Apollo Neuro
v2.0.0像神经系统一样选择最快路径:紧急的事快速处理,复杂的事慢慢想。
Security Scan
OpenClaw
Suspicious
medium confidencePurpose & Capability
The skill's name/description (routing tasks into fast/slow/intuition paths) aligns with the included classification script (neuro-check.sh) which implements keyword-based routing and reads a circadian phase file. However SKILL.md references other artifacts and scripts (e.g., scripts/neuro/neuro-classify.sh, .neuro/patterns.json, pattern accumulation/upgrade logic) that are not present in the code file. That discrepancy suggests the documentation and implementation are out of sync.
Instruction Scope
SKILL.md instructs classification, pattern accumulation, and integration with apollo-circadian; the script implements only keyword classification, circadian phase reading, and saving a route-state JSON. There are no instructions to access external network endpoints or unrelated system files, but the SKILL.md promises pattern storage/management behavior that is not in the shipped script.
Install Mechanism
No install spec (instruction-only plus a single shell script). No external downloads or package installs. This is low-risk from an install mechanism perspective.
Credentials
The skill requests no environment variables or external credentials. The script does, however, write to and read from absolute paths under /root/.openclaw/workspace/.neuro and /.circadian — persistent state is stored on disk in an agent workspace. While reasonable for a local skill, the use of absolute /root paths and on-disk persistence should be reviewed (it requires write access and will persist user task text).
Persistence & Privilege
always:false and no cross-skill config changes — good. The script persists per-run state to a JSON file under /root/.openclaw/workspace which is within the agent workspace but is an absolute path. This gives the skill the ability to accumulate and retain task descriptions and routing decisions over time; that persistence is expected for pattern accumulation but the actual pattern-storage implementation is missing.
What to consider before installing
What to check before installing:
- Confirm implementation vs docs: SKILL.md references scripts and features (neuro-classify.sh, pattern accumulation, .neuro/patterns.json) that are not present; ask the author whether those are intentionally unimplemented or missing files.
- Review persisted files: the script writes route-state.json under /root/.openclaw/workspace/.neuro — verify you are okay with the skill storing task descriptions and routing decisions on disk and that the path is appropriate for your environment (consider changing to a non-root workspace path).
- Confirm execution context: the script assumes Python3 and filesystem write permissions; ensure the agent will run it with the expected permissions and not as an overly-privileged account if you want least privilege.
- Data sensitivity: because the script saves task text, avoid using it for sensitive secrets unless you audit how/where patterns and states will be stored and rotated.
- If you expect pattern accumulation/auto-upgrade features, get a clarified or updated release that actually implements and documents them; otherwise treat this as a simple keyword classifier only.Like a lobster shell, security has layers — review code before you run it.
Runtime requirements
🧠 Clawdis
latest
Apollo Neuro — 分层响应决策系统
核心准则
AI的任务处理应该像人的神经系统:不同类型的任务走不同的路径,不需要每次都从零思考。
神经系统的分层响应:
- 交感神经:紧急情况,秒级响应,不经大脑
- 副交感神经:日常稳态,自动维持
- 脊髓反射:条件反射,不需要思考
- 皮层决策:复杂问题,深思熟虑
对应的AI路径:
- 快速路径:跳过多轮确认,直接执行
- 守护路径:低优先级队列,定期处理
- 直觉路径:匹配已知pattern,直接输出历史最优解
- 慢速路径:完整brainstorming→planning→verification
四个响应路径
路径1:快速路径(交感神经)
适用场景:
- 用户明确说"紧急/立即/马上"
- 或apollo-circadian判断当前是高效期
处理方式:
→ 跳过多轮确认
→ 直接执行
→ 简化验证(只验证结果,不验证过程)
→ 完成后简要报告
路径2:守护路径(副交感神经)
适用场景:
- 日常查询、简单操作
- 后台维护任务
- 低优先级队列中的任务
处理方式:
→ 低优先级处理
→ 批量执行
→ 不打断当前工作流
→ 完成后静默通知
路径3:直觉路径(脊髓反射)
适用场景:
- 匹配已知pattern(有历史案例)
- 相似任务多次执行后
处理方式:
→ 匹配历史最优解
→ 直接输出
→ 不重新推理
→ 如果结果满意则完成
→ 如果结果不满意则降级到慢速路径
路径4:慢速路径(皮层决策)
适用场景:
- 复杂分析(多维度、新领域)
- 多步骤决策
- 没有历史pattern的新任务
处理方式:
→ 完整apollo-workflow
→ brainstorming → planning → subagent → verification
→ 充分推理和多轮确认
→ 产出高质量结论
任务分类流程
收到任务 →
1. 关键词扫描:
"紧急/立即/马上" → 交感神经(快速路径)
"稍后/不急" → 副交感神经(守护路径)
"仔细想想/深入分析" → 皮层(慢速路径)
2. 上下文判断:
- 匹配已知pattern? → 直觉路径
- 新领域/复杂推理? → 慢速路径
- 简单查询/日常操作? → 守护路径
3. apollo-circadian协同:
- 高效期 + 复杂任务 → 慢速路径
- 低效期 + 复杂任务 → 建议推迟或降级
- 低效期 + 简单任务 → 守护路径
4. 输出分类结果
路径监察机制
快速路径监察:
→ 执行后评估结果
→ 如果失败 → 降级到慢速路径,重新处理
→ 如果成功 → 记录为可选的直觉pattern
直觉路径监察:
→ 如果结果不满意
→ 降级到慢速路径
→ 并记录:为什么直觉路径失败了
慢速路径监察:
→ 如果发现任务其实很简单
→ 记录为新的直觉pattern
→ 下次走直觉路径
直觉Pattern积累
直觉Pattern库(apollo-neuro的核心资产)
Pattern格式:
{
"task_type": "小红书爆款标题分析",
"trigger_keywords": ["标题", "爆款", "小红书"],
"success_cases": 5,
"average_output_quality": "8.5/10",
"last_used": "2026-04-06",
"downgrade_count": 0
}
积累规则:
- 同一类型任务慢速路径成功执行3次 → 记录为直觉pattern
- 直觉pattern连续失败2次 → 删除该pattern
- 直觉pattern每周至少使用一次 → 保留;否则降级
输出格式
🧠 [apollo-neuro] 响应路径决策
任务:[用户描述]
关键词触发:[紧急/稍后/分析/...]
历史匹配:✅ 匹配直觉pattern / ❌ 无历史匹配
分类结果:
路径:⚡快速 / 🌿守护 / 🎯直觉 / 🤔慢速
理由:[为什么选这条路径]
预期耗时:[X分钟 / 不知道(第一次遇到)]
apollo-circadian协同:
当前相位:[相位名]
任务与相位匹配:🟢匹配 / 🟡错配 / 🔴强烈不建议
路径监察:
状态:[执行中/完成/已降级]
直觉pattern积累:[N]个
紧急模式(肾上腺素模式)
当apollo-circadian的褪黑素窗口(21:00后)收到紧急任务时:
触发紧急模式 →
1. 通知主AI:紧急任务打断整理窗口
2. 暂停apollo-dream整理流程
3. 切换到快速路径处理紧急任务
4. 完成后询问:是否继续整理窗口?
5. 如果是 → 恢复apollo-dream
如果否 → 记录,下次再整理
与其他Skill的协同
apollo-circadian ──→ 判断"什么时候"(高效期/低效期)
apollo-neuro ──────→ 判断"来任务了该走哪条路"
apollo-dream ──────→ 慢速路径的一部分(深度整理)
apollo-renal ──────→ 守护路径可以批量的上下文过滤
验收标准
- 关键词触发分类(紧急/推迟/复杂分析)
- 四种路径正确识别(sympathetic/parasympathetic/cortical/intuition)
- 简单任务不走完整pipeline(可测量效率提升)
- 紧急任务不被复杂流程拖慢(响应时间可测量)
- 直觉pattern能够积累(同一类型任务第二次明显更快)
- 快速路径失败时自动降级到慢速路径
- 紧急打断apollo-circadian整理窗口时有确认机制
待决策事项(已全部确认 ✅)
✅ 2026-04-06 袁文同确认
- Q8 直觉升级阈值:3次慢速路径成功执行后升级为直觉pattern
- Q9 pattern上限:不设上限
- Q10 "紧急"定义:真正紧急的才走快速路径(生命/安全/有明确deadline的)
- Q11 直觉降级:连续2次失败则删除该pattern
实施文件
| 文件 | 作用 |
|---|---|
| scripts/neuro/neuro-classify.sh | 任务分类脚本 |
| .neuro/patterns.json | 直觉pattern存储 |
| skills/apollo-neuro/SKILL.md | 本文档 |
v1 实施规格
任务分类脚本
# 手动分类
bash scripts/neuro/neuro-classify.sh "帮我马上发这个文件给老板"
# 输出
🧠 [apollo-neuro] 任务分类
任务:帮我马上发这个文件给老板
路径:⚡快速路径|跳过多轮确认,直接执行
分类逻辑(v1)
| 触发关键词 | 路径 |
|---|---|
| 紧急/马上/立即 | ⚡交感(快速) |
| 稍后/不急 | 🌿副交感(守护) |
| 深入分析/仔细想想 | 🤔皮层(慢速) |
| 是什么/查一下/看看 | 🌿副交感(守护) |
| 规划/设计/策略/研究 | 🤔皮层(慢速) |
| 匹配直觉pattern | 🎯直觉路径 |
直觉Pattern存储
{
"patterns": [
{
"id": "小红书-标题生成",
"task_type": "小红书爆款标题分析",
"trigger_keywords": ["标题", "爆款", "小红书"],
"success_cases": 5,
"last_used": "2026-04-06",
"downgrade_count": 0
}
],
"meta": {"created": "2026-04-06", "total_uses": 0}
}
下一步
v1只实现了分类脚本和pattern存储。 以下功能待实施:
- 路径监察机制(执行后评估成功/失败)
- 直觉升级(3次成功→升级为pattern)
- 直觉降级(连续2次失败→删除pattern)
- 与apollo-circadian的协同(高效期/低效期路由)
Comments
Loading comments...
