菜单
List the skills available in this repository. Reads every .agents/skills/*/SKILL.md, extracts its name and description, and prints a compact index so an agent can pick the right skill before acting.
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
Drive one GitHub issue through the full self-driving lifecycle in an isolated worktree: investigate, plan, implement, verify, review, PR, wait for CI, and clean up. Every phase is gated and every state transition goes through the gh_issue_run.py state machine, never a hand edit. Dispatches the implementer and the four review subagents; does not merge its own PR.
The top-level self-driving loop. Orients against repo and run state, resumes an active run or selects the highest-priority unblocked GitHub issue, routes large issues through /gh-runbook and the rest straight to /gh-issue, drives the issue to a green PR against dev through gated subagent phases, mirrors the review locally with /gh-review, records lessons, and squash-merges only after rechecking the completed run and current GitHub state. One issue per invocation.
A local pull-request review that mirrors the CI review criteria so self-review and CI cannot drift: build, lint, tests, architecture, and the leak scan, run through the same scripts CI runs. Works on a PR number or on the local diff dev...HEAD. Sorts the diff into file buckets and ends with exactly one verdict.
Turn an audit, a scope, or a large issue into a parent runbook plus a set of grouped child issues. Produces a plan and a child-issue manifest (title, body, labels, milestone, blocked-by), then creates them with a script that dry-runs by default and only writes to GitHub when --apply is passed. Feeds /work.
Save and restore run state across a context compaction. Maintains a MEMORY.md whiteboard (current state, in progress, next, decisions, blockers) capped at 200 lines under the active run directory. Documents the protocol the save-session (PreCompact) and context-loader (SessionStart) hooks already run.
The in-session delegation engine /work uses to implement. One fresh subagent per task, parallel batches for independent tasks with disjoint file ownership dispatched in one message, and the fixed review order spec then quality then architecture then security. The orchestrator delegates and never implements.
| name | find-skills |
| version | 1.0.0 |
| description | List the skills available in this repository. Reads every .agents/skills/*/SKILL.md, extracts its name and description, and prints a compact index so an agent can pick the right skill before acting. |
| allowed-tools | ["Read","Glob"] |
Answer "what skills exist here and what does each do" without guessing. The skill directory is the source of truth; this skill reads it and reports.
.agents/skills/*/SKILL.md to enumerate every skill file.name and description.SKILL.md, because a skill without a
manifest is not invokable.name for a stable listing.A markdown list, one row per skill:
- `/<name>`: <first sentence of description>
Close with the count of skills found. When no SKILL.md files exist, say so
plainly rather than inventing entries.