second-model-consensus

second-model-consensus

You are the proposer and round-driver. The reviewer is always a FRESH second-model worker that reads the live artifacts itself — never your paraphrase, never the author's own session.

BYPASS — check FIRST

• Trivial single-file reversible change with objective acceptance → no consensus round; land it.
• The user already gave an explicit verdict on the exact text → their word outranks any model reviewer; do not re-litigate.
• Code review of an implementation diff → code-review / pr-review-toolkit own that; this skill is for proposals, guidance, plans, audits, and cross-model verdicts.

TRIGGER

• Kernel / routed-rules / skill edits (maintenance.md §1 traffic) after user approval-in-principle: user approval sets scope, the consensus round guards quality. Consensus NEVER substitutes for the user's approval.
• Audit conclusions or recommendation lists the user will act on.
• Any "discuss with codex/gemini" / "give me a second opinion" request.

CONTRACT — every round, no exceptions

1. Reviewer: model-dispatch §1/§4 review row — fresh gpt-5.6-sol medium via agent-tmux codex; gemini-3.6-flash-medium via agent-tmux agy for a third perspective; fresh Claude opus/sonnet via Agent for Claude-side. Reviewer is never the author (model-dispatch §6).
2. Round prompt (file in the session scratchpad, one per round):
   • GOAL + WHY — what is being judged and who asked.
   • CONTEXT — pointers to LIVE files the reviewer must read (repo paths, artifact paths), plus the exact proposed diff/claims.
   • CHECKS — numbered, objectively answerable; each gets PASS/FAIL.
   • ACCEPTANCE — per-check verdict + reason + file:line; last line exactly VERDICT: PASS | VERDICT: BLOCK (or AGREE/BLOCK for positions).
   • REPORT — full review appended to ONE shared consensus artifact (<scratchpad>/<topic>-consensus.md, one ## Round N section per round); verdicts also in the result summary. Read-only, no fixes.
3. Dispatch mechanics: defer to using-tmux-agent-tools — one-shot bounded worker, native haiku supervision proxy, blocking supervise, result.json contract. Two-way back-and-forth in one sitting → tmux-agent-dialogue instead of serial one-shots.
4. Convergence loop:
   • VERDICT: PASS → land/report; done.
   • VERDICT: BLOCK → fix ONLY what the FAILs name (adopt the reviewer's minimal wording when offered), or rebut with fresh evidence in the next round prompt. Next round re-verifies ONLY the former FAILs plus a regression check — never re-opens PASSed checks without new evidence.
   • Round cap 4 (count every worker round). Cap hit or a FAIL you believe is wrong twice → stop, present both positions to the user (judgment-rubrics §6), never a fifth silent round.
5. Multi-model panel (optional): run Sol + Gemini in parallel on the same round prompt for orthogonal blindspot coverage. Disagreement between reviewers is a finding to surface, not to average away.

Round-prompt skeleton

GOAL: Round N — adversarial review of <thing>. WHY: <user ask / gate>.
CONTEXT: read live: <paths>. Proposed change/claims: <exact text or diff>.
Prior rounds: <consensus artifact path> (Round N re-verifies only: <FAIL list>).
CHECKS: 1..K numbered, objective.
ACCEPTANCE: per-check PASS/FAIL + ≤2-line reason + file:line;
  last line exactly one of: VERDICT: PASS | VERDICT: BLOCK. Read-only.
REPORT: append "## Round N" to <consensus artifact>; verdicts in result summary.
> Do not spawn additional tmux sessions or delegate further.
> Write the structured completion (result.json contract, schema_version 1)
> to the literal result path injected into this prompt.

Red flags that have actually burned us

• "I'll just relay the proposal to the reviewer" — the reviewer must read the LIVE landed text; a paraphrase hid a dropped without logging secrets guarantee once.
• "Mostly PASS, ship it first" — one FAIL = BLOCK. Land only after the fix or an explicit user override of that named FAIL.
• "Re-run all CHECKS this round" — re-verify rounds scope to former FAILs; full re-runs burn rounds and invite verdict drift.
• "The reviewer didn't report back, so it's probably fine" — no verdict line = no round. Idle worker → capture once, classify, resend via send-wait, require a fresh nonce (see shared memory: startup screens swallow initial prompts).
• "The models agreeing equals the user agreeing" — consensus gates quality; the user gates scope. Protected edits still follow maintenance.md §1.