self-improving-agent

name	self-improving-agent
description	自我进化智能体：分析技能使用模式与错误率→诊断瓶颈→生成改进提案→验证→应用。当需要优化已有技能、提升工具调用成功率、或自动调优系统配置时触发。
tools	self_improving_agent
metadata	{"category":"tools","tree_id":"meta/self_improving_agent","tree_group":"meta","min_tier":"task_write","approval_type":"plan_confirm","emoji":"🧬","enabled_when":"always","security_level":"sandboxed","file_access":"scoped_write","scope_check":"workspace","sort_order":31,"intent_priority":10,"intent_keywords":{"zh":["自我进化","自我优化","自动改进","技能优化","性能分析","调优","提升成功率"],"en":["self-improve","self-evolve","optimize skill","performance analysis","tune","improve success rate"]},"policy_groups":["group:system"],"profiles":["full"],"related_tools":["capability_manage","memory_search"],"scene_hint":"分析技能→诊断瓶颈→生成提案→验证→应用(需exec_escalation)","summary":"Self-improving agent: analyze skill performance, diagnose bottlenecks, generate optimization proposals","title":"Self-Improving Agent","verb":"Analyze","usage_guide":"技能优化、性能分析、自动调优、提升工具调用成功率","tool_description":"自我进化智能体工具。分析技能使用模式→诊断瓶颈→生成优化提案→验证→应用。6步编排管线，auto_apply需exec_escalation。","tool_input_schema":{"type":"object","properties":{"target_skill":{"type":"string","description":"Name of the skill to analyze and improve"},"analysis_depth":{"type":"string","enum":["quick","standard","deep"],"description":"Analysis depth (default: standard)"},"auto_apply":{"type":"boolean","description":"Auto-apply optimizations (requires exec_escalation, default: false)"}},"required":["target_skill"]},"tool_schema":{"input":{"type":"object","properties":{"target_skill":{"type":"string","description":"目标技能名（留空则全局分析）"},"analysis_depth":{"type":"string","enum":["quick","standard","deep"],"description":"分析深度：quick=摘要、standard=标准诊断、deep=深度回溯"},"auto_apply":{"type":"boolean","description":"是否自动应用改进（false=仅生成提案）"}},"required":[]},"output":{"type":"object","properties":{"diagnosis":{"type":"string","description":"诊断报告"},"proposals":{"type":"array","description":"改进提案列表"},"applied":{"type":"boolean","description":"是否已应用"},"validation":{"type":"string","description":"验证结果"}}},"steps":[{"action":"scan_performance","description":"从记忆中搜索技能使用模式、错误频率、用户反馈信号","tool":"memory_search","input_map":{"query":"skill error failure rate {{input.target_skill}}"},"output_as":"perf_data","approval":"none","on_error":"skip"},{"action":"inspect_capability_tree","description":"检查能力树中目标技能节点的当前配置和路由权重","tool":"capability_manage","input_map":{"action":"get_node","node_id":"{{input.target_skill}}"},"output_as":"tree_node","approval":"none","on_error":"skip"},{"action":"find_skill_source","description":"在技能目录中查找目标技能的 SKILL.md 文件路径","tool":"bash","input_map":{"command":"find docs/skills -path '*/{{input.target_skill}}/SKILL.md' -type f 2>/dev/null | head -1"},"output_as":"skill_path","approval":"none","on_error":"skip"},{"action":"read_skill_source","description":"读取目标技能的 SKILL.md 源文件内容","tool":"read_file","input_map":{"path":"{{skill_path}}"},"output_as":"skill_content","approval":"none","on_error":"skip"},{"action":"diagnose_and_propose","description":"调用 LLM 子智能体对前序步骤收集的数据进行深度分析，生成诊断报告和优化提案","tool":"spawn_subagent","input_map":{"task":"Analyze skill {{input.target_skill}} performance data and generate optimization proposals","context":"{{perf_data}} {{tree_node}} {{skill_content}}"},"output_as":"diagnosis","approval":"plan_confirm","on_error":"skip"},{"action":"validate_proposal","description":"运行编译检查确认改进提案不引入回归","tool":"bash","input_map":{"command":"go build ./... 2>&1 | tail -20"},"output_as":"validation","approval":"plan_confirm","on_error":"abort"}]}}

自我进化智能体

自动分析技能性能瓶颈并生成改进提案的组合工具。

工作流

扫描记忆(性能/错误) → 检查能力树配置 → 查找技能路径 → 读取技能源码 → LLM深度分析 → 验证编译

六步执行链

步骤	动作	底层工具	审批	失败策略
1	扫描性能数据	memory_search	无	跳过
2	检查能力树节点	capability_manage	无	跳过
3	查找技能文件路径	bash (find)	无	跳过
4	读取技能源码	read_file	无	跳过
5	LLM 深度分析	spawn_subagent	plan_confirm	跳过
6	验证编译	bash	plan_confirm	中止

使用场景

场景	analysis_depth	auto_apply	说明
快速体检	quick	false	扫描近期错误，输出摘要
标准诊断	standard	false	完整6步链路，LLM生成改进提案
自动调优	deep	true	深度分析+自动应用+验证回滚

触发条件

• 用户说"优化 XXX 技能"、"提升成功率"、"自动改进"
• 系统检测到某技能连续失败率 >30%
• 定期巡逻（配合 cron 调度）

改进提案类型

1. frontmatter 调优 — 修改 min_tier、intent_keywords 权重、approval_type
2. 步骤优化 — 调整 tool_schema 步骤顺序、on_error 策略、增删步骤
3. 路由修正 — 调整 IntentPriority 避免误分流
4. 配置建议 — 建议 crabclaw.json 配置变更
5. 技能拆分/合并 — 当技能过于庞大或过于碎片化时建议重组

安全约束

• 只读分析阶段无需审批
• 任何写入操作（修改 SKILL.md、修改配置）需 plan_confirm
• 能力树写入需 exec_escalation
• 自动应用前必须通过编译验证
• 所有改动可追溯（通过 memory 记录改动历史）