A Codex plugin that reads your local Codex session history and grades how well you collaborate with the agent. 9 dimensions across 3 categories — Communication, Engineering, Outcome. Your own Codex model scores the sessions against a transparent rubric and writes a free-form diagnosis, plus paste-ready improvement prompts — all in a professional, readable single-file HTML dashboard. 100% local.
🌏 中文版 · 📖 Methodology · 🖥 Live preview · ⚖️ License
中文简介： 一款 Codex 插件，读取你本地的 Codex 会话记录，从「沟通力 / 工程力 / 成效」三个层面、9 个维度评估你和 Codex 的协作质量。由你自己的 Codex 模型对照公开 rubric 打分并撰写诊断，输出一组可直接粘贴的改写 prompt 和一个专业可读的单文件 HTML dashboard。全程本地。完整中文文档 → README_zh.md
🎯 Reads your actual Codex sessions, not synthetic prompts. Codex Radar parses the real JSONL your agent already writes — every prompt you sent, every shell command and its exit code, every apply_patch edit, every plan update, MCP / web-search / image / subagent call. The score reflects how you actually worked, not how you answer a quiz.
💬 Your Codex writes you a coaching note, not just a scoreboard. A deterministic parser extracts the facts; then your own Codex model scores them against a transparent rubric — a formula baseline plus a bounded, evidence-cited ±15 adjustment — and writes a free-form collaboration diagnosis. Every claim points back to your real messages.
📋 Every suggestion is a paste-ready prompt. No "be more thoughtful" advice. Each improvement comes with a concrete prompt you can copy straight into your next thread, plus the dimension it lifts and the expected impact.
⚖️ Project-type-aware weighting. A 3-message fix isn't judged on the same scale as a 50-session feature build. Codex Radar auto-classifies each project (one-shot / feature-build / long-running / learning) and applies different category weights. When a signal genuinely can't be evaluated — e.g. Architecture when the project directory no longer exists on this machine — the dimension is marked N/A instead of faked.
🛠 Scores how you drive the platform, not just how you talk. The Engineering category reads shell-command success rate, patching, plan updates, MCP servers, web search, image tools, and subagent spawns. Architecture rewards durable setup — AGENTS.md, repo scaffolding, tests, manifests — the leverage most users underuse.
🔒 100% local, zero telemetry. Read-only access to ~/.codex/sessions. No API key, no cloud, no network calls. Reports are escaped single-file HTML written to ~/.codex-radar/reports/.
中文要点：
- 🎯 评估你真实的 Codex 会话记录，直接解析 prompt、shell 命令与退出码、
apply_patch编辑、plan 更新、MCP / 搜索 / 图像 / 子代理调用- 💬 你的 Codex 亲自写诊断：解析器提取事实，模型对照公开 rubric 打分（公式基线 + 引用证据的 ±15 微调）并撰写自由格式诊断，每条结论都指向你的真实消息
- 📋 每条建议都是可粘贴 prompt，附带它能提升的维度和预期影响
- ⚖️ 按项目类型公平加权：一次性任务 / 功能开发 / 长期项目 / 学习探索 各用不同权重；信号无法评估时维度标记 N/A，绝不假装给 50 分
- 🛠 专门评估你怎么用平台：shell 成功率、补丁、plan、MCP、搜索、图像、子代理；Architecture 奖励 AGENTS.md 与仓库脚手架
- 🔒 100% 本地、零上传：只读访问
~/.codex/sessions，无 API key，无网络请求完整中文版 → README_zh.md
Ask Codex to run the plugin and you get a single-file, professional, readable HTML dashboard:
Overall grade — S through D, paired with your project profile so you know what scale you're being judged on.
Three category scorecards — each scoring 0-100 with its 3 dimensions:
| Category | Dimensions |
|---|---|
| Communication | Lock-On 瞄准力 · Scene Setting 铺场力 · Steering 校准力 |
| Engineering | Toolcraft 工具调度 · Architecture 工程脚手架 · Tempo 推进节奏 |
| Outcome | Efficiency 产出效率 · Proof Check 验证意识 · Completion 闭环完成 |
Diagnosis — your strongest signal, your main bottleneck, a collaboration summary, and a cross-category pattern read — all computed from your own session, with per-dimension evidence shown inline.
Improvement prompts — a prioritized set of paste-ready prompt rewrites (up to 7), one for each dimension with clear room to improve, each tagged with its expected score impact.
Signals & metrics — top tools, tool categories, project assets (AGENTS.md, .git, tests, manifests), command success rate, context compactions, and a key-messages table with the signals detected in each.
Step 1 — Add the marketplace (this repo is itself a Codex marketplace):
codex plugin marketplace add LeifDiao/codex-radarStep 2 — Install the plugin:
codex plugin add codex-radar@codex-radar-marketplaceAlternative (local checkout):
git clone https://github.com/LeifDiao/codex-radar.git ~/codex-radar
codex plugin marketplace add ~/codex-radar
codex plugin add codex-radar@codex-radar-marketplaceStart a new thread after installing so Codex picks up the new skill.
Codex Radar runs through natural language — there's no slash command to memorize. In a new thread, ask:
Run Codex Radar on this project
- Codex Radar detects your current working directory in local session history and asks whether to analyze it.
- Confirm, or pick from the recent-projects list.
- It parses and scores the sessions — a few seconds, pure Node, no waiting on a model.
- The dashboard is written to
~/.codex-radar/reports/and the path is printed for you to open.
Starter prompts like "Analyze my Codex collaboration" or "Create a Codex Radar report" work too.
- Codex CLI with plugin support
- Node.js 18+
- No
npm install, no build step, no server
Your session data stays on your machine:
- Everything runs locally — no network calls, no API key, no telemetry
~/.codex/sessions,~/.codex/archived_sessions, and~/.codex/session_index.jsonlare read-only- Reports write to
~/.codex-radar/reports/(temp JSON to~/.codex-radar/temp/) - Project-asset detection only checks whether files like
AGENTS.md,.git, or a tests folder exist — it never reads their contents - Reports may include short snippets of your own prompts as evidence — treat the HTML as private unless you intentionally share it
See PRIVACY.md for the full data map.
Two layers, mirroring ClaudeRadar:
- Facts —
parse-codex-project.mjsreads the matching session files and extracts countable signals (message patterns, tool categories, command exit codes, patch events, proof commands, project assets). Deterministic: same input → same facts. - Scoring + diagnosis — your own Codex model reads those facts and
data/rubric.json, computes each dimension's formula baseline, applies a bounded ±15 evidence-based adjustment, then writes a free-form diagnosis and paste-ready prompts. The rubric is the public scoring constitution. - Render —
render-report.mjsturns the report JSON into a single self-contained HTML file.
The analysis runs inside your own Codex session — the plugin makes no network calls of its own and needs no separate API key.
All scoring inputs live in plugins/codex-radar/data/rubric.json:
- 9 dimension definitions (English + 中文) and their categories
- 4 project profiles with per-profile category weight tables
- Grade thresholds (S / A / B / C / D)
Want scoring to match your team's standards? Edit this file and the dimension formulas in parse-codex-project.mjs — the parser re-reads the rubric on every run.
This repo is itself a Codex plugin marketplace.
codex-radar/
├── .agents/plugins/marketplace.json # Codex marketplace manifest
├── docs/
│ ├── index.html # GitHub Pages landing page
│ ├── METHODOLOGY.md / METHODOLOGY_zh.md # scoring spec (EN + 中文)
└── plugins/
└── codex-radar/
├── .codex-plugin/plugin.json # plugin manifest
├── data/rubric.json # 9-dim definitions + profile weights
└── skills/analyze/
├── SKILL.md # orchestration: detect → parse → render
└── scripts/
├── lib.mjs # shared Codex session helpers
├── list-codex-projects.mjs # scan ~/.codex/sessions + cwd match
├── parse-codex-project.mjs # signal extraction → facts JSON (no scoring)
└── render-report.mjs # report JSON → single-file HTML
Zero runtime dependencies.
Codex Radar is released under CC BY-NC 4.0:
- ✅ Free for personal, educational, research, and any non-commercial use
- ✅ Forking, modifying, sharing is welcomed — please attribute the original repo and indicate any changes
- ❌ Commercial use (bundling into paid products, internal use beyond individual scope in for-profit companies, paid SaaS hosting, selling reports/analyses based on the scoring) requires a separate license
For commercial licensing, contact: leifdiao@gmail.com
See LICENSE for the full terms.
Built for people who care about the quality of AI collaboration.