agentsclimarketplace

Audit pr

Skill UMEP-dev/SUEWS/.claude/skills/audit-pr

Review SUEWS PRs. Drafts comments for approval before posting.From its SKILL.md

Install
npx -y skills add UMEP-dev/SUEWS --skill audit-pr

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

SKILL.md

10.8 KB, as published. Nobody here has run it

Audit PR

Comprehensive PR review with GitHub integration.

Quick Start

gh pr view {pr} --json number,title,files,commits
gh pr diff {pr} --name-only

Size Gate (run first)

Before the detailed review, assess the review surface. If the PR is too large to review well in one pass, or bundles unrelated concerns, lead with a split recommendation instead of spending the review on an unreviewable diff:

  • Check diff scope: gh pr diff {pr} --name-only plus additions/deletions per file. Watch for a large meaningful diff, many touched subsystems, or a mix of refactor + behaviour change + formatting.
  • Apply .claude/rules/work-sizing.md. If the PR fails the right-sized-PR test, recommend splitting it (a stacked series of small PRs, or carving out mechanical/generated changes first) before review. PR size relates to function/file/module granularity and the language conventions.
  • State the proposed split concretely: which files/concerns go in which PR, and the suggested order. This is a recommendation; do not rewrite the branch.
  • A best-effort review may still follow if the author prefers, but the size recommendation comes first.
  • To carry out the split, hand off to split-pr; audit-pr only recommends.

Workflow

  • Context: Gather files, classify changes
  • Size gate (first): see "Size Gate" above -- recommend a split before reviewing an oversized or bundled PR
  • Code Style: lint-code
  • Scientific: Physics validation (if applicable)
  • Testing: Coverage, FIRST principles
  • Docs: CHANGELOG, PR description
  • Governance: Check feature status tags (see below)
  • Suggested reviewers: Derive reviewer suggestions from changed files, dev-ref/SCIENTIFIC_REVIEWERS.md, and dev-ref/REVIEW_PROCESS.md; include them in the draft summary without treating them as automatic blockers.
  • Schema version audit: If the PR touches src/supy/data_model/, apply the triggers in .claude/rules/python/schema-versioning.md. Field rename, removal, type change, or required-field addition should be accompanied by a bump to CURRENT_SCHEMA_VERSION, a new SCHEMA_VERSIONS entry, a matching handler in src/supy/util/converter/yaml_upgrade.py, and a sample_config.yml resync. Flag any of these that are missing. CI gate .github/workflows/schema-version-audit.yml runs the automated check; if the PR carries the 0-ci:schema-audit-ok bypass label, verify from the diff yourself that the reason is genuinely cosmetic before approving.
  • Physics-change evidence audit: If the diff touches physics source (suews_phys_*.f95, Rust physics backend), moves a reference fixture (test/fixtures/data_test/sample_output.csv.gz, .../stebbs_test/sample_output_stebbs.csv, test/fixtures/benchmark1/*.pkl), or changes a physics-affecting data-model default, apply .claude/rules/physics-change-evidence.md. Confirm the 0-physics:change label is present (apply or flag if missing), the PR body has a ## Scientific evidence section (quantities + mechanism, before/after comparison, sign/magnitude reasoning -- a bare "outputs changed" is blocking), the owner sign-off is present for any owned subsystem (STEBBS -> @yiqing1021), any moved fixture is refreshed in this PR (or a linked PR), and the full -m physics tier (including slow) has run green.
  • Build: meson.build inclusion for new source files
  • CI gate audit: see "CI Gate Audit" below -- every non-green check gets a named remedy in the draft, not just a status
  • Draft: Comments for approval
  • Approval: Wait for human confirmation
  • Post: Only after approval

CI Gate Audit

A review that audits the diff but ignores the checks hands back a PR that still cannot move. Every non-green check gets a diagnosis and a named remedy in the draft.

  1. Read the checks, the ruleset's required contexts, and the failing job logs:

    gh pr checks {pr} --repo UMEP-dev/SUEWS --json name,state,bucket,workflow,link
    gh run view --repo UMEP-dev/SUEWS --job {job-id} --log-failed
    

    Query which contexts are actually required rather than assuming, and read the <!-- ci-build-plan --> bot comment for which lanes this PR's paths selected.

  2. Classify each non-green check by who can clear it -- author-fixable (a specific edit, in this PR), maintainer-gated (a documented bypass label, with the reason the diff qualifies), re-trigger mechanics (correct already, run is stale), or infrastructure/flake (unrelated to the diff). The classification is the actionable part; state it. Remedies compound: gh#1642 needed maintainer-gated -> re-trigger, and neither half alone cleared the gate.

  3. Report it as a finding with a severity: red required or convention-blocking check -> [blocking]; a bypass label applied for a reason the diff does not support -> [major]; red informational check -> [minor] or a note; pending required checks -> a note in the summary.

  4. Propose, never perform. Applying a bypass label, re-running a workflow, closing/reopening, or pushing a fix onto a PR you do not own are author or maintainer actions. Name them; do not do them.

Two gates (schema-version-audit, knowledge-pack-audit) read their bypass label from the static event payload, so a label added after the run started leaves the gate red and gh run rerun replays the stale payload. Push a commit or close/reopen instead.

Gate-by-gate remedies, the required-vs-blocking distinction, the green/near-green/red vocabulary (Part C of triage-pr's rubric, plus its one named amendment), and the payload race: references/ci-gates.md.

Governance Check

For PRs that add new features or breaking changes:

  • Verify CHANGELOG entries have appropriate status tags
  • New [feature] entries should have [experimental] tag by default
  • Flag if [stable] tag is used without governance approval
  • Check PR description mentions if feature is experimental

Questions to ask:

  • Is this feature ready for public announcement?
  • Does it have associated publication/evaluation?
  • Has the governance panel reviewed it?

Details: references/workflow-steps.md

Suggested Reviewers

In every audit summary, include suggested reviewers based on the PR diff:

  1. Start from changed files (gh pr diff {pr} --name-only).
  2. For physics files, use the manual scientific reviewer routing documented in dev-ref/SCIENTIFIC_REVIEWERS.md.
  3. For modules without named domain reviewers, suggest the default scientific owners documented in dev-ref/SCIENTIFIC_REVIEWERS.md.
  4. For coding style, linting, naming, issue-triage, or review-process changes, suggest the code-governance maintainers documented in dev-ref/REVIEW_PROCESS.md.

These are reviewer suggestions, not permission to request reviews or block a PR by themselves. If a physics-changing PR lacks the needed scientific review or sign-off, report that separately under the scientific evidence audit.

Change Classification

TypePatternReview
Physicssuews_phys_*.f95Scientific + code
Utilitysuews_util_*.f95Code only
Controlsuews_ctrl_*.f95Code only
Pythonsrc/supy/*.pyCode only

Code Style Quick Reference

Fortran: File naming, module naming, IMPLICIT NONE, precision, units Python: Variable prefixes, logging, pathlib, type hints, docstrings

Details: references/style-checks.md

Draft Output Format

=== DRAFT LINE COMMENTS ===
1. file.f95:42
   > Issue description

=== DRAFT SUMMARY ===
**Status**: Needs attention / Ready
Verdict: clean | needs-attention
Size gate: pass | split-recommended
CI gate: green | near-green | red

| Category | Status |
|----------|--------|
| Code Style | PASS/FAIL |
| Scientific | PASS/FAIL/N/A |
| Testing | PASS/FAIL |
| Docs | PASS/FAIL |
| CI | PASS/FAIL/PENDING |

### Key Findings
- [severity] finding (severity = blocking | major | minor | false-positive)

### CI Findings
- [severity] {check name} -- enforces {what}; red because {cause for this diff}.
  Remedy: {concrete action} ({author-fixable | maintainer-gated | re-trigger |
  infrastructure}; compound remedies chain, e.g. `maintainer-gated -> re-trigger`,
  with the actor named at each step)

Omit the CI Findings section when every check is green. Never list a red check
without a remedy line.

The `Verdict`, `Size gate`, `CI gate`, and per-finding `[severity]` tags are the
fields a consuming skill (e.g. `fix-issue` step 5) branches on; keep them explicit.

### Suggested Reviewers
- @reviewer (source: dev-ref reviewer guide; reason: changed area)

Approval Options

approve | approve with edits | skip line comments | cancel

Privacy scrub gate

Before drafting any line comment or summary, scan the model-authored finding text (problem descriptions, summary, reviewer notes) for internal tracker IDs (PER-\d+ and similar), absolute local/home paths (/Users/..., worktree/internal-tooling paths), and private host names (internal compute servers and the like). Redact these in distilled text. If a private token appears inside a verbatim quote pulled from the diff, CI log, or stack trace that must be preserved to make the finding actionable, replace it with neutral phrasing or escalate rather than post as-is. This operationalises the "Privacy is a hard invariant on every outward write" rule in .claude/rules/autonomous-workflow.md.

Autonomous Mode

audit-pr is a read-only review stage in the issue -> PR -> merge pipeline. Under an autonomous orchestrator it NEVER posts to GitHub; its terminal output is the drafted review plus the machine-readable handoff block (Verdict, Size gate, CI gate, per-finding [severity]).

  • Auto-applicable (read-only / self-scoped): gather context, run the Size Gate, read the checks and failing job logs, perform the review, and emit the verdict block. This review-and-emit tier is the only one an opt-in standing approval can cover.
  • Human-gated, always escalate (never auto-applied, in any mode): posting any line comment or review onto the PR, and every CI remedy that mutates the PR -- applying a bypass label, re-running a workflow, closing/reopening to refresh the event payload, or pushing a fix. Naming such a remedy in the draft is part of the read-only review tier; carrying it out never is. The "Wait for human confirmation" step is NOT satisfied by a batch standing approval; there is no standing approval for posting. The autonomous terminal is escalate-with-draft, the same shape as triage-issue's needs-discussion and queue-pr's merge escalation.

References

  • references/review-checklist.md - Detailed checklist
  • references/workflow-steps.md - Full workflow
  • references/ci-gates.md - CI gate diagnosis, remedies, and bypass-label policy
  • references/style-checks.md - Style rules
  • dev-ref/CODING_GUIDELINES.md
  • dev-ref/REVIEW_PROCESS.md

Keep looking

Skills are one crate of 614,194. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.