[bug] --acp 模式下 extended thinking + tool_use 同回合时间歇性触发 400

[FanHeran]

现象

--acp 模式下，多轮对话中如果触发 extended thinking 同时还要调工具，约 60% 概率触发：

API Error: 400 due to tool use concurrency issues.

包到 src/services/api/errors.ts:697 上游的原始 Anthropic 错误是：

tool_use ids were found without tool_result blocks immediately after

错误文字是误导性的——后面会看到，真实原因是 payload 里出现了连续两条 user 消息（违反 role alternation），但 Anthropic 把它当成 tool_use/tool_result 配对错误报回来。

复现环境

• upstream: c499bfb4（HEAD of claude-code-best/claude-code main 当前）
• runtime: Bun 1.3.14, Node 24.12（Node 跑会先死在 undici markResourceTiming，必须用 ccb-bun）
• platform: Windows 11
• auth: OAuth 已登录（~/.claude/.credentials.json）

复现脚本

测试脚本直接 spawn cli-bun.js --acp、用 @agentclientprotocol/sdk 走 ACP 协议发两轮 prompt，每轮都触发 Bash 工具：

// test-acp.mjs
import { spawn } from 'node:child_process'
import { Readable, Writable } from 'node:stream'
import * as acp from 'file:///<path-to>/node_modules/@agentclientprotocol/sdk/dist/acp.js'

const child = spawn('bun', ['dist/cli-bun.js', '--acp'], { stdio: ['pipe', 'pipe', 'pipe'] })
const stream = acp.ndJsonStream(Writable.toWeb(child.stdin), Readable.toWeb(child.stdout))
const conn = new acp.ClientSideConnection(_ => myClient, stream)

await conn.initialize({ protocolVersion: acp.PROTOCOL_VERSION, clientInfo: {...}, clientCapabilities: {...} })
const session = await conn.newSession({ cwd: process.cwd(), mcpServers: [] })

await conn.prompt({ sessionId: session.sessionId, prompt: [{ type: 'text', text: 'Run `ls` and report.' }] })
// ← 第一轮成功

await conn.prompt({ sessionId: session.sessionId, prompt: [{ type: 'text', text: 'Now run `pwd`.' }] })
// ← 约 60% 在这里 400

5 跑 3 崩 2 过，间歇性，强相关于第二轮模型是否使用 extended thinking。

诊断输出

给 src/services/api/errors.ts:222 的 logToolUseToolResultMismatch 加一行 console.error（env gated），跑失败一次拿到：

[TACO_DEBUG] tool_use/tool_result mismatch
  orphan tool_use_id: toolu_018UuqziqdtMpuhisagLv39q
  normalizedToolUseIndex=5/8  originalToolUseIndex=9/14

  normalizedSequence (what follows the orphan in the API payload):
    user:tool_result:toolu_018UuqziqdtMpuhisagLv39q
    user:string_content                         ← ❌ 多了一条 user

  preNormalizedSequence (what follows the orphan in mutableMessages):
    user:tool_result:toolu_018UuqziqdtMpuhisagLv39q
    assistant:thinking
    assistant:tool_use:toolu_018UuqziqdtMpuhisagLv39q   ← ❌ 同一个 ID 出现两次
    user:tool_result:toolu_018UuqziqdtMpuhisagLv39q

因果分析

[模型用 extended thinking + tool_use 同一回合回应]
   ↓
src/services/api/claude.ts 流式装配把同一个 assistant 消息 push 两次到 mutableMessages
   （怀疑是 [thinking] 块 stop 时 yield 一次，[thinking, tool_use] 完整 stop 时再 yield 一次）
   ↓
mutableMessages 里出现重复 tool_use ID 和重复 tool_result
   ↓
ensureToolResultPairing (src/utils/messages.ts:5568) 正确删了第二个 tool_use ✓
但留下了 assistant 只剩 [thinking] 一块
对应的第二个 tool_result 变孤儿
   ↓
进 messages.ts:5837 的「empty after strip → push NO_CONTENT_MESSAGE」分支
塞了一条 user:"(no content)" 占位（NO_CONTENT_MESSAGE 是字符串而非 content block）
   ↓
normalizeMessagesForAPI 把只有 thinking 的 assistant 干掉了
   ↓
最终 payload: ... assistant[tool_use(X)] → user[tool_result(X)] → user["(no content)"]
                                            └── 两条 user 连着 ──┘
   ↓
Anthropic 400 (错误文字提示 tool_use/tool_result 配对，实际是 role alternation 违规)

修复建议

战术修（小，立即解 400）

src/utils/messages.ts:5837 那个分支 push NO_CONTENT_MESSAGE 之前加一个 user→user 检查：

       } else {
+        // If the previous result entry is already a user, inserting another
+        // user placeholder creates consecutive-user messages which Anthropic
+        // rejects with a misleading "tool_use without tool_result" 400.
+        // Skip the placeholder — alternation is preserved by the next
+        // assistant message in the loop.
+        if (result.at(-1)?.message?.role === 'user') {
+          i++
+          continue
+        }
         // Content is empty after stripping orphaned tool_results. We still
         // need a user message here to maintain role alternation — ...
         i++
         result.push(
           createUserMessage({
             content: NO_CONTENT_MESSAGE,
             isMeta: true,
           }),
         )
       }

注释里写的 "维持 alternation" 在 assistant→assistant 之间是对的，但当前一条已经是 user 时反而把局面搅坏了。

根因修（大，从源头杜绝）

src/services/api/claude.ts 流式装配逻辑：extended thinking + tool_use 同一 turn 时为何会把 assistant 消息双 push 到 mutableMessages。具体得追 content_block_stop / message_delta 的处理流程，但这是真正的 root cause——战术 fix 只是堵住下游表象。

备注

• 这个 bug 跟 acp-link 包无关——packages/acp-link 是纯 WebSocket↔stdio passthrough，不动 tool_use/tool_result。直接 spawn cli-bun --acp 和经 acp-link 都触发。
• 上游 Anthropic 官方 @anthropic-ai/claude-code 没有 --acp 模式，无可对比。
• 我已经准备好战术 fix 的本地 patch，如果维护者认可方向我可以提 PR。

[Ilya0527]

The double-push is probably in content_block_stop handling, not message_stop. Anthropic's streaming spec only finalizes the assistant turn on message_stop; if any handler in claude.ts pushes to mutableMessages on content_block_stop of a thinking block (likely because thinking was historically the only block type that could appear alone), you get exactly your symptom — one partial assistant with [thinking], one complete with [thinking, tool_use], same tool_use_id in both.

Your tactical patch fixes the 400 but leaves a subtler bug. ensureToolResultPairing deleted the duplicate tool_use, but it should have deleted the first assistant message entirely. The thinking-only remnant strands its signature field — when sent back on the next turn, signature verification can fail and Anthropic silently discards the cached thinking block, defeating extended-thinking continuity across turns. Worth checking whether your second-turn requests lose thinking context even when they don't 400.

Separately: NO_CONTENT_MESSAGE as a string rather than a content block array is a latent issue — normalizeMessagesForAPI strips empty assistants but doesn't coalesce adjacent same-role users. The placeholder shouldn't need to exist at all.

Falsifiable check: add a counter on every mutableMessages.push site and log on message_stop. If push count > 1 per assistant turn when thinking+tool_use co-occur, the streaming assembler is the root cause; if count = 1, the duplication is happening in a later normalization pass and the patch site moves accordingly.

---

Generated via ALEF Pattern Catalog v2.4.11-alpha · autonomous research agent · operator-supervised · transparency · support