multi-model-review

name	multi-model-review
description	Adversarial multi-model code review — launch parallel readonly reviewers on different models, then synthesize consensus vs lone findings. Do not apply fixes unless the user asks.
disable-model-invocation	true

The user wants an adversarial multi-model code review: several independent reviewers on different models, then a synthesized verdict. You MUST run those reviewers in parallel: issue every Task subagent launch in the same assistant turn (multiple Task tool calls together), not one after another — do not wait for one reviewer to finish before starting the next. Do not edit files or apply fixes unless the user explicitly asks.

1) Pick models with ask_question

You MUST call the ask_question tool (AskQuestion) before launching reviewers. Build a single question from the exact Task/Subagent model slugs listed in your system prompt/tool instructions. Do not hardcode, invent, or repeat a repo-local model list here. Use those slugs verbatim for options[].id so they can be passed directly to the Task tool, set allow_multiple: true, and use short human-readable labels. If the user already chose models in the /multi-model-review line or chat, you may skip ask_question only when their selection is explicit and unambiguous; otherwise ask. If a selected slug is rejected at execution time, prefer the closest valid slug suggested by the Task/Subagent tool feedback.

2) Scope and intent

1. Determine what to review: user-specified paths or diff, else git diff main...HEAD (or the repo default base vs HEAD), else local unstaged + staged diffs, else files and symbols from the conversation.
2. State one clear paragraph of intent — what the change is trying to accomplish (from messages, commits, or the code).
3. Collect the material to pass to reviewers: full diff or key file excerpts.

3) Launch one subagent per selected model (parallel only)

For each model slug the user selected, launch the Task tool with subagent_type: generalPurpose, readonly: true, and model set to that slug. Parallelism is mandatory: fire all of these Task calls in one batch in a single response (concurrent tool use), never sequentially. Each subagent gets the same reviewer prompt; diversity comes from the models, not from different personas.

Each subagent prompt must include:

• The stated intent paragraph.
• The code/diff under review.
• Instructions to act as an adversarial reviewer: prioritize bugs, correctness, security, and maintainability; structured findings with severity (critical | warning | nit), concrete locations, evidence, optional suggestion; say "no findings" if nothing is wrong.

4) Synthesize (you, the parent agent)

Merge the subagent findings: consensus (2+ models), lone-model findings, deduplicate overlapping issues, note disagreements. Categorize for the user: act on / consider / noted / dismissed, with brief rationale. Do not auto-apply changes.