agent-roast
Offline CLI, agent hook and GitHub Action that flags changes weakening tests, coverage gates or CI instead of fixing the code.
At a Glance
agent-roast CLI, agent hook, pre-commit check and GitHub Action under the MIT License; runs offline with no tokens or telemetry.
Engagement
Available On
Alternatives
Listed Nov 2026
About agent-roast
agent-roast is an open-source, MIT-licensed tool by Nari-Ab that reviews code changes for shortcuts that turn a red build green by weakening the checks rather than fixing the code. It compares the before and after of each change and runs as an agent hook, a pre-commit check and a GitHub Action. The latest published release is v0.28.0.
What It Is
agent-roast is a deterministic, offline review aid for Git projects. It looks for loosened or removed assertions, skipped, focused or deleted tests, lowered coverage thresholds, relaxed lint or strictness settings, and CI steps allowed to fail. It also reports empty catch handlers, type escapes and not-implemented stubs as warnings. Each finding cites the file and line and suggests a next step. The project states that it needs no model, tokens or network, and that findings need human review and do not prove AI authorship or intent.
Guarding coding agents
Running npx agent-roast hook install registers hooks for agents such as Claude Code, Codex, Cursor, Gemini CLI, GitHub Copilot and Factory Droid. The README also lists Kiro and Antigravity CLI. When an agent is about to finish, agent-roast reviews everything it changed. If a check got weaker, the agent is sent back once with the file, line and before/after, and told to fix the code. A second attempt only produces a warning, and anything it cannot check lets the agent finish. An advisory mode only warns. A Claude Code plugin and a read-only MCP server are also available. The README notes the hooks have been tested with simulated input, not yet in long live sessions.
Pull request and commit checks
The GitHub Action compares full changed source at the merge-base and head, using policy read from the trusted base commit, so a pull request cannot switch off the rules reviewing it. It can post one updated PR comment, annotations, a job summary and SARIF output. Only new findings block, and existing ones are counted separately. Defaults block skipped tests, focused tests, trivial assertions, weakened assertions and weakened guards. A replay command shows what it would have said about recent commits, and baselines and expiring exceptions help adopt it in existing projects. --staged checks before a commit.
Scope and limits
The full before/after comparison covers JavaScript/TypeScript, Python and Go, while skip and placeholder-assertion checks also cover Rust, Java, Kotlin, C#, Ruby, PHP and Swift. CI and config files are checked regardless of language. It reads Git objects and never executes project code. Matching is heuristic, and complex syntax or aliases can cause missed findings or false positives. The project's own benchmarks report that a share of blocked commits on unseen repositories were false alarms.
Community Discussions
Be the first to start a conversation about agent-roast
Share your experience with agent-roast, ask questions, or help others learn from your insights.
Pricing
Open Source (MIT)
agent-roast CLI, agent hook, pre-commit check and GitHub Action under the MIT License; runs offline with no tokens or telemetry.
- MIT-licensed core CLI
- Agent hooks for Claude Code, Codex, Cursor and others
- Pre-commit check
- GitHub Action PR quality gate
- Offline, no tokens, no telemetry
Capabilities
Key Features
- Detects weakened assertions, skipped, focused and removed tests
- Detects lowered coverage thresholds and CI steps allowed to fail
- Compares full before/after source of changed files
- Agent stop hook that sends the agent back once to fix the code
- Pre-commit and --staged checks
- GitHub Action with PR comment, annotations, job summary and SARIF
- Policy read from HEAD or base branch
- Baselines and expiring exceptions
- Replay of recent commits
- Read-only MCP server
- Offline, no telemetry, no API keys
