What problem does it solve?
Pull request CI on the SGLang repository requires persistent attention: workflows like lint.yml, pr-test.yml, and pr-test-extra.yml can fail for flaky infrastructure reasons, missing opt-in labels, or genuine code regressions, and someone must keep watching, diagnosing, retrying, and fixing until the latest head commit passes.
Core Features & Use Cases
- Persistent CI monitoring: Creates a durable goal that tracks only the selected workflows on the PR's latest head SHA, polling runs and ignoring unselected workflows entirely.
- Failure triage and bounded retries: Classifies each failure as PR-related, unrelated/infrastructure, or uncertain, and reruns failed jobs only (never full workflows) up to 10 times per workflow and head SHA.
- Guarded auto-fixing: Implements, validates, commits, and pushes fixes only when they are clean, non-tricky, and at most 100 changed lines; otherwise it stops for user review.
- SGLang gate awareness: Understands label gates such as run-ci and run-ci-extra, draft PR blocks, and AMD workflow variants, and can add authorized opt-in labels.
- Use Case: After pushing changes to PR 12345, ask the agent to babysit lint.yml and pr-test-extra.yml; it adds missing labels, retries a flaky GPU runner failure twice, and reports green runs with links once both workflows pass on the final head SHA.
Quick Start
Ask the agent to babysit the current pull request until lint.yml and pr-test.yml pass, optionally naming extra workflows such as pr-test-extra.yml or an AMD workflow.