an agent that picks open LiteLLM GitHub issues, tries to reproduce them, and writes a report with a verdict score and before/after screenshots.
when AUTO_FIX=true, easy/medium issues also get a fix attempt and a draft PR.
- picks open issues automatically, no manual triage
- reproduces bugs via browser + curl against a live LiteLLM proxy
- scores each issue 0–5 and classifies difficulty
- posts a report comment with screenshots when
POST_COMMENTS=true - opens a draft fix PR when
AUTO_FIX=true
nvm use 20
npm install
cp .env.example .env
# fill in LITELLM_BASE_URL, LITELLM_API_KEY, GITHUB_TOKEN, GITHUB_BOT_USERNAMEPrerequisites: gh CLI (authenticated), git, uv, ImageMagick.
# one-shot against a specific issue
npm run once -- --issue 9876
# continuous daemon
npm run devBoth default to false — first runs are local-only so you can review output before anything touches GitHub.
POST_COMMENTS— post the report as a GitHub issue commentAUTO_FIX— attempt a fix and open a draft PR
| Score | Meaning |
|---|---|
| 5 | Fully reproduced, root cause confirmed, fix validated |
| 4 | Reproduced, root cause with file:line evidence |
| 3 | Similar symptoms, not the exact reported flow |
| 2 | Partial signal — env starts but flow didn't trigger |
| 1 | Setup failed |
| 0 | Unreproducible (missing info, feature request, question) |
