| name | pr-review |
| description | Opinionated pull-request / diff review for the Mycelium repo. Enforces CONTRIBUTING.md — the honesty rule (per-op guarantee tags), append-only ADR/RFC discipline, grounding citations, and the "which FR/NFR/VR/SC + how verified" expectation — and does a hallucination/consistency pass. Auto-scales depth to the change (editorial → standard → deep) and can run an exhaustive all-severities pass. |
| when_to_use | Use when asked to review a PR, a branch, a commit, or "the diff"; to sanity-check changes before opening/merging a PR; or to catch hallucinated/unsupported claims. |
| argument-hint | [PR#|<base>...<head>|<commit>] [--tier T0|T1|T2] [--all] |
| allowed-tools | Bash(git diff:*), Bash(git show:*), Bash(git log:*), Bash(git status:*), Read, Grep, Glob |
pr-review
Opinionated, repo-grounded PR review. Shared triage/severity/report contract:
review-rubric — read it first, classify the change into a
tier (or honor --tier / --all), and emit the standard report.
1. Get the change
- PR number → read it via the GitHub MCP
pull_request_read (metadata + diff + linked issue).
<base>...<head> / branch / commit → git diff <base>...<head> or git show <ref>.- Note the change's stated intent (PR body / commit message) so you can check intent‑vs‑diff.
2. Triage (see rubric §1)
Score size, change-kind, path-fragility, and honesty-rule surface → choose T0/T1/T2.
Report the tier and the signals.
3. Review checklist (depth per tier)
Correctness & intent
- Does the diff do what its message claims — nothing more (scope creep), nothing less?
- For code: invariants hold; error paths explicit (no silent failure — esp. swaps/conversions
must never be silent; out-of-range returns
Option/error, never a quiet value).
- Logic/edge cases; concurrency; resource cleanup. New behavior has tests.
Mycelium honesty rule (non-negotiable — CONTRIBUTING)
- Any guarantee/bound carries a per-model/per-op tag
Exact ⊐ Proven ⊐ Empirical ⊐ Declared.
A Proven tag is valid only if it cites a theorem whose side-conditions are checked;
otherwise it must be Empirical or Declared. Flag any upgrade lacking a checked basis.
- A shipped approximate operation must include its bound + guarantee tag + a property test
that exercises the bound (SC-2). Missing any → High.
- No black boxes: representation selection / conversions must be reified +
EXPLAIN-able.
Decision discipline & grounding
- ADR/RFC changes are append-only: status moves forward (Draft→Accepted→Superseded); to
change an Accepted decision you supersede it, never silently rewrite. Flag in-place edits to
Accepted normative content.
- Every normative claim cites its grounding (
G1–G11, tensions A–E, R1–R8, T0.x/T1.x/T2.x)
or is explicitly marked an open question. Ungrounded assertions → High.
Hallucination / consistency pass
- Verify every factual/citation claim in the diff against the actual source file/section/line.
Cross-references must resolve (run/trust
just links). Numbers, IDs (FR/NFR/VR/SC/KC, RFC §),
file paths, and API names must be real. A claim contradicting its source → Critical.
Hygiene
- Conventional, imperative commit subject referencing the issue/task. PR states which
FR/NFR/VR/SC it advances (or which ADR/RFC it implements) and how it was verified;
editorial-only PRs say so.
- Engineering house style (SOLID/DRY/KISS/YAGNI/Demeter/SoC, composition over inheritance,
small auditable kernel). Changelog updated when a doc status moved or a notable change landed
(defer to the
changelog skill).
4. Report
Emit the rubric §4 format. Posture is advisory — recommend, don't gate. Be specific:
file:line → what → why → fix. End with summary, analysis, and prioritized recommendations
(always in --all mode).
