Files
HomeAgent/internal/agent/io/channel.go
JianFeeeee 4707b05498 fix(scheduler): inputch 划给子后输入只流向子 —— 补上「进内核之前」的输入路由
用户指出的语义(设计稿 §4.1 早已写明):
**inputch 是可分配资源**,「路由发生在**进内核之前**」—— 划给某个 agent 后,
该通道的输入**只流向那个 agent**;outputch 不同,授权是**非独占**的,
父依旧可以通过它发送内容。

而代码里 inputch 划拨只做了**登记**,没有做**路由**:
- 插件注入输入的 io 是**根 agent 的**(`cmd/homed` 里 `pluginReg.SetIOManager(iom)`);
- 唯一消费输入的是「该 io 自己的调度器」(`scheduler.go` 读 `a.io.InputChan()`);
- `ChannelRegistry.Assign` 只把 Owner 写进登记表,**没有任何转发动作**。

⇒ 现场表现(用户线上联调):子挂 `inputch=[timer]`,**timer 的输入却打在父身上**
(日志 `[agent] interrupt from timer/timer`),子侧 `轮次=0` 永远不动。
登记表里的 Owner 于是沦为标签。

改法(按 §4.1 把路由放回"进内核之前"):
- `IOManager` 增加 `InputRouter`(`SetInputRouter`),并把**五处直接入队**收口到
  `deliverInput`:`InjectInput` / `InjectInputSync` / `InjectInputTo` /
  `InjectInputSyncTo` / `InjectInterrupt`(排队与中断两条路都过路由)。
- 内核注入路由器 `Agent.routeInputByOwner`:查 inputch 的 Owner —— 归自己/未分配 ⇒
  本内核处理;归自己的某个驻留子 ⇒ `DeliverRouted` 交给它(**不再进父的队列**);
  归一个不存在的 agent ⇒ **不吞输入**,父兜底 + 留痕(吞掉输入比多处理一条更糟)。
- `DeliverRouted` 是"已路由"的投递口,不再二次路由(避免成环)。
- 同步输入的 `ResponseCh` 随事件一起走 ⇒ 回答由持有者写回同一回程(§4.3)。

判据(新增 6 条):
- io 层:被接管时排队/中断都**不入本内核队列**(且中断确实经过路由)/ 放行与未设
  路由器时与历史行为一致 / `DeliverRouted` 不再触发路由
- 内核层:划给子的 inputch 输入进**子**(子 Executed>0)且**父 Enqueued 不变** /
  归属到不存在的 agent 时父兜底(不吞)/ 未分配的 inputch 仍归父
2026-09-13 15:35:54 +08:00

962 lines
31 KiB
Go
Raw Blame History

This file contains ambiguous Unicode characters

This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.

package io
import (
"fmt"
"log"
"runtime/debug"
"sync"
"time"
pubsdk "gitcode.com/JianFeeeee/homeagent-sdk/sdk"
)
// ChannelDef 描述通道在记忆计算层的行为,与 ToolDef.NoMemory/Cleaner 语义一致。
type ChannelDef = pubsdk.ChannelDef
type DeviceType int
const (
DeviceInput DeviceType = 0
DeviceOutput DeviceType = 1
DeviceIO DeviceType = 2
)
// OutputCapability 定义通道支持的输出格式
type OutputCapability int
const (
CapText OutputCapability = 1 << iota // 文本
CapFile // 文件
CapImage // 图片
CapAudio // 音频
CapStructured // 结构化数据JSON/卡片)
)
func (c OutputCapability) Supports(cap OutputCapability) bool {
return c&cap != 0
}
func (c OutputCapability) String() string {
var flags []string
if c&CapText != 0 {
flags = append(flags, "text")
}
if c&CapFile != 0 {
flags = append(flags, "file")
}
if c&CapImage != 0 {
flags = append(flags, "image")
}
if c&CapAudio != 0 {
flags = append(flags, "audio")
}
if c&CapStructured != 0 {
flags = append(flags, "structured")
}
return fmt.Sprintf("%v", flags)
}
type Device interface {
Name() string
Type() DeviceType
Description() string
Tools() []ToolDef
Execute(tool string, args map[string]interface{}) (interface{}, error)
Start() error
Stop() error
OutputCapabilities() OutputCapability
ChannelDef() ChannelDef
}
type ToolHandler func(args map[string]interface{}) (interface{}, error)
type ToolDef struct {
Name string `json:"name"`
Description string `json:"description"`
Parameters map[string]interface{} `json:"parameters"`
Handler ToolHandler `json:"-"` // 可选插件工具的直接处理器Device 通过 Execute() 分发
}
type InputEvent struct {
RequestID string `json:"request_id"`
Source string `json:"source"`
Type string `json:"type"`
Payload map[string]interface{} `json:"payload"`
ResponseCh chan<- *OutputEvent `json:"-"`
OutputChannel string `json:"output_channel"` // 默认输出通道(不传则等于 Source
}
type OutputEvent struct {
RequestID string `json:"request_id"`
Target string `json:"target"`
Type string `json:"type"`
Payload map[string]interface{} `json:"payload"`
Done bool `json:"done,omitempty"`
OutputChannel string `json:"output_channel"` // 路由到此通道
}
type IOManager struct {
mu sync.RWMutex
devices map[string]Device
inputCh chan *InputEvent
interruptCh chan *InputEvent
outputCh chan *OutputEvent
nextReqID int64
channelReg *ChannelRegistry
// parent 是"上级 IOManager"(驻留子的轻量内核指向父的内核)。
//
// 为什么需要:**输出通道在 io 层就是 Device**,而它们是由插件登记在**父**的
// io 上的。驻留子有自己的 IOManager自己的输入入口、自己的 outputCh
// 若只看自己那张空表,`output_send__<通道>` 会被判"通道不存在或不可用"
// `output_list_channels` 是空的,`output_send__*` 工具也不会生成
// —— 现场表现就是"驻留子不会说话/不会发消息"(联调实录:父侧通道装载完整、
// 子侧 childIO 空壳)。
//
// 用**实时回退**而不是创建时复制快照:设备会随资源生灭(远程设备上线/掉线
// 以分钟计),复制出来的表转瞬就过期。授权由各自的 AllowedOutputs 白名单把关,
// 回退只解决"看得见",不解决"能不能用"。
parent *IOManager
// inputRouter 决定一条输入是否被"别的 agent"接管(返回 true = 已接管)。
//
// 为什么放在 ioinputch 是**最基本的输入路由单位**,而**路由发生在进内核之前**
// docs/zh/resident-subagent-design.md §4.1)。插件注入输入的收口就在这里,
// 所以路由必须在这里生效 —— inputch 划给某个 agent 后,输入**只流向那个 agent**
// 本内核根本看不到它。io 层不认识 agent路由器由内核注入
// (见 core.Agent.routeInputByOwner
inputRouter InputRouter
// toolBlocks插件工具注入多模态内容块process.go 在下一条 tool message 时消费。
// 用 interface{}[] 避免 import api.ContentBlock 导致的循环依赖。
toolBlocksMu sync.Mutex
toolPendingBlocks []interface{}
}
func NewIOManager() *IOManager {
return &IOManager{
devices: make(map[string]Device),
inputCh: make(chan *InputEvent, 256),
interruptCh: make(chan *InputEvent, 64),
outputCh: make(chan *OutputEvent, 256),
channelReg: NewChannelRegistry(),
}
}
// InputRouter 是输入路由器的签名。
//
// evt 待投递的输入事件OutputChannel 即它的 inputch
// isInterrupt 该输入是中断还是排队(两者都要按归属路由)
// 返回 true = 已被别的 agent 接管,本内核不再处理
type InputRouter func(evt *InputEvent, isInterrupt bool) bool
// SetInputRouter 注入输入路由器nil = 不路由,行为与以前完全一致)。
func (m *IOManager) SetInputRouter(r InputRouter) {
m.mu.Lock()
m.inputRouter = r
m.mu.Unlock()
}
// deliverInput 是**本内核**接收一条外部输入的收口:先按 inputch 归属路由,
// 被别的 agent 接管就不进本内核队列(划给子的 inputch父不再收到 —— 这是「划拨」
// 的语义,不是"父也顺便看一眼")。
func (m *IOManager) deliverInput(evt *InputEvent, isInterrupt bool) {
m.mu.RLock()
router := m.inputRouter
m.mu.RUnlock()
if router != nil && router(evt, isInterrupt) {
return
}
m.pushLocal(evt, isInterrupt)
}
// DeliverRouted 把**已被路由**的事件放进本内核队列(不再二次路由)。
// 由路由器实现调用:父把输入交给持有该 inputch 的子。
func (m *IOManager) DeliverRouted(evt *InputEvent, isInterrupt bool) {
m.pushLocal(evt, isInterrupt)
}
func (m *IOManager) pushLocal(evt *InputEvent, isInterrupt bool) {
if isInterrupt {
m.interruptCh <- evt
return
}
m.inputCh <- evt
}
// SetParentIO 设置上级 IOManagernil 表示无上级,行为与以前完全一致)。
// 见 parent 字段的说明:用于驻留子继承父的输出通道/设备视图。
func (m *IOManager) SetParentIO(p *IOManager) {
m.mu.Lock()
m.parent = p
m.mu.Unlock()
}
// lookupDevice 查设备:自己的登记优先,其次回退到上级。
//
// 先在自己锁内取快照再查上级,**不跨锁调用**(避免锁序问题)。
func (m *IOManager) lookupDevice(name string) Device {
m.mu.RLock()
dev, ok := m.devices[name]
parent := m.parent
m.mu.RUnlock()
if ok {
return dev
}
if parent != nil {
return parent.GetDevice(name)
}
return nil
}
func (m *IOManager) UnregisterDevice(name string) {
m.mu.Lock()
defer m.mu.Unlock()
delete(m.devices, name)
}
func (m *IOManager) nextRequestID() string {
m.mu.Lock()
defer m.mu.Unlock()
m.nextReqID++
return fmt.Sprintf("req_%d_%d", time.Now().UnixNano(), m.nextReqID)
}
// AtomicSwapDevices 原子化替换全部 IO 设备
// 1. 新设备必须在调用前已完成 Start()
// 2. 调用后旧设备立即摘除,新请求走向新设备
// 3. 返回旧设备列表,由调用方负责 Stop()
func (m *IOManager) AtomicSwapDevices(newDevices map[string]Device) map[string]Device {
m.mu.Lock()
defer m.mu.Unlock()
oldDevices := m.devices
m.devices = newDevices
return oldDevices
}
func (m *IOManager) RegisterDevice(dev Device) error {
m.mu.Lock()
defer m.mu.Unlock()
if _, ok := m.devices[dev.Name()]; ok {
return fmt.Errorf("device %s already registered", dev.Name())
}
m.devices[dev.Name()] = dev
return nil
}
func (m *IOManager) GetDevice(name string) Device {
return m.lookupDevice(name)
}
func (m *IOManager) StartAll() error {
m.mu.RLock()
devices := make([]Device, 0, len(m.devices))
for _, dev := range m.devices {
devices = append(devices, dev)
}
m.mu.RUnlock()
for _, dev := range devices {
if err := dev.Start(); err != nil {
return fmt.Errorf("start device %s: %w", dev.Name(), err)
}
}
return nil
}
func (m *IOManager) StopAll() {
m.mu.RLock()
devices := make([]Device, 0, len(m.devices))
for _, dev := range m.devices {
devices = append(devices, dev)
}
m.mu.RUnlock()
for _, dev := range devices {
if err := dev.Stop(); err != nil {
log.Printf("[io] stop device %s error: %v", dev.Name(), err)
}
}
}
func (m *IOManager) InjectInput(source string, eventType string, payload map[string]interface{}) {
m.deliverInput(&InputEvent{
RequestID: m.nextRequestID(),
Source: source,
Type: eventType,
Payload: payload,
OutputChannel: source,
}, false)
}
func (m *IOManager) InjectInputSync(source string, eventType string, payload map[string]interface{}) *OutputEvent {
ch := make(chan *OutputEvent, 1)
m.deliverInput(&InputEvent{
RequestID: m.nextRequestID(),
Source: source,
Type: eventType,
Payload: payload,
ResponseCh: ch,
OutputChannel: source,
}, false)
// 被路由走时,回答由持有该 inputch 的 agent 写进同一个 ResponseCh
//§4.3:同步输入的回程是事前定好的)——所以这里照常等待。
return <-ch
}
// InjectInputTo 注入输入事件并指定输出通道
func (m *IOManager) InjectInputTo(source, outputChannel, eventType string, payload map[string]interface{}) {
m.deliverInput(&InputEvent{
RequestID: m.nextRequestID(),
Source: source,
Type: eventType,
Payload: payload,
OutputChannel: outputChannel,
}, false)
}
// InjectInputSyncTo 注入输入事件(同步等待)并指定输出通道
func (m *IOManager) InjectInputSyncTo(source, outputChannel, eventType string, payload map[string]interface{}) *OutputEvent {
ch := make(chan *OutputEvent, 1)
m.deliverInput(&InputEvent{
RequestID: m.nextRequestID(),
Source: source,
Type: eventType,
Payload: payload,
ResponseCh: ch,
OutputChannel: outputChannel,
}, false)
return <-ch
}
// InjectOptions 声明一次注入在记忆层与上下文层的表现。
//
// 零值 = 记入记忆 + 不裁剪上下文,与历史的三参数注入方法完全一致。
// 别名到公共 SDK 而非另建一套:内置插件与外部插件必须用同一套结构,
// 否则内核要认两种类型,而漏认会静默丢失标志位。
type InjectOptions = pubsdk.InjectOptions
// applyInjectOpts 把注入标志位写进事件 payload。
//
// 只在非零时写:零值与旧 payload 逐字节一致,事件订阅方与旧内核
// (不认识这两个键)都不会受影响。
//
// 为什么不把标志位当独立参数传到底eventloop 与各注入路径都按 payload 取字段
// no_memory 本来就是这么走的payload 是这里唯一已有的携带面。
func applyInjectOpts(payload map[string]interface{}, opts InjectOptions) {
if opts.NoMemory {
payload["no_memory"] = true
}
if opts.ContextPolicy != "" {
payload["context_policy"] = opts.ContextPolicy
}
if opts.CleanerName != "" {
payload["cleaner_name"] = opts.CleanerName
}
// priority 只对中断注入有意义;排队路径会忽略它(内核侧只读不写)。
if opts.Priority != "" {
payload["priority"] = opts.Priority
}
}
func (m *IOManager) InjectInputOpts(source, eventType string, payload map[string]interface{}, opts InjectOptions) {
applyInjectOpts(payload, opts)
m.InjectInput(source, eventType, payload)
}
func (m *IOManager) InjectInputToOpts(source, outputChannel, eventType string, payload map[string]interface{}, opts InjectOptions) {
applyInjectOpts(payload, opts)
m.InjectInputTo(source, outputChannel, eventType, payload)
}
func (m *IOManager) InjectInputSyncToOpts(source, outputChannel, eventType string, payload map[string]interface{}, opts InjectOptions) *OutputEvent {
applyInjectOpts(payload, opts)
return m.InjectInputSyncTo(source, outputChannel, eventType, payload)
}
func (m *IOManager) InjectInterruptOpts(source, channel string, payload map[string]interface{}, opts InjectOptions) {
applyInjectOpts(payload, opts)
m.InjectInterrupt(source, channel, payload)
}
func (m *IOManager) InjectText(source string, text string) {
m.InjectInput(source, "text", map[string]interface{}{
"content": text,
})
}
func (m *IOManager) InjectTextSync(source string, text string) *OutputEvent {
return m.InjectInputSync(source, "text", map[string]interface{}{
"content": text,
})
}
// InjectTextTo 注入文本输入并指定输出通道
func (m *IOManager) InjectTextTo(source, outputChannel, text string) {
m.InjectInputTo(source, outputChannel, "text", map[string]interface{}{
"content": text,
})
}
// InjectTextNoMemoryTo 注入文本输入(不产生记忆)并指定输出通道
func (m *IOManager) InjectTextNoMemoryTo(source, outputChannel, text string) {
m.InjectInputTo(source, outputChannel, "text", map[string]interface{}{
"content": text,
"no_memory": true,
})
}
// InjectTextSyncNoMemoryTo 注入文本输入(同步等待,不产生记忆)并指定输出通道
func (m *IOManager) InjectTextSyncNoMemoryTo(source, outputChannel, text string) *OutputEvent {
return m.InjectInputSyncTo(source, outputChannel, "text", map[string]interface{}{
"content": text,
"no_memory": true,
})
}
// InjectInterrupt 向中断通道发送输入
func (m *IOManager) InjectInterrupt(source, channel string, payload map[string]interface{}) {
if payload == nil {
payload = map[string]interface{}{}
}
evtType, _ := payload["type"].(string)
m.deliverInput(&InputEvent{
RequestID: m.nextRequestID(),
Source: source,
Type: evtType,
Payload: payload,
OutputChannel: channel,
}, true)
}
func (m *IOManager) InjectInterruptText(source, channel, text string) {
m.InjectInterruptTextOpts(source, channel, text, InjectOptions{})
}
// InjectInterruptTextOpts 注入中断文本,并声明本次注入的记忆/裁剪行为。
//
// 中断也允许声明 ContextPolicyPrune中断同样携带内容进入上下文。
func (m *IOManager) InjectInterruptTextOpts(source, channel, text string, opts InjectOptions) {
m.InjectInterruptOpts(source, channel, map[string]interface{}{
"type": "text",
"content": text,
}, opts)
}
// InjectTextOpts 注入排队文本,并声明本次注入的记忆/裁剪行为。
func (m *IOManager) InjectTextOpts(source, channel, text string, opts InjectOptions) {
m.InjectInputToOpts(source, channel, "text", map[string]interface{}{
"content": text,
}, opts)
}
// InjectTextSyncOpts 同步注入文本并声明记忆/裁剪行为。
func (m *IOManager) InjectTextSyncOpts(source, outputChannel, text string, opts InjectOptions) *OutputEvent {
return m.InjectInputSyncToOpts(source, outputChannel, "text", map[string]interface{}{
"content": text,
}, opts)
}
func (m *IOManager) InputInterruptChan() <-chan *InputEvent { return m.interruptCh }
// InjectTextSyncTo 注入文本输入(同步等待)并指定输出通道
func (m *IOManager) InjectTextSyncTo(source, outputChannel, text string) *OutputEvent {
return m.InjectInputSyncTo(source, outputChannel, "text", map[string]interface{}{
"content": text,
})
}
// ---- 带标志位的注入(记忆/裁剪行为由调用点声明)----
// InjectInputMediaOpts 注入带媒体块的输入,并声明记忆/裁剪行为。
func (m *IOManager) InjectInputMediaOpts(source, outputChannel, text string, blocks []pubsdk.ContentBlock, opts InjectOptions) {
m.InjectInputToOpts(source, outputChannel, "text", map[string]interface{}{
"content": text,
"media_blocks": blocks,
}, opts)
}
// InjectInputMediaSyncOpts 注入带媒体块的输入并同步等待回复,同时声明记忆/裁剪行为。
func (m *IOManager) InjectInputMediaSyncOpts(source, outputChannel, text string, blocks []pubsdk.ContentBlock, opts InjectOptions) *OutputEvent {
return m.InjectInputSyncToOpts(source, outputChannel, "text", map[string]interface{}{
"content": text,
"media_blocks": blocks,
}, opts)
}
// InjectInterruptMediaOpts 注入带媒体块的中断,并声明记忆/裁剪行为。
func (m *IOManager) InjectInterruptMediaOpts(source, channel, text string, blocks []pubsdk.ContentBlock, opts InjectOptions) {
m.InjectInterruptOpts(source, channel, map[string]interface{}{
"type": "text",
"content": text,
"media_blocks": blocks,
}, opts)
}
func (m *IOManager) EmitOutput(target string, outputType string, payload map[string]interface{}) {
m.outputCh <- &OutputEvent{
RequestID: "",
Target: target,
Type: outputType,
Payload: payload,
Done: true,
}
}
// EmitOutputTo 通过指定输出通道发送
func (m *IOManager) EmitOutputTo(target, outputChannel, outputType string, payload map[string]interface{}) {
m.outputCh <- &OutputEvent{
RequestID: "",
Target: target,
Type: outputType,
Payload: payload,
Done: true,
OutputChannel: outputChannel,
}
}
func (m *IOManager) EmitText(target string, text string) {
m.EmitOutput(target, "text", map[string]interface{}{
"content": text,
})
}
// EmitTextTo 通过指定输出通道发送文本
func (m *IOManager) EmitTextTo(target, outputChannel, text string) {
m.EmitOutputTo(target, outputChannel, "text", map[string]interface{}{
"content": text,
})
}
func (m *IOManager) InputChan() <-chan *InputEvent { return m.inputCh }
func (m *IOManager) OutputChan() <-chan *OutputEvent { return m.outputCh }
// RegisterInputChannel 注册一个 inputch不带插件归属兼容旧调用
//
// inputch 是**最基本的输入路由单位**;一个插件可以注册多个。
// 新代码请用 RegisterInputChannelFrom 以便登记归属插件(可追溯)。
func (m *IOManager) RegisterInputChannel(name string, def ChannelDef) {
_ = m.RegisterInputChannelFrom("", name, def)
}
// RegisterInputChannelFrom 注册一个 inputch 并登记归属插件。
func (m *IOManager) RegisterInputChannelFrom(plugin, name string, def ChannelDef) error {
return m.channelReg.Register(InputChannel{Name: name, Plugin: plugin, Def: def})
}
// UnregisterInputChannel 注销一个 inputch。
func (m *IOManager) UnregisterInputChannel(name string) {
m.channelReg.Unregister(name)
}
// AssignInputChannel 把一个 inputch 划给某个 agent见 ChannelRegistry.Assign
func (m *IOManager) AssignInputChannel(name, agentID string, capacity int) error {
return m.channelReg.Assign(name, agentID, capacity)
}
// SetChannelRegistry 注入一份**共享的**登记表(根 agent 与驻留子共用同一份)。
func (m *IOManager) SetChannelRegistry(r *ChannelRegistry) {
if r == nil {
return
}
m.mu.Lock()
defer m.mu.Unlock()
m.channelReg = r
}
// ChannelRegistry 返回底层登记表(只读用途;可直接读 Views
func (m *IOManager) ChannelRegistry() *ChannelRegistry { return m.channelReg }
// InputChannels 返回全部已注册 inputch按名字排序
func (m *IOManager) InputChannels() []InputChannel { return m.channelReg.List() }
// LookupInputChannel 查询单个 inputch 的完整登记记录。
func (m *IOManager) LookupInputChannel(name string) (InputChannel, bool) {
return m.channelReg.Lookup(name)
}
// GetInputChannelDef 查询 inputch 的记忆行为定义。
func (m *IOManager) GetInputChannelDef(name string) (ChannelDef, bool) {
ch, ok := m.channelReg.Lookup(name)
if !ok {
return ChannelDef{}, false
}
return ch.Def, true
}
func (m *IOManager) GetAllTools() []ToolDef {
m.mu.RLock()
defer m.mu.RUnlock()
var tools []ToolDef
for _, dev := range m.devices {
tools = append(tools, dev.Tools()...)
}
return tools
}
// DeviceOfTool 返回提供该工具的**设备/输出通道名**(设备类工具才有)。
//
// 用途设备类工具device_ctl_*/screensee/computeruse/...)需要按"目标设备"
// 做授权判断,调用方得先知道这个工具属于哪个设备通道。
func (m *IOManager) DeviceOfTool(name string) (string, bool) {
m.mu.RLock()
defer m.mu.RUnlock()
for _, dev := range m.devices {
for _, t := range dev.Tools() {
if t.Name == name {
return dev.Name(), true
}
}
}
return "", false
}
func (m *IOManager) ExecuteTool(name string, args map[string]interface{}) (ret interface{}, err error) {
m.mu.RLock()
type nameDevice struct {
name string
dev Device
}
var candidates []nameDevice
for _, dev := range m.devices {
for _, t := range dev.Tools() {
if t.Name == name {
candidates = append(candidates, nameDevice{name: dev.Name(), dev: dev})
break
}
}
}
m.mu.RUnlock()
if len(candidates) == 0 {
// 自己没这个设备工具 → 看上级(驻留子的设备工具都在父的 io 上)。
m.mu.RLock()
parent := m.parent
m.mu.RUnlock()
if parent != nil {
if ret, err := parent.ExecuteTool(name, args); err == nil {
return ret, nil
}
}
return nil, fmt.Errorf("tool %s not found", name)
}
defer func() {
if r := recover(); r != nil {
log.Printf("[io] tool %s execute panic: %v\n%s", name, r, debug.Stack())
err = fmt.Errorf("tool %s execute panic: %v", name, r)
}
}()
return candidates[0].dev.Execute(name, args)
}
func (m *IOManager) ListDevices() []Device {
m.mu.RLock()
defer m.mu.RUnlock()
list := make([]Device, 0, len(m.devices))
for _, d := range m.devices {
list = append(list, d)
}
return list
}
// ChannelInfo 返回 IOManager 中已注册的所有通道信息
type ChannelInfo struct {
Name string `json:"name"`
Type DeviceType `json:"type"`
Description string `json:"description"`
Tools []ToolDef `json:"tools"`
OutputCaps OutputCapability `json:"output_capabilities"`
}
func (m *IOManager) ListChannels() []ChannelInfo {
m.mu.RLock()
own := make(map[string]Device, len(m.devices))
for name, dev := range m.devices {
own[name] = dev
}
parent := m.parent
m.mu.RUnlock()
// 自己的登记优先(子侧可覆盖/屏蔽同名通道),随后并入上级的可见通道。
// 去重按**名字**:同名即视为同一个通道,不重复列举。
seen := make(map[string]bool, len(own))
var list []ChannelInfo
appendDev := func(dev Device) {
if seen[dev.Name()] {
return
}
seen[dev.Name()] = true
list = append(list, ChannelInfo{
Name: dev.Name(),
Type: dev.Type(),
Description: dev.Description(),
Tools: dev.Tools(),
OutputCaps: dev.OutputCapabilities(),
})
}
for _, dev := range own {
appendDev(dev)
}
if parent != nil {
for _, ch := range parent.ListChannels() {
if seen[ch.Name] {
continue
}
seen[ch.Name] = true
list = append(list, ch)
}
}
return list
}
func (m *IOManager) GetChannelCapabilities(channel string) OutputCapability {
if dev := m.lookupDevice(channel); dev != nil {
return dev.OutputCapabilities()
}
return 0
}
// Microphone
type Microphone struct {
name string
sampleRate int
io *IOManager
}
func NewMicrophone(name string, sampleRate int, io *IOManager) *Microphone {
return &Microphone{name: name, sampleRate: sampleRate, io: io}
}
func (d *Microphone) Name() string { return d.name }
func (d *Microphone) Type() DeviceType { return DeviceInput }
func (d *Microphone) OutputCapabilities() OutputCapability { return 0 } // 纯输入
func (d *Microphone) Description() string {
return fmt.Sprintf("麦克风 (%s, %dHz)", d.name, d.sampleRate)
}
func (d *Microphone) Start() error { return nil }
func (d *Microphone) Stop() error { return nil }
func (d *Microphone) ChannelDef() ChannelDef { return ChannelDef{} }
func (d *Microphone) Tools() []ToolDef {
return []ToolDef{{
Name: d.name + "_capture",
Description: fmt.Sprintf("从 %s 录制音频", d.name),
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"duration": map[string]interface{}{"type": "number", "description": "录制时长(秒)", "default": 3},
},
},
}}
}
func (d *Microphone) Execute(tool string, args map[string]interface{}) (interface{}, error) {
return map[string]interface{}{"device": d.name, "status": "recorded", "format": "wav", "sample_rate": d.sampleRate}, nil
}
// Speaker
type Speaker struct {
name string
io *IOManager
}
func NewSpeaker(name string, io *IOManager) *Speaker {
return &Speaker{name: name, io: io}
}
func (d *Speaker) Name() string { return d.name }
func (d *Speaker) Type() DeviceType { return DeviceOutput }
func (d *Speaker) OutputCapabilities() OutputCapability { return CapText | CapAudio }
func (d *Speaker) Description() string { return fmt.Sprintf("扬声器 (%s)", d.name) }
func (d *Speaker) Start() error { return nil }
func (d *Speaker) Stop() error { return nil }
func (d *Speaker) ChannelDef() ChannelDef { return ChannelDef{} }
func (d *Speaker) Tools() []ToolDef {
return []ToolDef{{
Name: d.name + "_speak",
Description: fmt.Sprintf("通过 %s 播放语音", d.name),
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"text": map[string]interface{}{"type": "string", "description": "播放文本"},
},
"required": []string{"text"},
},
}}
}
func (d *Speaker) Execute(tool string, args map[string]interface{}) (interface{}, error) {
text, _ := args["text"].(string)
return map[string]interface{}{"device": d.name, "status": "playing", "text": text}, nil
}
// Camera
type Camera struct {
name string
io *IOManager
}
func NewCamera(name string, io *IOManager) *Camera {
return &Camera{name: name, io: io}
}
func (d *Camera) Name() string { return d.name }
func (d *Camera) Type() DeviceType { return DeviceInput }
func (d *Camera) OutputCapabilities() OutputCapability { return CapImage } // 可返回图片
func (d *Camera) Description() string { return fmt.Sprintf("摄像头 (%s)", d.name) }
func (d *Camera) Start() error { return nil }
func (d *Camera) Stop() error { return nil }
func (d *Camera) ChannelDef() ChannelDef { return ChannelDef{} }
func (d *Camera) Tools() []ToolDef {
return []ToolDef{
{
Name: d.name + "_capture",
Description: fmt.Sprintf("使用 %s 拍照", d.name),
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"quality": map[string]interface{}{"type": "integer", "description": "质量1-100", "default": 90},
},
},
},
{
Name: d.name + "_stream",
Description: fmt.Sprintf("控制 %s 视频流", d.name),
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"action": map[string]interface{}{"type": "string", "enum": []interface{}{"start", "stop"}},
},
"required": []string{"action"},
},
},
}
}
func (d *Camera) Execute(tool string, args map[string]interface{}) (interface{}, error) {
return map[string]interface{}{"device": d.name, "status": "captured"}, nil
}
// RobotArm
type RobotArm struct {
name string
io *IOManager
}
func NewRobotArm(name string, io *IOManager) *RobotArm {
return &RobotArm{name: name, io: io}
}
func (d *RobotArm) Name() string { return d.name }
func (d *RobotArm) Type() DeviceType { return DeviceIO }
func (d *RobotArm) OutputCapabilities() OutputCapability { return CapStructured }
func (d *RobotArm) Description() string { return fmt.Sprintf("机械臂 (%s)", d.name) }
func (d *RobotArm) Start() error { return nil }
func (d *RobotArm) Stop() error { return nil }
func (d *RobotArm) ChannelDef() ChannelDef { return ChannelDef{} }
func (d *RobotArm) Tools() []ToolDef {
return []ToolDef{
{
Name: d.name + "_move",
Description: fmt.Sprintf("移动 %s 到坐标", d.name),
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"x": map[string]interface{}{"type": "number", "description": "X 轴"},
"y": map[string]interface{}{"type": "number", "description": "Y 轴"},
"z": map[string]interface{}{"type": "number", "description": "Z 轴"},
},
"required": []string{"x", "y", "z"},
},
},
{
Name: d.name + "_grip",
Description: fmt.Sprintf("控制 %s 夹爪", d.name),
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"action": map[string]interface{}{"type": "string", "enum": []interface{}{"open", "close"}},
},
"required": []string{"action"},
},
},
}
}
func (d *RobotArm) Execute(tool string, args map[string]interface{}) (interface{}, error) {
return map[string]interface{}{"device": d.name, "tool": tool, "status": "executed"}, nil
}
// GPIODevice
type GPIODevice struct {
name string
pins []int
io *IOManager
}
func NewGPIODevice(name string, pins []int, io *IOManager) *GPIODevice {
return &GPIODevice{name: name, pins: pins, io: io}
}
func (d *GPIODevice) Name() string { return d.name }
func (d *GPIODevice) Type() DeviceType { return DeviceIO }
func (d *GPIODevice) OutputCapabilities() OutputCapability { return CapStructured }
func (d *GPIODevice) Description() string { return "GPIO 通用引脚" }
func (d *GPIODevice) Start() error { return nil }
func (d *GPIODevice) Stop() error { return nil }
func (d *GPIODevice) ChannelDef() ChannelDef { return ChannelDef{} }
func (d *GPIODevice) Tools() []ToolDef {
return []ToolDef{
{
Name: d.name + "_gpio_write",
Description: "设置引脚电平",
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"pin": map[string]interface{}{"type": "integer"},
"value": map[string]interface{}{"type": "integer", "enum": []interface{}{0, 1}},
},
"required": []string{"pin", "value"},
},
},
{
Name: d.name + "_gpio_read",
Description: "读取引脚电平",
Parameters: map[string]interface{}{
"type": "object",
"properties": map[string]interface{}{
"pin": map[string]interface{}{"type": "integer"},
},
"required": []string{"pin"},
},
},
}
}
func (d *GPIODevice) Execute(tool string, args map[string]interface{}) (interface{}, error) {
return map[string]interface{}{"device": d.name, "tool": tool, "status": "ok"}, nil
}
// SetToolBlocks 插件工具调用时注入多模态内容块image_url/audio_url 等),
// 下一条 tool message 追加这些块到 content 数组OpenAI 多模态格式)。
func (m *IOManager) SetToolBlocks(blocks []interface{}) {
m.toolBlocksMu.Lock()
m.toolPendingBlocks = blocks
m.toolBlocksMu.Unlock()
}
// ConsumeToolBlocks 返回并清空 pending blocksprocess.go 在 append tool message 时调用。
func (m *IOManager) ConsumeToolBlocks() []interface{} {
m.toolBlocksMu.Lock()
blocks := m.toolPendingBlocks
m.toolPendingBlocks = nil
m.toolBlocksMu.Unlock()
return blocks
}