Code Review

Code review

Review is source-read-only unless the user separately authorises fixes. A reviewer may write compressed findings and traces only to an assigned artifact directory. The diff is the entry point, not the boundary. Judge the resulting system, not the changed lines alone.

Review

1. Read project instructions and preserve the requested scope and output

contract.

1. Establish the change intent from the spec, issue, tests, commit message, or

user request. Mark unclear intent instead of inventing it.

1. Trace the affected dependency cone: full touched files, live callers/

consumers, exports, canonical owners, tests, configuration, migrations, persistence boundaries, dependency/lock changes, and generated artefacts. Record inspected/excluded surfaces; unavailable context stays unknown.

1. Select review lenses from the task/risk profile. Correctness/spec alignment

is mandatory; add security/privacy, data/concurrency, performance/reliability, tests, architecture, readability, UX/accessibility or operations only when activated by the dependency cone.

1. For substantial+ or wide-surface work, use independent targeted agents with

distinct primary lenses and deliberate overlap on the riskiest invariant. orchestrate adds the required other-primary and distinct-family review pressure from HARNESS.md. Then run an anonymised claim challenge and a fresh reduction. Never rank prose or majority-vote findings.

1. Review the design delta. Look for duplicated ownership, scattered flags,

nullable modes, casts hiding invariants, thin wrappers, parallel flows, and abstractions adding concepts without reducing complexity.

1. Ask whether a proven reframe could delete branches, state, layers, or

duplicated flows. Size is a signal, not a defect. Past 1000 lines a source file is a finding; tests only warn.

1. Check tests and verification against the acceptance criteria. Confirm the

trajectory: relevant deterministic checks ran and their results are available. Never infer coverage from a green summary alone.

1. Report only high-confidence, actionable findings. Return clean when no

genuine defect exists.

Load [review-lenses.md](references/review-lenses.md) for the detailed inspection questions, [multi-agent-review.md](references/multi-agent-review.md) for lens fan-out/council reduction, and [finding-contract.md](references/finding-contract.md) before writing findings.

Boundaries

• orchestrate chooses reviewer topology, families, and waves. This skill

defines what each reviewer inspects and the independence/reduction contract.

• diagnose owns reproduction and root cause for known broken behaviour.
• implement owns authorised fixes and bounded re-review; tdd or refactor

may supply the method. Do not mutate source during review.

• Artifact-only authority permits named outputs under the assigned run

directory; it does not permit arbitrary repo-root scratch. Use the system temporary directory when no run directory exists. Never redirect a command over a wildcard/list that can include its own growing output; keep captures bounded.

• Language, framework, UI, security, and project skills add specialised lenses;

do not repeat their doctrine here.

• A structural alternative blocks only when tied to a present defect or

material regression with a safer, validated design. Attractive redesign alone is not a finding.

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mblauberg
• Source: mblauberg/provenant
• License: MIT

Install and usage instructions live in the source repository linked above.