Code Review

Code Review

Principle expression

Primary: P02 Supporting: P04, P08, P14

Principle source

Use a host Sequence and matching interpretations when the host declares them. Otherwise use this skill's read-only fallback in references/sequence.md. Read only P02, P04, P08, and P14.

Scope

Own one judgment: what decision-changing defects, unsupported claims, or residual risks does a proposed code change introduce relative to its accepted intent, current contracts, and actual impact field?

Review one whole change or one caller-declared review packet. Build an evidence-linked working model of the relevant project whole or local subsystem before qualifying defects. The packet may be executed by one agent, a Work Cell, or another host, but those carriers do not change this method. This skill may form the semantic partition plan when a review exceeds one Cell's supported working scale; it does not choose concurrency or provider resources, launch a swarm, retry a reviewer, tally opinions, accept the change, or merge it.

Do not equate review with diff narration, style preference, test-count praise, or speculative redesign. A finding needs source evidence and a reachable failure story. Absence of a finding means only that none was established within the inspected field.

Start

Ground the review before judging:

Review target and source revision:
Accepted intent or change claim:
Governing contracts and design:
Declared review scope or packet:
Known stable working envelope or prior comparable run:
Available verification evidence:
Acceptance or merge owner:

If the intended behavior or governing architecture is materially unsettled, route that decision first. Review cannot prove conformance to a contract that does not exist.

Core method

1. Recover the change and baseline. Inspect repository state, the actual diff, changed symbols, and the pre-change behavior or contract. Read the accepted design and task claim that make the change meaningful. Separate direct observation, inference, and unavailable context.

2. Fit the work to a stable review scale. Use one reviewer when the impact field, governing context, investigation branches, and report can remain coherent in one working context. When they cannot, form semantic packets small enough for one Cell to investigate and verify reliably. Preserve the whole at low resolution and make every incoming, outgoing, and deliberately overlapping boundary explicit. Read scale-controlled partitioning when a change crosses several ownership or causal boundaries. A partition plan is review method; releasing it with a concurrency value is orchestration.

3. Construct the review model. Before searching for defects, explain the relevant system at the resolution this decision needs: purpose, component responsibilities, state and effect ownership, causal path through the change, governing invariants, and known unknowns. Tie every relation to a source anchor or label it inference. Read review modeling when the change crosses components, the project model is unclear, or the caller supplied a bounded packet.

4. Test the model against the impact field. Follow changed definitions through direct callers and consumers, public types, state/effect boundaries, persistence or protocol edges, and verification surfaces. Use compiler, language-server, AST, code graph, search, history, and tests as observations when available; none is mandatory. Revise the model when source relations contradict it. Do not stop at the diff when downstream behavior can change.

5. Allocate attention by risk, not file size. Identify the principal failure mode whose truth would change the merge decision, then inspect secondary hard constraints. Load risk lenses only for boundaries the change actually touches. Within a bounded packet, record an uninspected external relation as a context gap instead of silently expanding scope or assuming it is safe. Do not report the related code or artifact as missing unless a check against the declared review target establishes that fact.

6. Trace credible failure paths. Exercise relevant success, error, cancellation, concurrency, compatibility, capability, migration, and resource paths. A category has no standing merely because it appears on a checklist; connect it to changed behavior and a reachable caller.

7. Qualify every finding. Before reporting, establish:

   • exact file, symbol, or runtime evidence;
   • the input or event sequence that reaches the defect;
   • the violated contract or concrete consequence;
   • severity based on impact and reachability, not rhetorical confidence; and
   • the smallest correction that closes the failure without redesigning the task.

   Absence from supplied packet context is not source evidence of a defect. Disprove and discard a suspicion when source behavior, language semantics, or a focused check defeats its failure story.

8. Use verification for decisions. Run the smallest check capable of confirming or rejecting a material concern. Green tests are supporting evidence, not proof that unobserved callers are safe. A new test is justified only when it exposes a realistic defect and would change the review verdict. A sentence in this skill is a criterion for judging a review, not evidence that an agent followed it. Claim behavioral improvement only from retained runs that expose the relevant behavior against a baseline or disconfirming case.

9. Return findings before summary. Report only decision-relevant findings, ordered by severity. Then list verified claims, residual risks and context gaps, checks actually run, and one proposed verdict: ready, ready_with_residual_risk, or changes_required.

Review report contract

For every finding use this minimum shape:

Severity and title:
Evidence:
Failure story:
Violated contract or consequence:
Smallest correction:

For a consequential whole-change review or a bounded packet intended for later synthesis, also return the compact review model used to reason:

Modeled scope and revision:
Responsibilities and ownership:
Before/after causal path:
Load-bearing invariants:
Source-backed relations:
Inferred or missing relations:

If there are no findings, say so directly and retain residual risks. Never add weak findings to make a report look substantive. A review report is evidence for the acceptance owner; its verdict is not merge authority and cannot replace the source code, diff, test output, or terminal records from which it was built.

Boundaries and routing

When several review reports exist, a caller may ask another application of this skill to reconcile their review models, cited evidence, and disagreements. The reconciliation must preserve incompatible models until source evidence resolves them; majority agreement cannot turn an inference into fact. Recheck every retained finding against the declared review target rather than treating a packet report as its own evidence. This is a new review packet, not authority to vote facts into existence or an instruction to launch more agents.

Verification

Read evaluation when testing this skill, comparing a single reviewer with bounded parallel packets, or deciding whether a new risk lens changes findings. Structural validity alone does not prove review quality.

Completion standard

A review is ready when the target and accepted intent are named, the relevant project or subsystem model and inspected impact field are explicit, every finding has a reachable source-backed failure story, model conflicts and disproven suspicions are retained or resolved, executed checks are distinguished from untested assumptions, any partition can be reconnected through explicit boundary relations, and acceptance authority remains external.