code-review

Files (4)

• SKILL.md
• openai.yaml · agents/
• performance-baseline.md · references/
• standards-smells.md · references/

SKILL.md (first lines)

# Five-axis code review

Review `HEAD` against one fixed point. Keep **Standards**, **Spec**, **Performance**,
**Observability**, and **Accessibility** independent: each applicable axis gets an
isolated reviewer, and findings retain their originating axes.

## 1. Pin the comparison

Use the user's fixed point. Otherwise infer the target from repository evidence:
PR base, documented target, then default branch. Use its merge-base with `HEAD`;
without a target use `HEAD^`, or the empty tree for a root commit. State the inferred
point and evidence; do not ask the user to select it.

Resolve the ref with `git rev-parse <fixed-point>` and capture once:

```text
git diff <fixed-point>...HEAD
git log <fixed-point>..HEAD --oneline
```

Three-dot diff compares `HEAD` with the merge-base. For the root-commit exception,
obtain the empty tree with `git hash-object -t tree /dev/null`, then use
`git diff <empty-tree> HEAD` and `git log HEAD --oneline`; a tree has no merge-base.
Stop on an invalid ref or empty diff.

## 2. Require repository configuration

Read repository instructions and resolve applicable sources, including external
project files, through [project configuration locations](../setup-beto-frega-skills/references/project-configuration.md).
Require the tracker, observability, and accessibility configuration; defaults:

- `docs/agents/issue-tracker.md`
- `OBSERVABILITY.md`
- `ACCESSIBILITY.md`

If any is absent, stop before dispatch, list gaps, and suggest explicit
`$setup-beto-frega-skills`. Do not invoke setup autonomously or invent a baseline.
Read the resolved domain-layout document when present and its relevant context/ADR
pointers. Repository rules override generic heuristics.

Read optional `PERFORMANCE.md` or its normative replacement when present. It can add
critical paths, workloads, budgets, benchmarks, and acceptance rules; its absence
never blocks Performance review or removes the universal baseline.

## 3. Resolve the spec

Follow the configured tracker. Search in order:

1. Issue/MR references in the captured commit list.
2. User-supplied spec path or issue.
3. Matching files in `docs/`, `specs/`, or `.scratch/`.
4. PR metadata when the branch identifies a PR.

Ask only after exhausting repository evidence. User-confirmed absence makes Spec
`N/A — no spec available`.
For product behavior, read canonical functionality, shared rules, accepted decisions,
and approval evidence through [consolidate](../consolidate/SKILL.md). Check the spec
against these contracts and approval evidence. Tickets/specs do not themselves
establish release or policy authority.

## 4. Resolve standards

Read applicable `AGENTS.md`, `CLAUDE.md`, `CONTRIBUTING.md`, scoped standards, and
[the smell baseline](references/standards-smells.md). Repository rules override smells;
omit findings for tooling-enforced rules.

## 5. Dispatch isolated reviewers

Read [review-triage](../review-triage/SKILL.md) for classification, follow-up tickets,
and approval. Use [model-selection](../model-selection/SKILL.md) for each assignment;
disclose each concrete model, effort, and substitution before or alongside dispatch.
Run reviewers concurrently within capacity; queue others. Never combine axes.

Give every reviewer the fixed point, exact diff/log commands, changed files,
repository instructions, and authoritative sources. Paste review-triage's
classification and feature-isolation evidence requirements in full. Require:

- classification, location, evidence, impact, smallest credible correction;
- at most 400 words, ordered by triage classification;
- only actionable findings introduced or materially exposed by the diff; no praise,
  pass lists, or speculative hardening;
- findings, `No findings`, or `N/A — <evidence-backed reason>`.

For runner, gate, or diagnostic changes, give Spec and Observability reviewers the
canonical repository command, its effective configuration, and observed output.
Verify required behavior through that command; probes with overridden configuration
qualify only the fixture. If canonical evidence is missing, report that gap rather
than infer compliance from passing probes.

### Standards reviewer

Provide all standards and paste the smell baseline in full. Cite file and controlling
rule for breaches. Smells are judgement calls, never hard requirements: name the
smell and cite the changed hunk.

When the diff adds or changes tests, also paste
[Tests Earn Their Runtime](../beto-mode/principles/tests-earn-their-runtime.md) in
full. For each added test, require the failure only it catches; report as findings
tests with none, tests at a slower level than their failure needs, one case per value
where one per class suffices, real waits or network, and tests the diff made
redundant but kept. Cite the test and the rule it breaks.

### Spec reviewer

Provide spec path and contents. Find missing/partial requirements, unrequested scope,
and incorrect apparent compliance; quote or precisely cite the controlling passage.

### Performance reviewer

Provide the spec, [universal baseline](references/performance-baseline.md) in full,
repository performance sources, and available author acknowledgements from chat,
PR, issue, or spec. Return `N/A` only if runtime, build, tests, distribution, and
feedback loops are all unaffected; `No findings` requires an applicable review.
Apply relevant rules, cite controlling rule and hunk, preserve baseline evidence and
acknowledgement labels, and explain the complete causal mechanism. Use
`Author acknowledgement: Not found` only after checking supplied and other accessible
acknowledgement sources.

### Observability reviewer

From BetoFrega/skills, MIT.

More from BetoFrega/skills