菜单
Launch the unattended claims burndown for auths-site — sequential fresh agents close RED gaps under the fleet gate until done, capped, or parked out
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。
基于 SOC 职业分类
Run exactly ONE improvement cycle for auths-site — triage the highest-value RED claim, then either close it or (if it's too big) break it down RED-first, gate, commit, stop
Run the improvement loop for auths-site from *inside this session* — orchestrate cycles by spawning one fresh sub-agent per cycle, gated, until done / capped / parked out
Run the adversarial review protocol on a review-gated (security-tradeoff) gap for auths-site — try to BREAK the change, never to confirm it
| name | burndown |
| description | Launch the unattended claims burndown for auths-site — sequential fresh agents close RED gaps under the fleet gate until done, capped, or parked out |
You were invoked to run the unattended improvement loop. Each cycle makes one move on the
highest-value RED claim: close it if it's small, or break it down RED-first (its
sub-claims + a sufficiency-checked assembly, children linked by covers_claim:) if it's
too big — see the cycle skill. So a hard claim fans out into children automatically and
the loop recurses; an easy backlog just closes claim by claim. You launch it and relay the
result; the loop picks the move.
.recurve/RUN-AUTO.md (operator runbook) — verify its preconditions:
recurve validate, recurve matrix --gate, recurve lock status all
green/unlocked. If any fails, STOP and report; never start unattended on a
broken baseline..recurve/workflows/burndown.js via the Workflow tool (knobs via args:
cap, maxConsecFails, parked seed list) — or, without a workflow
runtime, run .recurve/workflows/burndown.sh with AGENT_CMD set.