Requesting Code Review

Requesting code review

Dispatch a fresh reviewer with pointers, not session history.

This skill covers AD-HOC review: work not driven by a plan file.

Executing a plan? Stop and use executing-plans.

Use ad-hoc review before merge, after a major unplanned feature, or when stuck.

Dispatch

1. Run the project's full gate and confirm it passes. Do not dispatch on failure
2. Collect:

git rev-parse HEAD
git status --porcelain

• BASE_REF - commit before the complete review scope. Resolve it from the target branch or supplied starting commit; use current HEAD only for purely uncommitted work
• CHANGED_FILES - deduplicated, newline-delimited exact-path union of committed changes since BASE_REF and current staged, unstaged, deleted, and untracked changes. Use rename destinations. These are entry points, not a boundary

3. Dispatch a fresh generic subagent:

MUST read instructions at <skill_dir>/code-reviewer.md FIRST.
Do not act until you have read it. Then apply:
  plan_or_requirements = <requirements or plan reference>
  base_ref             = <commit before the work>
  baseline_snapshot    = none
  changed_files        = <newline-delimited exact paths>
  solved_defects       = none

Resolve <skill_dir> to the directory containing this SKILL.md.

Ad-hoc review has no baseline snapshot and no task scope, so baseline_snapshot is always none and task_id and scope_mode are omitted. The reviewer reviews the full changed_files delta. Plan execution supplies those values instead.

4. PASS: finish
5. Findings: fix them, re-run the full gate, and dispatch a fresh reviewer

Reject incorrect findings only with code or test counter-evidence. Use one clarification round for ambiguity; escalate unresolved disputes. Keep reviewer responses out of the parent context except for PASS or the short finding list. Allow 3 total attempts for empty, errored, or malformed output. Limit findings to 3 fix/re-review rounds; escalate before a fourth.