← ClaudeAtlas

patchleaguelisted

Run, inspect, and compare multiple coding-agent CLIs on the same repository task with PatchLeague. Use when Codex should benchmark agents such as Codex, Claude Code, Gemini CLI, or OpenCode in isolated Git worktrees; verify their patches with tests; inspect saved runs or HTML reports; or preview and apply a selected winning patch.
269394628/PatchLeague · ★ 1 · Code & Development · score 72
Install: claude install-skill 269394628/PatchLeague
# PatchLeague Prefer the installed `patchleague` command. From the PatchLeague source repo, fall back to `node bin/patchleague.js`. ## Start safely Verify the environment and discover configured agents: ```bash patchleague --json doctor patchleague --json agents list git status --short ``` PatchLeague reuses each provider CLI's existing authentication. It does not store API keys. Require a clean repository unless the user explicitly accepts that agents will start from `HEAD` without uncommitted changes. ## Run a competition Put the exact task and acceptance criteria in a Markdown file. Preview first when selecting agents or checks is uncertain: ```bash patchleague --json run issue.md --agents codex,claude --check "npm test" --dry-run patchleague --json run issue.md --agents codex,claude --check "npm test" ``` Use `--check` repeatedly for deterministic validation. If omitted, PatchLeague detects common project checks. Treat a Git worktree as change isolation, not as a security sandbox; only run trusted repositories and configured agent commands. ## Inspect results ```bash patchleague --json runs list --limit 10 patchleague --json runs show <run-id> patchleague runs report <run-id> ``` Read the saved score components, check results, risk signals, patch size, and diff before choosing a winner. Do not select solely by total score when acceptance criteria require human judgment. ## Apply a patch Preview applicability first: ```bash patchleague --json winners apply