菜单
Use when the autopilot orchestrator has finished Phase 7 archive and needs to aggregate Episode records into the L1/L2/L3 learning tiers, cluster failure/success patterns, and emit L3 promotion candidates. ONLY for autopilot orchestrator; not for direct user invocation.
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 the user invokes /slim-task or requests structured multi-phase task execution (structured task, SOP execution, parallel sub-agent dispatch) with blind-audit quality review.
Use when autopilot orchestrator enters Phase 1 to understand requirements, drive multi-round clarification, and produce the structured requirement packet checkpoint.
Use when the autopilot orchestrator enters Phase 5 after Phase 4 test case design has been approved and is about to choose an implementation path (parallel, serial, or TDD) and dispatch implementation work.
Use when the autopilot orchestrator reaches Phase 7 after Phase 6 tri-path reports are available and must finalize the delivery by collecting summaries, extracting knowledge, previewing Allure, and evaluating archive readiness.
Translates Figma designs into production-ready application code with 1:1 visual fidelity. Use when implementing UI code from Figma files, when user mentions "implement design", "generate code", "implement component", "figma to code", "figma 转代码", "figma 还原", "实现设计稿", provides Figma URLs, or asks to build components matching Figma specs. For Figma canvas writes via `use_figma`, use `figma-use`. For full-screen design generation in Figma, use `figma-generate-design`.
Use when the autopilot orchestrator main thread is about to dispatch a phase-specific sub-agent via the Task tool and needs the JSON envelope contract, explicit path injection, and parameterized prompt templates. Not for direct user invocation; skip when the current context is not the autopilot orchestrator.
| name | autopilot-learn |
| description | Use when the autopilot orchestrator has finished Phase 7 archive and needs to aggregate Episode records into the L1/L2/L3 learning tiers, cluster failure/success patterns, and emit L3 promotion candidates. ONLY for autopilot orchestrator; not for direct user invocation. |
| user-invocable | false |
前置条件自检：本 Skill 仅在 autopilot 编排主线程的 Phase 7 归档之后使用。非 autopilot 流程请立即停止并忽略。
| 层级 | 粒度 | 存储 | 生命周期 |
|---|---|---|---|
| L1 Episode | 单次 Phase 执行轨迹 | docs/reports/{version}/episodes/phase{N}.json + claude-mem create_observations | 永久 |
| L2 Pattern | 跨 Episode 聚类得到的成功/失败模式 | claude-mem build_corpus(name="autopilot-lessons") | 按需重建 |
| L3 Skill-Rule | 命中次数 ≥ 3 且无反例的晋升规则 | docs/learned/<pattern_id>.md 或 CLAUDE.md "习得规则"区块 | 人工审核后长期持有 |
| 参数 | 来源 |
|---|---|
| phase | Phase 7 传入（默认 "phase7"） |
| checkpoint | openspec/changes/{slug}/phase-results/phase{N}.json |
| version | 当前发布版本号（用于 docs/reports/{version}/episodes/ 路径） |
bash "${CLAUDE_PLUGIN_ROOT}/runtime/scripts/learn-episode-write.sh" \
--phase "$phase" \
--checkpoint "$checkpoint" \
--version "$version"
脚本负责：
gate_result ∈ {blocked, failed}，强制生成 Reflexion 风格的自然语言反思（失败归因 + 可复用教训）learn-episode-schema-validate.sh 校验 schemadocs/reports/{version}/episodes/phase{N}.jsoncreate_observations(obs_type=phase_reflection|success_pattern|failure_pattern)Phase 7 之后触发一次聚合：
# dry-run：MCP 不可用时输出规范化 JSON 产物
claude-mem build_corpus \
--name "autopilot-lessons" \
--types "phase_reflection,failure_pattern,success_pattern" \
--limit 200
聚类规则参见 references/pattern-clustering.md：
failure_trace.root_cause + phase 做 hash clusteringpattern_idbash "${CLAUDE_PLUGIN_ROOT}/runtime/scripts/learn-promote-candidate.sh" \
--episodes-root docs/reports \
--out-dir docs/learned/candidates
脚本负责：
docs/reports/*/episodes/*.json（若 MCP 可用，切换到 autopilot-lessons corpus）pattern_id 聚合，命中次数 ≥ 3 且无反例（无成功 fingerprint 抵消）→ 输出候选到 docs/learned/candidates/{pattern_id}.mdstatus: pending_review，需人工审核后晋升到正式 docs/learned/<skill>.md 或 CLAUDE.md "习得规则"区块bash "${CLAUDE_PLUGIN_ROOT}/runtime/scripts/learn-inject-top-lessons.sh" \
--raw-requirement "$raw_requirement"
输出 JSON 数组（top-3 教训），主线在 Phase 0 banner 之后注入到 dispatch prompt。
{
"status": "ok|warning|blocked|failed",
"summary": "episode 写入 + 候选扫描结果",
"artifacts": [
"docs/reports/{version}/episodes/phase{N}.json",
"docs/learned/candidates/{pattern_id}.md"
]
}
references/episode-schema.md — L1 Episode 字段定义references/pattern-clustering.md — L2 聚类策略references/promotion-rules.md — L3 晋升门槛与反例判定