技能编译器
- 作者仓库星标 494
- 作者仓库 Athena-Public
- 领域
- 人工智能
- 作者声明的智能体
- 未找到明确声明;不据此推断已兼容或已测试
- 静态检查
- 88 / 100 · 启发式扫描,不代表运行安全
- 运行环境证据
- 未执行;当前没有隔离运行测试证据
- 作者 / 版本 / 许可证
- @winstonkoh87 · 未声明许可证
- 上下文用量评级
- 低消耗
- 接入复杂程度
- 需简单配置
- 是否需要外部接口密钥
- 不需要
- 兼容的系统
- 未声明(默认跨平台)
- 运行环境要求
- Python
- 检测到的文件与系统行为
-
- 只读
- 允许写入 / 修改
- 命令执行
- 网络行为
- 仅限本地
- 安装命令数
- 无(仅作为资料)
需要注意: 未声明工具权限范围;实际权限由运行环境决定。
第一阶段:模式提取(分析)
在`<analysis>`标签中执行私有分析(不写入文件):
第二阶段:技能草案生成
使用 Athena 标准 5W1H frontmatter 生成完整的 SKILL.md:
第三阶段:整合
将技能写入`.agent/skills/[name]/SKILL.md`(或`examples/skills/[category]/[name]/SKILL.md`,对于公共)使用新条目更新技能索引更新`AGENTS.md`技能表(如果存在context_trigger)通知用户:“📦已编译新技能:来自此会话的`[name]`。在[路径]处查看。”
# Skill Compiler — Solved-to-Skill Automation
> **Source:** Hermes Agent by Nous Research (May 2026)
> **Core Claim:** "It's the only agent with a built-in learning loop — it creates skills from experience."
> **Athena Adaptation:** Hermes does this via Python (`agent/curator.py` + `tools/skill_usage.py`). Athena does it via workflow-level pattern detection + markdown SKILL.md generation.
## The Problem This Solves
Athena currently relies on **manual insight filing** during `/end` sessions. The `[S]` and `[V]` markers in session logs capture learnings, but they remain trapped in session logs — they don't **become** reusable skills automatically.
Hermes solved this: when the agent completes a novel task, it automatically creates a new skill from the solution, so the same problem class never requires re-derivation.
## When to Use
### Automatic Trigger (Post-Task Detection)
After any task completion where **ALL** of the following are true:
1. **Novelty**: The task required a solution path not covered by any existing skill
2. **Complexity**: Task took ≥5 agent turns OR involved ≥3 tool calls
3. **Success**: User confirmed the solution worked (explicit or implicit — no corrections in final 2 turns)
4. **Reusability**: The solution generalizes beyond this specific instance
### Manual Trigger
User says: "compile this into a skill", "save this as a skill", "I want to remember how we did this"
## Execution Flow
### Phase 1: Pattern Extraction (Analysis)
Perform private analysis in `<analysis>` tags (not written to files):
```
<analysis>
1. What was the PROBLEM CLASS? (Not the specific instance)
- e.g., "Pairs trading dashboard with cointegration analysis"
- NOT "Dashboard 4-decimal rounding fix"
2. What was the SOLUTION ARCHITECTURE?
- Key steps in order
- Tools/APIs used
- Decision points and their resolution criteria
3. What were the FAILURE MODES encountered?
- What went wrong initially?
- What heuristics resolved it?
4. What is the REUSE SURFACE?
- When would someone encounter this problem class again?
- What context_trigger keywords would match?
5. OVERLAP CHECK
- Which existing skills partially cover this?
- Is this better as a new skill or a subsection of an existing one?
</analysis>
```
### Phase 2: Skill Draft Generation
Generate a complete SKILL.md with Athena-standard 5W1H frontmatter:
```markdown
---
name: [kebab-case-name]
description: "[One-line description of what the skill does]"
vibe: "[One-line emotional hook]"
context_trigger: "[comma-separated trigger keywords]"
auto-invoke: false
model: default
source: "Compiled from session [SESSION_ID] on [DATE]"
compiled_from: "[session log path]"
---
# [Skill Name] — [Subtitle]
> **Compiled**: [DATE] from session [SESSION_ID]
> **Problem Class**: [Description of the general problem this solves]
## When to Use
[Trigger conditions — when should the agent invoke this skill?]
## Solution Architecture
### Step 1: [Phase Name]
[What to do, with specifics]
### Step 2: [Phase Name]
[What to do, with specifics]
## Failure Modes & Mitigations
| Failure | Mitigation |
|---------|------------|
| [What can go wrong] | [How to recover] |
## Validated Patterns
- [V] [Pattern]: [Why it works] | Reapply: [When]
## References
- [Session log](path/to/session/log)
- [Related skill](path/to/related/skill)
```
### Phase 3: Integration
1. **Write** the skill to `.agent/skills/[name]/SKILL.md` (or `examples/skills/[category]/[name]/SKILL.md` for public)
2. **Update** skill index with the new entry
3. **Update** `AGENTS.md` skill table (if context_trigger present)
4. **Notify** user: "📦 Compiled new skill: `[name]` from this session. Review at [path]."
## Curator Integration (Stolen: Hermes `agent/curator.py`)
### Lifecycle States
Compiled skills follow a lifecycle identical to Hermes' curator model:
| State | Criteria | Action |
|-------|----------|--------|
| **active** | Created or used within 30 days | Normal operation |
| **stale** | No invocation for 30+ days | Flag for review at next `/audit` |
| **archived** | No invocation for 90+ days | Move to archive directory |
### Invariants (from Hermes)
- **Never auto-delete** — maximum destructive action is archive
- **Pinned skills are exempt** — manual pin via `pinned: true` in frontmatter
- **Only touch compiled skills** — bundled/manual skills are off-limits
- **Archive is recoverable** — archive directory with successor mapping in README.md
### Umbrella Consolidation Rule (Stolen: Hermes Curator Prompt)
> "A collection of hundreds of narrow skills where each one captures one session's specific bug is a FAILURE of the library — not a feature."
When compiling a new skill, first check if it belongs as a **subsection** of an existing umbrella skill rather than a standalone entry:
1. **PREFIX CLUSTER CHECK**: Does the new skill share a first word or domain keyword with 2+ existing skills?
2. **CLASS-LEVEL CHECK**: Would a maintainer write this as one skill with labeled subsections, or N separate skills?
3. If the answer is "one skill", absorb into the existing umbrella and add a `references/` entry instead.
## Anti-Patterns
- ❌ Compiling trivial tasks (< 5 turns, single tool call)
- ❌ Compiling tasks that are already covered by existing skills
- ❌ Creating overly specific skills tied to one instance (use umbrella pattern instead)
- ❌ Compiling without user confirmation of success
- ❌ One-session-one-skill micro-entries — consolidate into class-level umbrellas
## Hermes Comparison
| Feature | Hermes | Athena |
|---------|--------|--------|
| Skill creation | Automatic (Python `skill_manage`) | Workflow-triggered (this SKILL.md) |
| Skill lifecycle | `curator.py` (7-day review cycle) | `/audit` + archive directory |
| Skill evolution | DSPy + GEPA (self-evolution repo) | Manual via `/steal` + session learnings |
| Skill storage | `~/.hermes/skills/` (SQLite telemetry) | `.agent/skills/` (git-tracked) |
| Consolidation | LLM-driven umbrella-ification pass | Manual during `/audit` |
**Athena advantage**: Skills are version-controlled in git, not SQLite. Every skill change has a commit hash, blame, and diff history. Hermes can't `git blame` a skill edit.
## References
- [Hermes Agent Curator](https://github.com/NousResearch/hermes-agent/blob/main/agent/curator.py) — Source of lifecycle states + umbrella consolidation pattern
- [Hermes Self-Evolution](https://github.com/NousResearch/hermes-agent-self-evolution) — DSPy + GEPA evolutionary optimization (future steal candidate)
章节与文件索引
按说明文件中的顺序排列。
这解决的问题
自动触发(任务后检测)
手动触发
第一阶段:模式提取(分析)
第二阶段:技能草案生成
文件/命令 -> agent/curator.py · tools/skill_usage.py · /end · [S] · [V] · <analysis> · .agent/skills/[name]/SKILL.md · examples/skills/[category]/[name]/SKILL.md
内容校验值 -> f9a7d2ff0d5e
当前原文的自动摘译,尚未人工审校。
这解决的问题
Athena 目前在
/end会话期间依赖于手动洞察归档。会话日志中的[S]和[V]标记捕获学习内容,但它们仍然被困在会话日志中——它们不会自动成为可重用的技能。 Hermes 解决了这个问题:当智能体完成一项新任务时,它会自动从解决方案中创建一项新技能,因此同一问题类别永远不需要重新推导。自动触发(任务后检测)
满足以下所有条件的任何任务完成后: 新颖性:任务需要任何现有技能未涵盖的解决方案路径 复杂性:任务花费了 ≥5 个代理轮次或涉及 ≥3 个工具调用 成功:用户确认解决方案有效(显式或隐式 - 最后 2 轮中没有更正) 可重用性:解决方案可推广到此特定实例之外
手动触发
用户说:“将其编译为技能”、“将其保存为技能”、“我想记住我们是如何做到这一点的”