Review

Review

Review the code for a task, fix the problems worth fixing, and leave behind a report a human reviewer can act on. A standalone utility — run it whenever code needs a second pass, not only inside a larger flow.

The orchestrator drives the run: it talks to the user, dispatches sub-agents for analysis and for fixes, and pauses only twice — to choose which findings to fix, and to choose what happens next. Every other step runs to completion without asking.

---

Input

$ARGUMENTS — a task-slug or a path to a task-file. With no argument, determine the slug from the current branch or the latest .yoke/ai/*/ directory. This input contract is stable, so callers that pass a slug keep working.

---

Procedure

The full pipeline — scope resolution, the code-reviewer dispatch, finding classification, the fix loop, validation, and report assembly — lives in reference/review-procedure.md. Read that file and execute it end to end. The report follows the template in reference/review-format.md.

---

Rules

• Standalone utility. Review runs on its own; it never assumes a larger flow around it.
• Fix, then report. Findings the user accepts are fixed and committed; everything found — fixed or skipped — lands in the report.
• Report location. Write the report to .yoke/ai/<slug>/<slug>-review.md.
• Every fork is one question. For any decision — fix scope, next action — ask a single AskUserQuestion with the recommended option listed first.
• Delegate the work. File edits, bash, and analysis go to sub-agents; each receives only its own data.
• Read the git memory. Pull decision context from the affected files' commit history per ${CLAUDE_PLUGIN_ROOT}/skills/gca/reference/history-reading.md — a finding that contradicts a Constraint: or re-proposes a Rejected: approach must cite the commit it argues with.
• Language: match the ticket/input language, or follow the project-level definition in CLAUDE.md / AGENTS.md.