Commit Graph
4 Commits
Author SHA1 Message Date
xlplboandClaude Fable 5 7942338328 fix(apicompat): 回程还原 namespace 子工具调用,修复 Codex MCP 工具 unsupported call
Codex 0.14x 将 MCP 工具声明为 namespace 工具,chat 桥去程摊平为
"<namespace>__<name>" function 工具后,回程仅原样回传平铺名的
function_call 项;codex 按 namespace+name 路由查不到该名字,所有
MCP 工具调用被判为 unsupported call。

- NamespaceToolNames 构建摊平名 →(namespace, 子工具名)反查表
  (摊平名超长带截断哈希,无法按字符串切分还原)
- 非流式/流式回程命中映射时还原为裸子工具名 + namespace 字段,
  ResponsesOutput 新增 Namespace 字段并同步 wire 层白名单
- 回退桥入口将映射与 CustomTools/ToolSearchDeclared 一并穿入

已在测试机经 codex exec + MCP server 端到端验证。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NWQyEgFKKbdve67G6qCoAU
2026-07-10 17:38:17 +08:00
xlplbo 27e29f0562 feat(apicompat): 添加 tool_search 支持,增强工具调用处理逻辑 2026-07-07 19:19:35 +08:00
xlplbo 75fb3c41c2 fix(apicompat): responses→chat 桥支持 custom 工具,修复 Codex exec 丢失
Codex 0.14x 的核心执行工具 exec 是 type=custom(freeform)工具,
responsesToolsToChatTools 丢弃所有非 function 工具,导致模型工具列表中
没有 exec、无法执行任何命令;SSE 序列化层 responsesItemWire 也缺少
custom_tool_call 分支,即使产出该类型 item,call_id/name/input 也会被
白名单剥离。

- custom 工具降级为单一 input:string 参数的 function 工具
- 回程(非流式+流式)按名字还原 custom_tool_call 项/事件
- 历史 custom_tool_call / custom_tool_call_output 项转换(此前被跳过)
- responsesItemWire 补 custom_tool_call 分支
- tools 全部被丢弃时不再转发 tool_choice(修复上游 400
  "'tool_choice' is only allowed when 'tools' are specified")

feat(apicompat): chat 桥补齐 tool_search/namespace/字符串工具声明

对齐参考实现(cc-switch transform_codex_chat.rs)的工具类型覆盖:
- tool_search → 固定 function 代理工具(query 必填/limit 可选),
  历史项 tool_search_call/tool_search_output 一并转换
- namespace → 子 function 工具摊平为 <namespace>__<name>,
  超 64 字符截断加 sha256 短哈希;历史项带 namespace 字段时同规则摊平
- tools: ["name"] 字符串简写按 custom 工具解析

fix(apicompat): 工具调用宣告延迟到类型可判定,兼容 name 迟到的上游

output_item.added 的 custom/function 判定依赖工具名。名字晚于首个分片
到达时,此前会按 function_call 宣告造成 added/done 项类型不一致。
改为存在 custom 工具且名字未到时延迟宣告,宣告时补发已累积的参数增量;
请求无 custom 工具时行为不变。

PR 标题与正文(可直接粘贴)

标题:fix(apicompat): responses→chat 桥丢弃 custom 工具导致 Codex 无法执行命令

正文:

## 问题

OpenAI APIKey 账号接入仅支持 /v1/chat/completions 的第三方上游
(DeepSeek/GLM/Kimi 等)时,Codex 0.14x 客户端文本对话正常,但无法执行
任何命令:exec 调用全部 "aborted",spawn_agent 派出的子智能体永远卡在
running,多轮后模型自述"环境无法执行文件系统命令"放弃。

## 根因

Codex 0.14x 的核心执行工具 exec 是 Responses API 的 custom(freeform)
工具,而非普通 function:

    {"type": "custom", "name": "exec", "format": {"type": "grammar", ...}}

两处缺陷叠加:

1. `responsesToolsToChatTools` 丢弃所有非 function 工具
   (chatcompletions_responses_bridge.go),模型工具列表中从未出现 exec;
2. SSE 序列化 `responsesItemWire` 按 item 类型白名单重组字段,无
   custom_tool_call 分支,call_id/name/input 会被剥离
   (responses_stream_event_wire.go)。

次生缺陷:工具全部被丢弃时 tool_choice 仍被转发,部分上游报 400
"'tool_choice' is only allowed when 'tools' are specified"。

## 改动

(三个提交的要点清单,见各提交说明)

设计与 cc-switch 的 Codex chat 转换层(transform_codex_chat.rs /
streaming_codex_chat.rs)对齐:custom 工具降级为 {"input": string}
function 工具;custom 调用的参数增量不可流式还原为自由文本,缓冲至
收尾一次性下发 custom_tool_call_input.delta/done。

## 测试

- 新增 17 个单元测试:请求工具转换、tool_choice 兜底、历史项转换、
  非流式/流式回程还原、wire 序列化字段完整性、name 迟到宣告、命名摊平
- `go test ./internal/pkg/apicompat/...` 与 service 相关测试全绿
- 真机验证:Codex CLI 0.142.5 → 本桥 → GLM/DeepSeek 系 chat-only 上游,
  exec 实际执行命令并多轮续作正常

## 影响面

仅影响 "APIKey 账号 + 判定不支持 Responses" 的 chat 直转路径:

- OAuth 账号与透传路径不经过本桥(透传为字节级原文转发)
- CC→Responses 反方向仅构造 ResponsesRequest,新增字段带 omitempty
  且不赋值,序列化输出不变
- 探测逻辑未动,路由判定不变
- 共享类型改动均为增量容错(新字段 omitempty;UnmarshalJSON 只新增
  字符串工具的解析成功路径)
2026-07-07 16:23:44 +08:00
visa2andClaude Opus 4.8 f10bca8155 refactor(apicompat): redesign the Codex Responses ↔ Chat Completions bridge
Codex CLI speaks the OpenAI Responses protocol (streaming, store:false), while
many upstreams (e.g. DeepSeek in thinking mode) only expose Chat Completions.
The bridge that translates between the two had grown field by field and leaned
on Go's serialization defaults, which both the Responses client (Codex) and the
Chat upstream reject in ways the official OpenAI endpoints tolerate.

Problems this fixes (all observed running Codex CLI against a DeepSeek upstream):
  - Streaming reasoning was never shown in the Codex TUI (the answer appeared
    with no visible thinking): reasoning deltas were emitted before the reasoning
    item was opened, so the strict client discarded them.
  - A tool-using turn could wedge the session into a "no response" state: the
    function_call stream was never closed (no function_call_arguments.done /
    output_item.done), so Codex never saw the tool call complete.
  - Parallel tool calls were rejected upstream (400/502): each function_call
    became its own assistant message, producing consecutive assistant messages
    with mismatched tool replies.
  - A tool turn was rejected with "reasoning_content in the thinking mode must be
    passed back": the reasoning that produced the tool call was dropped instead
    of being returned on the assistant message.
  - Items with no Chat equivalent (web_search_call, ...) and Codex's
    command-approval notice landed between an assistant tool_calls message and
    its tool reply, triggering "An assistant message with 'tool_calls' must be
    followed by tool messages responding to each 'tool_call_id'".
  - Interrupt/reconnect left an unanswered or dangling tool_call in the history,
    triggering the same 400.

The shared root cause is reliance on serialization defaults — omitempty dropping
protocol-required zero values, and unrecognized item types falling through a
generic path — rather than deliberately reproducing the target protocol. The
bridge is reworked into two explicit layers.

Request direction (Responses input -> Chat messages): a parse -> build ->
normalize pipeline.
  - reasoning_content is carried back on the assistant message that produced a
    tool call (DeepSeek thinking mode requires it to continue the same thought)
  - consecutive function_call items (parallel tool calls) are merged into a
    single assistant message's tool_calls array
  - item types with no Chat equivalent are skipped instead of leaking through a
    generic path
  - normalizeChatMessages is the single invariant gate: it guarantees every
    assistant tool_calls message is immediately followed by one tool reply per
    tool_call_id — reordering any intervening message (such as a command-approval
    notice) to after the replies, dropping unanswered tool_calls and orphan tool
    replies, and preserving bare passthrough tool messages.

Response direction (Chat SSE -> Responses SSE): ResponsesStreamEvent.MarshalJSON
constructs each streamed event explicitly so protocol-required fields are always
present (output_index/content_index/summary_index at 0, message content:[],
reasoning summary:[], function_call call_id/name/arguments, output_text part
text/annotations/logprobs). This is a single source of truth that removes any
post-hoc JSON patching. Reasoning is emitted as its own output item, opened
before its deltas, and tool calls are fully closed
(function_call_arguments.done + output_item.done with complete arguments).

Tests cover request-direction message invariants against golden Codex request
shapes (parallel calls, unknown items, intervening messages, partial/dangling
calls), per-event wire completeness, and streaming lifecycle ordering.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-05-31 16:14:58 +08:00