Skip to content

ce-babysit-pr: persist review invariant rounds before non-convergence fixes #1575

Description

@mura0319

Problem

ce-babysit-pr persists coarse review trajectory (backlog trend, new threads, heads since progress), but it does not persist the semantic invariant shared by feedback across heads. A resolver can therefore treat repeated manifestations of the same root as separate fixes until the coarse trajectory finally rises.

Observed on a downstream PR: one Golden-adoption boundary produced 40 review threads and 16 commits before the workflow switched from regex patches to a fail-closed boundary. The existing non-convergence prose described this exact treadmill, but the durable state could not count rounds by root invariant.

Expected behavior

  • The resolver classifies a valid feedback cluster with a stable invariant key when it returns a fix outcome.
  • ce-babysit-pr persists the key by review round across head changes.
  • A third round for the same invariant is surfaced as one approach-level needs-human decision before another fix/commit/push.
  • Semantic classification remains with the resolver; the snapshot script only validates and counts structured data.
  • Unrelated invariants and changed observations remain independently actionable.

Safety boundary

This must not infer semantics from paths, regexes, or comment text in the snapshot script, and must not turn ordinary one- or two-round review progress into a blocker.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions