Epic: Build a standalone review-fix-loop skill

[shaug]

Outcome

Ship a standalone review-fix-loop skill that repeatedly obtains a complete review, applies selected fixes in an isolated attempt, and converges to a trustworthy terminal result without publishing intermediate fixes.

The authoritative design is design/review-fix-loop.md.

Scope

• Define the invocation, checkpoint, grant, cycle-budget, and terminal-result contracts.
• Implement local common-directory locking, isolated attempts, recovery, and clean promotion.
• Default reviews to fresh read-only subagents, with an explicit in-agent override.
• Support the local_commit and update_pr modes described by the design.
• Publish only after convergence, using expected-old Git publication for update_pr.
• Add deterministic and result-blind evaluation coverage.
• Package and document the skill through the repository's normal discovery surfaces.

Non-goals

• A shared ownership or distributed coordinator service.
• CI orchestration, PR lifecycle management, or acceptance ownership.
• Migrating implement-ticket, babysit-pr, carve-changesets, or other callers in this epic.
• A review-only or zero-cycle execution mode.

Delivery graph

The native GitHub subissue and dependency relationships are authoritative. The intended critical path is:

1. Contracts
2. Local execution and reviewer orchestration, in parallel
3. local_commit
4. update_pr
5. Evaluation
6. Packaging and documentation

Caller migrations are separate fast-follow issues, not subissues of this epic.

Epic acceptance criteria

• Every in-scope child is closed with acceptance evidence tied to its delivered commit.
• The packaged skill can execute both local_commit and update_pr workflows.
• Intermediate fixes remain unpushed; update_pr publishes only after convergence.
• Non-converged outcomes clearly report retained unpushed commits and the action needed from the operator.
• Fresh read-only reviewer subagents are the default wherever supported; in-agent review requires an explicit override.
• Local locking, isolated attempts, recovery, and expected-old publication behave as specified.
• Deterministic tests and result-blind evaluations cover success, failure, budget exhaustion, interruption, recovery, and publication races.
• just format, just lint, and just test pass on the exact accepted head.
• An exact-head repository review has no remaining worthwhile findings.
• Existing caller skills remain behaviorally unchanged by this epic.
• Separate dependency-linked fast-follow tickets exist for caller migrations.

Closeout evidence

Epic closeout must summarize the accepted commit, validation commands and results, evaluation evidence, exact-head review result, packaging/discovery verification, and the live state of the fast-follow migration tickets.