review

review

Read .agents/skills/review/SKILL.md and follow its workflow steps.

Claude Code-specific additions

Argument parsing

Parse $ARGUMENTS:

• A number or #N → treated as a PR number
• --axes=<axis,...> → explicit specification of perspectives to apply
• --deep → run in deep mode (launch Agent({subagent_type: "code-reviewer"}) per perspective in parallel, via multiple tool calls in the same message)

Parallel launch in deep mode

In deep mode, launch a worker per perspective in parallel. Prefer the Workflow approach when the Workflow tool is available, falling back to the Agent tool approach when it is not.

Workflow approach (recommended)

Pass the perspectives via args.axes, and structurally validate findings with schema (mismatches are automatically retried, so there is no dropped output from parse failures):

export const meta = {
  name: 'review-deep',
  description: 'review: 観点別 worker を並列起動して findings を収集する',
  phases: [{ title: 'Review' }],
}

// canonical の findings JSON schema（Schema v1）を JSON Schema 化したもの
const FINDINGS_SCHEMA = { /* axis / findings[]: {file, line, severity, message, ...} */ }

const results = await parallel(args.axes.map(axis => () =>
  agent(
    `axis: ${axis}\nmode: deep\ncontext:\n  base: ${args.base}\n  head: ${args.head}\n  pr_number: ${args.pr_number ?? ''}\n\n${args.diff}`,
    { label: axis, phase: 'Review', agentType: 'code-reviewer', schema: FINDINGS_SCHEMA }
  )
))
return { findings: results.filter(Boolean) }

• Reuse the existing subagent definition via agentType: 'code-reviewer'
• Aggregation (duplicate merging, cross-perspective conflicts, grouping) is not done inside the workflow; after return, the caller passes the findings to review.mjs render and leaves it to the engine (canonical step 3). Do not add an aggregation agent inside the workflow
• Pass the diff via args (do not have the worker fetch it)

Agent tool approach (fallback)

Agent({
  subagent_type: "code-reviewer",
  prompt: "axis: <axis>\nmode: deep\ncontext:\n  base: <base>\n  head: <head>\n  pr_number: <N>\n\n<diff>"
})

• Independent subagents within the same wave are launched in parallel via multiple tool calls in a single message
• Aggregation is done by the caller passing findings to review.mjs render (do not add an LLM call)
• Merge the subagents' return values (JSON) and push them into findings[]

After the completion report

Call AskUserQuestion immediately after the completion report (do not set the answers parameter).

When there are findings:

• "Fix the findings" → Fix the code based on Critical / Warning findings (Info is out of scope)
• "Proceed as-is" → exit

When there are no findings:

• "Run through commit and PR in one go" → Read .claude/skills/ship/SKILL.md and follow its steps
• "Proceed as-is" → exit