Tear apart a written artifact and iterate it to a pass. **Use it when something is already written** — an OpenSpec change, a proposal, a spec, a diff; a pure-prose proposal belongs here too. **Nothing written yet, and the question is "which way should we go" (technology selection / architecture / whether to build)? → that is `council`, not this.** Each round three reviewers run in parallel (Code Reviewer + Reality Checker + an Independent Reviewer — on Claude Code, a cross-family one), their returns are pasted verbatim, findings merge into one triage list, a fixer applies the accepted ones, and the loop re-reviews until a terminal token. Every check reads an artifact the checked party did not author — that is the whole design. Standing counter-pressures: a scope fence (a fix that would add features nobody asked for stops and asks), a simplicity lens (the loop only inserts, so something must count the bloat), a legibility lens (a cold reader, so an iterated review cannot turn a readable document into a patch p
Tear apart a written artifact and iterate it to a pass. **Use it when something is already written** — an OpenSpec change, a proposal, a spec, a diff; a pure-prose proposal belongs here too. **Nothing written yet, and the question is "which way should we go" (technology selection / architecture / whether to build)? → that is `council`, not this.** Each round three reviewers run in parallel (Code Reviewer + Reality Checker + Codex — a cross-family reviewer, which is the one distribution the other two cannot supply), their returns are pasted verbatim, findings merge into one triage list, a fixer applies the accepted ones, and the loop re-reviews until a terminal token. Every check reads an artifact the checked party did not author — that is the whole design. Standing counter-pressures: a scope fence (a fix that would add features nobody asked for stops and asks), a simplicity lens (the loop only inserts, so something must count the bloat), a legibility lens (a cold reader, so an iterated review cannot turn a re
Tear apart a written artifact and iterate it to a pass. **Use it when something is already written** — an OpenSpec change, a proposal, a spec, a diff; a pure-prose proposal belongs here too. **Nothing written yet, and the question is "which way should we go" (technology selection / architecture / whether to build)? → that is `council`, not this.** Each round three reviewers run in parallel (Code Reviewer + Reality Checker + an Independent Reviewer — on Claude Code, a cross-family one), their returns are pasted verbatim, findings merge into one triage list, a fixer applies the accepted ones, and the loop re-reviews until a terminal token. Every check reads an artifact the checked party did not author — that is the whole design. Standing counter-pressures: a scope fence (a fix that would add features nobody asked for stops and asks), a simplicity lens (the loop only inserts, so something must count the bloat), a legibility lens (a cold reader, so an iterated review cannot turn a readable document into a patch p
Use for open decisions with NOTHING WRITTEN yet — technology selection, architecture, whether to build something. Pulls real specialists from the agency-agents catalog (one seat per named viewpoint gap, 4 or more), gets independent round-one positions on ONE decision, numbers the disagreements into cruxes, and makes the seats argue them out. The expert debate is the core protocol. Hosts with canonical dispatch, return and tool records add an audited assurance tier and may emit CONVERGED after audit, human confirmation and a confirmation attestation; hosts with fresh-context seats but weaker provenance still run the full debate and emit a non-certifying, non-authorizing ADVISORY with every missing guarantee disclosed. Triggers - convene a council, multi-expert debate, get several perspectives on this decision, have some experts argue this out, 开个评议会, 让几个专家吵一下, 多视角讨论这个方案, 找专家评议, 这个选型该怎么定. Something already written — a proposal, a spec, a diff — that needs tearing apart is review-loop, not council.
以编排模式实现 OpenSpec 变更——主 agent 把任务按范围分组、每组派 subagent 开发、自己只做状态管理与 review。当用户想用 subagent 分组方式开始/继续实现任务时使用。这是 openspec-apply-change 的并存定制版，不覆盖原 skill。需要 openspec-cn CLI（对齐 1.6.0）与 git（review 的地面真相来自它）。平台中立版：实现专家先读 catalog 再解析（读不到即停机），不依赖任何宿主专属的 subagent API。
Use for open decisions with NOTHING WRITTEN yet — technology selection, architecture, whether to build something. Pulls real specialists from the agency-agents catalog (one seat per named viewpoint gap, 4 or more), gets independent round-one positions on ONE decision, numbers the disagreements into cruxes, and makes the seats argue them out. The expert debate is the core protocol. Hosts with canonical dispatch, return and tool records add an audited assurance tier and may emit CONVERGED after audit, human confirmation and a confirmation attestation; hosts with fresh-context seats but weaker provenance still run the full debate and emit a non-certifying, non-authorizing ADVISORY with every missing guarantee disclosed. Triggers - convene a council, multi-expert debate, get several perspectives on this decision, have some experts argue this out, 开个评议会, 让几个专家吵一下, 多视角讨论这个方案, 找专家评议, 这个选型该怎么定. Something already written — a proposal, a spec, a diff — that needs tearing apart is review-loop, not council.
以编排模式实现 OpenSpec 变更——主 agent 把任务按范围分组、每组派 subagent 开发、自己只做状态管理与 review。当用户想用 subagent 分组方式开始/继续实现任务时使用。这是 openspec-apply-change 的并存定制版，不覆盖原 skill。需要 openspec-cn CLI（对齐 1.6.0）与 git（review 的地面真相来自它）。
Use for open decisions with NOTHING WRITTEN yet — technology selection, architecture, whether to build something. Pulls real specialists from the agency-agents catalog (one seat per named viewpoint gap, 4 or more), gets independent round-one positions on ONE decision, numbers the disagreements into cruxes, and makes the seats argue them out. The expert debate is the core protocol. Hosts with canonical dispatch, return and tool records add an audited assurance tier and may emit CONVERGED after audit, human confirmation and a confirmation attestation; hosts with fresh-context seats but weaker provenance still run the full debate and emit a non-certifying, non-authorizing ADVISORY with every missing guarantee disclosed. Triggers - convene a council, multi-expert debate, get several perspectives on this decision, have some experts argue this out, 开个评议会, 让几个专家吵一下, 多视角讨论这个方案, 找专家评议, 这个选型该怎么定. Something already written — a proposal, a spec, a diff — that needs tearing apart is review-loop, not council.