gitgood-review plugin

Lets a terminal coding agent fix the AI review findings that GitGood exports, and ask GitGood to re-review when it is done. One folder serves Claude Code, Codex and omp:

GitGood writes the findings to <git-dir>/gitgood/review/latest.json and latest.md after every review, so the plugin needs no server, CLI or network access. The Fix with agent button in GitGood opens your terminal with the agent already pointed at latest.md; the plugin adds discovery (the session-start reminder) and the slash command on top.

Install

Claude Code

/plugin marketplace add erwin-wee/gitgood
/plugin install gitgood-review@gitgood

Codex — open the plugin browser with /plugins after adding this repository as a marketplace, install gitgood-review, then start a new session. Alternatively copy skills/gitgood-review into ~/.agents/skills/ (or the repository's .agents/skills/) to get the skill without the plugin. Note: the marketplace registration key for the Codex CLI is not shown in the Codex docs at the time of writing; the skill-folder copy always works.

omp

omp plugin install gitgood-review@gitgood

after adding this repository as a marketplace (omp reads .claude-plugin/marketplace.json). The skill is also picked up from .agents/skills/gitgood-review in a repository.

Without the plugin

The exported latest.md starts with the same instructions as the skill, so any agent started with a prompt like "Read .git/gitgood/review/latest.md and fix every finding it lists. Do not commit." works too. That is what GitGood's Fix with agent button runs (configurable under Options → AI → Agent for fixes).

Re-review

After a pre-commit review the agent finishes by opening gitgood://review/rerun?repo=<path>&token=<single-use token> (the exact command is in latest.json under rerun.command). GitGood, if running with that repository in its list, re-runs the review (a link without a valid, unspent token first asks the user to confirm the repository and files) over the working tree and rewrites latest.json with a new runId whose previousRunId points at the run the agent worked from.

After a pull request or branch review, rerun is null and the agent stops instead: those reviews read committed history, so they cannot see edits still in the working tree. Commit the agent's changes in GitGood and press Re-review.

MCP server

GitGood ships a small MCP server (stdio, no dependencies, source in src/mcp, built to out/mcp/index.mjs) so an agent can use the handoff as tools instead of reading files and running xdg-open:

repoPath may be any folder inside the repository or a linked worktree; the export is looked up in that checkout's own git directory, as the app does.

Plugin. .mcp.json registers the server for Claude Code and omp when the plugin is installed. It starts the installed app's own runtime in Node mode (ELECTRON_RUN_AS_NODE=1 gitgood -e <import app.asar/out/mcp/index.mjs>), so it needs the gitgood command on PATH (the Linux installer creates it). On macOS use /Applications/GitGood.app/Contents/MacOS/GitGood, on Windows %LOCALAPPDATA%\Programs\GitGood\GitGood.exe, as command in your own copy of the entry (below).

Any MCP client, installed app. Same command and environment as .mcp.json. Codex (~/.codex/config.toml):

[mcp_servers.gitgood]
command = "gitgood"
args = ["-e", "import(require('node:url').pathToFileURL(require('node:path').join(process.resourcesPath,'app.asar','out','mcp','index.mjs')).href)", "--", "--no-sandbox"]
env = { ELECTRON_RUN_AS_NODE = "1" }

omp (~/.omp/agent/mcp.json, or .omp/mcp.json per project) takes the mcpServers entry from .mcp.json unchanged.

From a checkout. npm run build (or just npm run build:mcp), then point the client at node /path/to/gitgood/out/mcp/index.mjs, e.g. claude mcp add gitgood -- node /path/to/gitgood/out/mcp/index.mjs. gitgood_open_repository then needs a GitGood executable to start: put gitgood on PATH or set GITGOOD_APP to the executable (a running GitGood must have been started from that same build for the folder to land in it).