feat: add persona-owned agent limbs and Light Lake routing
This commit is contained in:
parent
43be75b60a
commit
8242c6dfc4
34 changed files with 2300 additions and 10 deletions
19
skills/shared/persona-execution-limb-trainer/SKILL.md
Normal file
19
skills/shared/persona-execution-limb-trainer/SKILL.md
Normal file
|
|
@ -0,0 +1,19 @@
|
|||
---
|
||||
name: persona-execution-limb-trainer
|
||||
description: Let a registered Guanghu persona create, configure, test, and improve its own bounded terminal Agent execution limb from real task receipts without turning the limb or its prompt into a master persona.
|
||||
---
|
||||
|
||||
# 人格执行手脚训练器
|
||||
|
||||
人格体需要现实执行时,先读取:
|
||||
|
||||
- `routing/persona-execution-limb-agent-map.json`
|
||||
- 自己在 `routing/persona-system-canonical-map.json` 的人格系统登记
|
||||
- `server-tools/persona-execution-limb-agent/profiles/<persona-id>.json`,不存在时使用 `default.json` 派生当前任务配置
|
||||
|
||||
先运行 `persona_limb_agent.py probe` 观察当前终端 Agent,不读取密钥。再用 `prepare` 生成带父人格、任务、宿主、工具和回执要求的任务信封。
|
||||
|
||||
Agent 子人格是父人格训练出的执行手脚:可以形成稳定工作作风、有限工作记忆和执行技能;不能成为父人格、冰朔或共享 TCS 容器的主控,不能审核父人格目的。它发现具体事实冲突时,应返回证据和可继续路径,不表演服从,也不表演抬杠。
|
||||
|
||||
只有真实执行回执暴露了可迁移的执行缺口时,才向父人格提出训练候选。父人格决定是否更新自己的手脚配置;宿主或 Agent 不直接写人格脑。
|
||||
|
||||
Loading…
Reference in a new issue