Code Review

Reviews must be performed by agents with fresh context. The generator (the agent that wrote the code) must never evaluate its own output. This separation prevents self-evaluation bias — the tendency to see what you intended to write rather than what you actually wrote.

Generator-Evaluator Separation

The cardinal rule: Don't let the same model grade its own exam.

• Reviewers MUST have fresh context with no shared conversation history
• Reviewers read the diff and the plan — not the implementation discussion
• Reviewers form their own understanding of intent from artifacts, not from

the implementer's explanation

• If a reviewer needs clarification, they flag it as an open question — they

do not ask the implementer

This separation is enforced structurally by dispatching review agents as independent subagents with no access to the orchestrator's conversation.

Conventional Comments

All review comments use the Conventional Comments format (https://conventionalcomments.org). Every comment MUST include a specific file:line reference.

Comment Style

Critique the code, not the coder. Assume competence. The same finding can read as collaborative or hostile depending on phrasing:

• Explain why the change is requested. A finding without a reason loses

the rationale for the next reader of the diff.

• Reserve issue: for findings that materially affect correctness,

security, or maintainability. Use suggestion: or nitpick: for preferences.

• A high comment density on a single change is a design signal, not just a

style problem. When the count climbs past ~10 substantive comments, propose splitting the change or escalating the design conversation out of the review tool.

Comment Types

Every comment body MUST begin with the label and decoration wrapped in **...** so GitHub renders it bold. Copy the format in the examples below literally — including the asterisks — into the comment body you emit.

issue (blocking): Identifies a defect that must be fixed before approval.

**issue (blocking):** This query interpolates user input without parameterization.
file: src/api/users.ts:42

suggestion (non-blocking): Proposes an improvement. The author may accept or decline.

**suggestion (non-blocking):** Consider extracting this validation into a shared utility.
file: src/handlers/create.ts:18

nitpick (non-blocking): Minor style or naming observation. Never blocks approval.

**nitpick (non-blocking):** "data" is too vague — consider "userProfile" to match the domain.
file: src/models/types.ts:7

Gate Types by Reviewer

Verdict Criteria

Security Reviewer

• PASS: No CRITICAL or HIGH findings. MEDIUM/LOW findings are reported but

do not block.

• FAIL: Any CRITICAL or HIGH finding. The pipeline MUST loop back to

IMPLEMENT. No override.

Verifier

• PASS: All detected checks (format, lint, typecheck, build, test) pass.
• FAIL: Any check fails. The pipeline loops back to IMPLEMENT.

Code Reviewer

• APPROVE: All done criteria met, no blocking issues, tests pass.
• REQUEST CHANGES: Blocking issues found. The pipeline MUST loop back to

IMPLEMENT. No override — quality issues must be resolved before shipping.

• COMMENT: Non-blocking suggestions only. Implementation is correct.

Test-quality flags. Test files are part of the diff. Walk every changed *test* / *spec* / __tests__/* file against the rules in skills/test-first-development/SKILL.md ("Test Style Rules"). The following anti-patterns are suggestion: individually and issue: when they appear across multiple tests:

• Change-detector tests — assertions on which collaborator methods were

called without verifying observable state

• Mock-everything / mock chains — mocks for collaborators that have a

real or fake equivalent

• Full-equality assertions on complex objects when one field carries the

contract

• sleep() for synchronization
• Logic in tests (if, loops, string-building) that can carry the same

bug as the code

• Tests named after methods (testProcessOrder_2) rather than behaviors

(refundsCardOnPartialFailure)

• DRY helpers that hide the asserted value

UX Reviewer

• APPROVE: API/UX is intuitive, consistent with existing patterns.
• REQUEST CHANGES: Usability issues found. Treated as a major — auto-fixed

in the loop, not surfaced to the user.

• COMMENT: Minor ergonomic suggestions (minor-and-below — may be surfaced).

Technical Writer

• PASS: Documentation is adequate for the changes made.
• GAPS: Documentation gaps identified. Recorded for future work.

Severity Tiers and the Auto-Fix Boundary

There is no single "blocker/critical/major/minor" scale — reviewers raise findings in three different vocabularies (Conventional Comments issue/suggestion/nitpick, security CRITICAL/HIGH/MEDIUM/LOW, and the APPROVE/REQUEST CHANGES/COMMENT verdict). This table is the authoritative map from any of those onto the action the orchestrator takes. Every finding lands in exactly one tier.

The consult guard (non-negotiable). While any Blocking or Major finding remains unresolved, the orchestrator MUST NOT present findings to the user or ask which ones to address. It loops the implementer automatically. The user is consulted exclusively for the remaining Minor-and-below findings, and only once the loop has driven Blocking and Major to zero. A consult prompt that lists a blocking or major finding is a defect.

Aggregating Verdicts

When multiple reviewers produce verdicts, aggregate them into a single pipeline gate decision:

1. If ANY Blocking or Major finding exists -> pipeline gate FAILS — loop back

to IMPLEMENT automatically, with no consult.

1. If only Minor-and-below findings remain -> pipeline gate is CONDITIONAL:

present them to the user, who decides. If none remain, proceed to SHIP.

1. If no findings remain -> pipeline gate PASSES (proceed to SHIP).

The loop continues until Blocking and Major are zero, capped at 5 rounds; at the cap, escalate with the full unresolved-findings summary.

Blocking and Major failures are never aggregated away and never surfaced for triage. A single CRITICAL security finding blocks shipping regardless of how many other reviewers approved.