop`,
+ * `tool-calls`, `max-tokens`, or a future merge-extensible kind. Only
+ * `max-tokens` carries forward as a distinct {@link TurnEndReason}: a step that
+ * hit the output-token ceiling ended the turn cut-short rather than by the
+ * model's choice. `stop`/`tool-calls`/unknown kinds contribute nothing beyond
+ * the default `completed`. {@link runTurn} applies this with the rule "any
+ * `max-tokens` step in the turn makes the turn end `max-tokens`".
+ */
+function stepFinishReason(finish: FinishReason): TurnEndReason | undefined {
+  switch (finish.kind) {
+    case 'max-tokens':
+      return { kind: 'max-tokens' }
+    // stop / tool-calls / plugin-added kinds → no turn-end contribution
+    // beyond the default `completed`. FinishReason is merge-extensible, so a
+    // default (not assertNever) handles unknown kinds as ordinary success.
+    default:
+      return undefined
+  }
+}
+
 /**
  * Ambient handles the loop driver receives from the agent. Decouples the
  * pure function `runLoop` from the mutable LoopAgent fields, making the
@@ -294,7 +319,7 @@ async function runTurn(ctx: Context, agent: LoopAgent, handle: LoopHandle, turn:
       const abort = new AbortController()
       handle.setAbort(abort)
 
-      let stepOutcome: { hadToolCalls: boolean } | { error: Error }
+      let stepOutcome: { hadToolCalls: boolean; finish: FinishReason } | { error: Error }
       try {
         stepOutcome = await runStep(ctx, agent, turn, step, abort.signal)
       } catch (error: unknown) {
@@ -320,6 +345,16 @@ async function runTurn(ctx: Context, agent: LoopAgent, handle: LoopHandle, turn:
         break
       }
 
+      // The successful step's finish reason carries forward: a `max-tokens`
+      // step makes the whole turn end `max-tokens` (RFC 010's rule "any
+      // max-tokens step surfaces as max-tokens"). `stepFinishReason` returns
+      // `max-tokens` or `undefined`, so a later ordinary step never resets a
+      // max-tokens turn back to completed, and a never-truncated turn keeps the
+      // default `completed`. The disposal/abort/error branches above and the
+      // continuation-window disposal check below override this — they win.
+      const stepReason = stepFinishReason(stepOutcome.finish)
+      if (stepReason) reason = stepReason
+
       // Steering that arrived during streaming/tool execution.
       const steered = drainSteering(ctx, agent, turn)
 
@@ -423,7 +458,7 @@ async function runStep(
   turn: number,
   step: number,
   signal: AbortSignal,
-): Promise<{ hadToolCalls: boolean }> {
+): Promise<{ hadToolCalls: boolean; finish: FinishReason }> {
   const { session, options } = agent
 
   // --- Request assembly ---
@@ -517,7 +552,7 @@ async function runStep(
     /* v8 ignore stop */
   }
 
-  return { hadToolCalls: toolCalls.length > 0 }
+  return { hadToolCalls: toolCalls.length > 0, finish: assembler.finish }
 }
 
 /** The last turn number in a (possibly seeded) session log, or 0. */

packages/agent-loop/tests/loop.spec.ts

@@ -6,7 +6,7 @@ import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
 import ToolRegistry, { defineTool } from '@deepseek-ai/dsh-tools'
 import AgentRegistry from '@deepseek-ai/dsh-agent'
 import AgentLoop, { LoopAgent } from '@deepseek-ai/dsh-agent-loop'
-import { MockAdapter, textResponse, toolCallResponse } from './mock-adapter.ts'
+import { MockAdapter, maxTokensResponse, textResponse, toolCallResponse } from './mock-adapter.ts'
 
 async function harness(adapter: MockAdapter) {
   const ctx = new Context()
@@ -336,6 +336,76 @@ describe('agent loop', () => {
     expect(reasons).toEqual([{ kind: 'aborted', reason: 'user interrupt' }])
   })
 
+  it('surfaces max-tokens as the turn-end reason when the last step is cut off', async () => {
+    // A single step that ends with a max-tokens finish (no tool calls): the
+    // turn stops by default and ends max-tokens, not completed.
+    const adapter = new MockAdapter([maxTokensResponse('truncat')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create('a1', { model: 'mock' })
+
+    const reasons: TurnEndReason[] = []
+    ctx.on('agent/turn-end', (_agent, _turn, reason) => void reasons.push(reason))
+
+    send(agent, 'go')
+    await waitForIdle(ctx, agent)
+
+    expect(adapter.requests).toHaveLength(1)
+    expect(reasons).toEqual([{ kind: 'max-tokens' }])
+    // and the reason is recorded in the log's turn/end event
+    const turnEnd = agent.session.events.findLast(e => e.type === 'turn/end')
+    expect(turnEnd!.data.reason).toEqual({ kind: 'max-tokens' })
+  })
+
+  it('a max-tokens step earlier in a turn still surfaces as max-tokens after a later completed step', async () => {
+    // Step 1 is cut off (max-tokens, no tool calls → would stop by default), so
+    // continuation must be FORCED to reach step 2 which finishes normally
+    // (stop). The rule "any max-tokens step surfaces as max-tokens" means the
+    // turn ends max-tokens even though the LAST step completed cleanly.
+    const adapter = new MockAdapter([
+      maxTokensResponse('first half'),
+      textResponse('second half'),
+    ])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create('a1', { model: 'mock' })
+
+    let steps = 0
+    ctx.on('agent/step-end', () => void steps++)
+    // Force exactly one continuation (step 1 → step 2), then defer to default
+    // (step 2 is a plain stop with no tool calls → stops).
+    ctx.on('agent/turn-continuation', async (_agent, _turn, _defaultDecision, next) => {
+      if (steps < 2) return true
+      return next()
+    })
+
+    const reasons: TurnEndReason[] = []
+    ctx.on('agent/turn-end', (_agent, _turn, reason) => void reasons.push(reason))
+
+    send(agent, 'go')
+    await waitForIdle(ctx, agent)
+
+    expect(steps).toBe(2)
+    expect(adapter.requests).toHaveLength(2)
+    expect(reasons).toEqual([{ kind: 'max-tokens' }])
+  })
+
+  it('a completed step after no max-tokens keeps the turn completed (max-tokens does not leak across turns)', async () => {
+    // Two consecutive turns: turn 1 is cut off (max-tokens), turn 2 is a clean
+    // stop. The per-turn reason must be independent — turn 2 ends completed.
+    const adapter = new MockAdapter([maxTokensResponse('cut'), textResponse('clean')])
+    const ctx = await harness(adapter)
+    const agent = ctx.agentLoop.create('a1', { model: 'mock' })
+
+    const reasons: TurnEndReason[] = []
+    ctx.on('agent/turn-end', (_agent, _turn, reason) => void reasons.push(reason))
+
+    send(agent, 'first')
+    await waitForIdle(ctx, agent)
+    send(agent, 'second')
+    await waitForIdle(ctx, agent)
+
+    expect(reasons).toEqual([{ kind: 'max-tokens' }, { kind: 'completed' }])
+  })
+
   it('chains queued messages into consecutive turns', async () => {
     const adapter = new MockAdapter([textResponse('first'), textResponse('second')])
     const ctx = await harness(adapter)

packages/agent-loop/tests/mock-adapter.ts

@@ -12,6 +12,21 @@ export function textResponse(text: string): StreamChunk[] {
   ]
 }
 
+/**
+ * Like {@link textResponse} but the stream ends with a `max-tokens` finish —
+ * the model was cut off at the output-token ceiling (DeepSeek's `length`).
+ * Used to exercise the turn-end `max-tokens` surfacing rule.
+ */
+export function maxTokensResponse(text: string): StreamChunk[] {
+  return [
+    { type: 'block-start', index: 0, blockType: 'text' },
+    ...Array.from(text, (char): StreamChunk => ({ type: 'text-delta', index: 0, text: char })),
+    { type: 'block-end', index: 0, block: { type: 'text', text } },
+    { type: 'usage', usage: { inputTokens: 10, outputTokens: text.length } },
+    { type: 'finish', reason: { kind: 'max-tokens' } },
+  ]
+}
+
 export function toolCallResponse(rawCallId: string, name: string, args: object, text?: string): StreamChunk[] {
   const callId = CallId(rawCallId)
   const argumentsJson = JSON.stringify(args)