matteobortolazzo Actionable Comment Filter
This is the single source of truth for "is this PR comment actionable?". address-review (Phase 1D) applies this filter in-session, evaluating every comment against the Include/Exclude lists below directly — never restate the lists inline in a calling skill; reference this skill instead.
babysit is a thin adapter over the client-neutral cenci babysit Go supervisor (watch/internal/babysit/), which forbids in-session polling. The supervisor itself applies only the mechanical subset of the Exclude list below — bot logins, resolved threads, outdated comments — nothing semantic. It cannot judge whether a comment is "purely informational" or a genuine change request; that semantic judgment is applied by the model session the supervisor dispatches when new PR activity is detected. This skill remains the single source of truth: if the Include/Exclude lists below ever change, edit this file first — the supervisor and every calling skill read from it, never the reverse.
Include:
- Unresolved comments/threads
- Comments requesting changes (not approvals or neutral comments with no actionable content)
- Inline code review comments with suggestions
Exclude:
- Bot-generated comments (author is a known bot:
github-actions[bot],dependabot[bot], etc.) - Already-resolved threads
- Comments that are purely informational with no action requested
- Outdated comments on code that no longer exists (GitHub
outdatedflag)
If no actionable comments remain after filtering, the calling skill reports that and proceeds accordingly (address-review stops; babysit treats it as a quiet tick).