Files
HomeAgent/internal/sdk/plugin.go
JianFeeeee b777322b95 feat(multimodal): 内置多模态感知插件 + process.go 原生支持 tool message 多模态块
【新插件 internal/plugins/multimodal】
- see_picture(path): 读取本地图片/URL,base64 注入 image_url block,
  模型在下一轮 LLM 请求的 tool message 里直接看到图(1024×1024 图约 8500 token)。
  自动识别 MIME,限 3MB 防爆 context。
- see_video(path, frames): ffmpeg 提取关键帧,多帧作为 image_url block 注入。
  默认 4 帧,最大 10 帧,每帧限 2MB。
- listen(path): 读取音频文件,转为 audio_url block 注入,支持 mp3/wav/ogg/m4a。
  限 5MB。

【内核多模态 tool message 支持】
- agent/api 新增 ToolOutput 类型(为后续 handler 直接返回 blocks 预留)
- SDK 公共层新增 ContentBlock/ImageURL/AudioURL(OpenAI 多模态格式)
- IOManager 新增 SetToolBlocks/ConsumeToolBlocks(interface{} 避免循环依赖)
- PluginSDK.SetToolBlocks(blocks) 插件工具调用后注入 blocks
- ioAdapter 桥接 IOInjector.SetToolBlocks
- process.go 工具执行后消费 pending blocks → 追加到 tool message 的 Blocks 字段
  → MarshalJSON 输出 content 数组格式 → LLM 看到图/音频

【验证】
multimodal_see_picture 注入 1024×1024 PNG 后 llmsproxy 统计:
  prompt_tokens=44407(含 ~8500 image token),模型正确描述了图片内容。
2026-08-27 08:39:21 +08:00

366 lines
11 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 sdk
import (
"log"
"sync"
agentIO "gitcode.com/JianFeeeee/HomeAgent/internal/agent/io"
"gitcode.com/JianFeeeee/HomeAgent/internal/events"
pubsdk "gitcode.com/JianFeeeee/homeagent-sdk/sdk"
)
// SDKVersion 是对外 SDK 版本号,与核心 meta.Version 保持一致。
var SDKVersion = pubsdk.SDKVersion
type Plugin interface {
Name() string
Start(sdk *PluginSDK) error
Stop() error
}
type ToolHandler = pubsdk.ToolHandler
type StageHandler = pubsdk.StageHandler
type Stage = pubsdk.Stage
const (
StageOnInput = pubsdk.StageOnInput
StagePreAction = pubsdk.StagePreAction
StagePostAction = pubsdk.StagePostAction
StageBeforeToolcall = pubsdk.StageBeforeToolcall
StageAfterToolcall = pubsdk.StageAfterToolcall
StageBeforeOutput = pubsdk.StageBeforeOutput
StageAfterOutput = pubsdk.StageAfterOutput
)
type StageContext = pubsdk.StageContext
type MemItem = pubsdk.MemItem
type ToolCall = pubsdk.ToolCall
type ToolResult = pubsdk.ToolResult
type ToolDef = pubsdk.ToolDef
type IOInjector = pubsdk.IOInjector
type ToolRegistrar = pubsdk.ToolRegistrar
type StageRegistrar = pubsdk.StageRegistrar
type StageScope = pubsdk.StageScope
const (
StageScopeGlobal = pubsdk.StageScopeGlobal
StageScopeOwnTools = pubsdk.StageScopeOwnTools
)
type APIRegistrar = pubsdk.APIRegistrar
type OutputChannelRegistrar = pubsdk.OutputChannelRegistrar
type InputChannelRegistrar = pubsdk.InputChannelRegistrar
type ChannelDef = pubsdk.ChannelDef
type DisabledPluginInfo struct {
Name string `json:"name"`
DisabledAt string `json:"disabled_at"`
DisabledBy string `json:"disabled_by"`
}
// PluginMeta is the display-name metadata for a plugin (from plg.json / RegisterPluginMeta).
type PluginMeta struct {
NameZh string `json:"name_zh"`
NameEn string `json:"name_en"`
}
type PluginManager interface {
ListLoadedPlugins() []string
ListDisabledPlugins() []DisabledPluginInfo
IsPluginDisabled(name string) bool
// IsBuiltinPlugin 判断插件是否为内置插件编译期工厂init() 自注册)。
// 内置插件只能禁用/启用,不能卸载。
IsBuiltinPlugin(name string) bool
DisablePlugin(name, by string) error
EnablePlugin(name string) error
// RemovePlugin 卸载插件先停止stop handlers + Stop再执行插件注册的
// onRemove 回调RegisterOnRemoveHandler最后从注册表移除并清理配置表。
// 目录删除由调用方负责。
RemovePlugin(name string) error
// StopAndUnload 停止并从注册表移除插件但保留配置表,供更新/升级流程使用:
// 换产物不动配置,重装后配置原样生效。不触发 onRemove 回调。
StopAndUnload(name string) error
ReloadPlugins() (string, error)
// ReloadOne 重载单个插件(停止后重新加载,处理 dlclose/dynamic 句柄)。
ReloadOne(name string) error
PluginMetas() map[string]PluginMeta
PluginDir() string
}
type PluginSDK struct {
*pubsdk.PluginSDK
settings SettingsAPI
memory MemoryAPI
textMem TextMemoryAPI
docMem DocMemoryAPI
know KnowledgeAPI
llm LLMAPI
iom *agentIO.IOManager
eventBus *events.Bus
logger *log.Logger
pluginMgr PluginManager
status StatusAPI
supervisor SupervisorAPI
adapter AdapterAPI
tracker TrackerAPI
config ConfigAPI
tool ToolAPI
indexer IndexerAPI
selftestMu sync.Mutex
selftest *VirtualInstance
}
func (s *PluginSDK) PluginMgr() PluginManager { return s.pluginMgr }
// 以下访问器遮蔽公共 SDK 的同名方法,返回内置插件可用的全量接口。
func (s *PluginSDK) Settings() SettingsAPI { return s.settings }
func (s *PluginSDK) Memory() MemoryAPI { return s.memory }
func (s *PluginSDK) TextMemory() TextMemoryAPI { return s.textMem }
func (s *PluginSDK) DocMemory() DocMemoryAPI { return s.docMem }
func (s *PluginSDK) Knowledge() KnowledgeAPI { return s.know }
func (s *PluginSDK) LLM() LLMAPI { return s.llm }
// ioAdapter 桥接 IOManager 到公共 SDK 的 IOInjector 接口,
// 确保外部插件通过 s.InjectText() 等方法的调用能被路由到内核 IO 层。
type ioAdapter struct{ iom *agentIO.IOManager }
func (a ioAdapter) InjectInterruptText(source, channel, text string) {
if a.iom != nil {
a.iom.InjectInterrupt(source, channel, map[string]interface{}{"type": "text", "content": text})
}
}
// InjectInputSync 同步注入输入并等待回复(阻塞直至 agent 处理完成),返回回复文本。
func (a ioAdapter) InjectInputSync(source, channel, text string) string {
if a.iom == nil {
return ""
}
out := a.iom.InjectInputSyncTo(source, channel, "text", map[string]interface{}{
"content": text,
})
if out == nil {
return ""
}
reply, _ := out.Payload["content"].(string)
return reply
}
func (a ioAdapter) InjectText(source, channel, text string) {
if a.iom != nil {
a.iom.InjectInputTo(source, channel, "text", map[string]interface{}{"content": text})
}
}
func (a ioAdapter) InjectTextNoMemory(source, channel, text string) {
if a.iom != nil {
a.iom.InjectInputTo(source, channel, "text", map[string]interface{}{"content": text, "no_memory": true})
}
}
// SDKConfig holds all dependencies for creating a PluginSDK.
type SDKConfig struct {
IOManager *agentIO.IOManager
EventBus *events.Bus
Memory MemoryAPI
TextMemory TextMemoryAPI
DocMemory DocMemoryAPI
Knowledge KnowledgeAPI
LLM LLMAPI
Settings SettingsAPI
RegTool ToolRegistrar
RegStage StageRegistrar
RegAPI APIRegistrar
RegOutput OutputChannelRegistrar
RegInput InputChannelRegistrar
PluginMgr PluginManager
Status StatusAPI
Supervisor SupervisorAPI
Adapter AdapterAPI
Tracker TrackerAPI
Config ConfigAPI
Tool ToolAPI
Indexer IndexerAPI
}
func New(name string, cfg SDKConfig) *PluginSDK {
base := pubsdk.New(name, cfg.Settings, cfg.RegTool, cfg.RegStage, cfg.RegAPI, cfg.RegOutput)
if cfg.IOManager != nil {
base.SetIOInjector(ioAdapter{iom: cfg.IOManager})
}
if cfg.RegInput != nil {
base.SetInputChannelRegistrar(cfg.RegInput)
}
base.SetMemoryAPI(cfg.Memory)
base.SetTextMemoryAPI(cfg.TextMemory)
base.SetDocMemoryAPI(cfg.DocMemory)
base.SetKnowledgeAPI(cfg.Knowledge)
base.SetLLMAPI(cfg.LLM)
return &PluginSDK{
PluginSDK: base,
settings: cfg.Settings,
memory: cfg.Memory,
textMem: cfg.TextMemory,
docMem: cfg.DocMemory,
know: cfg.Knowledge,
llm: cfg.LLM,
iom: cfg.IOManager,
eventBus: cfg.EventBus,
logger: log.Default(),
pluginMgr: cfg.PluginMgr,
status: cfg.Status,
supervisor: cfg.Supervisor,
adapter: cfg.Adapter,
tracker: cfg.Tracker,
config: cfg.Config,
tool: cfg.Tool,
indexer: cfg.Indexer,
}
}
// Selftest 返回一个隔离的虚拟自检实例healthcheck 等内置插件用),
// 完全独立于生产存储,不产生任何污染。首次调用创建,复用已存在实例;
// 每轮自检前调用 SelftestReset 重建以清空上轮测试数据。
func (s *PluginSDK) Selftest(scope string) (*VirtualInstance, error) {
s.selftestMu.Lock()
defer s.selftestMu.Unlock()
if s.selftest == nil {
vi, err := NewVirtualInstance(scope)
if err != nil {
return nil, err
}
s.selftest = vi
}
return s.selftest, nil
}
// SelftestReset 清理并重建隔离自检实例,用于每轮健康检查前重置状态。
func (s *PluginSDK) SelftestReset(scope string) error {
s.selftestMu.Lock()
defer s.selftestMu.Unlock()
if s.selftest == nil {
vi, err := NewVirtualInstance(scope)
if err != nil {
return err
}
s.selftest = vi
return nil
}
vi, err := s.selftest.Reset(scope)
if err != nil {
return err
}
s.selftest = vi
return nil
}
func (s *PluginSDK) Status() StatusAPI { return s.status }
func (s *PluginSDK) Supervisor() SupervisorAPI { return s.supervisor }
func (s *PluginSDK) Adapter() AdapterAPI { return s.adapter }
func (s *PluginSDK) Tracker() TrackerAPI { return s.tracker }
func (s *PluginSDK) Config() ConfigAPI { return s.config }
func (s *PluginSDK) Tool() ToolAPI { return s.tool }
func (s *PluginSDK) Indexer() IndexerAPI { return s.indexer }
func (s *PluginSDK) InjectInput(source, channel, eventType string, payload map[string]interface{}) {
if s.iom != nil {
s.iom.InjectInputTo(source, channel, eventType, payload)
}
}
func (s *PluginSDK) InjectInputSync(source, channel, eventType string, payload map[string]interface{}) *agentIO.OutputEvent {
if s.iom != nil {
return s.iom.InjectInputSyncTo(source, channel, eventType, payload)
}
return nil
}
func (s *PluginSDK) InjectInterrupt(source, channel, eventType string, payload map[string]interface{}) {
if s.iom != nil {
if payload == nil {
payload = map[string]interface{}{}
}
payload["type"] = eventType
s.iom.InjectInterrupt(source, channel, payload)
}
}
func (s *PluginSDK) InjectTextSync(source, channel, text string) *agentIO.OutputEvent {
return s.InjectInputSync(source, channel, "text", map[string]interface{}{"content": text})
}
func (s *PluginSDK) InjectTextSyncNoMemory(source, channel, text string) *agentIO.OutputEvent {
return s.InjectInputSync(source, channel, "text", map[string]interface{}{"content": text, "no_memory": true})
}
func (s *PluginSDK) OutputChan() <-chan *agentIO.OutputEvent {
if s.iom != nil {
return s.iom.OutputChan()
}
return nil
}
func (s *PluginSDK) RegisterChannel(name string, dev agentIO.Device) error {
if s.iom != nil {
return s.iom.RegisterDevice(dev)
}
return nil
}
func (s *PluginSDK) UnregisterChannel(name string) {
if s.iom != nil {
s.iom.UnregisterDevice(name)
}
}
func (s *PluginSDK) ListChannels() []agentIO.ChannelInfo {
if s.iom != nil {
return s.iom.ListChannels()
}
return nil
}
func (s *PluginSDK) Publish(evt *events.Event) {
if s.eventBus != nil {
s.eventBus.Publish(evt)
}
}
func (s *PluginSDK) Subscribe(eventType events.EventType, handler events.Handler) func() {
if s.eventBus != nil {
return s.eventBus.Subscribe(eventType, handler)
}
return func() {}
}
// SetToolBlocks 桥接到 IOManager插件工具注入多模态块process.go 消费。
func (a ioAdapter) SetToolBlocks(blocks []pubsdk.ContentBlock) {
if a.iom == nil {
return
}
ifaces := make([]interface{}, len(blocks))
for i, b := range blocks {
ifaces[i] = b
}
a.iom.SetToolBlocks(ifaces)
}
// SetToolBlocks 注入多模态内容块(图片/音频),内核在下一条 tool message
// 的 content 数组里带上这些块,让模型在后续轮次看到图/听到音频。
func (s *PluginSDK) SetToolBlocks(blocks []pubsdk.ContentBlock) {
if s.iom != nil {
ifaces := make([]interface{}, len(blocks))
for i, b := range blocks {
ifaces[i] = b
}
s.iom.SetToolBlocks(ifaces)
}
}