Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/andyzengmath/copilot-plugin-cc

Use GitHub Copilot CLI from Claude Code for code review and task delegation (port of codex-plugin-cc over ACP v1)
https://github.com/andyzengmath/copilot-plugin-cc

Last synced: about 2 months ago
JSON representation

Use GitHub Copilot CLI from Claude Code for code review and task delegation (port of codex-plugin-cc over ACP v1)

Awesome Lists containing this project

README 

          
# Copilot plugin for Claude Code



Use **GitHub Copilot CLI** from inside Claude Code for code reviews or to

delegate tasks to Copilot.



This plugin is for Claude Code users who already have Copilot CLI installed

and want to invoke Copilot from the same workflow, without leaving Claude.



It is a port of [`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc)

that swaps the underlying runtime: `codex app-server` is replaced by

`copilot --acp`, which speaks the standard

[Agent Client Protocol v1](https://agentclientprotocol.com/).



## What you get



- `/copilot:review` — standard code review of your current work

- `/copilot:adversarial-review` — steerable challenge review

- `/copilot:rescue`, `/copilot:status`, `/copilot:result`, `/copilot:cancel` — delegate work and manage background jobs

- `/copilot:setup` — installer and auth check; toggle the optional stop-time review gate



## Requirements



- **GitHub Copilot subscription.** Usage counts against your Copilot allowance.

- **Copilot CLI 1.0.11 or later** (for native `--effort` passthrough; anything that speaks ACP v1).

- **Node.js 18.20.2+, 20.12.2+, or 22.0.0+** — `engines.node` enforces this so the CVE-2024-27980 (`.cmd` argv-injection) patch is guaranteed present. Earlier 18.x lines are explicitly rejected.



## Install



Add the marketplace in Claude Code:



```bash

/plugin marketplace add andyzengmath/copilot-plugin-cc

```



Install the plugin:



```bash

/plugin install copilot@copilot-plugin-cc

```



Reload plugins:



```bash

/reload-plugins

```



Then run:



```bash

/copilot:setup

```



`/copilot:setup` will tell you whether Copilot CLI is ready. If Copilot is

missing and npm is available, it can offer to install Copilot for you.



If you prefer to install Copilot yourself:



```bash

npm install -g @github/copilot

```



If Copilot is installed but not signed in:



```bash

!copilot login

```



## Usage



### `/copilot:review`



Runs a standard code review on your current work.



Use it when you want:



- a review of your uncommitted changes

- a review of your branch compared to a base branch like `main`



Flags: `--base `, `--wait`, `--background`, `--model `,

`--effort `, optional focus text.



```bash

/copilot:review

/copilot:review --base main

/copilot:review --background

/copilot:review --effort high

```



### `/copilot:adversarial-review`



A **steerable** review that challenges the implementation approach and

design choices rather than just scanning for defects. Accepts the same

flags as `/copilot:review`.



```bash

/copilot:adversarial-review

/copilot:adversarial-review --base main challenge whether the caching design is right

/copilot:adversarial-review --background look for race conditions

/copilot:adversarial-review --effort high focus on the retry semantics

```



### `/copilot:rescue`



Hands a task to Copilot through the `copilot:copilot-rescue` subagent.



Use it when you want Copilot to investigate a bug, try a fix, continue a

previous Copilot task, or take a different pass at the problem.



Flags: `--background`, `--wait`, `--resume`, `--fresh`, `--model `, `--effort `.



```bash

/copilot:rescue investigate why the tests started failing

/copilot:rescue fix the failing test with the smallest safe patch

/copilot:rescue --resume apply the top fix from the last run

/copilot:rescue --model claude-opus-4.6 --effort high investigate the regression

/copilot:rescue --background investigate the flaky integration test

```



**Model aliases**: `fast` → `claude-opus-4.6-fast`, `opus` →

`claude-opus-4.7`, `sonnet` → `claude-sonnet-4.6`, `haiku` →

`claude-haiku-4.5`, `gpt` → `gpt-5.4`, `codex` → `gpt-5.3-codex`,

`auto` → Copilot's auto-model-selection (GA 2026-04-17).

Any concrete Copilot model name works too — e.g.

`--model claude-opus-4.6` or `--model gpt-5.5` or

`--model gpt-5.4-mini` — even if it doesn't have a short alias.



> Note: `claude-opus-4.7` currently sits at a 7.5x Copilot premium-

> request multiplier (through 2026-04-30). The `opus` alias tracks

> it anyway; type `--model claude-opus-4.6` explicitly if you need

> the pre-4.7 billing rate.



**`--effort`** passes straight through to Copilot CLI's native

`--effort=` flag (Copilot CLI 1.0.11+). The

plugin's `none`/`minimal` aliases collapse to `low` at spawn time

for codex-plugin-cc command parity. When `--effort` is set without

`--model`, the user's default model from

[`~/.copilot/settings.json`](#default-model--effort) is preserved

— effort and model are independent knobs.



`--model` and `--effort` can be passed together; both flow through

verbatim and Copilot's runtime applies them independently. There is

no longer an "either/or" override: in v0.0.16 the plugin dropped the

internal effort-to-model mapping that used to silently force a Claude

model whenever `--effort` was set.



### Default model + effort



The plugin reads its default model and reasoning effort from

`~/.copilot/settings.json` (the standard Copilot CLI settings file —

location is `$COPILOT_HOME/settings.json` if you've set

`COPILOT_HOME`). Selection precedence is:



1. `/model` slash command inside an interactive `copilot` session

2. `--model` / `--effort` CLI flags on the plugin command

3. `COPILOT_MODEL` / `COPILOT_EFFORT_LEVEL` environment variables

4. `model` / `effortLevel` keys in `~/.copilot/settings.json`

5. Copilot CLI's built-in default (`claude-sonnet-4.5` today)



The fastest way to change your default for every plugin command is:



```bash

/copilot:setup --default-model gpt-5.5 --default-effort high

```



That writes the two keys into `~/.copilot/settings.json` atomically,

preserving leading `// ...` comments and every other key the file

already had. After that, `/copilot:rescue`, `/copilot:review`, and

the others will inherit GPT-5.5 + high reasoning unless a per-call

flag overrides them.



Every plugin call also emits an `[copilot] Using model: ...` line

to stderr before doing work, so you can confirm the active model +

effort + source on each invocation.



### `/copilot:status`, `/copilot:result`, `/copilot:cancel`



Same semantics as codex-plugin-cc. Track, fetch, and cancel background

Copilot jobs scoped to this Claude session.



### `/copilot:setup`



Checks install, ACP health, and auth state. The output also shows the

**active model** line — your inherited default model, effort, and where

the value came from (settings.json, env, or Copilot CLI default).



Toggle the optional stop-time review gate:



```bash

/copilot:setup --enable-review-gate

/copilot:setup --disable-review-gate

```



Persist a default model and reasoning effort into

`~/.copilot/settings.json`:



```bash

/copilot:setup --default-model gpt-5.5

/copilot:setup --default-effort high

/copilot:setup --default-model auto --default-effort medium

```



Aliases (`fast`, `opus`, `sonnet`, `haiku`, `gpt`, `codex`, `auto`)

resolve to canonical Copilot model identifiers before being written.

The plugin's `none`/`minimal` effort tiers collapse to `low` (Copilot's

own `effortLevel` only accepts `low`/`medium`/`high`/`xhigh`). Other

keys in `settings.json` are preserved.



Probe Copilot for model availability against a fixed list of common

models (Claude + GPT) before you hit an availability error mid-run:



```bash

/copilot:setup --probe-models

```



## How it works



- Spawns one shared `copilot --acp` broker process per Claude session.

- Routes all plugin commands to the broker over a Unix socket (or Windows

  named pipe).

- Tracks jobs in `${CLAUDE_PLUGIN_DATA}/state//` with per-job

  JSON + log files, session-scoped via a `SessionStart` hook.

- Cancels a live run with ACP `session/cancel`; falls back to killing the

  process tree if the session ID is unknown.

- Cross-Claude-session resume falls back to `copilot -p --continue` when no

  in-broker session is available (streaming progress degrades to coarse

  phase transitions).



See [`docs/plans/2026-04-17-copilot-plugin-cc-design.md`](docs/plans/2026-04-17-copilot-plugin-cc-design.md)

for the full design, including the Codex-RPC ↔ ACP-v1 mapping table and

the per-command porting decisions.



## Status (v0.0.18)



- Core runtime, broker, companion, and hooks all ported and under test.

- Standard and adversarial review commands share one prompt-engineered

  path (Copilot has no native `review/start` RPC). The `copilot-agents/`

  files are the canonical review methodology; the runtime prompt

  templates include them via `{{AGENT:}}` so there's no drift

  between the plugin prompt and `copilot --agent copilot-code-review`.

- Per-call `--model` and `--effort` flags pass straight through to

  Copilot CLI's native flags (1.0.11+) via a one-shot

  `copilot -p ""` subprocess that bypasses the shared ACP

  broker. Applied uniformly across `/copilot:rescue`, `/copilot:review`,

  and `/copilot:adversarial-review`. The legacy effort→model mapping

  was removed in v0.0.16 so `--effort` no longer overrides your

  configured default model.

- The active model + effort + source is echoed to stderr before every

  plugin call (v0.0.15) and shown in `/copilot:setup` output. Defaults

  can be persisted to `~/.copilot/settings.json` via

  `/copilot:setup --default-model X --default-effort Y` (v0.0.15).

- Multi-session broker coordination via a workspace-scoped

  `broker.lock` file (v0.0.8). Two Claude sessions starting on the

  same workspace simultaneously share one `copilot --acp` broker

  instead of racing to spawn duplicates. Fast path (live `broker.json`

  + reachable endpoint) skips the lock entirely so single-session

  reuse stays zero-overhead.

- Review-output enforcement walks every constraint in

  `plugins/copilot/schemas/review-output.schema.json` end-to-end

  (v0.0.8 + v0.0.9 tightening). Violations render as a bulleted

  `Schema violations:` section, not a short-circuited "first error"

  message.

- All three Copilot spawn sites use `lib/safe-spawn.mjs` (v0.0.18) — a

  cross-spawn-style helper that pre-resolves Windows `.cmd`/`.bat`

  launchers via PATHEXT and applies argv escaping with

  `windowsVerbatimArguments: true`, so `shell: false` everywhere. This

  closed the CVE-2024-27980 ("BatBadBut") class without the

  hand-rolled deny-list that v0.0.3-v0.0.17 carried.

- Defense-in-depth on the prompt-injection path: `--no-ask-user`

  (v0.0.17) removes the `ask_user` tool at the CLI level so a

  prompt-injected ask can't stall the run; `--secret-env-vars=COPILOT_GITHUB_TOKEN,GH_TOKEN,GITHUB_TOKEN`

  (v0.0.17) redacts auth tokens from any debug-style shell output the

  agent surfaces; `--deny-tool=shell(:*)` for `curl`/`wget`/`nc`/`ncat`/`ssh`

  (v0.0.18) cuts off exfiltration commands even with the broker's

  `--allow-all-tools` (denial > allow per `copilot help permissions`).

- Test coverage is end-to-end against a spawnable fake-ACP fixture

  (`tests/fake-copilot.mjs`). 173 tests across runtime suites, unit

  tests (safe-spawn, prompt loader, firstAllowOption, broker

  endpoint, schema validator, settings.json reader/writer, etc.) and

  the protocol-agnostic set. `tests/safe-spawn.test.mjs` (v0.0.18) is

  the first CI test that exercises the production Windows

  `.cmd`-launcher spawn path. CI runs on every PR on Ubuntu + Windows.



See

[`docs/plans/2026-04-20-v08-handoff.md`](docs/plans/2026-04-20-v08-handoff.md)

for the running backlog and per-release details.



## Security



See [SECURITY.md](SECURITY.md) for the threat model, the Windows ACL

caveat on `broker.json`, the v0.0.18 cross-spawn helper plus

`--deny-tool` exfiltration denials, and how to report vulnerabilities

via a private GitHub Security Advisory.



## License



Apache-2.0. See [LICENSE](LICENSE) and [NOTICE](NOTICE).