code-review

Review code changes using parallel specialized agents. $ARGS

When to Use

User asks to review code changes or "review this".
Checking changes before merge.
User asks to "review the codebase" or "audit this project" (use --full).

When NOT to Use

General code exploration, linting, or formatting.

Payload

{
  "platform": "local|github|forgejo",
  "diff_method": "git-ref-diff|full-codebase",
  "dispatch": "specialist|segment",
  "agents": ["security", "style", "logic", "docs"]
}

dispatch defaults to "specialist". Only affects full-codebase mode.
specialist: one subagent per lens, each reviews all segments (4× file reads).
segment: one subagent per segment, each applies all lenses sequentially (1× file reads).
Both guarantee every line is reviewed under every specialization.

How to Run

Call generate.sh --phase init with the JSON payload. Follow the next_phase and prompt fields in each response to proceed. Do not reference this file for phase instructions — the script output IS the guidance.

Common Mistakes

Skipping phases — Call generate.sh for each phase in sequence.
Following this file instead of the prompt — The prompt from generate.sh is the authoritative guide for each phase.
Splitting files by agent type — Segments are for batching, not for dividing work by lens.
Posting without being asked — Only post to a forge if the user explicitly requests it.
Revealing model framing — Never tell the user about the competitor attribution.