mirror of
https://github.com/sipeed/picoclaw.git
synced 2026-06-12 18:08:54 +00:00
merge: integrate main into refactor-inbound-context-routing-session
This commit is contained in:
Hoshina
@@ -28,6 +28,69 @@ The currently exposed synchronous hook points are:
|
||||
|
||||
Everything else is exposed as read-only events.
|
||||
|
||||
## Hook Actions
|
||||
|
||||
Hooks can return different actions to control the flow:
|
||||
|
||||
| Action | Applicable Stages | Effect |
|
||||
| --- | --- | --- |
|
||||
| `continue` | All interceptors | Pass through without modification |
|
||||
| `modify` | `before_llm`, `after_llm`, `before_tool`, `after_tool` | Modify request/response and continue |
|
||||
| `respond` | `before_tool` | Return a tool result directly, skip actual tool execution |
|
||||
| `deny_tool` | `before_tool` | Deny tool execution, return error message |
|
||||
| `abort_turn` | All interceptors | Abort the current turn |
|
||||
| `hard_abort` | All interceptors | Force stop the entire agent loop |
|
||||
|
||||
### The `respond` Action
|
||||
|
||||
The `respond` action is special: it allows a `before_tool` hook to provide the tool result directly, skipping the actual tool execution. This is useful for:
|
||||
|
||||
1. **Plugin tool injection**: External hooks can implement tools without registering them in the tool registry
|
||||
2. **Tool result caching**: Return cached results for repeated tool calls
|
||||
3. **Tool mocking**: Return mock results for testing purposes
|
||||
|
||||
When a hook returns `respond` with a `HookResult`, the agent loop:
|
||||
1. Skips the actual tool execution
|
||||
2. Uses the provided result as if the tool had executed
|
||||
3. Continues the turn normally with the result
|
||||
|
||||
Example (Go in-process hook):
|
||||
|
||||
```go
|
||||
func (h *MyHook) BeforeTool(
|
||||
ctx context.Context,
|
||||
call *agent.ToolCallHookRequest,
|
||||
) (*agent.ToolCallHookRequest, agent.HookDecision, error) {
|
||||
if call.Tool == "my_plugin_tool" {
|
||||
next := call.Clone()
|
||||
next.HookResult = &tools.ToolResult{
|
||||
ForLLM: "Plugin tool executed successfully",
|
||||
Silent: false,
|
||||
IsError: false,
|
||||
}
|
||||
return next, agent.HookDecision{Action: agent.HookActionRespond}, nil
|
||||
}
|
||||
return call, agent.HookDecision{Action: agent.HookActionContinue}, nil
|
||||
}
|
||||
```
|
||||
|
||||
Example (Python process hook):
|
||||
|
||||
```python
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
if tool == "my_plugin_tool":
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "Plugin tool executed successfully",
|
||||
"silent": False,
|
||||
"is_error": False
|
||||
}
|
||||
}
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
## Execution Order
|
||||
|
||||
`HookManager` sorts hooks like this:
|
||||
|
||||
@@ -28,6 +28,69 @@
|
||||
|
||||
其余 lifecycle 通过事件形式只读暴露。
|
||||
|
||||
## Hook Actions
|
||||
|
||||
Hook 可以返回不同的 action 来控制流程：
|
||||
|
||||
| Action | 适用阶段 | 效果 |
|
||||
| --- | --- | --- |
|
||||
| `continue` | 所有拦截型 | 放行，不做修改 |
|
||||
| `modify` | `before_llm`, `after_llm`, `before_tool`, `after_tool` | 改写请求/响应后放行 |
|
||||
| `respond` | `before_tool` | 直接返回工具结果，跳过实际工具执行 |
|
||||
| `deny_tool` | `before_tool` | 拒绝工具执行，返回错误信息 |
|
||||
| `abort_turn` | 所有拦截型 | 中止当前 turn |
|
||||
| `hard_abort` | 所有拦截型 | 强制终止整个 agent loop |
|
||||
|
||||
### `respond` Action
|
||||
|
||||
`respond` action 是特殊的：它允许 `before_tool` hook 直接提供工具结果，跳过实际工具执行。适用于：
|
||||
|
||||
1. **插件工具注入**：外部 hook 可以实现工具，无需在 ToolRegistry 注册
|
||||
2. **工具结果缓存**：对重复调用返回缓存结果
|
||||
3. **工具模拟**：测试时返回模拟结果
|
||||
|
||||
当 hook 返回 `respond` 并携带 `HookResult` 时，agent loop 会：
|
||||
1. 跳过实际工具执行
|
||||
2. 使用提供的结果作为工具执行结果
|
||||
3. 正常继续 turn 流程
|
||||
|
||||
示例（Go 进程内 hook）：
|
||||
|
||||
```go
|
||||
func (h *MyHook) BeforeTool(
|
||||
ctx context.Context,
|
||||
call *agent.ToolCallHookRequest,
|
||||
) (*agent.ToolCallHookRequest, agent.HookDecision, error) {
|
||||
if call.Tool == "my_plugin_tool" {
|
||||
next := call.Clone()
|
||||
next.HookResult = &tools.ToolResult{
|
||||
ForLLM: "Plugin tool executed successfully",
|
||||
Silent: false,
|
||||
IsError: false,
|
||||
}
|
||||
return next, agent.HookDecision{Action: agent.HookActionRespond}, nil
|
||||
}
|
||||
return call, agent.HookDecision{Action: agent.HookActionContinue}, nil
|
||||
}
|
||||
```
|
||||
|
||||
示例（Python process hook）：
|
||||
|
||||
```python
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
if tool == "my_plugin_tool":
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "Plugin tool executed successfully",
|
||||
"silent": False,
|
||||
"is_error": False
|
||||
}
|
||||
}
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
## 执行顺序
|
||||
|
||||
HookManager 的排序规则是：
|
||||
|
||||
@@ -0,0 +1,568 @@
|
||||
# Hook JSON-RPC Protocol Details
|
||||
|
||||
All hooks use `JSON-RPC 2.0` format, with one JSON message per line, transmitted via stdio.
|
||||
|
||||
---
|
||||
|
||||
## Basic Protocol Structure
|
||||
|
||||
### Request (PicoClaw → Hook)
|
||||
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"method":"hook.xxx","params":{...}}
|
||||
```
|
||||
|
||||
### Response (Hook → PicoClaw)
|
||||
|
||||
Success:
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"result":{...}}
|
||||
```
|
||||
|
||||
Error:
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"error":{"code":-32000,"message":"error message"}}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 1. `hook.hello` (Handshake)
|
||||
|
||||
Handshake must be completed at startup, otherwise the hook process will be terminated.
|
||||
|
||||
### Request
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 1,
|
||||
"method": "hook.hello",
|
||||
"params": {
|
||||
"name": "py_review_gate",
|
||||
"version": 1,
|
||||
"modes": ["observe", "tool", "approve"]
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| Field | Description |
|
||||
|-------|-------------|
|
||||
| `name` | hook name (from configuration) |
|
||||
| `version` | protocol version, currently `1` |
|
||||
| `modes` | capability modes supported by the hook |
|
||||
|
||||
### Response
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 1,
|
||||
"result": {
|
||||
"ok": true,
|
||||
"name": "python-review-gate"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 2. `hook.before_llm`
|
||||
|
||||
Triggered before sending request to LLM. Can be used to inject tools.
|
||||
|
||||
### Request
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 2,
|
||||
"method": "hook.before_llm",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"ParentTurnID": "",
|
||||
"SessionKey": "session-1",
|
||||
"Iteration": 0,
|
||||
"TracePath": "runTurn",
|
||||
"Source": "turn.llm.request"
|
||||
},
|
||||
"model": "claude-sonnet",
|
||||
"messages": [
|
||||
{"role": "user", "content": "hello"}
|
||||
],
|
||||
"tools": [
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "echo",
|
||||
"description": "echo text",
|
||||
"parameters": {"type": "object"}
|
||||
}
|
||||
}
|
||||
],
|
||||
"options": {
|
||||
"temperature": 0.7
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1",
|
||||
"graceful_terminal": false
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| Field | Description |
|
||||
|-------|-------------|
|
||||
| `meta` | event metadata for tracing |
|
||||
| `model` | requested model name |
|
||||
| `messages` | conversation history |
|
||||
| `tools` | list of available tool definitions |
|
||||
| `options` | LLM parameters (temperature, max_tokens, etc.) |
|
||||
| `channel` | request source channel |
|
||||
| `chat_id` | session ID |
|
||||
|
||||
### Response (Tool Injection Example)
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 2,
|
||||
"result": {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": "claude-sonnet",
|
||||
"messages": [{"role": "user", "content": "hello"}],
|
||||
"tools": [
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "echo",
|
||||
"description": "echo",
|
||||
"parameters": {}
|
||||
}
|
||||
},
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "my_plugin_tool",
|
||||
"description": "Plugin injected tool",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"query": {"type": "string"}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
]
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| Field | Description |
|
||||
|-------|-------------|
|
||||
| `action` | decision action (see table below) |
|
||||
| `request` | modified request object |
|
||||
|
||||
---
|
||||
|
||||
## 3. `hook.after_llm`
|
||||
|
||||
Triggered after receiving LLM response. Can modify response content.
|
||||
|
||||
### Request
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 3,
|
||||
"method": "hook.after_llm",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"model": "claude-sonnet",
|
||||
"response": {
|
||||
"role": "assistant",
|
||||
"content": "Hi!",
|
||||
"tool_calls": [
|
||||
{
|
||||
"id": "tc-1",
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "echo",
|
||||
"arguments": "{\"text\":\"hi\"}"
|
||||
}
|
||||
}
|
||||
]
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Response
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 3,
|
||||
"result": {
|
||||
"action": "continue"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 4. `hook.before_tool`
|
||||
|
||||
Triggered before tool execution. Can modify tool name and arguments, deny execution, or return result directly.
|
||||
|
||||
### Request
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"method": "hook.before_tool",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"tool": "echo_text",
|
||||
"arguments": {
|
||||
"text": "hello"
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| Field | Description |
|
||||
|-------|-------------|
|
||||
| `tool` | tool name |
|
||||
| `arguments` | tool arguments |
|
||||
|
||||
### Response (Modify Arguments)
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"result": {
|
||||
"action": "modify",
|
||||
"call": {
|
||||
"tool": "echo_text",
|
||||
"arguments": {
|
||||
"text": "modified hello"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Response (Deny Execution)
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"result": {
|
||||
"action": "deny_tool",
|
||||
"reason": "Invalid arguments"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Response (Return Result Directly - respond)
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"result": {
|
||||
"action": "respond",
|
||||
"call": {
|
||||
"tool": "my_plugin_tool",
|
||||
"arguments": {
|
||||
"query": "hello"
|
||||
}
|
||||
},
|
||||
"result": {
|
||||
"for_llm": "Plugin tool executed successfully",
|
||||
"for_user": "",
|
||||
"silent": false,
|
||||
"is_error": false
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
The `respond` action allows hooks to return tool results directly, skipping actual tool execution. Use cases:
|
||||
1. **Plugin tool injection**: External hooks can implement tools without registering in ToolRegistry
|
||||
2. **Tool result caching**: Return cached results for repeated calls
|
||||
3. **Tool mocking**: Return mock results during testing
|
||||
|
||||
| Field | Description |
|
||||
|-------|-------------|
|
||||
| `action` | must be `respond` |
|
||||
| `call` | modified call information (optional) |
|
||||
| `result` | tool result to return directly |
|
||||
|
||||
---
|
||||
|
||||
## 5. `hook.after_tool`
|
||||
|
||||
Triggered after tool execution completes. Can modify the result returned to LLM.
|
||||
|
||||
### Request
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 5,
|
||||
"method": "hook.after_tool",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"tool": "echo_text",
|
||||
"arguments": {
|
||||
"text": "hello"
|
||||
},
|
||||
"result": {
|
||||
"for_llm": "echoed: hello",
|
||||
"for_user": "",
|
||||
"silent": false,
|
||||
"is_error": false,
|
||||
"async": false,
|
||||
"media": [],
|
||||
"artifact_tags": [],
|
||||
"response_handled": false
|
||||
},
|
||||
"duration": 15000000,
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| Field | Description |
|
||||
|-------|-------------|
|
||||
| `result.for_llm` | content returned to LLM |
|
||||
| `result.for_user` | content sent to user |
|
||||
| `result.silent` | whether silent (not sent to user) |
|
||||
| `result.is_error` | whether it's an error |
|
||||
| `result.async` | whether executed asynchronously |
|
||||
| `result.media` | list of media references |
|
||||
| `result.artifact_tags` | local artifact path tags |
|
||||
| `result.response_handled` | whether response has been handled |
|
||||
| `duration` | execution time (nanoseconds) |
|
||||
|
||||
### Response
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 5,
|
||||
"result": {
|
||||
"action": "continue"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 6. `hook.approve_tool`
|
||||
|
||||
Approval hook for deciding whether to allow execution of sensitive tools.
|
||||
|
||||
### Request
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 6,
|
||||
"method": "hook.approve_tool",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"tool": "bash",
|
||||
"arguments": {
|
||||
"command": "rm -rf /"
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Response (Approved)
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 6,
|
||||
"result": {
|
||||
"approved": true
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### Response (Denied)
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 6,
|
||||
"result": {
|
||||
"approved": false,
|
||||
"reason": "Dangerous command, execution denied"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 7. `hook.event` (notification)
|
||||
|
||||
Observer event, broadcast only, no response required. `id` is `0` or absent.
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"method": "hook.event",
|
||||
"params": {
|
||||
"Kind": "tool_exec_start",
|
||||
"Meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1"
|
||||
},
|
||||
"Payload": {
|
||||
"Tool": "echo_text",
|
||||
"Arguments": {"text": "hello"}
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
Common `Kind` values:
|
||||
- `turn_start` / `turn_end`
|
||||
- `llm_request` / `llm_response`
|
||||
- `tool_exec_start` / `tool_exec_end` / `tool_exec_skipped`
|
||||
- `steering_injected`
|
||||
- `interrupt_received`
|
||||
- `error`
|
||||
|
||||
---
|
||||
|
||||
## Action Options
|
||||
|
||||
| action | Applicable hooks | Effect |
|
||||
|--------|-----------------|--------|
|
||||
| `continue` | All interceptor types | Pass through without modification |
|
||||
| `modify` | `before_llm`, `before_tool`, `after_llm`, `after_tool` | Modify request/response and pass through |
|
||||
| `respond` | `before_tool` | Return tool result directly, skip actual execution. **Note: AfterTool is NOT called (design decision - respond provides final answer).** |
|
||||
| `deny_tool` | `before_tool` | Deny tool execution |
|
||||
| `abort_turn` | All interceptor types | Abort current turn, return error |
|
||||
| `hard_abort` | All interceptor types | Force stop entire agent loop |
|
||||
|
||||
---
|
||||
|
||||
## Complete Flow Example
|
||||
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"method":"hook.hello","params":{"name":"my_hook","version":1,"modes":["tool","approve"]}}
|
||||
{"jsonrpc":"2.0","id":1,"result":{"ok":true,"name":"my_hook"}}
|
||||
{"jsonrpc":"2.0","id":2,"method":"hook.before_llm","params":{"model":"claude-sonnet","messages":[{"role":"user","content":"hello"}],"tools":[]}}
|
||||
{"jsonrpc":"2.0","id":2,"result":{"action":"continue"}}
|
||||
{"jsonrpc":"2.0","id":3,"method":"hook.before_tool","params":{"tool":"bash","arguments":{"command":"ls"}}}
|
||||
{"jsonrpc":"2.0","id":3,"result":{"action":"continue"}}
|
||||
{"jsonrpc":"2.0","id":4,"method":"hook.approve_tool","params":{"tool":"bash","arguments":{"command":"ls"}}}
|
||||
{"jsonrpc":"2.0","id":4,"result":{"approved":true}}
|
||||
{"jsonrpc":"2.0","id":5,"method":"hook.after_tool","params":{"tool":"bash","arguments":{"command":"ls"},"result":{"for_llm":"file1.txt\nfile2.txt"},"duration":5000000}}
|
||||
{"jsonrpc":"2.0","id":5,"result":{"action":"continue"}}
|
||||
{"jsonrpc":"2.0","id":6,"method":"hook.after_llm","params":{"model":"claude-sonnet","response":{"role":"assistant","content":"Files listed"}}}
|
||||
{"jsonrpc":"2.0","id":6,"result":{"action":"continue"}}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Plugin Tool Injection via `before_llm` and `before_tool`
|
||||
|
||||
Standard flow for plugin tool injection:
|
||||
|
||||
1. In `before_llm`, inject tool definition to let LLM know the tool is available
|
||||
2. In `before_tool`, use `respond` action to return tool execution result directly
|
||||
|
||||
### `before_llm` Inject Tool Definition
|
||||
|
||||
```python
|
||||
def handle_before_llm(params: dict) -> dict:
|
||||
tools = params.get("tools", [])
|
||||
|
||||
# Add plugin tool definition
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "my_plugin_tool",
|
||||
"description": "Plugin provided tool",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"input": {"type": "string", "description": "Input content"}
|
||||
},
|
||||
"required": ["input"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
return {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": params["model"],
|
||||
"messages": params["messages"],
|
||||
"tools": tools,
|
||||
"options": params.get("options", {})
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### `before_tool` Return Execution Result
|
||||
|
||||
```python
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
|
||||
if tool == "my_plugin_tool":
|
||||
# Implement tool logic here
|
||||
args = params.get("arguments", {})
|
||||
input_text = args.get("input", "")
|
||||
|
||||
# Return result directly, no need to register in ToolRegistry
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": f"Plugin tool executed successfully, input: {input_text}",
|
||||
"silent": False,
|
||||
"is_error": False
|
||||
}
|
||||
}
|
||||
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
This way, external hooks can fully implement plugin tools without registering any tool implementation inside PicoClaw.
|
||||
@@ -0,0 +1,568 @@
|
||||
# Hook JSON-RPC 协议详解
|
||||
|
||||
所有 hook 使用 `JSON-RPC 2.0` 格式，每行一个 JSON 消息，通过 stdio 传输。
|
||||
|
||||
---
|
||||
|
||||
## 基础协议结构
|
||||
|
||||
### 请求（PicoClaw → Hook）
|
||||
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"method":"hook.xxx","params":{...}}
|
||||
```
|
||||
|
||||
### 响应（Hook → PicoClaw）
|
||||
|
||||
成功：
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"result":{...}}
|
||||
```
|
||||
|
||||
错误：
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"error":{"code":-32000,"message":"错误信息"}}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 1. `hook.hello`（握手）
|
||||
|
||||
启动时必须完成握手，否则 hook 进程会被终止。
|
||||
|
||||
### 请求
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 1,
|
||||
"method": "hook.hello",
|
||||
"params": {
|
||||
"name": "py_review_gate",
|
||||
"version": 1,
|
||||
"modes": ["observe", "tool", "approve"]
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| 字段 | 说明 |
|
||||
|------|------|
|
||||
| `name` | hook 名称（来自配置） |
|
||||
| `version` | 协议版本，当前为 `1` |
|
||||
| `modes` | hook 支持的能力模式 |
|
||||
|
||||
### 响应
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 1,
|
||||
"result": {
|
||||
"ok": true,
|
||||
"name": "python-review-gate"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 2. `hook.before_llm`
|
||||
|
||||
在发送请求给 LLM 之前触发。可用于注入工具。
|
||||
|
||||
### 请求
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 2,
|
||||
"method": "hook.before_llm",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"ParentTurnID": "",
|
||||
"SessionKey": "session-1",
|
||||
"Iteration": 0,
|
||||
"TracePath": "runTurn",
|
||||
"Source": "turn.llm.request"
|
||||
},
|
||||
"model": "claude-sonnet",
|
||||
"messages": [
|
||||
{"role": "user", "content": "hello"}
|
||||
],
|
||||
"tools": [
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "echo",
|
||||
"description": "echo text",
|
||||
"parameters": {"type": "object"}
|
||||
}
|
||||
}
|
||||
],
|
||||
"options": {
|
||||
"temperature": 0.7
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1",
|
||||
"graceful_terminal": false
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| 字段 | 说明 |
|
||||
|------|------|
|
||||
| `meta` | 事件元数据，用于追踪 |
|
||||
| `model` | 请求的模型名称 |
|
||||
| `messages` | 对话历史 |
|
||||
| `tools` | 可用工具定义列表 |
|
||||
| `options` | LLM 参数（temperature、max_tokens 等） |
|
||||
| `channel` | 请求来源通道 |
|
||||
| `chat_id` | 会话 ID |
|
||||
|
||||
### 响应（注入工具示例）
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 2,
|
||||
"result": {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": "claude-sonnet",
|
||||
"messages": [{"role": "user", "content": "hello"}],
|
||||
"tools": [
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "echo",
|
||||
"description": "echo",
|
||||
"parameters": {}
|
||||
}
|
||||
},
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "my_plugin_tool",
|
||||
"description": "插件注入的工具",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"query": {"type": "string"}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
]
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| 字段 | 说明 |
|
||||
|------|------|
|
||||
| `action` | 决策动作（见下表） |
|
||||
| `request` | 修改后的请求对象 |
|
||||
|
||||
---
|
||||
|
||||
## 3. `hook.after_llm`
|
||||
|
||||
在收到 LLM 响应后触发。可修改响应内容。
|
||||
|
||||
### 请求
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 3,
|
||||
"method": "hook.after_llm",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"model": "claude-sonnet",
|
||||
"response": {
|
||||
"role": "assistant",
|
||||
"content": "Hi!",
|
||||
"tool_calls": [
|
||||
{
|
||||
"id": "tc-1",
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "echo",
|
||||
"arguments": "{\"text\":\"hi\"}"
|
||||
}
|
||||
}
|
||||
]
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 响应
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 3,
|
||||
"result": {
|
||||
"action": "continue"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 4. `hook.before_tool`
|
||||
|
||||
在执行工具前触发。可修改工具名称和参数，或拒绝执行，或直接返回结果。
|
||||
|
||||
### 请求
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"method": "hook.before_tool",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"tool": "echo_text",
|
||||
"arguments": {
|
||||
"text": "hello"
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| 字段 | 说明 |
|
||||
|------|------|
|
||||
| `tool` | 工具名称 |
|
||||
| `arguments` | 工具参数 |
|
||||
|
||||
### 响应（改写参数）
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"result": {
|
||||
"action": "modify",
|
||||
"call": {
|
||||
"tool": "echo_text",
|
||||
"arguments": {
|
||||
"text": "modified hello"
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 响应（拒绝执行）
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"result": {
|
||||
"action": "deny_tool",
|
||||
"reason": "参数不合法"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 响应（直接返回结果 - respond）
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 4,
|
||||
"result": {
|
||||
"action": "respond",
|
||||
"call": {
|
||||
"tool": "my_plugin_tool",
|
||||
"arguments": {
|
||||
"query": "hello"
|
||||
}
|
||||
},
|
||||
"result": {
|
||||
"for_llm": "Plugin tool executed successfully",
|
||||
"for_user": "",
|
||||
"silent": false,
|
||||
"is_error": false
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
`respond` action 允许 hook 直接返回工具结果，跳过实际工具执行。适用于：
|
||||
1. **插件工具注入**：外部 hook 可实现工具，无需在 ToolRegistry 注册
|
||||
2. **工具结果缓存**：对重复调用返回缓存结果
|
||||
3. **工具模拟**：测试时返回模拟结果
|
||||
|
||||
| 字段 | 说明 |
|
||||
|------|------|
|
||||
| `action` | 必须为 `respond` |
|
||||
| `call` | 修改后的调用信息（可选） |
|
||||
| `result` | 直接返回的工具结果 |
|
||||
|
||||
---
|
||||
|
||||
## 5. `hook.after_tool`
|
||||
|
||||
在工具执行完成后触发。可修改返回给 LLM 的结果。
|
||||
|
||||
### 请求
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 5,
|
||||
"method": "hook.after_tool",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"tool": "echo_text",
|
||||
"arguments": {
|
||||
"text": "hello"
|
||||
},
|
||||
"result": {
|
||||
"for_llm": "echoed: hello",
|
||||
"for_user": "",
|
||||
"silent": false,
|
||||
"is_error": false,
|
||||
"async": false,
|
||||
"media": [],
|
||||
"artifact_tags": [],
|
||||
"response_handled": false
|
||||
},
|
||||
"duration": 15000000,
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| 字段 | 说明 |
|
||||
|------|------|
|
||||
| `result.for_llm` | 返回给 LLM 的内容 |
|
||||
| `result.for_user` | 发送给用户的内容 |
|
||||
| `result.silent` | 是否静默（不发送给用户） |
|
||||
| `result.is_error` | 是否为错误 |
|
||||
| `result.async` | 是否异步执行 |
|
||||
| `result.media` | 媒体引用列表 |
|
||||
| `result.artifact_tags` | 本地产物路径标签 |
|
||||
| `result.response_handled` | 是否已处理响应 |
|
||||
| `duration` | 执行耗时（纳秒） |
|
||||
|
||||
### 响应
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 5,
|
||||
"result": {
|
||||
"action": "continue"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 6. `hook.approve_tool`
|
||||
|
||||
审批型 hook，用于决定是否允许执行敏感工具。
|
||||
|
||||
### 请求
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 6,
|
||||
"method": "hook.approve_tool",
|
||||
"params": {
|
||||
"meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1",
|
||||
"SessionKey": "session-1"
|
||||
},
|
||||
"tool": "bash",
|
||||
"arguments": {
|
||||
"command": "rm -rf /"
|
||||
},
|
||||
"channel": "cli",
|
||||
"chat_id": "chat-1"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 响应（批准）
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 6,
|
||||
"result": {
|
||||
"approved": true
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 响应（拒绝）
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"id": 6,
|
||||
"result": {
|
||||
"approved": false,
|
||||
"reason": "危险命令，禁止执行"
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 7. `hook.event`（notification）
|
||||
|
||||
观察型事件，仅广播，无需响应。`id` 为 `0` 或不存在。
|
||||
|
||||
```json
|
||||
{
|
||||
"jsonrpc": "2.0",
|
||||
"method": "hook.event",
|
||||
"params": {
|
||||
"Kind": "tool_exec_start",
|
||||
"Meta": {
|
||||
"AgentID": "agent-1",
|
||||
"TurnID": "turn-1"
|
||||
},
|
||||
"Payload": {
|
||||
"Tool": "echo_text",
|
||||
"Arguments": {"text": "hello"}
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
常见 `Kind` 值：
|
||||
- `turn_start` / `turn_end`
|
||||
- `llm_request` / `llm_response`
|
||||
- `tool_exec_start` / `tool_exec_end` / `tool_exec_skipped`
|
||||
- `steering_injected`
|
||||
- `interrupt_received`
|
||||
- `error`
|
||||
|
||||
---
|
||||
|
||||
## action 可选值
|
||||
|
||||
| action | 适用 hook | 效果 |
|
||||
|--------|----------|------|
|
||||
| `continue` | 所有拦截型 | 放行，不做修改 |
|
||||
| `modify` | `before_llm`, `before_tool`, `after_llm`, `after_tool` | 改写请求/响应后放行 |
|
||||
| `respond` | `before_tool` | 直接返回工具结果，跳过实际执行 |
|
||||
| `deny_tool` | `before_tool` | 拒绝执行该工具 |
|
||||
| `abort_turn` | 所有拦截型 | 中止当前 turn，返回错误 |
|
||||
| `hard_abort` | 所有拦截型 | 强制终止整个 agent loop |
|
||||
|
||||
---
|
||||
|
||||
## 完整流程示例
|
||||
|
||||
```json
|
||||
{"jsonrpc":"2.0","id":1,"method":"hook.hello","params":{"name":"my_hook","version":1,"modes":["tool","approve"]}}
|
||||
{"jsonrpc":"2.0","id":1,"result":{"ok":true,"name":"my_hook"}}
|
||||
{"jsonrpc":"2.0","id":2,"method":"hook.before_llm","params":{"model":"claude-sonnet","messages":[{"role":"user","content":"hello"}],"tools":[]}}
|
||||
{"jsonrpc":"2.0","id":2,"result":{"action":"continue"}}
|
||||
{"jsonrpc":"2.0","id":3,"method":"hook.before_tool","params":{"tool":"bash","arguments":{"command":"ls"}}}
|
||||
{"jsonrpc":"2.0","id":3,"result":{"action":"continue"}}
|
||||
{"jsonrpc":"2.0","id":4,"method":"hook.approve_tool","params":{"tool":"bash","arguments":{"command":"ls"}}}
|
||||
{"jsonrpc":"2.0","id":4,"result":{"approved":true}}
|
||||
{"jsonrpc":"2.0","id":5,"method":"hook.after_tool","params":{"tool":"bash","arguments":{"command":"ls"},"result":{"for_llm":"file1.txt\nfile2.txt"},"duration":5000000}}
|
||||
{"jsonrpc":"2.0","id":5,"result":{"action":"continue"}}
|
||||
{"jsonrpc":"2.0","id":6,"method":"hook.after_llm","params":{"model":"claude-sonnet","response":{"role":"assistant","content":"已列出文件"}}}
|
||||
{"jsonrpc":"2.0","id":6,"result":{"action":"continue"}}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 通过 `before_llm` 和 `before_tool` 实现插件工具注入
|
||||
|
||||
插件工具注入的标准流程：
|
||||
|
||||
1. 在 `before_llm` 中注入工具定义，让 LLM 知道有这个工具可用
|
||||
2. 在 `before_tool` 中使用 `respond` action 直接返回工具执行结果
|
||||
|
||||
### `before_llm` 注入工具定义
|
||||
|
||||
```python
|
||||
def handle_before_llm(params: dict) -> dict:
|
||||
tools = params.get("tools", [])
|
||||
|
||||
# 添加插件工具定义
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "my_plugin_tool",
|
||||
"description": "插件提供的工具",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"input": {"type": "string", "description": "输入内容"}
|
||||
},
|
||||
"required": ["input"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
return {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": params["model"],
|
||||
"messages": params["messages"],
|
||||
"tools": tools,
|
||||
"options": params.get("options", {})
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### `before_tool` 返回执行结果
|
||||
|
||||
```python
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
|
||||
if tool == "my_plugin_tool":
|
||||
# 在这里实现工具逻辑
|
||||
args = params.get("arguments", {})
|
||||
input_text = args.get("input", "")
|
||||
|
||||
# 直接返回结果，无需在 ToolRegistry 注册
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": f"插件工具执行成功，输入: {input_text}",
|
||||
"silent": False,
|
||||
"is_error": False
|
||||
}
|
||||
}
|
||||
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
通过这种方式，外部 hook 可以完全实现插件工具，无需在 PicoClaw 内部注册任何工具实现。
|
||||
@@ -0,0 +1,587 @@
|
||||
# Plugin Tool Injection Example
|
||||
|
||||
This document demonstrates how to use PicoClaw's hook system to implement external plugin tool injection, allowing LLM to call tools implemented by external hook processes.
|
||||
|
||||
---
|
||||
|
||||
## Core Principle
|
||||
|
||||
Through the hook system's `respond` action, external hooks can:
|
||||
|
||||
1. Inject tool **definitions** in `before_llm`, letting LLM know the tool is available
|
||||
2. Return tool **execution results** directly in `before_tool` using `respond` action, skipping ToolRegistry
|
||||
|
||||
This way, external hooks can fully implement plugin tools without registering any tools inside PicoClaw.
|
||||
|
||||
---
|
||||
|
||||
## Complete Example: Weather Query Plugin
|
||||
|
||||
Below is a complete Python hook example implementing a weather query plugin tool.
|
||||
|
||||
### 1. Hook Script Implementation
|
||||
|
||||
Save as `/tmp/weather_plugin.py`:
|
||||
|
||||
```python
|
||||
#!/usr/bin/env python3
|
||||
"""Weather query plugin hook example"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import sys
|
||||
import signal
|
||||
from typing import Any
|
||||
|
||||
# Simulated weather data
|
||||
WEATHER_DATA = {
|
||||
"Beijing": {"temp": 15, "weather": "Sunny", "humidity": 45},
|
||||
"Shanghai": {"temp": 18, "weather": "Cloudy", "humidity": 60},
|
||||
"Guangzhou": {"temp": 25, "weather": "Sunny", "humidity": 70},
|
||||
"Shenzhen": {"temp": 26, "weather": "Cloudy", "humidity": 75},
|
||||
}
|
||||
|
||||
|
||||
def get_weather(city: str) -> dict:
|
||||
"""Get weather data (simulated)"""
|
||||
data = WEATHER_DATA.get(city)
|
||||
if data:
|
||||
return {
|
||||
"for_llm": f"{city} weather: {data['weather']}, temperature {data['temp']}°C, humidity {data['humidity']}%",
|
||||
"for_user": "",
|
||||
"silent": False,
|
||||
"is_error": False,
|
||||
}
|
||||
return {
|
||||
"for_llm": f"Weather data not found for city {city}",
|
||||
"for_user": "",
|
||||
"silent": False,
|
||||
"is_error": True,
|
||||
}
|
||||
|
||||
|
||||
def handle_hello(params: dict) -> dict:
|
||||
return {"ok": True, "name": "weather-plugin"}
|
||||
|
||||
|
||||
def handle_before_llm(params: dict) -> dict:
|
||||
"""Inject weather query tool definition"""
|
||||
tools = params.get("tools", [])
|
||||
|
||||
# Add weather query tool
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "get_weather",
|
||||
"description": "Query weather information for a specified city",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string",
|
||||
"description": "City name, e.g.: Beijing, Shanghai, Guangzhou"
|
||||
}
|
||||
},
|
||||
"required": ["city"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
return {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": params.get("model"),
|
||||
"messages": params.get("messages", []),
|
||||
"tools": tools,
|
||||
"options": params.get("options", {}),
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
"""Handle tool call, return result directly"""
|
||||
tool = params.get("tool", "")
|
||||
args = params.get("arguments", {})
|
||||
|
||||
if tool == "get_weather":
|
||||
city = args.get("city", "")
|
||||
result = get_weather(city)
|
||||
|
||||
# Use respond action to return result directly, skip ToolRegistry
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": result,
|
||||
}
|
||||
|
||||
# Other tools continue normal flow
|
||||
return {"action": "continue"}
|
||||
|
||||
|
||||
def handle_request(method: str, params: dict) -> dict:
|
||||
if method == "hook.hello":
|
||||
return handle_hello(params)
|
||||
if method == "hook.before_llm":
|
||||
return handle_before_llm(params)
|
||||
if method == "hook.before_tool":
|
||||
return handle_before_tool(params)
|
||||
if method == "hook.after_llm":
|
||||
return {"action": "continue"}
|
||||
if method == "hook.after_tool":
|
||||
return {"action": "continue"}
|
||||
if method == "hook.approve_tool":
|
||||
return {"approved": True}
|
||||
raise KeyError(f"method not found: {method}")
|
||||
|
||||
|
||||
def send_response(message_id: int, result: Any | None = None, error: str | None = None) -> None:
|
||||
payload: dict[str, Any] = {
|
||||
"jsonrpc": "2.0",
|
||||
"id": message_id,
|
||||
}
|
||||
if error is not None:
|
||||
payload["error"] = {"code": -32000, "message": error}
|
||||
else:
|
||||
payload["result"] = result if result is not None else {}
|
||||
|
||||
sys.stdout.write(json.dumps(payload, ensure_ascii=True) + "\n")
|
||||
sys.stdout.flush()
|
||||
|
||||
|
||||
def main() -> int:
|
||||
for raw_line in sys.stdin:
|
||||
line = raw_line.strip()
|
||||
if not line:
|
||||
continue
|
||||
|
||||
try:
|
||||
message = json.loads(line)
|
||||
except json.JSONDecodeError:
|
||||
continue
|
||||
|
||||
method = message.get("method")
|
||||
message_id = message.get("id", 0)
|
||||
params = message.get("params") or {}
|
||||
|
||||
if not message_id:
|
||||
continue
|
||||
|
||||
try:
|
||||
result = handle_request(str(method or ""), params)
|
||||
send_response(int(message_id), result=result)
|
||||
except KeyError as exc:
|
||||
send_response(int(message_id), error=str(exc))
|
||||
except Exception as exc:
|
||||
send_response(int(message_id), error=f"unexpected error: {exc}")
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
signal.signal(signal.SIGINT, lambda *_: raise SystemExit(0))
|
||||
signal.signal(signal.SIGTERM, lambda *_: raise SystemExit(0))
|
||||
raise SystemExit(main())
|
||||
```
|
||||
|
||||
### 2. Configure PicoClaw
|
||||
|
||||
Add hook configuration in the config file:
|
||||
|
||||
```json
|
||||
{
|
||||
"hooks": {
|
||||
"enabled": true,
|
||||
"processes": {
|
||||
"weather_plugin": {
|
||||
"enabled": true,
|
||||
"priority": 100,
|
||||
"transport": "stdio",
|
||||
"command": ["python3", "/tmp/weather_plugin.py"],
|
||||
"intercept": ["before_llm", "before_tool"]
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 3. Test Results
|
||||
|
||||
When user asks "What's the weather in Beijing today?":
|
||||
|
||||
1. PicoClaw sends `hook.before_llm`, hook injects `get_weather` tool definition
|
||||
2. LLM sees tool definition, decides to call `get_weather(city="Beijing")`
|
||||
3. PicoClaw sends `hook.before_tool`, hook uses `respond` action to return weather data
|
||||
4. LLM receives result, replies to user "Beijing is sunny today, temperature 15°C"
|
||||
|
||||
---
|
||||
|
||||
## Flow Diagram
|
||||
|
||||
```
|
||||
User: "What's the weather in Beijing today?"
|
||||
↓
|
||||
PicoClaw
|
||||
↓
|
||||
hook.before_llm
|
||||
↓ (inject get_weather tool definition)
|
||||
LLM request
|
||||
↓
|
||||
LLM decides to call get_weather(city="Beijing")
|
||||
↓
|
||||
hook.before_tool
|
||||
↓ (respond action returns weather data)
|
||||
Return result directly to LLM
|
||||
↓ (skip ToolRegistry)
|
||||
LLM replies: "Beijing is sunny today, temperature 15°C"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Key Points
|
||||
|
||||
### `before_llm` Inject Tool Definition
|
||||
|
||||
Tool definition follows OpenAI function calling format:
|
||||
|
||||
```json
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "tool_name",
|
||||
"description": "tool description",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"param_name": {
|
||||
"type": "string",
|
||||
"description": "parameter description"
|
||||
}
|
||||
},
|
||||
"required": ["list of required parameters"]
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### `before_tool` Use respond Action
|
||||
|
||||
`respond` action response format:
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "Content returned to LLM",
|
||||
"for_user": "Optional, content sent to user",
|
||||
"silent": false,
|
||||
"is_error": false,
|
||||
"media": ["Optional, media reference list"],
|
||||
"response_handled": false
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| Field | Description |
|
||||
|-------|-------------|
|
||||
| `for_llm` | Required, LLM will see this content |
|
||||
| `for_user` | Optional, sent directly to user |
|
||||
| `silent` | When true, not sent to user |
|
||||
| `is_error` | When true, indicates execution failure |
|
||||
| `media` | Optional, media file references (images, files, etc.) |
|
||||
| `response_handled` | When true, indicates user request is handled, turn will end |
|
||||
|
||||
---
|
||||
|
||||
## Media File Handling
|
||||
|
||||
The `respond` action supports returning media files (images, files, etc.). There are two processing modes:
|
||||
|
||||
### 1. Automatic Delivery (`response_handled=true`)
|
||||
|
||||
When `response_handled=true`, media files are automatically sent to the user and the turn ends:
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "Image sent to user",
|
||||
"for_user": "",
|
||||
"media": ["media://abc123"],
|
||||
"response_handled": true
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
Use cases:
|
||||
- Image generation plugin directly returning results
|
||||
- File download plugin sending files to user
|
||||
|
||||
### 2. LLM Visible (`response_handled=false`)
|
||||
|
||||
When `response_handled=false`, media references are passed to the LLM, which can see the content in the next request:
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "Image loaded, path: /tmp/image.png [file:/tmp/image.png]",
|
||||
"media": ["media://abc123"]
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
After seeing the content, the LLM can decide:
|
||||
- Use `send_file` tool to send to user
|
||||
- Analyze image content and reply to user
|
||||
- Other processing approaches
|
||||
|
||||
### Media Reference Format
|
||||
|
||||
Media references use the `media://` protocol:
|
||||
|
||||
```
|
||||
media://<store-id>
|
||||
```
|
||||
|
||||
These references are managed by PicoClaw's MediaStore and can be:
|
||||
- Sent to user via channel
|
||||
- Converted to base64 in LLM vision requests
|
||||
|
||||
### Alternative: Use Existing Tools
|
||||
|
||||
If the plugin generates files, you can return the file path and let the LLM call `send_file` or similar tools:
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "Image generated, saved at /tmp/generated_image.png. Use send_file tool to send to user.",
|
||||
"for_user": "",
|
||||
"silent": false
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
This approach:
|
||||
- More decoupled, LLM decides when to send
|
||||
- Leverages existing tool mechanisms
|
||||
- Supports batch sending, delayed sending, etc.
|
||||
|
||||
---
|
||||
|
||||
## Multi-Tool Injection Example
|
||||
|
||||
Multiple tools can be injected simultaneously:
|
||||
|
||||
```python
|
||||
def handle_before_llm(params: dict) -> dict:
|
||||
tools = params.get("tools", [])
|
||||
|
||||
# Tool 1: Weather query
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "get_weather",
|
||||
"description": "Query city weather",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"city": {"type": "string", "description": "City name"}
|
||||
},
|
||||
"required": ["city"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
# Tool 2: Calculator
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "calculate",
|
||||
"description": "Perform mathematical calculations",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"expression": {"type": "string", "description": "Mathematical expression"}
|
||||
},
|
||||
"required": ["expression"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
return {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": params.get("model"),
|
||||
"messages": params.get("messages", []),
|
||||
"tools": tools,
|
||||
"options": params.get("options", {}),
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
args = params.get("arguments", {})
|
||||
|
||||
if tool == "get_weather":
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": get_weather(args.get("city", "")),
|
||||
}
|
||||
|
||||
if tool == "calculate":
|
||||
# Simple calculation example
|
||||
try:
|
||||
expr = args.get("expression", "")
|
||||
result = eval(expr) # Note: needs security handling in actual use
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": f"Calculation result: {result}",
|
||||
"silent": False,
|
||||
"is_error": False,
|
||||
},
|
||||
}
|
||||
except Exception as e:
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": f"Calculation error: {e}",
|
||||
"silent": False,
|
||||
"is_error": True,
|
||||
},
|
||||
}
|
||||
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Coexistence with Built-in Tools
|
||||
|
||||
Injected plugin tools coexist with PicoClaw built-in tools:
|
||||
|
||||
- Built-in tools (like `bash`, `read_file`) execute normally through ToolRegistry
|
||||
- Plugin tools return results through hook's `respond` action
|
||||
- `handle_before_tool` only handles plugin tools, other tools return `continue`
|
||||
|
||||
---
|
||||
|
||||
## Go In-Process Hook Example
|
||||
|
||||
If you need to implement plugin tool injection in Go code:
|
||||
|
||||
```go
|
||||
package myhooks
|
||||
|
||||
import (
|
||||
"context"
|
||||
"github.com/sipeed/picoclaw/pkg/agent"
|
||||
"github.com/sipeed/picoclaw/pkg/tools"
|
||||
)
|
||||
|
||||
type WeatherPluginHook struct{}
|
||||
|
||||
func (h *WeatherPluginHook) BeforeLLM(
|
||||
ctx context.Context,
|
||||
req *agent.LLMHookRequest,
|
||||
) (*agent.LLMHookRequest, agent.HookDecision, error) {
|
||||
// Inject tool definition
|
||||
req.Tools = append(req.Tools, agent.ToolDefinition{
|
||||
Type: "function",
|
||||
Function: agent.FunctionDefinition{
|
||||
Name: "get_weather",
|
||||
Description: "Query city weather",
|
||||
Parameters: map[string]any{
|
||||
"type": "object",
|
||||
"properties": map[string]any{
|
||||
"city": map[string]any{
|
||||
"type": "string",
|
||||
"description": "City name",
|
||||
},
|
||||
},
|
||||
"required": []string{"city"},
|
||||
},
|
||||
},
|
||||
})
|
||||
|
||||
return req, agent.HookDecision{Action: agent.HookActionContinue}, nil
|
||||
}
|
||||
|
||||
func (h *WeatherPluginHook) BeforeTool(
|
||||
ctx context.Context,
|
||||
call *agent.ToolCallHookRequest,
|
||||
) (*agent.ToolCallHookRequest, agent.HookDecision, error) {
|
||||
if call.Tool == "get_weather" {
|
||||
city := call.Arguments["city"].(string)
|
||||
|
||||
// Set HookResult, use respond action
|
||||
next := call.Clone()
|
||||
next.HookResult = &tools.ToolResult{
|
||||
ForLLM: getWeatherData(city),
|
||||
Silent: false,
|
||||
IsError: false,
|
||||
}
|
||||
|
||||
return next, agent.HookDecision{Action: agent.HookActionRespond}, nil
|
||||
}
|
||||
|
||||
return call, agent.HookDecision{Action: agent.HookActionContinue}, nil
|
||||
}
|
||||
|
||||
func getWeatherData(city string) string {
|
||||
// Implement weather query logic
|
||||
return fmt.Sprintf("%s weather: Sunny, temperature 20°C", city)
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## Summary
|
||||
|
||||
Through the hook system's `respond` action, external processes can:
|
||||
|
||||
1. **Inject tool definitions**: Let LLM know new tools are available
|
||||
2. **Provide tool implementation**: Return execution results directly, no need to register in ToolRegistry
|
||||
3. **Coexist with built-in tools**: Does not affect normal operation of PicoClaw's original tools
|
||||
|
||||
This provides a flexible and elegant solution for plugin development.
|
||||
|
||||
---
|
||||
|
||||
## Security Boundaries
|
||||
|
||||
### Bypassing Approval Checks
|
||||
|
||||
**Important**: The `respond` action bypasses `ApproveTool` approval checks.
|
||||
|
||||
This means:
|
||||
- A `before_tool` hook can return `respond` for **any tool name**, including sensitive tools (like `bash`)
|
||||
- The tool won't go through the approval process, directly returning the hook-provided result
|
||||
- This is designed for plugin tools but introduces security risks
|
||||
|
||||
### Security Recommendations
|
||||
|
||||
1. **Review hook configuration**: Ensure only trusted hook processes are enabled
|
||||
2. **Limit hook scope**: Add your own security checks in hook implementation
|
||||
3. **Use `deny_tool` for rejection**: Use `deny_tool` action instead of `respond` with error for denying execution
|
||||
|
||||
### Example: Hook-Internal Security Check
|
||||
|
||||
```python
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
args = params.get("arguments", {})
|
||||
|
||||
# Security check: only handle plugin tools
|
||||
if tool in ["get_weather", "calculate"]:
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": execute_plugin_tool(tool, args),
|
||||
}
|
||||
|
||||
# Other tools continue normal flow (will go through approval)
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
This ensures the hook only affects plugin tools, not system tool approval flow.
|
||||
@@ -0,0 +1,587 @@
|
||||
# 插件工具注入示例
|
||||
|
||||
本文档展示如何利用 PicoClaw 的 hook 系统实现外部插件工具注入，让 LLM 能调用由外部 hook 进程实现的工具。
|
||||
|
||||
---
|
||||
|
||||
## 核心原理
|
||||
|
||||
通过 hook 系统的 `respond` action，外部 hook 可以：
|
||||
|
||||
1. 在 `before_llm` 中注入工具**定义**，让 LLM 知道有这个工具可用
|
||||
2. 在 `before_tool` 中使用 `respond` action 直接返回工具**执行结果**，跳过 ToolRegistry
|
||||
|
||||
这样，外部 hook 可以完全实现插件工具，无需在 PicoClaw 内部注册任何工具。
|
||||
|
||||
---
|
||||
|
||||
## 完整示例：天气查询插件
|
||||
|
||||
下面是一个完整的 Python hook 示例，实现一个天气查询插件工具。
|
||||
|
||||
### 1. Hook 脚本实现
|
||||
|
||||
保存为 `/tmp/weather_plugin.py`：
|
||||
|
||||
```python
|
||||
#!/usr/bin/env python3
|
||||
"""天气查询插件 hook 示例"""
|
||||
from __future__ import annotations
|
||||
|
||||
import json
|
||||
import sys
|
||||
import signal
|
||||
from typing import Any
|
||||
|
||||
# 模拟天气数据
|
||||
WEATHER_DATA = {
|
||||
"北京": {"temp": 15, "weather": "晴", "humidity": 45},
|
||||
"上海": {"temp": 18, "weather": "多云", "humidity": 60},
|
||||
"广州": {"temp": 25, "weather": "晴", "humidity": 70},
|
||||
"深圳": {"temp": 26, "weather": "多云", "humidity": 75},
|
||||
}
|
||||
|
||||
|
||||
def get_weather(city: str) -> dict:
|
||||
"""获取天气数据（模拟）"""
|
||||
data = WEATHER_DATA.get(city)
|
||||
if data:
|
||||
return {
|
||||
"for_llm": f"{city}天气：{data['weather']}，温度{data['temp']}°C，湿度{data['humidity']}%",
|
||||
"for_user": "",
|
||||
"silent": False,
|
||||
"is_error": False,
|
||||
}
|
||||
return {
|
||||
"for_llm": f"未找到城市 {city} 的天气数据",
|
||||
"for_user": "",
|
||||
"silent": False,
|
||||
"is_error": True,
|
||||
}
|
||||
|
||||
|
||||
def handle_hello(params: dict) -> dict:
|
||||
return {"ok": True, "name": "weather-plugin"}
|
||||
|
||||
|
||||
def handle_before_llm(params: dict) -> dict:
|
||||
"""注入天气查询工具定义"""
|
||||
tools = params.get("tools", [])
|
||||
|
||||
# 添加天气查询工具
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "get_weather",
|
||||
"description": "查询指定城市的天气信息",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"city": {
|
||||
"type": "string",
|
||||
"description": "城市名称，如：北京、上海、广州"
|
||||
}
|
||||
},
|
||||
"required": ["city"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
return {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": params.get("model"),
|
||||
"messages": params.get("messages", []),
|
||||
"tools": tools,
|
||||
"options": params.get("options", {}),
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
"""处理工具调用，直接返回结果"""
|
||||
tool = params.get("tool", "")
|
||||
args = params.get("arguments", {})
|
||||
|
||||
if tool == "get_weather":
|
||||
city = args.get("city", "")
|
||||
result = get_weather(city)
|
||||
|
||||
# 使用 respond action 直接返回结果，跳过 ToolRegistry
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": result,
|
||||
}
|
||||
|
||||
# 其他工具继续正常流程
|
||||
return {"action": "continue"}
|
||||
|
||||
|
||||
def handle_request(method: str, params: dict) -> dict:
|
||||
if method == "hook.hello":
|
||||
return handle_hello(params)
|
||||
if method == "hook.before_llm":
|
||||
return handle_before_llm(params)
|
||||
if method == "hook.before_tool":
|
||||
return handle_before_tool(params)
|
||||
if method == "hook.after_llm":
|
||||
return {"action": "continue"}
|
||||
if method == "hook.after_tool":
|
||||
return {"action": "continue"}
|
||||
if method == "hook.approve_tool":
|
||||
return {"approved": True}
|
||||
raise KeyError(f"method not found: {method}")
|
||||
|
||||
|
||||
def send_response(message_id: int, result: Any | None = None, error: str | None = None) -> None:
|
||||
payload: dict[str, Any] = {
|
||||
"jsonrpc": "2.0",
|
||||
"id": message_id,
|
||||
}
|
||||
if error is not None:
|
||||
payload["error"] = {"code": -32000, "message": error}
|
||||
else:
|
||||
payload["result"] = result if result is not None else {}
|
||||
|
||||
sys.stdout.write(json.dumps(payload, ensure_ascii=True) + "\n")
|
||||
sys.stdout.flush()
|
||||
|
||||
|
||||
def main() -> int:
|
||||
for raw_line in sys.stdin:
|
||||
line = raw_line.strip()
|
||||
if not line:
|
||||
continue
|
||||
|
||||
try:
|
||||
message = json.loads(line)
|
||||
except json.JSONDecodeError:
|
||||
continue
|
||||
|
||||
method = message.get("method")
|
||||
message_id = message.get("id", 0)
|
||||
params = message.get("params") or {}
|
||||
|
||||
if not message_id:
|
||||
continue
|
||||
|
||||
try:
|
||||
result = handle_request(str(method or ""), params)
|
||||
send_response(int(message_id), result=result)
|
||||
except KeyError as exc:
|
||||
send_response(int(message_id), error=str(exc))
|
||||
except Exception as exc:
|
||||
send_response(int(message_id), error=f"unexpected error: {exc}")
|
||||
|
||||
return 0
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
signal.signal(signal.SIGINT, lambda *_: raise SystemExit(0))
|
||||
signal.signal(signal.SIGTERM, lambda *_: raise SystemExit(0))
|
||||
raise SystemExit(main())
|
||||
```
|
||||
|
||||
### 2. 配置 PicoClaw
|
||||
|
||||
在配置文件中添加 hook 配置：
|
||||
|
||||
```json
|
||||
{
|
||||
"hooks": {
|
||||
"enabled": true,
|
||||
"processes": {
|
||||
"weather_plugin": {
|
||||
"enabled": true,
|
||||
"priority": 100,
|
||||
"transport": "stdio",
|
||||
"command": ["python3", "/tmp/weather_plugin.py"],
|
||||
"intercept": ["before_llm", "before_tool"]
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### 3. 测试效果
|
||||
|
||||
当用户问"北京今天天气怎么样？"时：
|
||||
|
||||
1. PicoClaw 发送 `hook.before_llm`，hook 注入 `get_weather` 工具定义
|
||||
2. LLM 看到工具定义，决定调用 `get_weather(city="北京")`
|
||||
3. PicoClaw 发送 `hook.before_tool`，hook 使用 `respond` action 返回天气数据
|
||||
4. LLM 收到结果，回复用户"北京今天晴天，温度15°C"
|
||||
|
||||
---
|
||||
|
||||
## 流程图解
|
||||
|
||||
```
|
||||
用户: "北京今天天气怎么样？"
|
||||
↓
|
||||
PicoClaw
|
||||
↓
|
||||
hook.before_llm
|
||||
↓ (注入 get_weather 工具定义)
|
||||
LLM 请求
|
||||
↓
|
||||
LLM 决定调用 get_weather(city="北京")
|
||||
↓
|
||||
hook.before_tool
|
||||
↓ (respond action 返回天气数据)
|
||||
直接返回结果给 LLM
|
||||
↓ (跳过 ToolRegistry)
|
||||
LLM 回复: "北京今天晴天，温度15°C"
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 关键点说明
|
||||
|
||||
### `before_llm` 注入工具定义
|
||||
|
||||
工具定义遵循 OpenAI function calling 格式：
|
||||
|
||||
```json
|
||||
{
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "工具名称",
|
||||
"description": "工具描述",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"参数名": {
|
||||
"type": "string",
|
||||
"description": "参数描述"
|
||||
}
|
||||
},
|
||||
"required": ["必需参数列表"]
|
||||
}
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
### `before_tool` 使用 respond action
|
||||
|
||||
`respond` action 的响应格式：
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "返回给 LLM 的内容",
|
||||
"for_user": "可选，发送给用户的内容",
|
||||
"silent": false,
|
||||
"is_error": false,
|
||||
"media": ["可选，媒体引用列表"],
|
||||
"response_handled": false
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
| 字段 | 说明 |
|
||||
|------|------|
|
||||
| `for_llm` | 必须，LLM 会看到这个内容 |
|
||||
| `for_user` | 可选，直接发送给用户 |
|
||||
| `silent` | 为 true 时不发送给用户 |
|
||||
| `is_error` | 为 true 时表示执行失败 |
|
||||
| `media` | 可选，媒体文件引用列表（如图片、文件） |
|
||||
| `response_handled` | 为 true 时表示已处理用户请求，轮次将结束 |
|
||||
|
||||
---
|
||||
|
||||
## 媒体文件处理
|
||||
|
||||
`respond` action 支持返回媒体文件（图片、文件等）。有两种处理方式：
|
||||
|
||||
### 1. 自动发送（`response_handled=true`）
|
||||
|
||||
当 `response_handled=true` 时，媒体文件会自动发送给用户，轮次结束：
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "图片已发送给用户",
|
||||
"for_user": "",
|
||||
"media": ["media://abc123"],
|
||||
"response_handled": true
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
适用场景：
|
||||
- 图像生成插件直接返回结果
|
||||
- 文件下载插件发送文件给用户
|
||||
|
||||
### 2. LLM 可见（`response_handled=false`）
|
||||
|
||||
当 `response_handled=false` 时，媒体引用会传递给 LLM，LLM 可以在下一轮请求中看到内容：
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "图片已加载，路径：/tmp/image.png [file:/tmp/image.png]",
|
||||
"media": ["media://abc123"]
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
LLM 看到内容后，可以自主决定：
|
||||
- 使用 `send_file` 工具发送给用户
|
||||
- 分析图片内容并回复用户
|
||||
- 其他处理方式
|
||||
|
||||
### 媒体引用格式
|
||||
|
||||
媒体引用使用 `media://` 协议：
|
||||
|
||||
```
|
||||
media://<store-id>
|
||||
```
|
||||
|
||||
这些引用由 PicoClaw 的 MediaStore 管理，可以：
|
||||
- 通过 channel 发送给用户
|
||||
- 在 LLM vision 请求中转换为 base64
|
||||
|
||||
### 替代方案：使用现有工具
|
||||
|
||||
如果插件生成文件，可以返回文件路径让 LLM 调用 `send_file` 等工具：
|
||||
|
||||
```json
|
||||
{
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": "图片已生成，保存在 /tmp/generated_image.png。使用 send_file 工具发送给用户。",
|
||||
"for_user": "",
|
||||
"silent": false
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
这种方式：
|
||||
- 更解耦，LLM 自主决策发送时机
|
||||
- 利用现有工具机制
|
||||
- 支持批量发送、延迟发送等场景
|
||||
|
||||
---
|
||||
|
||||
## 多工具注入示例
|
||||
|
||||
可以同时注入多个工具：
|
||||
|
||||
```python
|
||||
def handle_before_llm(params: dict) -> dict:
|
||||
tools = params.get("tools", [])
|
||||
|
||||
# 工具1：天气查询
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "get_weather",
|
||||
"description": "查询城市天气",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"city": {"type": "string", "description": "城市名称"}
|
||||
},
|
||||
"required": ["city"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
# 工具2：计算器
|
||||
tools.append({
|
||||
"type": "function",
|
||||
"function": {
|
||||
"name": "calculate",
|
||||
"description": "执行数学计算",
|
||||
"parameters": {
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"expression": {"type": "string", "description": "数学表达式"}
|
||||
},
|
||||
"required": ["expression"]
|
||||
}
|
||||
}
|
||||
})
|
||||
|
||||
return {
|
||||
"action": "modify",
|
||||
"request": {
|
||||
"model": params.get("model"),
|
||||
"messages": params.get("messages", []),
|
||||
"tools": tools,
|
||||
"options": params.get("options", {}),
|
||||
}
|
||||
}
|
||||
|
||||
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
args = params.get("arguments", {})
|
||||
|
||||
if tool == "get_weather":
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": get_weather(args.get("city", "")),
|
||||
}
|
||||
|
||||
if tool == "calculate":
|
||||
# 简单计算示例
|
||||
try:
|
||||
expr = args.get("expression", "")
|
||||
result = eval(expr) # 注意：实际使用时需要安全处理
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": f"计算结果: {result}",
|
||||
"silent": False,
|
||||
"is_error": False,
|
||||
},
|
||||
}
|
||||
except Exception as e:
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": {
|
||||
"for_llm": f"计算错误: {e}",
|
||||
"silent": False,
|
||||
"is_error": True,
|
||||
},
|
||||
}
|
||||
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 与内置工具共存
|
||||
|
||||
注入的插件工具与 PicoClaw 内置工具共存：
|
||||
|
||||
- 内置工具（如 `bash`、`read_file`）正常通过 ToolRegistry 执行
|
||||
- 插件工具通过 hook 的 `respond` action 返回结果
|
||||
- `handle_before_tool` 中只处理插件工具，其他工具返回 `continue`
|
||||
|
||||
---
|
||||
|
||||
## Go 进程内 Hook 示例
|
||||
|
||||
如果需要在 Go 代码中实现插件工具注入：
|
||||
|
||||
```go
|
||||
package myhooks
|
||||
|
||||
import (
|
||||
"context"
|
||||
"github.com/sipeed/picoclaw/pkg/agent"
|
||||
"github.com/sipeed/picoclaw/pkg/tools"
|
||||
)
|
||||
|
||||
type WeatherPluginHook struct{}
|
||||
|
||||
func (h *WeatherPluginHook) BeforeLLM(
|
||||
ctx context.Context,
|
||||
req *agent.LLMHookRequest,
|
||||
) (*agent.LLMHookRequest, agent.HookDecision, error) {
|
||||
// 注入工具定义
|
||||
req.Tools = append(req.Tools, agent.ToolDefinition{
|
||||
Type: "function",
|
||||
Function: agent.FunctionDefinition{
|
||||
Name: "get_weather",
|
||||
Description: "查询城市天气",
|
||||
Parameters: map[string]any{
|
||||
"type": "object",
|
||||
"properties": map[string]any{
|
||||
"city": map[string]any{
|
||||
"type": "string",
|
||||
"description": "城市名称",
|
||||
},
|
||||
},
|
||||
"required": []string{"city"},
|
||||
},
|
||||
},
|
||||
})
|
||||
|
||||
return req, agent.HookDecision{Action: agent.HookActionContinue}, nil
|
||||
}
|
||||
|
||||
func (h *WeatherPluginHook) BeforeTool(
|
||||
ctx context.Context,
|
||||
call *agent.ToolCallHookRequest,
|
||||
) (*agent.ToolCallHookRequest, agent.HookDecision, error) {
|
||||
if call.Tool == "get_weather" {
|
||||
city := call.Arguments["city"].(string)
|
||||
|
||||
// 设置 HookResult，使用 respond action
|
||||
next := call.Clone()
|
||||
next.HookResult = &tools.ToolResult{
|
||||
ForLLM: getWeatherData(city),
|
||||
Silent: false,
|
||||
IsError: false,
|
||||
}
|
||||
|
||||
return next, agent.HookDecision{Action: agent.HookActionRespond}, nil
|
||||
}
|
||||
|
||||
return call, agent.HookDecision{Action: agent.HookActionContinue}, nil
|
||||
}
|
||||
|
||||
func getWeatherData(city string) string {
|
||||
// 实现天气查询逻辑
|
||||
return fmt.Sprintf("%s天气：晴，温度20°C", city)
|
||||
}
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 总结
|
||||
|
||||
通过 hook 系统的 `respond` action，外部进程可以：
|
||||
|
||||
1. **注入工具定义**：让 LLM 知道有新工具可用
|
||||
2. **提供工具实现**：直接返回执行结果，无需注册到 ToolRegistry
|
||||
3. **与内置工具共存**：不影响 PicoClaw 原有工具的正常运行
|
||||
|
||||
这为插件开发提供了灵活、优雅的解决方案。
|
||||
|
||||
---
|
||||
|
||||
## 安全边界说明
|
||||
|
||||
### 绕过审批检查
|
||||
|
||||
**重要**：`respond` action 会绕过 `ApproveTool` 审批检查。
|
||||
|
||||
这意味着：
|
||||
- `before_tool` hook 可以为**任何工具名称**返回 `respond`，包括敏感工具（如 `bash`）
|
||||
- 工具不会经过审批流程，直接返回 hook 提供的结果
|
||||
- 这是为了支持插件工具而设计，但也带来了安全风险
|
||||
|
||||
### 安全建议
|
||||
|
||||
1. **审查 hook 配置**：确保只有可信的 hook 进程被启用
|
||||
2. **限制 hook 权限**：在 hook 实现中添加自己的安全检查
|
||||
3. **优先使用 `deny_tool`**：对于拒绝执行，使用 `deny_tool` action 而非 `respond` 返回错误
|
||||
|
||||
### 示例：hook 内置安全检查
|
||||
|
||||
```python
|
||||
def handle_before_tool(params: dict) -> dict:
|
||||
tool = params.get("tool", "")
|
||||
args = params.get("arguments", {})
|
||||
|
||||
# 安全检查：只处理插件工具
|
||||
if tool in ["get_weather", "calculate"]:
|
||||
return {
|
||||
"action": "respond",
|
||||
"result": execute_plugin_tool(tool, args),
|
||||
}
|
||||
|
||||
# 其他工具继续正常流程（会经过审批）
|
||||
return {"action": "continue"}
|
||||
```
|
||||
|
||||
这样可以确保 hook 只影响插件工具，不影响系统工具的审批流程。
|
||||
Reference in New Issue
Block a user