Peer review

Fresh eyes from another model family. Every reviewer gets the same brief. None sees the others. Agreement is the signal.

The reviewers on this machine, how each is spawned, and whether each works today live in references/agents.tsv. Open it before spawning. Update its state and checked columns when a lane is probed. Do not inline a lane's command anywhere else; this file is the one place it lives.

The pi-subagents external CLI adapters take no model or effort flags, so they are not lanes. There is no Gemini lane. A CLI that could run under a personal Google account is not worth the account.

Steps

1. Pick lanes. One for a routine second look. Two or three when the decision is contested or expensive to reverse.
2. Materialize the evidence. Named paths, the patch as text, command output, test results. Bash lanes have no Pi tools and cannot ask questions, so the brief carries everything they need. The reviewer lane may read the repo.
3. Write one neutral brief. The question, the scope, what to return, and the rule that a missing artifact yields MISSING EVIDENCE rather than a guess. No hint of your own view. The same text goes to every lane.
4. Fan out in one async workflowScript capped at 15 minutes (timeoutMs: 900000). This is the only fixed deadline any child gets in ppstack; a reviewer that has not answered in 15 minutes is not going to. Role lanes are runs.run("<lane>", { agent: "<role>", task }). Bash lanes run from the repo under review with the brief in a heredoc so quotes survive, BRIEF="$(cat <<'EOF' ... EOF)". Run all lanes in parallel.
5. Read every answer end to end. Do not pass summaries through. Where lanes agree, treat it as high signal. Where they disagree, name the question and settle it with evidence, not by vote.

Reply: per lane, verdict and key findings; the agreement map; your decision and the evidence that settled any disagreement.