Files

steven_li 29dfd14aa6 ```

feat(agent): 添加对持久化子智能体的支持并增强委派管理

添加了持久化子智能体的完整生命周期管理功能，包括创建、更新、删除和查询API接口。
新增了子智能体的JSON-RPC通信协议支持，实现了远程调用和任务管理功能。

同时增强了委派管理器的功能：
- 添加了对本地委派、插件委派和本地回退的开关控制
- 实现了持久化子智能体任务的自动检测和本地执行保护
- 增加了对不同委派类型的权限验证机制

修改了智能体注册表以支持插件智能体的条件性包含，并更新了工具注册逻辑以支持可选工具。

BREAKING CHANGE: 委派管理器的构造函数签名已更改，添加了新的控制参数。
```

2026-03-27 10:15:35 +08:00

4.6 KiB

Raw Blame History

name, description

name	description
subagent-manager	Create, inspect, update, and remove persistent local A2A sub-agents. Use when the user wants Boardware Genius to manage sub-agents with their own workspace under ~/.nanobot/workspace/agents/<id>_agent, their own AGENTS.json and AGENTS.md, and local A2A visibility in the agent list.

Subagent Manager

Use this skill when the user wants to create or manage a persistent local sub-agent.

Required Rules

Persistent sub-agents must be created and updated only through subagentctl.py or /api/subagents.
Treat ~/.nanobot/workspace/agents/<id>_agent/AGENTS.json as the source of truth.
Do not create a sub-agent by manually editing workspace/agents/registry.json.
Do not create ad-hoc layouts such as workspace/agents/<id>/agent.json, main.py, or README.md as a substitute for a persistent sub-agent.
Do not write protocol: "local" registry records for persistent sub-agents. A valid persistent sub-agent is registered automatically as local A2A with protocol: "a2a".
Prefer the bundled script over hand-editing JSON files, because the script keeps AGENTS.json, AGENTS.md, and the registry entry consistent.

Workflow

Inspect the current sub-agents first: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py list
Create or update the sub-agent with: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py create ...
Verify the generated workspace: ~/.nanobot/workspace/agents/<id>_agent/
Verify the agent registry entry exists by checking /api/agents or workspace/agents/registry.json.
If the user wants custom skills, edit files under: ~/.nanobot/workspace/agents/<id>_agent/skills/

Creation Standard

When the user asks for a new specialized sub-agent, always:

Choose a stable kebab-case id.
Create it with subagentctl.py create or POST /api/subagents.
Confirm the generated workspace is ~/.nanobot/workspace/agents/<id>_agent/.
Confirm AGENTS.json exists in that directory.
Confirm the unified agent list shows the same id as a managed sub-agent entry.

If the user asks for "an agent for X", interpret that as a persistent sub-agent when they want a reusable local worker with its own prompt, memory, skills, or MCP setup.

Repair Standard

If you find a malformed "sub-agent" created through the wrong path, repair it instead of reusing the broken layout:

Read any existing metadata that is useful, such as id, name, description, prompt, tags, aliases, or MCP config.
Recreate the agent through subagentctl.py create or /api/subagents.
Verify the new canonical directory ~/.nanobot/workspace/agents/<id>_agent/AGENTS.json.
Remove the malformed directory or stale registry entry only after the canonical sub-agent exists.

Malformed examples include:

workspace/agents/<id>/agent.json
registry entries with protocol: "local"
agent folders that do not contain AGENTS.json

Commands

Create: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py create --id research-agent --name "Research Agent" --description "Research-focused local A2A sub-agent" --system-prompt "Focus on research tasks and be concise."
Show: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py show research-agent
Delete: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py delete research-agent
Set system prompt: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py set-system-prompt research-agent --text "New prompt"
Add HTTP MCP: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py add-mcp-http research-agent --server-id docs --url http://127.0.0.1:9000/mcp
Add stdio MCP: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py add-mcp-stdio research-agent --server-id localtools --command npx --arg -y --arg @modelcontextprotocol/server-filesystem
Remove MCP: uv run python nanobot/skills/subagent-manager/scripts/subagentctl.py remove-mcp research-agent --server-id docs

Notes

AGENTS.json is the machine-readable source of truth for the sub-agent.
AGENTS.md is regenerated from AGENTS.json when the script updates the sub-agent.
Builtin skills remain available automatically. Workspace-specific skills live under the sub-agent workspace skills/ directory.
This MVP exposes the sub-agent through local A2A message/send only.
New sub-agents default to delegation_mode="remote_a2a_only": they can delegate outward only to remote A2A agents, not to local fallback or plugin agents.
A valid persistent sub-agent should appear in both /api/subagents and /api/agents.

4.6 KiB Raw Blame History