菜单
Use when driving a complex task through an agent team — parallel exploration across independent axes converging on a shared artifact, with an adversarial reviewer gate. NOT for single-axis or research-only work.
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
Use when this tick or session hit real, disproportionately costly friction with the harness (a pitfall it could have warned about, a slow gate, missing operational knowledge), touched structural debt it can't fix right now (a duplicate surface, a hand-roll a sanctioned crate already covers), or can ground an apply-ready diff — to the harness's operating layer, or advocating a process-phrase re-cut — in a named invariant. Files one terse capture to .flume/friction/, .flume/refactor/, or .flume/amendments/ in the documented format — exceptional, never a duty; most ticks file nothing.
Verifies the checklist summarizer end to end: runs it against TODO.md and compares the printed count with the file's items. Use after changing anything under src/ or TODO.md.
How to add knowledge to this harness: which kind owns a new fact, convention, procedure, or enforcement, and the trigger that says it is time. Use when adding or reorganizing a rule, skill, hook, or memory entry.
Use when driving a complex task through an agent team — parallel exploration across independent axes converging on a shared artifact, with an adversarial reviewer gate. NOT for single-axis or research-only work.
Use when importing a harness; not for single files.
Use when importing a harness; not for single files.
| name | coordinate |
| description | Use when driving a complex task through an agent team — parallel exploration across independent axes converging on a shared artifact, with an adversarial reviewer gate. NOT for single-axis or research-only work. |
| license | MIT |
| allowed-tools | ["Task","Read","Edit"] |
Drive a complex task through a team of agents. See the playbook for the full reference on phases, fan-out, and the reviewer gate.
Parallel exploration across three or more independent axes, converging on a shared artifact, with an adversarial reviewer gate before shipping.
