👾 Open Computer Use – Open-Source Alternative to Codex Computer Use
Swift accessibilityai-agentai-agentsclaude-codecodexcodex-computer-usecomputer-usecomputer-use-agentcross-platformcuadesktop-automationgemini-cligui-automationmacosmcpmodel-context-protocolopen-computer-useopen-computer-use-mcpopenai-codexopencode
chore: improve codex debugging workflow docs
Leo7f742488c82d
diff --git a/README.md b/README.mdindex 73178e6..b4fa6f8 100644--- a/README.md+++ b/README.md@@ -94,7 +94,7 @@ go run . call list_apps如果你想用 `mitmproxy` / `mitmweb` 观察 Codex 自己打到上游的请求,仓库内提供了一个默认脱敏的 addon:-详细的长期复用方法、后台启动方式和样本沉淀约定,见 [docs/references/codex-network-capture.md](/Users/bytedance/projects/github/open-codex-computer-use/docs/references/codex-network-capture.md)。+详细的长期复用方法、后台启动方式和样本沉淀约定,见 [docs/references/codex-network-capture.md](./docs/references/codex-network-capture.md)。如果你只是想快速后台启动一份抓包并让后续 Agent 复用,优先直接用:@@ -136,6 +136,12 @@ https://chatgpt.com/backend-api/codex/responses它会先做 `101 Switching Protocols`,随后在 WebSocket 帧里承载 `response.create`、`response.output_text.delta` 等消息。`scripts/codex_dump.py` 会把这些帧按 JSONL 持久化到 `websocket/` 目录里,同时把匹配到的 HTTP 请求写到 `http/` 目录。+现在脚本还会把当前抓包 `session_id` 对应的 `~/.codex/sessions/rollout-*.jsonl` 摘要一起写到 `local-sessions/`,所以同一个实验目录里可以直接对照看:++- 模型侧的 tool decision+- Codex 宿主侧的 `function_call`+- 本地 tool 的 `function_call_output`+仓库默认已经把 `artifacts/codex-dumps/` 加进 `.gitignore`,适合把真实抓包样本留在 repo 目录里反复分析,而不误提交到 Git。## 工程结构diff --git a/docs/REPO_COLLAB_GUIDE.md b/docs/REPO_COLLAB_GUIDE.mdindex b484a15..414841b 100644--- a/docs/REPO_COLLAB_GUIDE.md+++ b/docs/REPO_COLLAB_GUIDE.md@@ -15,6 +15,7 @@- `AGENTS.md` 只做路由,不要在里面堆一大坨规则。- `docs/` 是仓库级知识的正式来源。- 行为一旦变化,对应文档就要在同一次改动里同步更新。+- 仓库内文件、目录、脚本入口和文档链接一律使用相对路径,不要写任何机器相关的绝对路径。- 比起继续往大文档里堆内容,更推荐新增一份边界清楚的小文档。## Git 与评审diff --git a/docs/histories/2026-04/20260417-1510-nonintrusive-computer-use.md b/docs/histories/2026-04/20260417-1510-nonintrusive-computer-use.mdindex 1599a3c..e438317 100644--- a/docs/histories/2026-04/20260417-1510-nonintrusive-computer-use.md+++ b/docs/histories/2026-04/20260417-1510-nonintrusive-computer-use.md@@ -50,3 +50,17 @@- `plugins/open-computer-use/.codex-plugin/plugin.json`- `docs/ARCHITECTURE.md`- `docs/references/codex-network-capture.md`++### 🔬 Follow-up (2026-04-17 17:20)++在继续做 Codex 宿主侧调试时,把“隔离另一套 plugin 再跑 case”正式收进仓库流程:++- **[Isolated Exec Helper]**: 新增 `scripts/run-isolated-codex-exec.sh`,把 `computer-use` / `open-computer-use` / `all` 三种模式收成统一入口,底层通过单次 `codex exec -c 'plugins."...".enabled=false'` 做临时覆写,不改全局 `~/.codex/config.toml`。+- **[A/B Routing Guidance]**: 在 `docs/references/codex-network-capture.md` 增补“隔离 plugin 路径”章节,明确当需要比较官方和仓库插件时,默认应该关掉另一套,避免 prompt 锚点和共存插件一起污染结论。+- **[Verification Outcome]**: 这台机器上的隔离验证结果是:只开官方 `computer-use` 时 `list_apps` 正常;只开 `open-computer-use` 时 `list_apps` 仍然返回 `user cancelled MCP tool call`;而 direct JSON-RPC 调用仓库插件 launcher 是正常的,因此当前主瓶颈不是两套 plugin 互相干扰,而更像是 Codex host 对第三方 plugin 调用的 gate。++这一步的价值不只是“跑通一个 shell alias”。之前调试里,`prompt` 文案、插件共存状态和宿主策略混在一起,很容易把路由偏差误判成 runtime 行为差异。把隔离入口收进仓库后,后续做焦点行为对比、MITM 抓样本和 eval 时,至少能先把“到底调用的是谁”固定下来。++**Isolation Files:**+- `scripts/run-isolated-codex-exec.sh`+- `docs/references/codex-network-capture.md`diff --git a/docs/histories/2026-04/20260417-1619-add-codex-mitm-dump-script.md b/docs/histories/2026-04/20260417-1619-add-codex-mitm-dump-script.mdindex ff7bffb..d0bc49a 100644--- a/docs/histories/2026-04/20260417-1619-add-codex-mitm-dump-script.md+++ b/docs/histories/2026-04/20260417-1619-add-codex-mitm-dump-script.md@@ -30,3 +30,35 @@ Codex 当前主模型调用走 `chatgpt.com/backend-api/codex/responses` WebSock- `docs/references/README.md`- `docs/references/codex-network-capture.md`- `docs/histories/2026-04/20260417-1619-add-codex-mitm-dump-script.md`++## [2026-04-17 17:50] | Task: 改进 Codex dump script,补充本地 session 摘要++### 🤖 Execution Context+* **Agent ID**: `codex`+* **Base Model**: `GPT-5`+* **Runtime**: `Codex CLI`++### 📥 User Query+> 改进 `codex_dump.py`,让抓包目录里更容易直接看到正常请求、MCP tool 调用和 tool result。++### 🛠 Changes Overview+**Scope:** `scripts`、`README.md`、`docs/references`、`docs/histories`++**Key Actions:**+- **[本地摘要导出]**: 扩展 `scripts/codex_dump.py`,在保留 HTTP / WebSocket dump 的同时,把当前 websocket `session_id` 对应的 `~/.codex/sessions/rollout-*.jsonl` 导出成 `local-sessions/*.json`。+- **[结构化结果]**: `local-sessions/*.json` 只保留高信号字段,包括 `user_prompts`、`tool_calls`、`function_call_output` 解析结果和 `final_answer`,避免直接复制原始 session JSONL。+- **[精确匹配]**: 优先用 websocket 握手头里的 `session_id` 精确关联本地 session,避免把同时间窗口内的旧 session 一起扫进当前 dump 目录。+- **[启动健壮性]**: `scripts/start-codex-mitm-dump.sh` 现在会确认端口真的开始监听后才返回,并把 mitmdump 的 stdin 重定向到 `/dev/null`。+- **[文档顺序更新]**: 更新 runbook,明确新的默认排查顺序是 `websocket/` -> `local-sessions/` -> `logs_2.sqlite`。++### 🧠 Design Intent (Why)+用户真正要分析的通常不是“模型决定调用了哪个 tool”这一半,而是“宿主最终把什么 `function_call` 下发给本地 MCP、又收到了什么 `function_call_output`”。把这层摘要直接落进同一个 dump 目录,能显著降低来回切换 `mitm` 样本和 `~/.codex/sessions` 的成本,也更适合做长期 eval 留档。++### 📁 Files Modified+- `scripts/codex_dump.py`+- `scripts/start-codex-mitm-dump.sh`+- `README.md`+- `docs/references/README.md`+- `docs/references/codex-network-capture.md`+- `docs/references/codex-local-runtime-logs.md`+- `docs/histories/2026-04/20260417-1619-add-codex-mitm-dump-script.md`diff --git a/docs/histories/2026-04/20260417-1812-normalize-repo-path-links.md b/docs/histories/2026-04/20260417-1812-normalize-repo-path-links.mdnew file mode 100644index 0000000..b601fd8--- /dev/null+++ b/docs/histories/2026-04/20260417-1812-normalize-repo-path-links.md@@ -0,0 +1,25 @@+## [2026-04-17 18:12] | Task: 清理仓库内绝对路径链接++### 🤖 Execution Context+* **Agent ID**: `Codex`+* **Base Model**: `GPT-5`+* **Runtime**: `Codex CLI`++### 📥 User Query+> 把仓库里写成仓库根绝对路径的目录引用清理掉,改成相对路径,并在 `AGENTS.md` 对应的文档里强调后续只能写相对路径。++### 🛠 Changes Overview+**Scope:** `README.md`、`docs/REPO_COLLAB_GUIDE.md`、`docs/histories`++**Key Actions:**+- **[清理链接]**: 把 `README.md` 中指向仓库内 reference 文档的两处绝对路径链接改成相对路径。+- **[补充约束]**: 在 `docs/REPO_COLLAB_GUIDE.md` 的文档纪律里新增规则,明确仓库内路径引用不得使用机器相关的绝对路径。+- **[留存变更]**: 新增本次 history,记录路径规范清理的目的和范围。++### 🧠 Design Intent (Why)+仓库内文档如果写死本机绝对路径,会让链接不可移植,也会把环境偶然性带进仓库知识。把 repo-local 引用统一成相对路径,才能保证换机器、换用户目录、换协作者后仍然稳定可用。++### 📁 Files Modified+- `README.md`+- `docs/REPO_COLLAB_GUIDE.md`+- `docs/histories/2026-04/20260417-1812-normalize-repo-path-links.md`diff --git a/docs/references/README.md b/docs/references/README.mdindex b5a28da..8da9630 100644--- a/docs/references/README.md+++ b/docs/references/README.md@@ -16,8 +16,8 @@- `codex-computer-use-reverse-engineering/`- 官方 `Codex Computer Use.app` / `SkyComputerUseClient` 的持续逆向分析资料。- `codex-network-capture.md`- - 用 `mitmdump` + `scripts/codex_dump.py` 抓 Codex 上游 HTTP / WebSocket 流量,并把样本沉淀到 `artifacts/codex-dumps/` 做持续分析。+ - 用 `mitmdump` + `scripts/codex_dump.py` 抓 Codex 上游 HTTP / WebSocket 流量,并把对应 `session_id` 的本地 `function_call` / `function_call_output` 摘要一起沉淀到 `artifacts/codex-dumps/` 做持续分析。- `codex-local-runtime-logs.md`- - 当上游 LLM dump 不足以解释本地 tool / MCP 行为时,补查 Codex 本地 `logs_2.sqlite`;默认顺序仍然是先看上游抓包,再看本地日志。+ - 当抓包目录里的 `websocket/` + `local-sessions/` 仍不足以解释本地 tool / MCP 行为时,再补查 Codex 本地 `logs_2.sqlite`。- `codex-computer-use-cli.md`- 仓库内 `scripts/computer-use-cli/` 的用途、使用方法,以及为什么探测官方 bundled `computer-use` 时要优先走 `codex app-server` 代理而不是 direct stdio。diff --git a/docs/references/codex-local-runtime-logs.md b/docs/references/codex-local-runtime-logs.mdindex 97c9d99..f0b4d1e 100644--- a/docs/references/codex-local-runtime-logs.md+++ b/docs/references/codex-local-runtime-logs.md@@ -5,10 +5,11 @@默认排查顺序应该是:1. 先看 `docs/references/codex-network-capture.md` 对应的上游 HTTP / WebSocket dump。-2. 只有当上游样本不足以解释本地 tool 行为时,再查 Codex 本地日志。-3. 只有当本地日志仍然不够,且确实需要原始 `stdio` JSON-RPC 字节流时,才考虑 wrapper / shadow plugin / 更底层 hook。+2. 再看同一个 dump 目录里的 `local-sessions/*.json`,这里已经会把当前 `session_id` 对应的 `function_call` / `function_call_output` 摘要一起导出来。+3. 只有当这两层仍然不足以解释本地 tool 行为时,再查 Codex 本地日志。+4. 只有当本地日志仍然不够,且确实需要原始 `stdio` JSON-RPC 字节流时,才考虑 wrapper / shadow plugin / 更底层 hook。-大多数情况下,第 1 步已经足够;这份文档是补充路径,不是默认入口。+大多数情况下,前 2 步已经足够;这份文档是更深一层的补充路径,不是默认入口。## 适用场景@@ -19,19 +20,21 @@## 为什么需要这条补充路径-`mitmdump` 抓到的是 Codex 到上游模型服务的 HTTP / WebSocket 流量。它很适合回答:+`mitmdump` 抓到的是 Codex 到上游模型服务的 HTTP / WebSocket 流量;仓库内增强后的 `scripts/codex_dump.py` 还会顺手把同一个 `session_id` 的本地 session 摘要落到 `local-sessions/*.json`。这两层很适合回答:- 模型看到了什么上下文。- 模型何时决定调用某个 tool。- tool call 在模型协议层长什么样。+- Codex 宿主最终把什么 `function_call` 分发给了本地 MCP。+- 本地 tool 返回了什么 `function_call_output`。-但对于官方 `computer-use` 这种本地 `stdio` MCP,网络抓包看不到:+但如果你要看的问题更偏宿主内部日志视角,这两层仍然看不到:-- Codex 宿主与本地 MCP server 的 `stdio` 交换细节。-- 本地 tool 的宿主侧落盘结果。-- 某些本地 tool 调用后的完整输出摘要。+- Codex 宿主与本地 MCP server 的完整 `stdio` 交换字节流。+- `logs_2.sqlite` 里额外的 host 级事件、错误分类和埋点字段。+- 某些不在 session JSONL 摘要里的上下文。-这时补看 Codex 自己的本地日志,通常就足够了。+这时补看 Codex 自己更底层的本地日志,通常就足够了。## 主要日志位置@@ -217,8 +220,9 @@ LIMIT 100;## 推荐工作流1. 先按 `docs/references/codex-network-capture.md` 抓上游样本。-2. 如果上游样本已经能回答问题,就直接在仓库文档里沉淀结论。-3. 如果问题落在本地 `stdio` MCP / `computer-use` 行为上,再查 `logs_2.sqlite`。-4. 只有当本地日志仍然不够,才考虑 wrapper / shadow plugin 方案。+2. 先看同一个 dump 目录里的 `local-sessions/*.json`,确认 `function_call` 和 `function_call_output`。+3. 如果前两层样本已经能回答问题,就直接在仓库文档里沉淀结论。+4. 如果问题仍然落在本地 `stdio` MCP / `computer-use` 更深一层的宿主行为上,再查 `logs_2.sqlite`。+5. 只有当本地日志仍然不够,才考虑 wrapper / shadow plugin 方案。这个顺序能避免一上来就走高侵入、难维护的拦截路径。diff --git a/docs/references/codex-network-capture.md b/docs/references/codex-network-capture.mdindex ee4ff52..98be13d 100644--- a/docs/references/codex-network-capture.md+++ b/docs/references/codex-network-capture.md@@ -5,9 +5,10 @@默认排查顺序里,这份文档应该优先于 `docs/references/codex-local-runtime-logs.md`:- 先看上游 LLM call dump。-- 只有当上游样本不足以解释本地 tool / MCP 行为时,才补查 Codex 自己的本地日志。+- 再看同一个 dump 目录里的 `local-sessions/*.json`,确认本地 `function_call` / `function_call_output`。+- 只有当这两层仍然不足以解释本地 tool / MCP 行为时,才补查 Codex 自己更底层的本地日志。-尤其是官方 `computer-use` 这类本地 `stdio` MCP,网络抓包通常只能回答“模型侧发生了什么”,不能完整回答“本地 tool 最终收到了什么参数、回了什么结果”;这类问题再切到本地日志路径更合适。+现在的 `scripts/codex_dump.py` 会利用 websocket 握手里的 `session_id`,把对应的 `~/.codex/sessions/rollout-*.jsonl` 摘要一起落到当前 session 目录里,所以很多关于官方 `computer-use` 的问题,已经不需要第一时间跳去查 `logs_2.sqlite`。## 适用场景@@ -95,12 +96,12 @@ codex exec --skip-git-repo-check -C /tmp 'reply with one word: ok'session_dir="artifacts/codex-dumps/$(date +%Y%m%d-%H%M%S)-basic-ok"mkdir -p "$session_dir"-nohup mitmdump \+nohup setsid mitmdump \--listen-host 127.0.0.1 \--listen-port 8082 \-s scripts/codex_dump.py \--set codex_dump_dir="$session_dir" \- >"$session_dir/mitmdump.log" 2>&1 &+ </dev/null >"$session_dir/mitmdump.log" 2>&1 &echo $! >"$session_dir/mitmdump.pid"```@@ -163,6 +164,7 @@ https://chatgpt.com/backend-api/codex/responsesartifacts/codex-dumps/<session-name>/http/websocket/+ local-sessions/```其中:@@ -171,11 +173,22 @@ artifacts/codex-dumps/<session-name>/保存匹配到的 HTTP 请求和响应。- `websocket/*.jsonl`按事件逐行保存 WebSocket 开始、消息和结束。+- `local-sessions/*.json`+ 从 `~/.codex/sessions/rollout-*.jsonl` 导出的结构化摘要,只保留当前抓包 `session_id` 命中的 user prompt、tool call、tool result 和 final answer。- `mitmdump.log`如果用后台方式启动,会包含 mitmdump 自身日志。- `mitmdump.pid`如果用后台方式启动,会保存后台进程 PID。+把这三层串起来看,通常就能直接回答:++- `websocket/`+ 模型何时决定调用哪个 tool,以及调用参数是什么。+- `local-sessions/`+ Codex 宿主实际把哪个 `function_call` 分发给了本地 MCP,以及 `function_call_output` 返回了什么。+- `http/`+ 非 websocket 的补充请求,例如 `wham/apps`、plugin/config 初始化等。+## 脱敏默认值仓库内的 `scripts/codex_dump.py` 默认会对以下内容做脱敏:@@ -205,6 +218,8 @@ artifacts/codex-dumps/<session-name>/4. 结束后停止 `mitmdump`。5. 重点分析:- `websocket/` 里的 `response.create`+ - `websocket/` 里的 `response.output_item.done`,尤其是 `item.type=="function_call"`+ - `local-sessions/` 里的对应 `tool_calls[].output`- `response.output_text.delta`- `response.completed`6. 把结论、差异和评估结果沉淀到仓库文档,而不是直接提交原始抓包。@@ -219,6 +234,7 @@ artifacts/codex-dumps/<session-name>/- 是否设置了 `SSL_CERT_FILE=$HOME/.mitmproxy/mitmproxy-ca-cert.pem`- 是否在抓 `chatgpt.com/backend-api/codex/responses`- 是否只盯着 HTTP,而没有看 `websocket/*.jsonl`+- 如果要看本地 tool 结果,是否同时检查了 `local-sessions/*.json`### 2. 为什么不直接抓 `api.openai.com`@@ -267,3 +283,55 @@ printf '%s\n%s\n%s\n' \- 插件缓存或安装态没有同步到最新不要一开始就假设是 MCP server 逻辑本身坏了。先把“宿主问题”和“server 问题”分开,排查成本会低很多。++### 6. 做官方 `computer-use` 和 `open-computer-use` A/B 时,建议显式隔离另一个 plugin++如果两套 plugin 同时启用,prompt 里的 tool 名称锚点会显著影响模型路由。做更纯净的 A/B 时,建议在单次 `codex exec` 上临时关闭另一套 plugin,而不是直接共存跑。++仓库内已经提供 helper:++```bash+./scripts/run-isolated-codex-exec.sh computer-use --skip-git-repo-check -C /tmp \+ '使用computer-use列出正在运行的前三个应用'++./scripts/run-isolated-codex-exec.sh open-computer-use --skip-git-repo-check -C /tmp --json \+ '使用open-computer-use列出正在运行的前三个应用'+```++它本质上只是给 `codex exec` 增加临时 config override:++- `computer-use`+ 会加 `-c 'plugins."open-computer-use@open-computer-use-local".enabled=false'`+- `open-computer-use`+ 会加 `-c 'plugins."computer-use@openai-bundled".enabled=false'`++这比直接改 `~/.codex/config.toml` 更安全,因为:++- 只影响当前这一条命令。+- 不需要手工改全局配置后再恢复。+- 更适合批量 eval 或 Agent 后台脚本。++### 7. 当前这台机器上的隔离验证结论++2026-04-17 的隔离样本已经验证:++1. `-c 'plugins."...".enabled=false'` 这条覆写是生效的。+2. 只保留官方 `computer-use` 时,`computer-use/list_apps` 可以正常完成。+3. 只保留 `open-computer-use` 时,`open-computer-use/list_apps` 仍然直接返回 `user cancelled MCP tool call`。+4. 同时,直接对 `./plugins/open-computer-use/scripts/launch-open-computer-use.sh` 发 JSON-RPC 的 `tools/list` / `tools/call list_apps` 是正常的。++这说明在当前环境里:++- “两套 plugin 互相干扰”不是主要问题。+- `open-computer-use` MCP server 本身不是这一步的主故障点。+- 更可能的瓶颈仍然在 Codex host 对第三方 plugin 调用的 gate 或会话策略上。++### 8. 后台 `mitmdump` 显示启动成功,但代理端口很快就没了++先看 `mitmdump.log`。仓库内 `scripts/start-codex-mitm-dump.sh` 现在会额外检查端口是否真的开始监听,但在某些受控 runner / agent 宿主里,父进程结束时仍然可能把后台子进程一起清掉。++如果你遇到这种情况,优先用下面几种更稳的方式:++1. 在真实登录 shell、`tmux` 或单独终端里执行 `./scripts/start-codex-mitm-dump.sh`。+2. 或者直接前台运行 `mitmdump` / `mitmweb`,再在另一个终端执行 `codex exec`。+3. 不要把“脚本返回了 PID”误当成“代理一定还活着”,先用 `lsof -nP -iTCP:<port> -sTCP:LISTEN` 确认监听状态。diff --git a/scripts/codex_dump.py b/scripts/codex_dump.pyindex 6fb284b..996d3b3 100644--- a/scripts/codex_dump.py+++ b/scripts/codex_dump.py@@ -3,7 +3,6 @@from __future__ import annotationsimport json-import osimport reimport tempfilefrom datetime import datetime, timezone@@ -40,6 +39,12 @@ SENSITIVE_JSON_KEYS = {"cookie",}TRUNCATE_PREVIEW_BYTES = 240+SESSION_TEXT_PREVIEW_BYTES = 4000+SESSION_MATCH_WINDOW_SECONDS = 6 * 60 * 60+DEFAULT_SESSION_ROOT = Path.home() / ".codex" / "sessions"+IGNORED_SESSION_TEXT_PREFIXES = (+ "<environment_context>",+)def _utc_now() -> str:@@ -77,6 +82,43 @@ def _json_preview(value: Any) -> str:return text[:TRUNCATE_PREVIEW_BYTES] + "...<truncated>"+def _truncate_text(value: str, limit: int = SESSION_TEXT_PREVIEW_BYTES) -> str:+ if len(value) <= limit:+ return value+ return value[:limit] + "...<truncated>"+++def _truncate_nested(value: Any, limit: int = SESSION_TEXT_PREVIEW_BYTES) -> Any:+ if isinstance(value, dict):+ return {key: _truncate_nested(item, limit=limit) for key, item in value.items()}+ if isinstance(value, list):+ return [_truncate_nested(item, limit=limit) for item in value]+ if isinstance(value, str):+ return _truncate_text(value, limit=limit)+ return value+++def _normalize_text(value: str) -> str:+ return re.sub(r"\s+", " ", value).strip()+++def _extract_text_blocks(content: Any) -> list[str]:+ if isinstance(content, str):+ return [content]+ if not isinstance(content, list):+ return []++ texts: list[str] = []+ for item in content:+ if not isinstance(item, dict):+ continue+ for key in ("text", "input_text", "output_text"):+ value = item.get(key)+ if isinstance(value, str) and value:+ texts.append(value)+ return texts++class CodexDump:def __init__(self) -> None:self.output_dir: Path | None = None@@ -84,9 +126,16 @@ class CodexDump:self.path_patterns: list[re.Pattern[str]] = []self.include_http = Trueself.include_ws = True+ self.include_local_sessions = True+ self.local_session_root = DEFAULT_SESSION_ROOT+ self.capture_started_at = datetime.now(timezone.utc)self._http_counter = 0self._ws_counter = 0self._ws_files: dict[str, Path] = {}+ self._observed_session_ids: set[str] = set()+ self._observed_prompts: set[str] = set()+ self._local_session_exports: dict[Path, tuple[str, Path]] = {}+ self._local_session_counter = 0def load(self, loader) -> None:loader.add_option(@@ -119,6 +168,24 @@ class CodexDump:default=True,help="Whether to persist matching WebSocket messages.",)+ loader.add_option(+ name="codex_dump_include_local_sessions",+ typespec=bool,+ default=True,+ help="Whether to export recent ~/.codex/sessions JSONL summaries into the dump directory.",+ )+ loader.add_option(+ name="codex_dump_local_session_root",+ typespec=str,+ default=str(DEFAULT_SESSION_ROOT),+ help="Directory that contains Codex local rollout session JSONL files.",+ )+ loader.add_option(+ name="codex_dump_local_session_window_seconds",+ typespec=int,+ default=SESSION_MATCH_WINDOW_SECONDS,+ help="How far back to scan local session files when exporting summaries.",+ )def configure(self, updated) -> None:hosts = [item.strip() for item in ctx.options.codex_dump_hosts.split(",") if item.strip()]@@ -127,6 +194,17 @@ class CodexDump:self.path_patterns = [re.compile(pattern) for pattern in paths]self.include_http = bool(ctx.options.codex_dump_include_http)self.include_ws = bool(ctx.options.codex_dump_include_ws)+ self.include_local_sessions = bool(ctx.options.codex_dump_include_local_sessions)+ self.capture_started_at = datetime.now(timezone.utc)+ self._observed_session_ids.clear()+ self._observed_prompts.clear()+ self._local_session_exports.clear()+ self._local_session_counter = 0+ self._http_counter = 0+ self._ws_counter = 0+ self._ws_files.clear()+ self.local_session_root = Path(ctx.options.codex_dump_local_session_root).expanduser()+ self.local_session_window_seconds = int(ctx.options.codex_dump_local_session_window_seconds)if ctx.options.codex_dump_dir:base_dir = Path(ctx.options.codex_dump_dir).expanduser()@@ -138,6 +216,8 @@ class CodexDump:self.output_dir.mkdir(parents=True, exist_ok=True)(self.output_dir / "http").mkdir(exist_ok=True)(self.output_dir / "websocket").mkdir(exist_ok=True)+ if self.include_local_sessions:+ (self.output_dir / "local-sessions").mkdir(exist_ok=True)def request(self, flow: http.HTTPFlow) -> None:if self.include_http and self._matches(flow) and not self._is_ws_upgrade(flow):@@ -149,6 +229,7 @@ class CodexDump:self._ws_counter += 1request = flow.request+ self._observe_request_session(request)path_stem = _safe_stem(request.path.split("?", 1)[0].strip("/"))file_path = self.output_dir / "websocket" / f"{self._ws_counter:03d}-{path_stem}.jsonl"self._ws_files[flow.id] = file_path@@ -190,6 +271,7 @@ class CodexDump:redacted = _redact_json(parsed)payload["json"] = redactedpayload["preview"] = _json_preview(redacted)+ self._observe_ws_payload(redacted)else:payload["text"] = decodedpayload["preview"] = decoded[:TRUNCATE_PREVIEW_BYTES]@@ -212,6 +294,10 @@ class CodexDump:"close_reason": getattr(flow.websocket, "close_reason", None),}self._append_jsonl(file_path, payload)+ self._sync_local_sessions()++ def done(self) -> None:+ self._sync_local_sessions()def _matches(self, flow: http.HTTPFlow) -> bool:request = flow.request@@ -299,5 +385,244 @@ class CodexDump:with file_path.open("a", encoding="utf-8") as handle:handle.write(json.dumps(payload, ensure_ascii=False) + "\n")+ def _observe_ws_payload(self, payload: Any) -> None:+ if not isinstance(payload, dict):+ return++ if payload.get("type") == "response.create":+ for prompt in self._extract_observed_prompts(payload):+ normalized = _normalize_text(prompt)+ if normalized:+ self._observed_prompts.add(normalized)++ if payload.get("type") == "response.completed":+ self._sync_local_sessions()++ def _extract_observed_prompts(self, payload: dict[str, Any]) -> list[str]:+ prompts: list[str] = []+ for item in payload.get("input", []):+ if not isinstance(item, dict):+ continue+ if item.get("type") != "message" or item.get("role") != "user":+ continue+ prompts.extend(_extract_text_blocks(item.get("content")))+ return prompts++ def _sync_local_sessions(self) -> None:+ if not self.include_local_sessions or self.output_dir is None:+ return+ if self.local_session_root is None or not self.local_session_root.exists():+ return++ for session_path in sorted(self.local_session_root.rglob("rollout-*.jsonl")):+ if self._observed_session_ids and not any(session_id in session_path.name for session_id in self._observed_session_ids):+ continue+ try:+ stat = session_path.stat()+ except OSError:+ continue++ age_seconds = (self.capture_started_at.timestamp() - stat.st_mtime)+ if age_seconds > self.local_session_window_seconds:+ continue++ signature = f"{stat.st_mtime_ns}:{stat.st_size}"+ existing = self._local_session_exports.get(session_path)+ if existing is not None and existing[0] == signature:+ continue++ summary = self._summarize_local_session(session_path)+ if summary is None or not self._session_matches_capture(summary):+ continue++ if existing is None:+ self._local_session_counter += 1+ session_id = summary.get("session", {}).get("id") or session_path.stem+ export_path = self.output_dir / "local-sessions" / f"{self._local_session_counter:03d}-{_safe_stem(session_id)}.json"+ else:+ export_path = existing[1]++ export_path.write_text(json.dumps(summary, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")+ self._local_session_exports[session_path] = (signature, export_path)+ ctx.log.info(f"codex_dump local_session -> {export_path}")++ def _summarize_local_session(self, session_path: Path) -> dict[str, Any] | None:+ try:+ raw_lines = session_path.read_text(encoding="utf-8").splitlines()+ except OSError:+ return None++ session_meta: dict[str, Any] = {}+ user_prompts: list[str] = []+ tool_calls: list[dict[str, Any]] = []+ tool_call_index: dict[str, dict[str, Any]] = {}+ final_answers: list[str] = []++ for raw_line in raw_lines:+ try:+ entry = json.loads(raw_line)+ except json.JSONDecodeError:+ continue++ entry_type = entry.get("type")+ payload = entry.get("payload")++ if entry_type == "session_meta" and isinstance(payload, dict):+ session_meta = {+ "id": payload.get("id"),+ "timestamp": payload.get("timestamp"),+ "cwd": payload.get("cwd"),+ "originator": payload.get("originator"),+ "cli_version": payload.get("cli_version"),+ "source": payload.get("source"),+ "model_provider": payload.get("model_provider"),+ }+ continue++ if entry_type != "response_item" or not isinstance(payload, dict):+ continue++ payload_type = payload.get("type")++ if payload_type == "message":+ role = payload.get("role")+ texts = _extract_text_blocks(payload.get("content"))+ if role == "user":+ for text in texts:+ if not any(text.startswith(prefix) for prefix in IGNORED_SESSION_TEXT_PREFIXES):+ user_prompts.append(text)+ if role == "assistant" and payload.get("phase") == "final_answer":+ final_answers.extend(texts)+ continue++ if payload_type == "function_call":+ call = {+ "timestamp": entry.get("timestamp"),+ "call_id": payload.get("call_id"),+ "namespace": payload.get("namespace"),+ "name": payload.get("name"),+ "arguments": self._parse_function_arguments(payload.get("arguments")),+ }+ tool_calls.append(call)+ call_id = payload.get("call_id")+ if isinstance(call_id, str) and call_id:+ tool_call_index[call_id] = call+ continue++ if payload_type == "function_call_output":+ output_summary = {+ "timestamp": entry.get("timestamp"),+ "result": self._parse_function_output(payload.get("output")),+ }+ call_id = payload.get("call_id")+ existing_call = tool_call_index.get(call_id) if isinstance(call_id, str) else None+ if existing_call is not None:+ existing_call["output"] = output_summary+ else:+ tool_calls.append(+ {+ "timestamp": entry.get("timestamp"),+ "call_id": call_id,+ "output": output_summary,+ }+ )++ unique_prompts = []+ seen_prompts: set[str] = set()+ for prompt in user_prompts:+ normalized = _normalize_text(prompt)+ if not normalized or normalized in seen_prompts:+ continue+ seen_prompts.add(normalized)+ unique_prompts.append(prompt)++ final_answer = "\n".join(text for text in final_answers if text).strip()++ return {+ "captured_at": _utc_now(),+ "source_path": str(session_path),+ "session": session_meta,+ "user_prompts": [_truncate_text(text) for text in unique_prompts],+ "tool_calls": tool_calls,+ "final_answer": _truncate_text(final_answer) if final_answer else None,+ }++ def _session_matches_capture(self, summary: dict[str, Any]) -> bool:+ if self._observed_prompts:+ session_prompts = [_normalize_text(item) for item in summary.get("user_prompts", []) if isinstance(item, str)]+ for session_prompt in session_prompts:+ if not session_prompt:+ continue+ for observed_prompt in self._observed_prompts:+ if session_prompt in observed_prompt or observed_prompt in session_prompt:+ summary["matched_prompt"] = _truncate_text(session_prompt)+ return True+ return False++ session_meta = summary.get("session")+ if not isinstance(session_meta, dict):+ return False++ session_id = session_meta.get("id")+ if self._observed_session_ids:+ return isinstance(session_id, str) and session_id in self._observed_session_ids++ timestamp = session_meta.get("timestamp")+ if not isinstance(timestamp, str):+ return False++ try:+ session_time = datetime.fromisoformat(timestamp.replace("Z", "+00:00"))+ except ValueError:+ return False++ delta = abs((session_time - self.capture_started_at).total_seconds())+ return delta <= self.local_session_window_seconds++ def _observe_request_session(self, request: http.Request) -> None:+ for header_name in ("session_id", "x-client-request-id"):+ value = request.headers.get(header_name)+ if isinstance(value, str) and value:+ self._observed_session_ids.add(value)++ metadata = request.headers.get("x-codex-turn-metadata")+ if not isinstance(metadata, str) or not metadata:+ return++ parsed = self._try_parse_json(metadata)+ if not isinstance(parsed, dict):+ return++ session_id = parsed.get("session_id")+ if isinstance(session_id, str) and session_id:+ self._observed_session_ids.add(session_id)++ def _parse_function_arguments(self, arguments: Any) -> Any:+ if not isinstance(arguments, str):+ return _truncate_nested(_redact_json(arguments))+ parsed = self._try_parse_json(arguments)+ if parsed is not None:+ return _truncate_nested(_redact_json(parsed))+ return _truncate_text(arguments)++ def _parse_function_output(self, output: Any) -> Any:+ if not isinstance(output, str):+ return _truncate_nested(_redact_json(output))++ match = re.match(r"^Wall time:\s*([0-9.]+)\s+seconds\nOutput:\n(.*)$", output, flags=re.DOTALL)+ if match is None:+ return {"text": _truncate_text(output)}++ result: dict[str, Any] = {+ "wall_time_seconds": float(match.group(1)),+ }+ body = match.group(2)+ parsed = self._try_parse_json(body)+ if parsed is not None:+ result["output"] = _truncate_nested(_redact_json(parsed))+ else:+ result["text"] = _truncate_text(body)+ return result+addons = [CodexDump()]diff --git a/scripts/run-isolated-codex-exec.sh b/scripts/run-isolated-codex-exec.shnew file mode 100755index 0000000..063c09c--- /dev/null+++ b/scripts/run-isolated-codex-exec.sh@@ -0,0 +1,39 @@+#!/usr/bin/env bash++set -euo pipefail++if [[ $# -lt 2 ]]; then+ cat <<'EOF' >&2+Usage:+ ./scripts/run-isolated-codex-exec.sh <computer-use|open-computer-use|all> [codex exec args...]++Examples:+ ./scripts/run-isolated-codex-exec.sh computer-use --skip-git-repo-check -C /tmp '使用computer-use列出正在运行的前三个应用'+ ./scripts/run-isolated-codex-exec.sh open-computer-use --skip-git-repo-check -C /tmp --json '使用open-computer-use列出正在运行的前三个应用'+ ./scripts/run-isolated-codex-exec.sh all --skip-git-repo-check -C /tmp 'reply with one word: ok'+EOF+ exit 1+fi++mode="$1"+shift++declare -a overrides=()++case "${mode}" in+ computer-use)+ overrides=(-c 'plugins."open-computer-use@open-computer-use-local".enabled=false')+ ;;+ open-computer-use)+ overrides=(-c 'plugins."computer-use@openai-bundled".enabled=false')+ ;;+ all)+ ;;+ *)+ echo "Unsupported mode: ${mode}" >&2+ echo "Expected one of: computer-use, open-computer-use, all" >&2+ exit 1+ ;;+esac++exec codex exec "${overrides[@]}" "$@"diff --git a/scripts/start-codex-mitm-dump.sh b/scripts/start-codex-mitm-dump.shindex cbeaf39..903562e 100755--- a/scripts/start-codex-mitm-dump.sh+++ b/scripts/start-codex-mitm-dump.sh@@ -72,17 +72,31 @@ export NO_PROXY="127.0.0.1,localhost"export SSL_CERT_FILE="${ca_cert}"EOF-nohup mitmdump \+launcher=(nohup)+if command -v setsid >/dev/null 2>&1; then+ launcher+=(setsid)+fi+launcher+=(mitmdump)++"${launcher[@]}" \--listen-host "${listen_host}" \--listen-port "${listen_port}" \-s "${addon_script}" \--set codex_dump_dir="${session_dir}" \- >"${log_path}" 2>&1 &+ </dev/null >"${log_path}" 2>&1 &pid=$!echo "${pid}" >"${pid_path}"-sleep 1+for _ in $(seq 1 20); do+ if ! kill -0 "${pid}" >/dev/null 2>&1; then+ break+ fi+ if lsof -nP -iTCP:"${listen_port}" -sTCP:LISTEN >/dev/null 2>&1; then+ break+ fi+ sleep 0.25+doneif ! kill -0 "${pid}" >/dev/null 2>&1; thenecho "mitmdump 启动失败,最近日志:" >&2@@ -92,6 +106,14 @@ if ! kill -0 "${pid}" >/dev/null 2>&1; thenexit 1fi+if ! lsof -nP -iTCP:"${listen_port}" -sTCP:LISTEN >/dev/null 2>&1; then+ echo "mitmdump 进程已启动,但端口 ${listen_port} 尚未开始监听。" >&2+ if [[ -f "${log_path}" ]]; then+ tail -n 40 "${log_path}" >&2+ fi+ exit 1+fi+cat <<EOFmitmdump 已后台启动。