Model Debate

Review a decision using independent models and host-verified evidence. This is a small workflow, not an approval framework for ordinary edits. Honor user scope, project quorum, excluded providers and existing authorization. Use subscription CLIs; no implicit paid fallback, upgrades or global configuration changes.

Decide whether review is needed

Do not start a model review merely because work finished, files changed, a version changed, or a queue reached a checkpoint. Routine bug fixes with a reproducible cause and adequate regression tests, documentation, formatting, small reversible refactors and already-settled decisions normally need direct verification only. Record a brief skip reason; do not generate a separate review package.

Use independent review when the user explicitly asks, an applicable project hard gate requires it, or a concrete difficult issue remains after local investigation: competing consequential designs, repeated unexplained failures, ambiguous core contracts, or serious security/data-loss risk not covered by deterministic checks. State the unresolved question and how a review could change the action. One focused reviewer often suffices for a bounded implementation uncertainty; use multi-model debate only when differing independent judgments add value. Do not create a multi-model quorum for an ordinary single-review task. Never recursively review a review, or route skill maintenance through live debate by default.

Default: focused, one useful call per model

Use three independent models from different vendors unless the user or project specifies otherwise; never count aliases or repeated roles as distinct models. The host organizes and verifies, not an extra seat. Choose complementary roles with shared coverage of correctness, feasibility and counterevidence. Do not spawn additional helper/judge/synthesis seats by default.

1. Prepare once. Read the actual sources and settle cheaply checkable facts. Freeze the question, alternatives, constraints, source hashes and open claims in one compact packet. Prefer the relevant functions, contracts and direct consumers over complete files, repository histories or previous transcripts. Target roughly 4–8k input tokens per seat; this is a sizing target, not a truncation limit. Include essential context even if larger; record why.
2. Admit once per launch batch. Read quota.md for the account guard and runtime.md for adapter boundaries. Budget the whole remaining work, including implementation/review if promised, one shared contingency allowance, and any actually necessary standalone probe. Do not reserve a retry for every seat automatically. No inference ping by default: validate metadata locally and combine remaining capability checks with the first substantive call where runtime enforcement permits it.
3. Launch independent critiques together across distinct pools after fresh admission; serialize shared-pool starts. Each fresh session gets the same factual packet plus its own role, no peer output or host conversation. One call includes required research and critique. A local contract review needs local evidence, not ceremonial web searches. External changing claims need live primary sources; a verified host packet is valid with provenance marked research_origin=host-packet, not independent research.
4. Verify, then decide. The host checks load-bearing claims and every blocker against primary evidence. Agreement is not proof. Resolve factual objections locally and record design tradeoffs, minority concerns and unknowns. Do not send verified facts back to models for another acknowledgement.
5. Follow up only where it changes the decision. An unresolved material dispute may receive one compact delta question to the relevant model within the shared reserve. Include both opposing arguments faithfully. Additional seats/rounds need a concrete unresolved question and budget; never schedule all-to-all rebuttals automatically. If evidence remains insufficient, retain the blocker rather than manufacture consensus. No surviving dispute means the focused review is complete after round1.
6. Freeze and deliver. Keep model/runtime identities, exact prompts, raw results, completion records, verification, ruling, dissent and limitations. Use artifacts.md when assembling records. State what was applied; existing implementation authorization remains effective. Do not reopen permissions at routine stage boundaries.

Cost and scheduling discipline

Default reply contract: verdict, material blockers with evidence and minimal fixes, unresolved questions, and decisive tests. Aim for 600–1000 words, avoid restating the packet and enumerating already-settled claims. Record actual token usage when exposed; otherwise unknown, not a fabricated estimate. Include calls, failures, retries and elapsed time, distinguishing cached input and output usage.

Use one task-local launcher with per-seat completion records, deadlines and owned-process cleanup. Poll a batch together, typically every30–60seconds; do useful local verification between polls. No persistent scheduler, task queue, worktree or new runner framework solely for a short review. Prefer existing verified adapters. Do not repeatedly rediscover help, parse the same transcript, or reread unchanged instructions within a run.

When a result needs correction, send only the changed evidence, exact objection and directly affected context. Preserve valid independent answers for unchanged inputs; invalidate only dependent results. Never restart successful independent seats just because another seat failed. Retry only reconciled failures within existing authorization and limits.

Explicit debate profile

Use the longer debate profile when explicitly requested or when justified by an unresolved high-impact fork; disclose its extra calls before launch. It adds independent research if needed and a full cross-examination round. Do not label a focused review a completed two-round debate. For changes to this skill or an adapter, use acceptance.md; offline checks are the default, live smoke requires a concrete unresolved runtime risk and call budget.