{"repo": "agents365-ai/pi-plugin-cc", "files": {"README_CN.md": {"b6a5c2dd0c19d427b3c02c2e048db2347a499804": {"t": 1785674335.297288, "n": 11645, "h": "b5dc33d29c1501ae", "c": "# pi-plugin-cc — 在 Claude Code 和 Codex 里驱动 Pi 编码 agent 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#选模型)\n\n[English](README.md) · **中文**\n\n外部参考：[Pi 编码 agent](https://github.com/earendil-works/pi) · [Pi RPC 模式](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\n把代码评审与编码任务从 Claude Code 转交给 [Pi 编码 agent](https://github.com/earendil-works/pi) 的插件。改编自 [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc)，把底层 runtime 从 Codex 换成 Pi。也可以在 OpenAI Codex CLI 里使用 —— 见[在 Codex 里使用](#-在-codex-里使用)。\n\n**硬依赖是 pi 本身，不是某个具体的大模型。** Pi 可以配置成 DeepSeek、OpenAI、Anthropic、Google、Ollama、LM Studio，或者任何 OpenAI 兼容端点 —— 通过 `~/.pi/agent/models.json`。插件默认把模型选择完全交给 pi，除非你单次 `--model` 覆盖。\n\n- **代码评审**：针对 working tree 或基于分支的 diff，输出结构化 findings\n- **对抗式评审**：质疑设计本身，而不是只挑改动里的拼写错误\n- **任务转交**：诊断、重构、长流程救援，前台或后台执行\n- **并行分发**：`/pi:parallel-rescue` 通过 [`pi-subagents`](https://github.com/nicobailon/pi-subagents) 并发跑多个独立任务\n- **分片并行评审**：`/pi:review --shards ` 把大 diff 改动的文件拆成 N 个并行评审任务，再合并 findings\n- **后台作业控制**：`status`、`result`、`cancel`，以及可选的 stop-time 评审守门\n- **无 OAuth 登录**：Pi 用 provider 的 API key 认证，不需要 `codex login`\n\n深度集成 [`pi-subagents`](https://github.com/nicobailon/pi-subagents)（`pi install npm:pi-subagents`）：`/pi:setup` 会检测安装状态并列出 agent 档案，`/pi:rescue` 的 prompt 会告知 Pi 可用 `subagent` 工具，`/pi:parallel-rescue` 把多个任务分发给并行子代理（scout、researcher、planner、worker、reviewer 等）。\n\n## 🔄 工作流程\n\n\"pi-plugin-cc\n\nCodex 的 broker 层被去掉 —— Pi 是\"一进程一会话\"模型，插件给每个任务直接 spawn 一个新的 `pi --mode rpc`。后台作业用工作区作用域的状态文件追踪。Pi 没有 `outputSchema` 这种参数，所以评审 prompt 直接把 JSON schema 内联进去。\n\n## 斜杠命令\n\n| 命令 | 作用 |\n|---|---|\n| `/pi:setup` | 检查 `pi` 是否安装、provider 是否配置；切换 stop-time 评审守门 |\n| `/pi:review` | 针对本地 git 状态的标准代码评审 |\n| `/pi:adversarial-review` | 可指定 focus 的对抗式评审 —— 质疑实现方案本身 |\n| `/pi:rescue` | 把任务转交给 `pi:pi-companion-forwarder` 子代理跑一个 Pi 会话 |\n| `/pi:parallel-rescue` | 通过 pi-subagents 并行跑多个独立任务（`subagent({ tasks })` 分发） |\n| `/pi:status [job-id]` | 列出本仓库正在运行 / 最近完成的 Pi 作业 |\n| `/pi:result ` | 查看一个已完成作业的最终输出 |\n| `/pi:cancel ` | 终止一个正在运行的后台作业 |\n\n每个命令都接受 `--model ` 临时指定模型。不传 `--model`、也没配 env 覆盖（见[选模型](#选模型)）时，pi 用它自己默认的模型。\n\n## 快速开始\n\n```bash\n# 1. 装 pi（必需）\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. 配一个 provider —— 任选其一\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# 或本地跑：见 https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. 验证 pi 看到了模型\npi --list-models | head\n```\n\n在 Claude Code 里安装插件：\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` 给出就绪报告；如果 `pi` 没装但 `npm` 在 PATH，它会问你要不要替你装。\n\n## 使用方式\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:review --out-file review.md\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue 调研一下 Windows CI 为什么编译失败\n> /pi:rescue --background --model gpt-4o 重构 src/payments/\n> /pi:parallel-rescue \"审计 auth 模块\" \"给 db 查询做基准测试\" \"更新 API 文档\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` 会经 `set_thinking_level` 传给 Pi。不支持 thinking 的模型会静默忽略（插件会往 stderr 写一行提示）。\n\n`--out-file `（用于 `/pi:review`、`/pi:adversarial-review`、`/pi:rescue`、`/pi:result`）把 Pi 的完整输出写到文件，只返回一段简短摘要——verdict、findings 计数、每条一行。繁重推理本就跑在更便宜的模型上；这一步还把大段结果挡在调用方的上下文之外，大型评审就不会在\"转述\"上烧 Claude Code 的 token。要看全文就打开那个文件。\n\n`--incremental`（用于 `/pi:review`、`/pi:adversarial-review`）只评审当前分支上自上次评审以来的新提交，靠一份按分支缓存的\"上次评审提交\"记录来实现——跳过已经审过的代码，省下 Pi 的输入 token 和时间。没有可用缓存时会自动回退为全量评审。\n\n## 🧩 在 Codex 里使用\n\n本插件的核心是一个与宿主无关的 CLI（`pi-companion.mjs`）—— 任何能跑 shell 命令的 coding agent 都能驱动它。对 OpenAI 的 Codex CLI，安装自带的 [custom prompts](codex-prompts/)：\n\n```bash\n# 1. 克隆仓库（任意位置；prompts 默认假设在 ~/pi-plugin-cc）\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. 安装 Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. 如果克隆到了别处，把 prompts 指向它\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\n之后在 Codex 里用：`/pi-review`、`/pi-adversarial-review`、`/pi-rescue`、`/pi-parallel-rescue`、`/pi-status`、`/pi-result`、`/pi-cancel`、`/pi-setup`（注意是 `-` 不是 `:` —— Codex 的 prompt 名不能带冒号）。\n\nCodex 下不可用的部分：stop-time 评审守门和会话恢复询问（两者依赖 Claude Code 的 hooks / 子代理）。其余功能 —— 包括 pi-subagents 并行分发 —— 行为一致。\n\n## 🧑‍⚖️ 多模型评审团\n\n单个评审者有盲区，评审团的盲区不重叠。给任一评审命令传 `--models`，同一份 diff 会**并行**发给多个模型评审，findings 自动合并——被 2 个以上模型同时报告的问题排在最前，并带 `found by:` 标注：\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- 共识 findings（2+ 模型）排最前，单模型 findings 紧随其后。\n- 同文件 + 行区间带容差匹配去重；严重度取各模型报告的最高档，不同表述的标题会保留为备注。\n- 某个成员失败（provider 报错、返回非法 JSON）只按成员如实报告，不影响整体——只要有一个模型返回有效评审，评审团就算成功。\n- 评审团成员不走 `PI_PLUGIN_FALLBACK_MODELS` 降级链——评审团本身就是冗余机制。\n- 这个能力只有 Pi 这种多供应商 agent 才做得到：单一供应商的 CLI 无法召集跨厂商评审团。\n\n## 🛟 模型故障自动降级\n\n配置一次降级链，之后任何失败的运行——provider 宕机、鉴权错误、自动重试耗尽——都会自动换下一个模型重跑：\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\n评审和 rescue 任务都生效。当结果来自降级模型时，输出末尾会附 `Model fallback:` 说明（JSON payload 里带 `modelAttempts`）。`/pi:setup` 会显示当前配置的降级链。\n\n## 🏁 模型竞速\n\n硬骨头任务可以让多个模型**并行**跑同一个任务，然后选出赢家：\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **写竞速**（`--write`，`/pi:rescue` 默认带）：每个 racer 在独立的 git worktree（从 `HEAD` 创建）里干活——racer 之间以及和你的工作树完全隔离。每个 racer 的成果捕获为 patch；审阅后用 `git apply ` 应用其中一个。\n- **只读竞速**（排查类任务）：racer 分析同一份代码，输出并排展示各自答案——多模型结论一致是强信号。\n- 某个 racer 失败或没产生改动会被如实标注；只要有一个 racer 完成，竞速就算成功。\n- 不能与 `--model` 或 `--resume` 同用（每个 racer 都是全新 session）。racer 从 `HEAD` 出发，任务涉及未提交改动时先 commit 或 stash。\n\n## 选模型\n\n模型解析三层优先级：\n\n| 优先级 | 来源 | 例子 |\n|---|---|---|\n| 1 | 斜杠命令上的 `--model ` | `/pi:review --model gpt-4o` |\n| 2 | Env var（仅 review / adversarial-review） | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | pi 自己配置的默认模型 | 你 `~/.pi/agent/models.json` 里的，或 pi TUI `/model` 上次选的 |\n\n层 1 > 层 2 > 层 3。**三层都可以不设** —— 全空就让 pi 自己决定。\n\n### 各 provider 的建议默认\n\n只是建议，不强制。按延迟 / 成本 / 质量预算自己挑：\n\n| Provider | 日常 review（`/pi:review`） | 对抗式 review（`/pi:adversarial-review`） |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` 或 `gpt-5-mini` | `o1` 或 `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` 或 `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| 本地（Ollama） | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` 或 `deepseek-r1` |\n\n想让这些选择粘住：\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### 性价比档 —— 比 Claude/GPT 便宜、依然能打\n\n驱动 Pi 的意义就在于你不被 Claude 或 GPT 锁死。下面这些模型的成本只是 Claude/GPT 顶配的零头，性能虽然没那么强，但用于日常 review 和 rescue 绰绰有余。在 `~/.pi/agent/models.json` 里配置它们（大多是 OpenAI 兼容端点），并用你的 provider 实际暴露的 id —— 用 `pi --list-models` 核对。\n\n| Provider | 建议模型 |\n|---|---|\n| DeepSeek | `deepseek-v4-flash`（日常）· `deepseek-v4-pro`（对抗式） |\n| xAI | `grok-4.5` |\n| 智谱（GLM） | `glm-5.1` · `glm-5.2`（更重） |\n| 月之暗面（Kimi） | `kimi-k3`（或 `kimi-k2.6`） |\n| MiniMax | `MiniMax-M3` |\n| 字节（Seed） | `doubao-seed-2.1-pro` |\n| 小米（MiMo） | `mimo-v2.5`（日常）· `mimo-v2.5-pro`（更重） |\n| Meta | `muse-spark-1.1` |\n\n……以及任何你配置的 OpenAI 兼容端点。可用性和确切 id 取决于你的 `~/.pi/agent/models.json`。\n\n> 不是每个模型都能为结构化 review 命令吐出干净 JSON。如果某个模型在 `/pi:review` 上解析报错，它用于 `/pi:rescue`（自由格式）仍然没问题 —— 或者把它加进 `PI_PLUGIN_FALLBACK_MODELS` 链，坏成员会自动重试下一个。（已观察到：部分 MiniMax 和本地模型会回显 prompt 而非返回 JSON。）\n\n## 配置 pi\n\n最简就是一个环境变量装 API key。要细配就写 `~/.pi/agent/models.json`：\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\n内置 provider（`anthropic`、`openai`、`google`、`deepseek`、`ollama`、`lmstudio`）只要 `apiKey` 和可选 `baseUrl`，pi 自带模型清单。OpenAI 兼容端点用 `api: \"openai-completions\"` 并写自定义 `models`。\n\n完整字段见 [pi providers 文档](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)。\n\n## Stop-time 评审守门\n\n用 `/pi:setup --enable-review-gate` 开启。每次结束 Claude 会话时，插件会对上一轮 Claude 的输出做一次 Pi 对抗式评审；若发现重要问题会阻止退出。开了守门但 pi 不可用时 hook 会 block（**不会**静默放行）。用 `/pi:setup --disable-review-gate` 关闭。\n\n## 🔗 相关项目\n\n| 项目 | 定位 | 适用场景 |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | 同样的命令面，底层跑 Codex | 想用 OpenAI Codex agent + ChatGPT 账号时 |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | 本插件驱动的编码 agent 本体 | 想直接用 Pi、不经 Claude Code 时 |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi 扩展，加 `subagent` 工具 + `/run` / `/chain` / `/parallel` | 驱动 `/pi:parallel-rescue`，也让 `/pi:rescue` 能转交给专门的子代理 |\n\n## ❤️ 支持\n\n如果这个插件对你有帮助，欢迎打赏支持作者：\n\n\n \n \n \n \n \n \n
\n \"微信支付\"\n
\n 微信支付\n 		\n \"支付宝\"\n
\n 支付宝\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"打赏\"\n
\n 打赏\n
\n\n## 👤 作者\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE)。改编自 [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) —— 见 [NOTICE](NOTICE)。\n"}}, "README.md": {"b6a5c2dd0c19d427b3c02c2e048db2347a499804": {"t": 1785674337.1953242, "n": 16376, "h": "d67b7230bf2c4583", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:review --out-file review.md\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n`--out-file ` (on `/pi:review`, `/pi:adversarial-review`, `/pi:rescue`, `/pi:result`) writes Pi's full output to a file and returns only a short summary — verdict, finding counts, one line per finding. The heavy reasoning already runs on the cheaper model; this also keeps the large result out of the calling agent's context, so a big review doesn't burn Claude Code tokens on the relay. Open the file for the full detail.\n\n`--incremental` (on `/pi:review`, `/pi:adversarial-review`) reviews only the commits since the last review on the current branch, using a per-branch cache of the last-reviewed commit — saving Pi input tokens and time by skipping the code that was already reviewed. Falls back to a full review when there is no valid cache.\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm-5.1` · `glm-5.2` (heavier) |\n| Moonshot (Kimi) | `kimi-k3` (or `kimi-k2.6`) |\n| MiniMax | `MiniMax-M3` |\n| ByteDance (Seed) | `doubao-seed-2.1-pro` |\n| Xiaomi (MiMo) | `mimo-v2.5` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "c85ec1db22cb4eb4ce21cf86d4563f1d30076589": {"t": 1785766134.2490118, "n": 16376, "h": "d67b7230bf2c4583", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:review --out-file review.md\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n`--out-file ` (on `/pi:review`, `/pi:adversarial-review`, `/pi:rescue`, `/pi:result`) writes Pi's full output to a file and returns only a short summary — verdict, finding counts, one line per finding. The heavy reasoning already runs on the cheaper model; this also keeps the large result out of the calling agent's context, so a big review doesn't burn Claude Code tokens on the relay. Open the file for the full detail.\n\n`--incremental` (on `/pi:review`, `/pi:adversarial-review`) reviews only the commits since the last review on the current branch, using a per-branch cache of the last-reviewed commit — saving Pi input tokens and time by skipping the code that was already reviewed. Falls back to a full review when there is no valid cache.\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm-5.1` · `glm-5.2` (heavier) |\n| Moonshot (Kimi) | `kimi-k3` (or `kimi-k2.6`) |\n| MiniMax | `MiniMax-M3` |\n| ByteDance (Seed) | `doubao-seed-2.1-pro` |\n| Xiaomi (MiMo) | `mimo-v2.5` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "f6dcdea1a4730cc17c81cd933b4e5f2c6e2e1ebf": {"t": 1785766135.4894242, "n": 16357, "h": "5d8804b3f1770c1f", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:review --out-file review.md\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n`--out-file ` (on `/pi:review`, `/pi:adversarial-review`, `/pi:rescue`, `/pi:result`) writes Pi's full output to a file and returns only a short summary — verdict, finding counts, one line per finding. The heavy reasoning already runs on the cheaper model; this also keeps the large result out of the calling agent's context, so a big review doesn't burn Claude Code tokens on the relay. Open the file for the full detail.\n\n`--incremental` (on `/pi:review`, `/pi:adversarial-review`) reviews only the commits since the last review on the current branch, using a per-branch cache of the last-reviewed commit — saving Pi input tokens and time by skipping the code that was already reviewed. Falls back to a full review when there is no valid cache.\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm-5.1` · `glm-5.2` (heavier) |\n| Moonshot (Kimi) | `kimi-k2.6` |\n| MiniMax | `MiniMax-M3` |\n| ByteDance (Seed) | `doubao-seed-2.1-pro` |\n| Xiaomi (MiMo) | `mimo-v2.5` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "f75dde35051b071dfde3fa30cd44de8159ed148d": {"t": 1785766136.9154866, "n": 16033, "h": "74de8e9d414a7d4a", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:review --out-file review.md\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n`--out-file ` (on `/pi:review`, `/pi:adversarial-review`, `/pi:rescue`, `/pi:result`) writes Pi's full output to a file and returns only a short summary — verdict, finding counts, one line per finding. The heavy reasoning already runs on the cheaper model; this also keeps the large result out of the calling agent's context, so a big review doesn't burn Claude Code tokens on the relay. Open the file for the full detail.\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm-5.1` · `glm-5.2` (heavier) |\n| Moonshot (Kimi) | `kimi-k2.6` |\n| MiniMax | `MiniMax-M3` |\n| ByteDance (Seed) | `doubao-seed-2.1-pro` |\n| Xiaomi (MiMo) | `mimo-v2.5` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "258cd685e7032b8e7ec06ed53129933789a1b981": {"t": 1785766138.1770766, "n": 16019, "h": "3a971f2cc40c0eb3", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:review --out-file review.md\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n`--out-file ` (on `/pi:review`, `/pi:adversarial-review`, `/pi:rescue`) writes Pi's full output to a file and returns only a short summary — verdict, finding counts, one line per finding. The heavy reasoning already runs on the cheaper model; this also keeps the large result out of the calling agent's context, so a big review doesn't burn Claude Code tokens on the relay. Open the file for the full detail.\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm-5.1` · `glm-5.2` (heavier) |\n| Moonshot (Kimi) | `kimi-k2.6` |\n| MiniMax | `MiniMax-M3` |\n| ByteDance (Seed) | `doubao-seed-2.1-pro` |\n| Xiaomi (MiMo) | `mimo-v2.5` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "8ed484fb0c3e031f841338850204f716fd77ccd1": {"t": 1785766139.5717459, "n": 15569, "h": "6bd500e1c8efea93", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm-5.1` · `glm-5.2` (heavier) |\n| Moonshot (Kimi) | `kimi-k2.6` |\n| MiniMax | `MiniMax-M3` |\n| ByteDance (Seed) | `doubao-seed-2.1-pro` |\n| Xiaomi (MiMo) | `mimo-v2.5` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "2df961524e4ff395c93cf726dee8f5b1e6571d41": {"t": 1785766141.054888, "n": 15567, "h": "ee0070d556dd4dbb", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm5.1` · `glm5.2` (heavier) |\n| Moonshot (Kimi) | `kimi 2.6` |\n| MiniMax | `minimax-m3` |\n| ByteDance (Seed) | `seed-2.1-pro-preview` |\n| Xiaomi (MiMo) | `mimo-v2.5` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "7fec5eb927d428f2659d3d92dc46dd8e8095afb7": {"t": 1785766142.7484827, "n": 15574, "h": "1ddd5e73f5954b02", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm5.1` · `glm5.2` (heavier) |\n| Moonshot (Kimi) | `kimi 2.6` |\n| MiniMax | `minimax-m3` |\n| ByteDance (Seed) | `seed-2.1-pro-preview` |\n| Xiaomi (MiMo) | `mimo-v2.5-dflash` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "114dc2068d993704a9de7e84949119c98fcb6d51": {"t": 1785766144.4497705, "n": 15576, "h": "3de9895bf8462e79", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm5.1` · `glm5.2` (heavier) |\n| Moonshot (Kimi) | `kimi 2.6` |\n| MiniMax | `minimax-m3` |\n| ByteDance (Seed) | `seed-2.1-pro-preview` |\n| Xiaomi (MiMo) | `mimo-v2.5-standard` (everyday) · `mimo-v2.5-pro` (heavier) |\n| Meta | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "9c41526998eefddf79f3a5dab6a576cc4a01c828": {"t": 1785766145.655902, "n": 15532, "h": "9f68b670e03cecbb", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n### Value tier — cheaper than Claude/GPT, still competitive\n\nThe whole point of driving Pi is that you are not locked to Claude or GPT. These models cost a fraction of the frontier Claude/GPT tiers and, while not quite as strong, are more than good enough for everyday review and rescue work. Configure them in `~/.pi/agent/models.json` (most are OpenAI-compatible endpoints) and use the exact id your provider exposes — verify with `pi --list-models`.\n\n| Provider | Suggested model(s) |\n|---|---|\n| DeepSeek | `deepseek-v4-flash` (everyday) · `deepseek-v4-pro` (adversarial) |\n| xAI | `grok-4.5` |\n| Zhipu (GLM) | `glm5.1` · `glm5.2` (heavier) |\n| Moonshot (Kimi) | `kimi 2.6` |\n| MiniMax | `minimax-m3` |\n| ByteDance (Seed) | `seed-2.1-pro-preview` |\n| Xiaomi (MiMo) | `mimo-v2.5 pro` |\n| Muse | `muse-spark-1.1` |\n\n…and any other OpenAI-compatible endpoint you configure. Availability and the exact id depend on your `~/.pi/agent/models.json`.\n\n> Not every model emits clean JSON for the structured review commands. If `/pi:review` fails with a parse error on a given model, it is still fine for `/pi:rescue` (free-form) — or add it to a `PI_PLUGIN_FALLBACK_MODELS` chain so a flaky member is retried with the next one. (Observed: some MiniMax and local models echo the prompt back instead of returning JSON.)\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "4161825f6a997e8c1f1c8b5a84418f36627fdd2a": {"t": 1785766147.0463903, "n": 14218, "h": "4a69d521d8bdf733", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Sharded parallel review** — `/pi:review --shards ` splits a large diff's changed files across N review jobs that run in parallel, then merges the findings\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "c4048b250d41256e8b957492c032f0b3574ea384": {"t": 1785766148.622456, "n": 14056, "h": "bd9425325f2dd2cc", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## 🏁 Model racing\n\nFor hard problems, run the same rescue task with several models **in parallel** and pick the winner:\n\n```text\n> /pi:rescue --race deepseek-v4-pro,claude-sonnet-4-6 fix the flaky retry logic in src/queue.mjs\n> /pi:rescue --race deepseek-v4-flash,gemini-2.5-pro why does the Windows CI build fail?\n```\n\n- **Write races** (`--write`, the `/pi:rescue` default): each racer works in an isolated git worktree created from `HEAD` — racers can never touch your working tree or each other. Each racer's result is captured as a patch; review them and apply exactly one with `git apply `.\n- **Read-only races** (investigations): racers analyze the same tree; the output presents each answer side by side — agreement across models is a strong signal.\n- A racer that fails or produces no changes is reported as such; the race succeeds while at least one racer finishes.\n- Not combinable with `--model` or `--resume` (each racer starts a fresh session). Racers start from `HEAD`, so commit or stash first if the task concerns uncommitted work.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "42dc0140b1fb1a8b11e6d1660bb5e24552fce33e": {"t": 1785766150.023535, "n": 12999, "h": "7d28b35cd0287197", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## 🧑‍⚖️ Multi-model review panel\n\nOne reviewer has blind spots; a panel doesn't share them. Passing `--models` to either review command runs the same diff through several models **in parallel** and merges their findings — issues reported by 2+ models rank first with a `found by:` tag:\n\n```text\n> /pi:review --models deepseek-v4-flash,claude-sonnet-4-6,gpt-5-mini\n> /pi:adversarial-review --models deepseek-v4-pro,o1 focus on concurrency\n```\n\n- Consensus findings (2+ models) are listed first; single-model findings follow.\n- Duplicate findings are matched per file with line-range slack; severity escalates to the highest reported and alternate titles are kept.\n- A member that fails (provider error, invalid JSON) is reported inline and does not sink the panel — it succeeds as long as one model returns a valid review.\n- Panel members do not use the `PI_PLUGIN_FALLBACK_MODELS` chain — the panel itself is the redundancy.\n- This only exists because Pi is provider-agnostic: a single-vendor CLI cannot convene a cross-vendor panel.\n\n## 🛟 Automatic model fallback\n\nSet a fallback chain once, and any failed run — provider outage, auth error, exhausted retries — is automatically retried with the next model:\n\n```bash\nexport PI_PLUGIN_FALLBACK_MODELS=deepseek-v4-flash,MiniMax-M3\n```\n\nApplies to reviews and rescue tasks alike. When a fallback produced the result, the output ends with a `Model fallback:` note (and the JSON payload carries `modelAttempts`). `/pi:setup` shows the configured chain.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "af736691555c8b86e0e1789fd238cb3b88685021": {"t": 1785766151.372643, "n": 11499, "h": "eff1eb57e2bb8188", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code and Codex 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "227f73fcd4da1b8b1ff4dd00355e183f812287b6": {"t": 1785766152.7912116, "n": 11611, "h": "349316883bb8e902", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n[![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/79JF5Atuk)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi. Also usable from OpenAI's Codex CLI — see [Use from Codex](#-use-from-codex).\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## 🧩 Use from Codex\n\nThe core of this plugin is a harness-agnostic CLI (`pi-companion.mjs`) — any coding agent that can run shell commands can drive it. For OpenAI's Codex CLI, install the bundled [custom prompts](codex-prompts/):\n\n```bash\n# 1. Clone the repo (anywhere; ~/pi-plugin-cc is the default the prompts assume)\ngit clone https://github.com/Agents365-ai/pi-plugin-cc ~/pi-plugin-cc\n\n# 2. Install the Codex custom prompts\nmkdir -p ~/.codex/prompts\ncp ~/pi-plugin-cc/codex-prompts/*.md ~/.codex/prompts/\n\n# 3. If you cloned somewhere else, point the prompts at it\necho 'export PI_PLUGIN_ROOT=\"$HOME/path/to/pi-plugin-cc\"' >> ~/.zshrc\n```\n\nThen inside Codex: `/pi-review`, `/pi-adversarial-review`, `/pi-rescue`, `/pi-parallel-rescue`, `/pi-status`, `/pi-result`, `/pi-cancel`, `/pi-setup` (note `-` instead of `:` — Codex prompt names cannot contain colons).\n\nNot available under Codex: the stop-time review gate and session-resume prompts (both rely on Claude Code hooks / subagents). Everything else — including pi-subagents parallel fan-out — works the same.\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "58386acde4c59763d583b9695007a594f288b3d8": {"t": 1785766154.080476, "n": 10463, "h": "9b58e5fa9eaf2178", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n[![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/79JF5Atuk)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi.\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Parallel fan-out** — `/pi:parallel-rescue` runs multiple independent tasks concurrently via [`pi-subagents`](https://github.com/nicobailon/pi-subagents)\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nIntegrates with [`pi-subagents`](https://github.com/nicobailon/pi-subagents) (`pi install npm:pi-subagents`): `/pi:setup` detects it and lists the agent profiles, `/pi:rescue` prompts advertise the `subagent` tool to Pi, and `/pi:parallel-rescue` fans multiple tasks out to parallel child agents (scout, researcher, planner, worker, reviewer, …).\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-companion-forwarder` subagent |\n| `/pi:parallel-rescue` | Run multiple independent tasks in parallel via pi-subagents (`subagent({ tasks })` fan-out) |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:parallel-rescue \"audit the auth module\" \"benchmark the db queries\" \"update the API docs\"\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Powers `/pi:parallel-rescue` and lets `/pi:rescue` delegate to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "da8ae28a3637f572fdcf865d20c21e8178fe8cff": {"t": 1785766155.8937817, "n": 9854, "h": "a9ea100f84b67980", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n[![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/79JF5Atuk)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). Adapted from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi.\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nPlays well with [`pi-subagents`](https://github.com/nicobailon/pi-subagents): if installed, it works inside `/pi:rescue` runs without configuration.\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-rescue` subagent |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Let `/pi:rescue` delegate further to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "f1c2de6ec7499f0c3f60214a18533a94028f2cdb": {"t": 1785766157.75643, "n": 9853, "h": "900b05056dcb655a", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n[![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/79JF5Atuk)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). 1:1 fork of [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi.\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nPlays well with [`pi-subagents`](https://github.com/nicobailon/pi-subagents): if installed, it works inside `/pi:rescue` runs without configuration.\n\n## 🔄 How it works\n\n\"pi-plugin-cc\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-rescue` subagent |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Let `/pi:rescue` delegate further to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "3690a358289b706f869b9c470fa3cab2fc615614": {"t": 1785766159.1251938, "n": 9827, "h": "64f3bcd256800989", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n[![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/79JF5Atuk)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). 1:1 fork of [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi.\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nPlays well with [`pi-subagents`](https://github.com/nicobailon/pi-subagents): if installed, it works inside `/pi:rescue` runs without configuration.\n\n## 🔄 How it works\n\n![pi-plugin-cc workflow](docs/pi-plugin-cc-workflow.png)\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-rescue` subagent |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Let `/pi:rescue` delegate further to specialized child agents |\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "d10b89dde66f40b247f516052c0c4bd69add5cdb": {"t": 1785766160.4806986, "n": 10548, "h": "fae12598042cf58f", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![Model agnostic](https://img.shields.io/badge/Model-agnostic-555)](#pick-your-model)\n[![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/79JF5Atuk)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [Pi providers](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi). 1:1 fork of [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc), runtime swapped from Codex to Pi.\n\n**The hard dependency is pi, not any particular LLM.** Pi can be configured for DeepSeek, OpenAI, Anthropic, Google, Ollama, LM Studio, or any OpenAI-compatible endpoint via `~/.pi/agent/models.json`. The plugin defers all model selection to pi unless you override per command.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (provider-specific), no `codex login` required\n\nPlays well with [`pi-subagents`](https://github.com/nicobailon/pi-subagents): if installed, it works inside `/pi:rescue` runs without configuration.\n\n## 🔄 How it works\n\n```\nClaude Code ─► /pi:review · /pi:rescue · /pi:status ...\n │\n ▼\n pi-companion.mjs (Node)\n │ one pi --mode rpc subprocess per task\n ▼\n Pi coding agent ─► whichever provider/model pi is configured for\n │\n ▼\n JSONL events + final assistant message\n │\n ▼\n job state files (status / result / cancel)\n```\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | What it does |\n|---|---|\n| `/pi:setup` | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | Standard code review of local git state |\n| `/pi:adversarial-review` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | Delegate investigation or implementation to a Pi run via the `pi:pi-rescue` subagent |\n| `/pi:status [job-id]` | List active / recent Pi jobs in this repository |\n| `/pi:result ` | Show the stored final output for a finished job |\n| `/pi:cancel ` | Terminate a running background job |\n\nEvery command accepts `--model ` to pin a specific model just for that run. With no `--model` and no env override (see [Pick your model](#pick-your-model)), pi falls back to whatever it has configured by default.\n\n## Quick Start\n\n```bash\n# 1. Install pi (required)\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Configure a provider — pick one\nexport OPENAI_API_KEY=sk-... # OpenAI\nexport ANTHROPIC_API_KEY=sk-ant-... # Anthropic\nexport GOOGLE_API_KEY=... # Google\nexport DEEPSEEK_API_KEY=sk-... # DeepSeek\n# or run a local model: see https://github.com/earendil-works/pi (Ollama / LM Studio)\n\n# 3. Verify pi sees a model\npi --list-models | head\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report. If `pi` is missing and `npm` is on PATH, it offers to install it for you.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:review --model claude-sonnet-4\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background --model gpt-4o refactor src/payments/\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that do not support thinking silently ignore it (the plugin logs a one-line note to stderr when this happens).\n\n## Pick your model\n\nThe plugin keeps three layers of model resolution:\n\n| Priority | Source | Example |\n|---|---|---|\n| 1 | `--model ` on the slash command | `/pi:review --model gpt-4o` |\n| 2 | Env var (review / adversarial-review only) | `export PI_PLUGIN_REVIEW_MODEL=deepseek-v4-flash`
`export PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=deepseek-v4-pro` |\n| 3 | Pi's own configured default | whatever your `~/.pi/agent/models.json` has, or `/model` last picked in pi TUI |\n\nLayer 1 wins over layer 2 wins over layer 3. **None of the layers are required** — leave them all unset and pi picks for you.\n\n### Suggested settings by provider\n\nThese are opinions, not requirements. Pick what fits your latency / cost / quality budget.\n\n| Provider | Everyday review (`/pi:review`) | Adversarial review (`/pi:adversarial-review`) |\n|---|---|---|\n| DeepSeek | `deepseek-v4-flash` | `deepseek-v4-pro` |\n| OpenAI | `gpt-4o-mini` or `gpt-5-mini` | `o1` or `gpt-5` |\n| Anthropic | `claude-haiku-4-5` | `claude-sonnet-4-6` or `claude-opus-4-7` |\n| Google | `gemini-2.5-flash` | `gemini-2.5-pro` |\n| Local (Ollama) | `qwen2.5-coder:7b` | `qwen2.5-coder:32b` or `deepseek-r1` |\n\nTo make these defaults sticky:\n\n```bash\nexport PI_PLUGIN_REVIEW_MODEL=claude-haiku-4-5\nexport PI_PLUGIN_ADVERSARIAL_REVIEW_MODEL=claude-sonnet-4-6\n```\n\n## Configure pi\n\nThe minimum is a single env var with your API key. For richer setups, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": { \"apiKey\": \"sk-...\" },\n \"openai\": { \"apiKey\": \"sk-...\" },\n \"anthropic\":{ \"apiKey\": \"sk-ant-...\" },\n \"google\": { \"apiKey\": \"...\" },\n \"openrouter\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek via OpenRouter\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nBuilt-in providers (`anthropic`, `openai`, `google`, `deepseek`, `ollama`, `lmstudio`) only need `apiKey` (and optional `baseUrl`); pi ships their model lists. For custom OpenAI-compatible endpoints, set `api: \"openai-completions\"` and declare the `models` you want exposed.\n\nFull reference: [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md).\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. If pi is unavailable while the gate is enabled, the hook blocks (it does **not** silently let the session end). Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | You want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | You want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Let `/pi:rescue` delegate further to specialized child agents |\n\n## 💬 Community\n\n- **Discord:** https://discord.gg/79JF5Atuk\n- **WeChat:** scan the QR code below\n\n

\n \"WeChat\n

\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}, "d5ce89827ca058c156ed5e395d1c743bd0a9751e": {"t": 1785766161.7805314, "n": 8279, "h": "c08cdb161a604194", "c": "# pi-plugin-cc — drive the Pi coding agent from Claude Code 🥧\n\n[![License: Apache 2.0](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](LICENSE)\n[![GitHub stars](https://img.shields.io/github/stars/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/stargazers)\n[![GitHub forks](https://img.shields.io/github/forks/Agents365-ai/pi-plugin-cc?style=flat&logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/network/members)\n[![Latest Release](https://img.shields.io/github/v/release/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/releases/latest)\n[![Last Commit](https://img.shields.io/github/last-commit/Agents365-ai/pi-plugin-cc?logo=github)](https://github.com/Agents365-ai/pi-plugin-cc/commits/main)\n\n[![Claude Code Plugin](https://img.shields.io/badge/Claude%20Code-plugin-8a2be2)](https://docs.claude.com/en/docs/claude-code/plugins)\n[![Pi Coding Agent](https://img.shields.io/badge/Pi-coding%20agent-0a7d4a)](https://github.com/earendil-works/pi)\n[![DeepSeek](https://img.shields.io/badge/DeepSeek-V4-1f6feb)](https://platform.deepseek.com/)\n[![Discord](https://img.shields.io/badge/Discord-Join-5865F2?logo=discord&logoColor=white)](https://discord.gg/79JF5Atuk)\n\n**English** · [中文](README_CN.md)\n\nExternal references: [Pi coding agent](https://github.com/earendil-works/pi) · [Pi RPC mode](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/rpc.md) · [DeepSeek API](https://platform.deepseek.com/)\n\nA Claude Code plugin that delegates reviews and coding tasks to the [Pi coding agent](https://github.com/earendil-works/pi), defaulting to DeepSeek V4 (Flash for everyday review, Pro for adversarial review). 1:1 fork of [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) with the runtime swapped from Codex to Pi.\n\n- **Code review** against the working tree or a branch base, with structured findings\n- **Adversarial review** that challenges the design — not just spell-checks the diff\n- **Task delegation** for diagnoses, refactors, and longer rescues, foreground or background\n- **Background job control** — `status`, `result`, `cancel`, and stop-time review gate\n- **No OAuth** — pi authenticates by API key (`DEEPSEEK_API_KEY`, etc.), no `codex login` needed\n\nPlays well with [`pi-subagents`](https://github.com/nicobailon/pi-subagents): if installed, it works inside `/pi:rescue` runs without configuration.\n\n## 🔄 How it works\n\n```\nClaude Code ─► /pi:review · /pi:rescue · /pi:status ...\n │\n ▼\n pi-companion.mjs (Node)\n │ one pi --mode rpc subprocess per task\n ▼\n Pi coding agent ─► DeepSeek (or any pi-configured provider)\n │\n ▼\n JSONL events + final assistant message\n │\n ▼\n job state files (status / result / cancel)\n```\n\nCodex's broker layer is gone — Pi is one-conversation-per-process, so the plugin spawns a fresh `pi --mode rpc` for each task. Background jobs are tracked in workspace-scoped state files. Review prompts inline the JSON schema since Pi has no `outputSchema` knob.\n\n## Slash commands\n\n| Command | Default model | What it does |\n|---|---|---|\n| `/pi:setup` | — | Verifies `pi` is installed + a provider is configured; toggles the stop-time review gate |\n| `/pi:review` | `deepseek-v4-flash` | Standard code review of local git state |\n| `/pi:adversarial-review` | `deepseek-v4-pro` | Steerable challenge review — questions the approach itself |\n| `/pi:rescue` | user's pi default | Delegate investigation or implementation to a Pi run via the `pi:pi-rescue` subagent |\n| `/pi:status [job-id]` | — | List active / recent Pi jobs in this repository |\n| `/pi:result ` | — | Show the stored final output for a finished job |\n| `/pi:cancel ` | — | Terminate a running background job |\n\n## Quick Start\n\n```bash\n# 1. Install pi\nnpm install -g --ignore-scripts @earendil-works/pi-coding-agent\n\n# 2. Provide a provider key (DeepSeek by default)\nexport DEEPSEEK_API_KEY=sk-your-key-here\n\n# 3. Verify\npi --list-models deepseek # should show deepseek-v4-flash and -pro\n```\n\nInstall the plugin in Claude Code:\n\n```text\n> /plugin marketplace add Agents365-ai/pi-plugin-cc\n> /plugin install pi@agents365-pi\n> /reload-plugins\n> /pi:setup\n```\n\n`/pi:setup` returns a readiness report and offers to install Pi for you if it's missing and `npm` is on PATH.\n\n## Usage\n\n```text\n> /pi:review\n> /pi:review --base main\n> /pi:adversarial-review focus on the new auth middleware\n> /pi:rescue investigate why the Windows CI build is failing\n> /pi:rescue --background refactor src/payments/ to remove the polling loop\n> /pi:status\n> /pi:status task-mpgyiwb9-e3k641 --wait\n> /pi:result task-mpgyiwb9-e3k641\n> /pi:cancel task-mpgyiwb9-e3k641\n```\n\n`--effort ` is passed through to Pi via `set_thinking_level`. Models that don't support thinking silently ignore it.\n\n## Configure DeepSeek\n\nThe minimum config is just an env var — Pi ships built-in DeepSeek models:\n\n```bash\nexport DEEPSEEK_API_KEY=sk-your-key-here\n```\n\nFor a full custom config, write `~/.pi/agent/models.json`:\n\n```json\n{\n \"providers\": {\n \"deepseek\": {\n \"apiKey\": \"sk-your-deepseek-key\"\n },\n \"openai\": {\n \"apiKey\": \"sk-your-openai-key\"\n },\n \"openrouter-deepseek\": {\n \"api\": \"openai-completions\",\n \"apiKey\": \"sk-or-v1-...\",\n \"baseUrl\": \"https://openrouter.ai/api/v1\",\n \"models\": [\n {\n \"id\": \"deepseek/deepseek-chat\",\n \"name\": \"DeepSeek (via OpenRouter)\",\n \"contextWindow\": 128000,\n \"maxTokens\": 8192,\n \"input\": [\"text\"]\n }\n ]\n }\n }\n}\n```\n\nSee [pi providers docs](https://github.com/earendil-works/pi/blob/main/packages/coding-agent/docs/providers.md) for every supported provider and field.\n\n## Stop-time review gate\n\nOpt in with `/pi:setup --enable-review-gate`. When a Claude session ends, the plugin runs a Pi adversarial review of the previous turn and can block the stop if it finds material issues. Disable with `/pi:setup --disable-review-gate`.\n\n## 🔗 Related projects\n\n| Project | Niche | When to use |\n|---|---|---|\n| [codex-plugin-cc](https://github.com/openai/codex-plugin-cc) | Same surface, runs Codex | When you want OpenAI's Codex agent + ChatGPT auth |\n| [pi (earendil-works)](https://github.com/earendil-works/pi) | The coding agent this plugin drives | If you want to use Pi directly without Claude Code |\n| [pi-subagents](https://github.com/nicobailon/pi-subagents) | Pi extension adding `subagent` tool + `/run` / `/chain` / `/parallel` | Lets `/pi:rescue` delegate to specialized child agents |\n\n## 💬 Community\n\n- **Discord:** https://discord.gg/79JF5Atuk\n- **WeChat:** scan the QR code below\n\n

\n \"WeChat\n

\n\n## ❤️ Support\n\nIf this plugin helps you, consider supporting the author:\n\n\n \n \n \n \n \n \n
\n \"WeChat\n
\n WeChat Pay\n 		\n \"Alipay\"\n
\n Alipay\n 		\n \"Buy\n
\n Buy Me a Coffee\n 		\n \"Give\n
\n Give a Reward\n
\n\n## 👤 Author\n\n**Agents365-ai**\n\n- GitHub: https://github.com/Agents365-ai\n- Bilibili: https://space.bilibili.com/441831884\n\n## 📄 License\n\n[Apache License 2.0](LICENSE). Forked from [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) (Apache-2.0, OpenAI) — see [NOTICE](NOTICE).\n"}}}}