Skip to main content

Local Responses proxy for OpenAI Codex CLI: folds gpt-5.5 518n-2 reasoning truncation (516 degradation) via the official openai_base_url wiring — no provider change, WebSocket-first, no fallback noise.

Project description

codex-516-guard

Local Responses proxy for OpenAI Codex CLI: detects the gpt-5.5 "516" reasoning-truncation fingerprint (reasoning_tokens == 518*n - 2), auto-continues the model's thinking, and folds all rounds into one response — without changing model_provider, so session grouping, remote compaction and remote-control stay intact. WebSocket-first: no "Falling back from WebSockets" retry noise.

自研本地 Responses 代理,缓解 Codex gpt-5.5 的「516 降智」:思考在 reasoning_tokens == 518*n - 2(516、1034、1552…)处被截断,答案质量骤降 (上游 issue:openai/codex#30364,无官方修复)。 本代理检测该指纹后自动让模型继续思考,并把多轮续写折叠为单个下游响应

机制思路来自 neteroster/CodexCont(MIT), 实现为全新代码。与其关键差异:

codex-516-guard CodexCont
Codex 侧接线 顶层 openai_base_url不新建 provider 新建 [model_providers](会话按 provider 分组被隐藏、remote-control 不可用、丢远程压缩)
下游传输 WebSocket 第一传输(完整实现 responses_websockets 协议)+ SSE 兜底 仅 SSE(codex 先试 ws → 405 → 每会话约 5 次重连告警后回退)
zstd 请求压缩(0.142.x 内置 provider 默认开) 原生解压,无需改 codex 配置 [features] enable_request_compression = false
GET /v1/models 模型目录刷新 /v1/* 透传 未代理(静默失败,靠本地缓存)
续写方法 commentary 法(phase:"commentary" 消息 + encrypted reasoning 重放) commentary + tool_pair legacy + 跨轮 repair 等更多可配置项

原理

  1. 上游每轮结束时读取 usage.output_tokens_details.reasoning_tokens,命中 518n-2(n∈[1,6],最多续写 3 轮)即判定思考被截断;
  2. 丢弃该轮的暂定输出(message / tool calls——它们基于被截断的思考),把该轮 reasoning items(含 encrypted_content)+ 一条 Continue thinking...phase:"commentary" 助手消息追加进 input 重放,开下一轮;
  3. 思考流实时透传给 agent,只有干净收尾那一轮的最终输出被放行;terminal 事件重建为单响应口径的 usage(input 取第 1 轮防止「假爆上下文」,reasoning 求和),真实累计成本记在 metadata.proxy_billed_usage

安装

要求:uv(自带 Python 管理)、Codex CLI(ChatGPT OAuth 登录,0.142.x 实测)。

uv tool install codex-516-guard          # 从 PyPI 安装
# 或直接从源码仓库:
# uv tool install git+https://github.com/dzshzx/codex-516-guard

uv 会建一个隔离环境并把可执行文件放进 uv 的 bin 目录(Unix/macOS 默认 ~/.local/bin, Windows 用 where.exe codex-516-guard 查实际路径;uv tool update-shell 可把该目录加进 PATH)。 之后:

codex-516-guard                          # 前台跑起(默认 127.0.0.1:8787)
codex-516-guard --port 8790 --log-level debug   # 可选参数:--host/--port/--upstream/--log-level

升级 / 卸载:uv tool upgrade codex-516-guard / uv tool uninstall codex-516-guard

Codex 侧接线——~/.codex/config.toml 顶层(必须在第一个 [table] 之前)加一行:

openai_base_url = "http://127.0.0.1:8787/v1"

这是覆盖内置 openai provider base_url 的官方 config key#16719;同名 [model_providers.openai] 覆盖被维护者拒绝,OPENAI_BASE_URL 环境变量已移除)。provider id 保持 openai, 因此会话历史分组、远程压缩、remote-control 均不受影响。

关闭:注释掉 openai_base_url 行 + 停掉代理进程。代理停止而 key 在位时,Codex 会因上游不可达报错。

开机自启动(可选,默认不开)

安装本身不注册任何自启动——是否开机自启完全由你决定。要开时一条命令,不想开就别运行它。

codex-516-guard install-service     # 注册并立即启动(当前平台)
codex-516-guard uninstall-service   # 撤销

install-service 按平台选「随用户登录启动、跑在用户上下文」的方式(而非系统级服务——系统服务跑在无用户环境的 session 里,够不到用户 profile 下的 uv 可执行文件与代理设置):Linux/WSL → systemd user unit; macOS → launchd LaunchAgent(~/Library/LaunchAgents/);Windows → onlogon 计划任务(schtasks)。 自定义端口等参数会一并写进自启动条目:codex-516-guard install-service --port 8790

下面是各平台手动等价操作(想自己管、或 install-service 在你的环境跑不动时用)。先用 which codex-516-guard(Unix/macOS)/ where.exe codex-516-guard(Windows)拿到绝对路径。

Linux / WSL — systemd user unit

install-service 生成的等价 unit 见 systemd/codex-516-guard.service.example

cp systemd/codex-516-guard.service.example ~/.config/systemd/user/codex-516-guard.service
systemctl --user daemon-reload && systemctl --user enable --now codex-516-guard
loginctl enable-linger   # 可选:无需登录也在开机时启动

macOS — launchd LaunchAgent

macOS 用 launchd 管理后台任务。放在 ~/Library/LaunchAgents/ 的是 LaunchAgent,随用户登录启动、 跑在用户 GUI session 里(对回环代理正确的选择);/Library/LaunchDaemons/ 里的 LaunchDaemon 开机即起但无用户会话,本场景不适用。

把可执行文件绝对路径填进下面的 plist,存为 ~/Library/LaunchAgents/com.dzshzx.codex-516-guard.plist

<?xml version="1.0" encoding="UTF-8"?>
<!DOCTYPE plist PUBLIC "-//Apple//DTD PLIST 1.0//EN" "http://www.apple.com/DTDs/PropertyList-1.0.dtd">
<plist version="1.0">
<dict>
    <key>Label</key>            <string>com.dzshzx.codex-516-guard</string>
    <key>ProgramArguments</key>
    <array>
        <string>/Users/YOU/.local/bin/codex-516-guard</string>
    </array>
    <key>RunAtLoad</key>        <true/>
    <key>KeepAlive</key>        <true/>
    <key>StandardOutPath</key>  <string>/tmp/codex-516-guard.log</string>
    <key>StandardErrorPath</key><string>/tmp/codex-516-guard.log</string>
</dict>
</plist>

注意:launchd 不读 shell 配置,ProgramArguments 必须是绝对路径;崩溃后 KeepAlive 会重启(10 秒节流)。 加载 / 停用(现代 launchctlload/unload 已是 legacy):

launchctl bootstrap gui/$(id -u) ~/Library/LaunchAgents/com.dzshzx.codex-516-guard.plist
launchctl enable   gui/$(id -u)/com.dzshzx.codex-516-guard
launchctl kickstart -k gui/$(id -u)/com.dzshzx.codex-516-guard   # 立即(重)启动
launchctl bootout  gui/$(id -u)/com.dzshzx.codex-516-guard        # 卸载

Windows — 启动文件夹隐藏启动器(免管理员)

install-service 在启动文件夹(%APPDATA%\...\Startup)写一个 VBS 启动器,用 WScript.Shell.Run(cmd, 0, False) 在登录时隐藏窗口、免管理员地拉起代理。等价手动操作(<exe> = where.exe codex-516-guard 的结果):

# %APPDATA%\Microsoft\Windows\Start Menu\Programs\Startup\codex-516-guard.vbs
'CreateObject("WScript.Shell").Run """<exe>""", 0, False' |
  Set-Content "$env:APPDATA\Microsoft\Windows\Start Menu\Programs\Startup\codex-516-guard.vbs"
# 删除该 .vbs 即取消自启

为什么不用其它方式:系统 service 要求实现 SCM 协议(sc.exe 直指控制台程序会 1053 超时), 且跑在 session 0/SYSTEM,够不到用户 profile 下 uv 装的 exe 与代理设置;计划任务 onlogon 虽合适, 但在有策略限制的机器上 schtasks /create 需要提权(实测部分机器非管理员被拒)。启动文件夹启动器无这些约束。

验证

curl -sS http://127.0.0.1:8787/healthz            # {"ok":true,...}
journalctl --user -u codex-516-guard -f | grep -E 'round|done'   # Linux/WSL;mac 看 plist 日志文件

命中折叠时的日志(实测样例,连环双 516 被击破、答案正确):

round 1: in=21550 out=664 reason=516 total=22214 | n=1 buffered=['function_call'] -> continue
round 2: in=22078 out=652 reason=516 total=22730 | n=1 buffered=['function_call'] -> continue
round 3: in=22606 out=566 reason=291 total=23172 | n=None buffered=[...] -> clean
done: 3 round(s) | ... | status=completed stop=natural

开发

git clone https://github.com/dzshzx/codex-516-guard && cd codex-516-guard
uv sync
uv run python test_fold.py        # 折叠状态机自测,应输出 ALL PASS
uv run codex-516-guard            # 本地跑

发布走 PyPI Trusted Publishing(.github/workflows/release.yml,OIDC,无 token):推 v* tag 即自动构建上传。

结构

  • guard/fold.py — 指纹检测 + 折叠状态机(传输无关;test_fold.py 覆盖丢弃/放行、重编号、双口径 usage)
  • guard/server.py — starlette 传输层:ws / SSE 下游、SSE 上游、zstd/gzip 请求解压、/v1/* 透传
  • guard/cli.py — CLI 入口(codex-516-guard;仅监听回环;auth passthrough,不存储任何凭据)

安全与免责

  • 代理只做 auth passthrough:转发 Codex 发来的 Authorization 头,不读取、不落盘任何凭据。
  • 仅监听回环地址;不要暴露到非回环接口。
  • 非官方项目,依赖上游未公开的行为(截断指纹、ws 帧格式),OpenAI 侧变更可能使其失效;使用风险自负。
  • 续写会产生额外的真实 token 消耗(见 metadata.proxy_billed_usage),guard 以 n 窗口 + 3 轮上限约束。

License

MIT(见 LICENSE;机制思路 credit neteroster/CodexCont)。

Project details


Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distribution

codex_516_guard-0.2.4.tar.gz (50.0 kB view details)

Uploaded Source

Built Distribution

If you're not sure about the file name format, learn more about wheel file names.

codex_516_guard-0.2.4-py3-none-any.whl (19.9 kB view details)

Uploaded Python 3

File details

Details for the file codex_516_guard-0.2.4.tar.gz.

File metadata

  • Download URL: codex_516_guard-0.2.4.tar.gz
  • Upload date:
  • Size: 50.0 kB
  • Tags: Source
  • Uploaded using Trusted Publishing? Yes
  • Uploaded via: twine/6.1.0 CPython/3.13.13

File hashes

Hashes for codex_516_guard-0.2.4.tar.gz
Algorithm Hash digest
SHA256 676fed90aede574ae70927da19bd0981d3e2d612148eccee0d1618d16c9cbd4c
MD5 2984de01b19c64a6875b19de74c20427
BLAKE2b-256 08dd4170638b86090ce6af5643daefb73eb25721918bdaedc744d7767a1bc770

See more details on using hashes here.

Provenance

The following attestation bundles were made for codex_516_guard-0.2.4.tar.gz:

Publisher: release.yml on dzshzx/codex-516-guard

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

File details

Details for the file codex_516_guard-0.2.4-py3-none-any.whl.

File metadata

File hashes

Hashes for codex_516_guard-0.2.4-py3-none-any.whl
Algorithm Hash digest
SHA256 ea3d6d2743a19777ce539a561722c084fe67bce193793cda88c836899d6a9559
MD5 b6a10483202c51b48326b251e6cce352
BLAKE2b-256 268c532eefb5a11c64cebc4de33cbd8e2090d45e6d66c9faae265ac9b873519a

See more details on using hashes here.

Provenance

The following attestation bundles were made for codex_516_guard-0.2.4-py3-none-any.whl:

Publisher: release.yml on dzshzx/codex-516-guard

Attestations: Values shown here reflect the state when the release was signed and may no longer be current.

Supported by

AWS Cloud computing and Security Sponsor Datadog Monitoring Depot Continuous Integration Fastly CDN Google Download Analytics Pingdom Monitoring Sentry Error logging StatusPage Status page