内容
# agent-bridges
A fixed-role delegation pipeline that pairs **Claude Code** with two
delegation bridges, each with fixed jobs:
| Role | Bridge |
|---|---|
| Document Analyzer, Coder/Executor, Release-Changelog Writer | **Antigravity** (`agy-mcp` MCP server — async jobs, not blocking calls) |
| QA Engineer, Security Engineer | **Codex** (`codex-companion.mjs task` called directly via `Bash`, from the separately-installed [`openai/codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) plugin — not its `codex:codex-rescue` subagent; see `docs/ARCHITECTURE.md`) |
| Planner and Code Reviewer, final say on every bridge output | **Claude Code** |
Claude Code is the only agent with tool access, repo/session state, and
write permission — it plans, reviews first, and owns verification (a
delegated finding is never the sole basis for an edit until checked; see
"Parallelism, failures, and verification" in `templates/CLAUDE.md`).
## Why
Routing everything through one model is wasteful when another provider
fits a role better. Antigravity ingests specs, writes the code, then
drafts release notes; Codex QA always tests the diff; Codex Security
always reviews it for exploits; Claude Code always plans, reviews before
either Codex role sees a diff, and has final say — including over the
docs Antigravity drafts. Delegating also keeps Claude Code's own context
free of work — reading a large file, diffing a big repo, running QA/
security — where only the *answer* is needed, not the investigation.
For the full design rationale (topology, session continuity, memory
sharing across harnesses, learning-curator internals), see
[docs/ARCHITECTURE.md](docs/ARCHITECTURE.md).
## Install
No clone needed — Claude Code pulls this repo itself. Run these as two
separate commands, not as one pasted block — pasting both lines at once
can merge them into a single command and break the owner/repo argument:
```
/plugin marketplace add sanoval/agent-bridges
```
```
/plugin install agent-bridges@agent-bridges
```
This registers the `antigravity` MCP server, the `delegation-pipeline` and
`learning-curator` skills, and the Gate `PreToolUse` hook — and keeps them
current (`/plugin install agent-bridges@agent-bridges --update` to force a
check).
**Upgrading from a pre-0.2.0 install (this plugin used to run its own
Codex MCP server)?** Updating alone won't clear the old `codex-qa`/
`codex-security` entries cached in `~/.claude.json`, your stale
`~/.agents/skills/` symlinks, or your own already-copied `CLAUDE.md` —
see `docs/SETUP.md` step 0c.
**Upgrading from a pre-0.3.0 install (the Antigravity bridge used to be
`agy-bridge`, a blocking call with a timeout ceiling)?** You need the
`agy-mcp` binary installed separately (it's not bundled — see
`docs/SETUP.md` step 1), and updating the plugin alone may not be enough
to pick up the new `antigravity` server command: Claude Code caches each
project's *resolved* MCP server per name in `~/.claude.json` and doesn't
always re-resolve it just because `.mcp.json` changed. If `claude mcp
list` still shows `antigravity` running `npx ... agy-bridge` after
updating, see `docs/MIGRATION.md` for the fix and the full behavior
change (`agy_run` returns a `job_id` immediately instead of blocking —
delegation calls no longer time out, they wake you when done).
**Three-bridge mode (Codex QA/Security) is a separate install**, not part
of this plugin: it's OpenAI's own
[`openai/codex-plugin-cc`](https://github.com/openai/codex-plugin-cc)
plugin — see `docs/SETUP.md` step 2. Not on a Codex subscription, or
haven't installed that plugin? You're in two-bridge mode by default,
harmless, see "Which mode do I need?" below.
**Scope this to the projects that actually run the pipeline.** The Gate
hook fires on every `Edit`/`Write`/`NotebookEdit` in any project where the
plugin is enabled — enable it per-project (project `.claude/settings.json`
`enabledPlugins`) rather than at user scope, or you'll get Gate prompts in
unrelated repos. See [Configure team marketplaces](https://code.claude.com/docs/en/discover-plugins#configure-team-marketplaces).
Installing the plugin covers Claude Code. Antigravity/Codex CLIs, model
pins, and copying `CLAUDE.md`/`AGENTS.md` into your project are separate,
still-manual steps — see **[docs/SETUP.md](docs/SETUP.md)** for the full
walkthrough and **"Verifying it works"** in that doc once you're done.
## Usage
Once setup is verified, see **[docs/USAGE.md](docs/USAGE.md)** for what a
day-to-day session looks like: how to start a unit of work, what the Gate
hook will ask you to approve, what happens automatically (checkpointing,
learning), and a narrated end-to-end example.
## Which mode do I need?
- **Have a Codex subscription too?** Install `openai/codex-plugin-cc`
(`docs/SETUP.md` step 2) and use `templates/CLAUDE.md` as-is (three-bridge
mode). QA and Security each get their own model pin, passed per call.
- **Only Claude Code + Antigravity?** Copy `templates/CLAUDE.md` and
append `templates/CLAUDE-two-bridge-overlay.md` (two-bridge mode). QA
and Security fold into two separately-framed, separately-pinned
Antigravity `agy_run` calls instead of dedicated bridges — the overlay
explains what's still weaker than three-bridge mode (shared account/CLI
infrastructure with the Coder role) even with independent model pins,
and why your own Review step has to work harder.
## Acknowledgements
This is an original configuration/template repository — not a fork, no
source code from its MCP dependencies included.
The Antigravity integration uses
[agy-mcp](https://github.com/tphakala/agy-mcp) by tphakala as an external
MCP dependency (registered locally as `antigravity`), distributed under
the MIT License. Versions before 0.3.0 used
[agy-bridge](https://github.com/sshahzaiib/agy-bridge) by Shahzaib Akram
instead (same MIT license) — see `docs/MIGRATION.md` for why this project
moved off it and how to roll back if needed.
The Codex QA/Security integration (three-bridge mode) uses OpenAI's own
[`codex-plugin-cc`](https://github.com/openai/codex-plugin-cc) plugin as an
external dependency — installed separately, not bundled with this plugin.
## License
Copyright (c) 2026 Sanovalaw. Licensed under the [MIT License](LICENSE).
连接信息
猜你喜欢
暂无推荐服务器