docs: EN README for public — install, permissions, models discovery

This commit is contained in:
admin committed 2026-09-01 00:26:52 +00:00
1 parent d0ff0f6aa0
commit b8bb384d4b
2 files changed
+63 -45

No files matched your search

+10
View File
@@ -0,0 +1,10 @@
# Contributing
Keep PRs small and focused. Use English only.
1. Fork, branch from `main`, make minimal changes.
2. Run relevant tests (`tests/test-run.sh` if touching `scripts/`).
3. Bump `version` in `.claude-plugin/plugin.json` if publishing.
4. Ensure `codex CLI >=0.151` and `jq` requirements are respected.
5. Open PR with clear description and verification steps.
6. Be respectful and constructive in review.
+53 -45
View File
@@ -1,58 +1,66 @@
# codex-crew
Codex CLI as a first-class Claude Code subagent: task delegation, code
review, second opinion. Inherits the user's `~/.codex/config.toml` (no
sandboxes, no restrictions), sessions are resumable, hangs are detected by a
deterministic watchdog.
## Contents
- `agents/codex.md` — transport subagent (sonnet). Holds the codex session:
follow-ups via SendMessage continue the same dialogue. The task must
include `cwd:`; optional `codex options:` (model/sandbox/effort) and
`review scope:`.
- `skills/codex-agent` — delegate execution.
- `skills/codex-review` — review a branch/commit/uncommitted/files.
- `skills/codex-consult` — second opinion on a spec/plan, iterate to
consensus.
- `scripts/codex-run.sh` — wrapper around `codex exec --json` with a
watchdog (`CODEX_STALL_TIMEOUT`, default 600s of output stagnation → kill;
the session continues via resume). Tests: `tests/test-run.sh`.
## How it works
The orchestrator spawns the `codex` subagent with a task. The subagent
writes a prompt file (transport preamble + task), launches `codex-run.sh`
in the background, and polls the status file.
Review mode: the codex CLI can't combine a scope flag
(`--base`/`--commit`/`--uncommitted`) with a custom prompt, so with a
non-empty prompt file the script translates the scope into prompt text
automatically; with no prompt file the native `codex exec review <scope>`
path is used.
Prompt convention: wrapper instructions outside, the codex task between
`=== TASK FOR CODEX ===` / `=== END TASK ===` markers — the wrapper forwards
only the marked block, so it never mistakes the task for its own job, and
wrapper instructions never leak into the codex prompt.
Codex CLI as a first-class Claude Code subagent: delegate, review, consult. Respects your `~/.codex/config.toml`, sessions resumable, stall watchdog.
## Install
```
claude plugin marketplace add <path-or-github-url>
claude plugin install codex-crew@codex-crew
/plugin marketplace add smartass/codex-crew
/plugin install codex-crew@codex-crew
```
Requires codex CLI ≥0.142.
Requires `codex CLI >=0.151` and `jq`.
Update after push (Claude Code doesn't re-fetch the cache; `plugin update`
compares the version from plugin.json — no version bump, no content update):
## How it works
Orchestrator → `codex` subagent (`agents/codex.md`) → `scripts/codex-run.sh` → `codex exec --json` (`--skip-git-repo-check -o last.md`). Non-empty `prompt.md` is fed via stdin `-` so flags and prompt combine. Stall watchdog (`CODEX_STALL_TIMEOUT`, default 600s of event silence → kill; resume via `thread_id`). `cwd:` in task is mandatory; `codex options:` forwarded verbatim. Review scope (`--base`/`--commit`/`--uncommitted`) is translated to prompt text when a custom prompt is present (CLI forbids scope + prompt together), otherwise native `codex exec review <scope>`.
## Skills
| Skill | Trigger |
|---|---|
| `codex-agent` | `delegate to codex`, `have codex do` — execution, bugfix, research |
| `codex-review` | `codex review` — branch (`--base`), commit, uncommitted, or files |
| `codex-consult` | `second opinion`, `have codex review the spec/plan` — iterate to consensus (max 5 rounds) |
Details: `skills/references/common.md`.
## Permissions & Sandbox
Default inherits `~/.codex/config.toml` (no flags injected). Override per-task via `codex options:`:
- `-s read-only | workspace-write | danger-full-access`
- `--add-dir <dir>` — extra writable directory
- `--dangerously-bypass-approvals-and-sandbox` — dangerous, only in an externally sandboxed environment
Example: `codex options: -s workspace-write --add-dir /tmp`
Flags are saved per-job and re-applied on `resume`.
## Models & Reasoning
Default from `~/.codex/config.toml` (`model`, `model_reasoning_effort`). Discovery:
```
# 1. bump version in .claude-plugin/plugin.json, commit, push
claude plugin marketplace update codex-crew
claude plugin update codex-crew@codex-crew
./scripts/codex-models.sh
# or
codex debug models | jq '[.models[] | select(.visibility=="list") | {slug, levels: [.supported_reasoning_levels[].effort]}]'
```
then `/reload-plugins` or restart the session (agent/skill definitions are
snapshotted on reload; hot-patching the cache doesn't work).
Pass via `codex options: -m <model-from-discovery> -c model_reasoning_effort="<effort-from-discovery>"` (values from `scripts/codex-models.sh` output).
## Updating
Bump `version` in `.claude-plugin/plugin.json`, commit, push, then:
```
/plugin marketplace update codex-crew
/plugin update codex-crew@codex-crew
/reload-plugins
```
No version bump → no update (Claude Code caches by version).
## Contributing
See [CONTRIBUTING.md](CONTRIBUTING.md). PRs welcome — keep changes minimal and EN-only.