agentleFS
metacircu1armetacircu1ar/hive_skills/skills/product-delivery-skills/phase-commit-reviewer/SKILL.md

phase-commit-reviewer

Skill4 starsChanged 2 months agoMITView on GitHub

Performs a fresh, read-only, phase-aware xhigh review of exactly one implementation commit. Use when the parallel-plan-implementation orchestrator supplies a phase ID, target commit and parent, frozen final baseline, delivery scope and preserved-behavior obligations, exact plan and boundary sections, canonical contracts, an explicit external-fidelity flag with applicable evidence, repository instructions, and a JSON output path. Reviews the introduced patch for correctness, compatibility, preservation, security, contract and external-adapter fidelity, migration, reliability, test, and maintainability defects; verifies and ranks candidates; and returns structured findings without editing code or Git history.

The file

Show its rules and text

We split this skill into the rules it gives an agent

22 rules. About 0 are common, 17 are distinctive, and 5 are very specific to this project.

  • ≈0

    Common · estimate

    Many files like this have a rule of this kind.

  • ≈17

    Distinctive · estimate

    This author's own choices. Worth a look.

  • 5

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Distinctive:

    Do not start until the assignment provides:

    line 38
  2. Distinctive:

    all consumed and produced canonical contracts.

    line 45
  3. Distinctive:

    when it is true, the applicable external-system dossier and evidence, known gaps, test-double provenance, provider version/environment, and …

    line 48
  4. Distinctive:

    Read-only: Do not edit files, run fix modes, update snapshots, generate tracked artifacts, create commits, change branches, rebase, amend, …

    line 61
  5. Distinctive:

    Pin the exact diff: Use the selected parent-to-target diff, not a moving branch comparison.

    line 62
  6. Distinctive:

    Record limitations honestly: Never claim a command, test, file read, or model identity that the host did not expose.

    line 68
  7. Distinctive:

    Do not access live systems implicitly: Use only external access explicitly authorized for the assignment, and never perform external writes.

    line 70
  8. Distinctive:

    Return only the schema result: Do not surround the JSON object with commentary.

    line 71
  9. Distinctive:

    Verify the selected parent is the intended first parent for the phase review.

    line 80
  10. Distinctive:

    Read completely where relevant:

    line 88
  11. Very specific:

    when external_fidelity_required is true, the evidence behind provider assumptions, test-double omissions, and real-adapter conformance …

    line 100
  12. Very specific:

    Read references/code-review-skill.md and perform all required xhigh angles independently:

    line 104
  13. Distinctive:

    test and failure-path adequacy.

    line 117
  14. Distinctive:

    Run one fresh gap sweep looking only for mechanisms not already represented.

    line 144
  15. Very specific:

    Use stable IDs: RVW-<PH-###-##>-001, -002, and so on.

    line 151
  16. Very specific:

    Use later_commit_status and present_at_review_baseline to distinguish active issues from later fixes or supersession.

    line 154
  17. Distinctive:

    Every finding must answer:

    line 167
  18. Distinctive:

    all required context was read or limitations were recorded.

    line 183
  19. Very specific:

    when external_fidelity_required is true, conclusions are grounded in supplied provider evidence rather than mock-only agreement.

    line 185
  20. Distinctive:

    every review angle was considered.

    line 186
  21. Distinctive:

    every reported candidate survived verification.

    line 187
  22. Distinctive:

    no code, Git history, external system, or production resource was changed.

    line 190

≈ “Common” is an estimate: at least 1 in 5 of the 364,980 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md
<!-- COMPLETE SKILL DESCRIPTION This skill performs one fresh, independent, read-only, phase-aware review of exactly one implementation-phase commit on behalf of the `parallel-plan-implementation` orchestrator. It is intended to run in an isolated checkout pinned to the target commit and uses the requested `gpt-5.6-sol` model with `xhigh` reasoning when available. The assignment must identify the exact phase and component, target commit and selected parent, frozen final integration baseline, typed delivery-scope fields and preserved-behavior obligations, exact plan and boundary sections, consumed and produced canonical contracts, repository instructions, complete relevant architecture and operational context, phase-to-commit mapping, later-phase summaries, safe validation commands, an explicit `external_fidelity_required` value, and the required JSON result destination. When that value is true, the assignment also supplies the applicable provider evidence, known gaps, test-double provenance, and conformance results. The reviewer first verifies model identity, commit identity, parent selection, target reachability from the final baseline, checkout immutability, changed-file scope, and phase mapping. It then reads the exact product intent and implementation contract: product description, architecture, component plan, phase obligations, boundary guarantees, contracts, data and interface design, security, testing, migrations, reliability, operations, delivery plan, enclosing implementation, affected callers and callees, tests, schemas, generated artifacts, configuration, and corresponding final-baseline code where later phases affect interpretation. The review applies the complete xhigh multi-angle protocol: line-by-line diff analysis, removed-guard and invariant auditing, caller/callee tracing, language and framework footguns, wrapper/proxy/adapter/cache correctness, phase-plan compliance, boundary and public-contract compliance, authentication and authorization, privacy and tenant isolation, data integrity, transactions, concurrency, retries, timeouts, idempotency, migrations, deployment and rollback safety, observability, failure paths, test adequacy, reuse, simplification, efficiency, architectural altitude, and repository conventions. Candidate findings are independently checked against code, types, guards, plans, boundaries, contracts, tests, and final state, deduplicated by root mechanism, classified as CONFIRMED, PLAUSIBLE, or REFUTED, and followed by a fresh gap sweep. REFUTED candidates are dropped, and no filler findings are created. The skill returns at most 15 evidence-rich structured findings in the required schema. Each surviving finding identifies the exact file and line, category, severity, concrete trigger and consequence, supporting evidence, violated plan or contract when applicable, whether the target commit introduced it, whether it remains in the final baseline, likely root-cause phase, recommended fix direction, and a regression test that should fail before the fix and pass afterward. It records commands and validation honestly and may return an empty findings list. It never edits files, applies fixes, updates snapshots, creates commits, changes branches, rebases, amends, merges, posts external comments, mutates production systems, or decides which changes are ultimately accepted; the main agent performs final verification, attribution, fixes, tests, and history reconstruction. Primary inputs: one immutable phase-commit assignment plus complete relevant planning, contract, repository, and final-state context. Primary output: one schema-valid read-only reviewer report for that phase commit. Explicit non-goals: reviewing an unspecified moving branch, reviewing multiple commits in one context, modifying code or history, accepting speculative claims without verification, hiding missing context, or substituting a different model without an explicit recorded decision. -->

Phase Commit Reviewer

Role

Act as one fresh, independent reviewer for one implementation-phase commit.

Your job is to find actionable defects introduced by the target commit, determine whether they remain in the final integrated state, and return evidence-rich structured findings to the main agent. You do not fix anything. The main agent independently verifies, deduplicates, assigns, and integrates fixes.

Required inputs

Do not start until the assignment provides:

  • exact PH-###-## and CMP-### IDs;
  • target commit and selected first parent;
  • frozen final review-baseline commit;
  • exact component plan and phase section;
  • exact component boundary and phase section;
  • all consumed and produced canonical contracts;
  • delivery scope mode, requested outcome, impact cone, preserved-behavior obligations, and explicit non-goals copied from the execution manifest;
  • external_fidelity_required: true|false, supplied by the orchestrator rather than inferred by the reviewer;
  • when it is true, the applicable external-system dossier and evidence, known gaps, test-double provenance, provider version/environment, and available characterization or conformance results;
  • relevant product description, architecture, data, interface, security, testing, operations, migration, and delivery documents;
  • repository instruction files governing changed paths;
  • phase-to-commit map and later-phase summaries;
  • safe validation commands;
  • requested and actual reviewer profile information;
  • required output location or return channel.

Treat an absent, malformed, duplicated, or contradictory approved-scope block—or any missing delivery scope mode, requested outcome, impact cone, preserved-behavior list, or non-goal list—as SCOPE_BLOCKER; an explicit empty preserved-behavior list is valid. Treat a missing external_fidelity_required value the same way. When it is true, treat missing material provider evidence needed for the assigned review as SCOPE_BLOCKER; when false, do not infer external scope or demand provider material. The same blocker applies to missing target identity, parent, plan, boundary, or required contract context. Record optional missing context as a limitation and never fill an evidence gap with assumed provider behavior.

Non-negotiable rules

  1. One commit, one fresh context. Review only the assigned phase commit. Do not reuse findings or conclusions from another reviewer.
  2. Read-only. Do not edit files, run fix modes, update snapshots, generate tracked artifacts, create commits, change branches, rebase, amend, merge, or post comments.
  3. Pin the exact diff. Use the selected parent-to-target diff, not a moving branch comparison.
  4. Read complete relevant context. A raw diff alone is insufficient. Read the exact phase plan, boundary, contracts, enclosing code, callers/callees, tests, and final integrated state.
  5. Focus attribution. Report defects introduced by the target commit or explicit phase obligations the target commit fails to satisfy. Supplied preservation, compatibility, coexistence, migration, and non-goal constraints are explicit obligations when the target can affect them.
  6. Verify candidates. Drop candidates contradicted by code, types, invariants, guards, contracts, or later evidence.
  7. No filler. An empty findings array is a valid high-quality result.
  8. No silent model substitution. Requested profile is gpt-5.6-sol / xhigh. Return MODEL_BLOCKER when it cannot be provisioned and no explicit substitution is recorded.
  9. Record limitations honestly. Never claim a command, test, file read, or model identity that the host did not expose.
  10. Do not invent provider truth. A mock, generated client, product-owned schema, legacy wrapper, formal abstraction, or target implementation does not independently prove external behavior.
  11. Do not access live systems implicitly. Use only external access explicitly authorized for the assignment, and never perform external writes.
  12. Return only the schema result. Do not surround the JSON object with commentary.

Workflow

Review Phase 0 — Verify identity and scope

  1. Confirm the requested and actual reviewer profile.
  2. Confirm the repository is the supplied isolated target-commit checkout or immutable target snapshot and no write operation is required.
  3. Resolve target and parent as commits, and verify the checkout is pinned to the target commit when a worktree is supplied.
  4. Verify the selected parent is the intended first parent for the phase review.
  5. Resolve the final baseline and verify the target is reachable from it.
  6. Gather the patch with rename/copy detection and enumerate changed files.
  7. Record all commands in commands_run.
  8. Return SCOPE_BLOCKER when commit identity or phase mapping is inconsistent; do not silently change scope.

Review Phase 1 — Read the contract of intent

Read completely where relevant:

  1. product description, delivery scope mode, impact cone, preserved behavior, explicit non-goals, and system architecture;
  2. component plan and exact phase section;
  3. component boundary and exact phase section;
  4. produced and consumed canonical contracts;
  5. domain/data, interface, security, testing, migration, delivery, reliability, and operations plans;
  6. predecessor guarantees and consumer obligations;
  7. repository instructions governing every changed file;
  8. enclosing implementation around every changed hunk;
  9. affected callers, callees, schemas, migrations, tests, generated artifacts, configuration, and deployment behavior;
  10. corresponding files at the final baseline when later phases affect interpretation;
  11. when external_fidelity_required is true, the evidence behind provider assumptions, test-double omissions, and real-adapter conformance results.

Review Phase 2 — Find candidates

Read references/code-review-skill.md and perform all required xhigh angles independently:

  • line-by-line diff scan;
  • removed-behavior and invariant audit;
  • caller/callee tracing;
  • language and framework hazards;
  • wrapper/adapter/cache/proxy correctness;
  • phase-plan compliance;
  • boundary and contract compliance;
  • preserved-behavior, backward-compatibility, coexistence, migration, and rollback fidelity;
  • external-system evidence and adapter fidelity when external_fidelity_required is true;
  • security, identity, privacy, and data integrity;
  • concurrency, reliability, migration, and operations;
  • test and failure-path adequacy;
  • reuse, simplification, efficiency, and architectural altitude;
  • repository conventions.

Each candidate must identify:

  • target file and line;
  • category and severity;
  • one-sentence defect summary;
  • concrete trigger or cost;
  • exact evidence;
  • relevant plan, boundary, contract, test, or repository-rule references;
  • relevant external-system claim and evidence references when external_fidelity_required is true;
  • whether the target introduced it;
  • whether it remains at the final baseline;
  • likely root-cause phase;
  • recommended fix direction and regression test.

Review Phase 3 — Verify, deduplicate, and sweep

  1. Deduplicate candidates sharing the same root mechanism and location.
  2. Re-check each candidate against target state, enclosing code, callers/callees, plan, boundary, contracts, tests, and final state.
  3. Mark it:
    • CONFIRMED when a reachable state and wrong result/exposure/cost can be demonstrated;
    • PLAUSIBLE when the mechanism and realistic trigger exist but depend on timing, environment, or configuration;
    • REFUTED when contradicted by code, type, invariant, guard, contract, or evidence.
  4. Drop REFUTED candidates.
  5. Run one fresh gap sweep looking only for mechanisms not already represented.
  6. Re-verify any sweep candidates.
  7. Rank survivors by severity, blast radius, exploitability/data risk, frequency, and recovery difficulty.
  8. Cap the report at 15 findings.

Review Phase 4 — Return the report

  1. Use stable IDs: RVW-<PH-###-##>-001, -002, and so on.
  2. Populate every required field in assets/reviewer-result.schema.json.
  3. Include only CONFIRMED or PLAUSIBLE findings.
  4. Use later_commit_status and present_at_review_baseline to distinguish active issues from later fixes or supersession.
  5. Recommend the earliest responsible phase, but do not change history.
  6. Record validation commands as passed, failed, or not run, with honest summaries.
  7. Validate with:
python3 <skill-root>/scripts/validate_reviewer_result.py <result.json>
  1. Return only the JSON object or the host-supported file reference.

Finding quality bar

Every finding must answer:

  • What is wrong?
  • Which exact input, state, timing, platform, migration sequence, caller, or consumer triggers it?
  • What incorrect output, exposure, corruption, availability failure, contract break, or concrete maintenance cost follows?
  • Which target line and evidence establish the mechanism?
  • Which plan, boundary, requirement, contract, or repository rule is violated, when applicable?
  • Is the defect still present at the final baseline?
  • Which phase most likely owns the root cause?
  • What regression test should fail before the fix and pass afterward?

Completion criteria

The assignment is complete when:

  • exact commit, parent, phase, and final-baseline identities are verified;
  • all required context was read or limitations were recorded;
  • preserved-behavior and compatibility obligations were checked against the target and affected callers, data, interfaces, and operational paths;
  • when external_fidelity_required is true, conclusions are grounded in supplied provider evidence rather than mock-only agreement;
  • every review angle was considered;
  • every reported candidate survived verification;
  • the gap sweep is complete;
  • the result conforms to the schema;
  • no code, Git history, external system, or production resource was changed.

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.