panel

Run an adversarial review panel: every configured provider reviews the same diff in parallel.

Raw slash-command arguments: $ARGUMENTS

Parse $ARGUMENTS into the engine command (build as separate args, never splice the raw string into a shell line):

• first bare word not starting with - → base git ref, pass as --base <ref>
• --staged → review staged changes only
• --diff-only → restrict CLI reviewers to the diff (they otherwise read repo files read-only for context; API reviewers always see only the diff), pass as --diff-only
• --wait → execution mode (don't pass to engine; foreground)
• --background → execution mode (don't pass to engine; background)

Execution mode

A panel run waits on the slowest model and is bounded by the longest CLI latency, often many minutes on a real diff. Default to background.

1. If $ARGUMENTS contains --wait: foreground, skip the question.
2. If $ARGUMENTS contains --background: background, skip the question.
3. Otherwise, always recommend background (panel runs are slow even on small diffs because they fan out to N models). Call AskUserQuestion once:
   • Run in background (Recommended)
   • Wait for results

Foreground path

Run synchronously:

node "${CLAUDE_PLUGIN_ROOT}/scripts/llm-review.mjs" review --all [--base <ref>] [--staged] [--diff-only]

The script first pre-flights every configured provider and only invokes the ones verified to work in this local setup. It prints a header like:

Panel pre-flight — 3/5 model(s) verified working in this setup.
Invoking: gemini (cli), claude (cli), deepseek (api)
Skipping: minimax — no balance / quota exceeded, kimi — CLI not logged in

then one section per invoked model, each delimited by a ===== <provider> · <label> · <model> ===== header. After it returns:

1. Surface the pre-flight "Invoking / Skipping" lines first, verbatim, so the user knows exactly which models ran and why others were skipped.
2. Show each model's raw review verbatim, under its own heading, in the order returned. Do not drop or soften any findings.
3. Then add a final "## Panel synthesis" section that you write, doing only aggregation (not your own new review):
   • Consensus — issues flagged by 2+ models (these are highest-confidence; list them first with which models agreed).
   • Single-model findings — notable issues only one model raised (lower confidence, but may be real blind-spot catches; keep them, attributed).
   • Disagreements — where models reached opposite conclusions or conflicting verdicts.
   • Overall verdict — combine the per-model verdicts: if any model says DO NOT SHIP, surface that prominently with its reason.
4. If a section still shows ERROR: (rare — passed pre-flight then failed), note which model and continue. If the pre-flight invoked nothing, point the user to /llm-review:status.