Skip to content

pillumina/ascend-sleuth / skills/self-evolve

self-evolve

Verified

自演进深度轮(显式触发)+ 攒批聚合器。用户下内容目标(补 case / reference 沉淀 / 诊断 / 拉取)时,由对应内容 skill 执行,**收尾自动跑 evolve-check**(/skill:evolve-check, 伴随评估:有信号产卡、agent 自验证并判断采纳——用户不需要另说"改进系统")。 本 skill 只管两件事:①用户显式说"跑一轮自演进/看看有什么可改进"时的**全库深度 观测轮**(跨轮聚合信号:归因事件聚合 / 容量 / S2 校准集 / timeline → 产候选卡); ②把 evolve-check 与深度轮产出的卡**攒批聚合为一个 PR** 供人审。演进由数据触发 ——不是用户为"改进"单独立目标,而是系统做事时自动校准;本文自包含执行参数。

FileKindSize
SKILL.mdskill md16.3 KB
references/experiment.mdmarkdown3.4 KB

references/experiment.md

View on GitHub
# 自演进实验协议

maintainer 只调用 `/skill:self-evolve`;本文件是内部执行参考,不是独立 skill。

## 先选目标

- 先跑 `python3 scripts/evolve.py traces sync`,再跑 `python3 scripts/evolve.py rsi plan`;有 replay 结果时加 `--replay .s2-replay`;
- RSI plan 合并 trace/metrics/replay 信号,推荐 region,并标记 `actionable` / `needs_agent_attribution` / `needs_agent_selection`;
- 规则不确定时,让 agent 产出带 evidence 的 region 结论,用 `rsi attribute/select --state` 写回;不要在 skill 里手改 plan;
- 目标必须是具体 skill 的可变异区;骨架、phase、工具、schema、trace 字段、severity、halt 条件冻结;
- holdout issue 不进候选池;没有可变异区或归因不明确就停,记 `insufficient`。

## 执行

以 `rsi plan` 输出的 `stages` 为准;不要在本文件里手选算法或固定模式。

默认组合是:

- 归因、选 region 先走规则;规则不确定时才让 agent 补证据/在注册区中选;
- 提案默认由 agent 从失败反馈写可变异区;写之前用 `experiment request --target <t> --component <region>` 取 target-specific prompt、允许路径和输出 schema。`propose.agent` 是启发式提案节点,不是 GEPA;GEPA 是 `select.pareto + propose.agent` 的组合。baseline 只评估 seed,不产生候选。
- 候选选择默认单候选;只有 plan 里选择搜索策略时才进入候选池/Pareto;
- judge 独立于 proposer;
- gate 确定性/人工,不交给 agent;
- baseline 只评估 seed,不产生候选。

candidate 只能作为单个 region 的 Patch,在隔离 worktree 应用;apply/rollback/diff 走 `candidate` 命令,不改主检出,不写真实 traces。

执行顺序:

```text
task subagents → judge subagents
  → reflection/proposer(若 stages 需要)
  → candidate task/judge
  → gate → experiment record
```

DSH Desktop 的提案/实验由会话内 subagent 编排,不走 `agent_adapter --agent dsh`;Claude/Codex 只有在 CLI 是标准做法时才走外部命令,结果仍写回 ProposalPort/RSIStore。具体算法实现、state 命令和可选组合看机制文档与 `rsi plan` 的 `stages`;需要跨进程状态机时用对应 CLI 的 `--help`,不要把某个算法写成固定模式。

结果至少写 `target_kind`、`target_component`、`algorithm`、`algorithm_version`、`execution`、`decision`、`samples`、baseline/candidate 分数;subagent 模式写 `execution=subagent`、`eval_profile_id=subagent_judge`、`eval_ready=false`。同时保存实际 `stages`。

agent 归因只能给候选,不能写成 hard;agent 选点只能在注册区里选。

如果目标是 `diagnose.knowledge_usage`,记录命中、首命中步、候选数、全文读取、reference 查询、wrong_cases、token/时间;用 `knowledge-usage compare` 做准确率不降、效率不降的门禁,不能只看速度。

如果目标是 `diagnose.source_usage`,记录 repo/ref、files_read、tool_calls、fetch/grep、elapsed;用 `source-usage compare` 要求源码结论 correctness 不降、访问效率不降。演进评估 trace 只写 `.s2-replay/evolve/usage-traces/`,不写真实 `traces/`。

## 停点

- 预算用尽、种群满、没有未反射候选:停;
- baseline 全对:记 `insufficient`;
- 有回归:`reject`;
- 候选格式漂移/无法拼回完整 skill:`reject`;
- 记录后停;合入走 review、CODEOWNERS、分支保护,不自动合并。
reverse-engineer-anythingmorluto/reaReverse engineer native, managed, Electron/JavaScript, packaged, firmware, and browser targets with REA, and build agent environments (AgentEnv bundles) of web applications from a live instance or a capture, and of Linux desktop apps with a task from a request. Use shipped-artifact or requested runtime evidence to explain features, compare versions, decompile code, or guide a reconstruction. Skip REA for ordinary source-repository architecture analysis.Stars: 471kUpdated Claude CodeCodexVerifiedInstall
diagnosing-superpowersobra/superpowersUse when a superpowers session went wrong and your human partner wants to know why — repeated work, ignored plans, stumbles, poor results, a skill that didn't fire, "it took too long", "why is it so expensive", "what is it doing" — or wants to build a bug report for the superpowers maintainers, for the current session or a past one identified by id or path, on any harness.Stars: 322kUpdated Claude CodeCodexVerifiedInstall
dispatching-parallel-agentsobra/superpowersUse when facing 2+ independent tasks that can be worked on without shared state or sequential dependenciesStars: 322kUpdated Claude CodeCodexVerifiedInstall
executing-plansobra/superpowersUse when executing an implementation plan in the current session as the implementer yourself — your human partner chose inline execution, or no subagent tool is availableStars: 322kUpdated Claude CodeCodexVerifiedInstall