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", "memory", "project-memory-conformance", "overengineering"],
  "model_maker": "anthropic|openai|google|...",
  "model_identity": "claude-3.5-sonnet|gpt-4o|glm-5.1|..."
}

• 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.
• modelmaker and modelidentity are optional. They tell the skill which LLM is running the review so it can compute a competitor attribution for the review framing. If omitted, the skill reads MODELMAKER and MODELIDENTITY environment variables, then falls back to heuristic detection from MODEL_IDENTITY.

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.

Report Structure

The final report's shape is defined in templates/report.md.j2 (a Jinja-style shape contract, not a rendered template). The report phase passes this template to the report agent, which fills it in. Required sections include:

• Header (refs, platform, diff method, dispatch, date).
• Models — REQUIRED. Lists the orchestrator model (the model writing the report) and every subagent model that conducted a review (one row per specialist agent or segment, plus synthesis).
• Recommendation + summary.
• Per-agent or per-segment finding subsections.
• Finding counts table.

To support the Models section, the dispatch prompts in agents and synthesize phases instruct the orchestrator to record each subagent's chosen model to a sidecar file (/tmp/codereview<agent>model.txt, /tmp/codereviewsegment<id>model.txt, /tmp/codereviewsynthesis_model.txt). The report phase reads these sidecars; missing files become the literal string unknown.

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.
• Omitting the Models section — The final report MUST list the orchestrator model and every subagent model. If a sidecar file is missing, write unknown; never drop the row.
• Skipping the model sidecar write — When dispatching subagents in the agents and synthesize phases, always write the model identifier to the corresponding sidecar file. The report phase depends on it.