Agent plugins
The repository includes one plugin package, plugins/openreview, for Claude Code and for Codex. It provides:
- The OpenReview MCP server (the remote endpoint at
/api/mcp). - The review-local-changes skill: reviews the branch's committed changes with
openreview review --agentbefore you push, fixes real findings and reviews again, at most three times. It needs the CLI, signed in withopenreview login. - The review-status skill: shows a pull request's review status and open findings.
- The resolve-review-findings skill: fixes open findings, commits locally, requests a re-review, and repeats until the pull request is clean or three rounds have run. It asks before every push unless you told it to push, and it never force-pushes or rewrites commits.
Sign-in uses OAuth (see MCP). Your client opens the browser the first time it connects.
Claude Code
-
Set
OPENREVIEW_URLto your server, for exampleexport OPENREVIEW_URL=https://review.example.com. The plugin's MCP configuration reads it. -
Install the plugin from the repository that holds this marketplace:
/plugin marketplace add <owner>/<repo> /plugin install openreview@openreviewOn Claude Code 2.1.275 or later,
/plugin install openreview --marketplace <owner>/<repo>does both in one step. From a checkout,claude --plugin-dir ./plugins/openreviewloads it for one session. -
Run
/mcpto sign in to OpenReview. Then ask Claude to check the review, or run/openreview:resolve-review-findings.
To use an API key instead of OAuth, add the server yourself with an Authorization: Bearer or_… header (see MCP).
Codex
Codex reads the same package through its documented plugin format: the root plugin.json (Agent Plugins schema, OpenAI settings under extensions.com.openai), skills/ and mcp.json.
- Edit
plugins/openreview/mcp.jsonso thaturlpoints at your server's/api/mcp. The file ships withhttps://openreview.example.com/api/mcp. - Add the marketplace with
codex plugin marketplace add <owner>/<repo>, or use the repository marketplace file.agents/plugins/marketplace.jsonfrom a checkout. - Install OpenReview from the Plugins Directory, then start a new chat and ask Codex to fix the open OpenReview findings.
The resolve loop from the command line
openreview resolve runs the same loop without an editor, using a local coding agent to make the fixes:
openreview resolve --pr 42 --fix-agent claude
- Fixing. Each round lists the pull request's open findings over MCP and hands them to the agent.
- Claude Code runs as
claude -p … --permission-mode acceptEdits. - Codex runs as
codex exec --sandbox workspace-write …. - Both run in the repository root.
- Claude Code runs as
- Committing. The loop commits the agent's changes locally. It starts only from a clean working tree, so each commit holds nothing but the fixes.
- Pushing. Without
--pushit asks before each push, and in a non-interactive shell it stops after the local commit. With--pushit runs a plaingit push. It never force-pushes. - Re-review. It requests a review of the new commit with
trigger_reviewand waits up to 20 minutes for that commit's review to finish. - Stopping. It stops when no findings remain, after
--maxrounds (default 3, at most 10), when the agent changes nothing, or when the re-review fails. The exit code is non-zero unless the pull request ends clean.
It needs a write-scoped API key from openreview login, and the checkout's origin must be a repository that key can see.