maintainer-radar

Heuristics

Maintainer Radar uses deterministic rules. The score is not truth. It is a routing hint for maintainer attention.

Every analyzed PR includes a score_breakdown list. Each item names the heuristic, shows its risk_delta, and marks it as a signal or flag. Positive values raise risk. Negative values lower risk. The final risk score is clamped to the 0 to 100 range, then reviewability is calculated as 100 - risk.

Each analyzed PR also includes next_step, a deterministic sentence that turns the selected action and strongest flags into a concrete maintainer move.

Example:

[
  { "label": "CI passed", "risk_delta": -8, "kind": "signal" },
  { "label": "code changed without tests", "risk_delta": 10, "kind": "flag" }
]

Positive Signals

• CI passed
• PR is approved
• PR is mergeable
• review has been requested
• test plan language is present
• tests changed
• docs-only shape

Risk Flags

• draft PR
• large or very large diff
• no visible checks
• CI failing or pending
• changes requested
• merge conflicts
• branch behind base
• merge blocked by repository rules
• merge checks unstable
• stale update window
• maintainer blocker language
• maintainer blocking label
• no test plan found for code changes
• code changed without tests
• generated or lockfile changes

File evidence

docs-only shape requires a complete file list with every file classified as documentation. A README alongside a lockfile, test, or unclassified file does not receive the docs-only score adjustment or shorter review estimate.

When the supplied file list contains fewer entries than the reported changed file count, Radar shows incomplete file list. It still uses visible evidence, such as a test file that was returned, but does not infer that tests are absent. The incomplete-list warning itself does not add risk points.

Maintainer Blocker Language

Blocker detection intentionally looks for plain maintainer feedback patterns:

• not working
• broken
• failing or failed
• changes requested
• please fix
• regression
• cannot merge
• missing tests

This is not sentiment analysis. The goal is to catch comments that usually mean “do not spend full review time until the author follows up.”

Maintainer Blocking Labels

Label detection looks for labels that usually mean the PR is not ready for review:

• blocked
• do not merge
• changes requested
• needs tests
• waiting on author
• waiting for dependency
• blocked upstream

These labels route the PR to author follow-up even when CI is green.

Merge Readiness

Hydrated GitHub scans also read merge readiness fields such as mergeable, mergeStateStatus, and requested reviewers.

• Merge conflicts route the PR to author follow-up.
• A branch behind the base branch routes the PR to author follow-up.
• Repository rule blocking is shown as a risk flag, but it does not automatically blame the author.
• Requested reviewers are shown as positive queue context.

Fixture Corpus

The test suite includes tests/fixtures/blocker-prs.json to keep blocker detection concrete. Add new fixture cases when a real maintainer pattern should be detected.

This site is open source. Improve this page.