菜单
Fan out a 4-member council (1 Codex thinking-high + 1 Gemini Pro + 1 Claude Opus + 1 ChatGPT Extended Pro) on a stalled re-attack. Preserves all four independent memos; hands off to the regular Strategy Searcher for pure selection. Opt-in (attempt ≥2), ~3× the cost of a single search.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Harvest a completed ChatGPT chat's last assistant message into a response file. Use when a role finished but the response file wasn't captured (e.g., orchestrator lost the chat URL, or the submit script died after the model responded).
Create or update {PROOF_REPO}/referee_targets.yaml — the target-journal registry the paper_referee role consults to judge whether the result clears the bar at the journals you might submit to. Optional; without it the referee falls back to a generic publishability check.
Use when working in MathPipeProver on ChatGPT project preparation, durable source refresh, project-source inspection, or browser-backed external_agent submission. Prefer scripts/chatgpt_browser_agent.sh prepare/submit over ad hoc UI work, enforce the Extended Pro target, and keep durable sources separate from temporary attachments.
Read-only one-shot inspection of a live ChatGPT chat — returns generating flag, assistant turn count, last assistant message length and tail. Use during long-running roles to check whether the chat is still working or done.
Verify or set the ChatGPT composer to the Extended Pro target (reasoning Pro + model 5.5; pill reads "Extended Pro" or "Pro"). Required before any analytical submission.
Add or remove durable files in a ChatGPT project's Sources tab. Use after a verified result lands, after a route pivot, or when the durable proof-state file has been updated locally.
基于 SOC 职业分类
| name | search-council |
| description | Fan out a 4-member council (1 Codex thinking-high + 1 Gemini Pro + 1 Claude Opus + 1 ChatGPT Extended Pro) on a stalled re-attack. Preserves all four independent memos; hands off to the regular Strategy Searcher for pure selection. Opt-in (attempt ≥2), ~3× the cost of a single search. |
Canonical content: .claude/commands/search-council.md in this repo.
Codex and Claude Code share the same procedure — edit the Claude command
file, not this stub.
Use only on attempt ≥2 when first-attempt routes stalled or the gatekeeper flagged scope narrowing. Don't invoke on attempt 1 (the regular searcher is sufficient when the dossier is empty).
Mechanics: the skill fans out four adapters in parallel from
scripts/council/, one per model architecture:
dispatch_codex.sh × 1 (Codex GPT-5.5 thinking-high, ephemeral session)dispatch_gemini.sh × 1 (latest Gemini Pro — the adapter passes no model
flag, so it tracks the Gemini CLI's built-in latest-Pro default; different
pretraining/RLHF lineage; requires the gemini CLI on PATH, see below)dispatch_opus.sh × 1 (different RLHF priors, surfaces cross-domain leaps)dispatch_extended_pro.sh × 1 (slowest member, 8-20 min — determines
council wall-clock)All four adapters honor the same --packet-dir / --prompt / --out contract.
Each member writes an immutable memo (codex_memo.md, gemini_memo.md,
opus_memo.md, extended_pro_memo.md) to
{proof_repo}/runs/<run>/branches/<branch>/council/attempt-<N>/. Memos are
NEVER merged — they hand off as-is to the regular /submit-role searcher,
which runs in selection mode (pick 2-4 routes across all four memos, rank,
recommend).
If the Gemini CLI is not installed/authenticated, drop that member with
--skip-member gemini; the council proceeds with the remaining three. Install
with npm install -g @google/gemini-cli and authenticate once via gemini.
The council-member prompt is at
prompts/soft/03b_council_member_soft.md. Each member is asked for 2-3
routes to keep the searcher's total under ~12.
Read .claude/commands/search-council.md for the full step-by-step,
partial-completion handling, and cost-discipline guidance.