mirror of
https://gitcode.com/JianFeeeee/homeagent-sdk.git
synced 2026-10-03 23:54:12 +00:00
feat(sdk): 注入行为的记忆/裁剪标志位(纯追加)+ plugindev 退出码修复 + 示例 hmap 随发版
## 1. 注入标志位(公开 API 纯追加,无签名变更)
给注入行为补上工具早已有的两类声明位,并让通道定义也带上:
- `InjectOptions{NoMemory, ContextPolicy, CleanerName}`
- `IOInjector` 新增六个 `*Opts` 变体(排队/中断/同步 × 纯文本/带媒体)
- `ChannelDef.ContextPolicy`,并**补上 JSON tag**(Cleaner 标 `json:"-"`)
零值 InjectOptions 与旧的三参数方法完全等价(记入记忆 + 不裁剪),
存量插件不需要改一行、也不需要重编;旧方法保留为转发到零值 opts 的语法糖。
三条设计要点:
- **默认不裁剪**:裁剪会归档丢弃低相关事件,必须显式声明(ContextPolicy=prune)。
- **中断也允许声明 prune**(已确认):中断同样携带内容进上下文。
- `CleanerName`:注入的 source 未必是注册过的输入通道名,而注入内容常带
ANSI/JSON 包装;允许显式指定用哪个已注册 cleaner 清洗。
顺带修掉一个易静默丢字段的坑:`ChannelDef` 原来没有 JSON tag,只能手写字段
白名单跨进程传(`{"NoMemory": ...}`),新增字段会被丢掉。现在模板整体传 `def`。
## 2. 示例调用点统一写明意图
rss/memo/calendar/qq 的中断注入显式 `NoMemory: true`(行为等价,写清语义)。
## 3. plugindev 出错却 exit 0(真缺陷)
`buildBundle`/`buildTarget` 遇错只 Printf 后 return,`cmdBuild` 返回 void,
于是**构建失败也退 0**。实测中一个示例的 windows 目标编译失败,批量脚本却报
「17/17 全绿」,并因此少产出 16 个 .hmap。现在累计 `buildFailed` 并以非零退出。
## 4. 平台策略:插件目标去掉 windows
homed 已放弃 Windows 原生(见核心仓 cmd/homed/platform_windows.go:插件体系依赖
fd 继承 + 统一共享内存区的段内偏移解引用,Windows 句柄模型无法表达),
插件只运行在 homed 能跑的平台上,故 `allBundleTargets` 去掉 windows,
并对 windows 目标给出**可执行的报错**(指引 WSL2),而不是让它死在一句
`undefined: attachUnifiedShm` 上。
## 5. 发版附带各示例插件的 .hmap
新增 `package/build-examples.sh` 并接入 `package/build.sh`(组件 all|plugindev|examples):
- 用**刚构建出来的**那把工具链构建示例,保证与本次发版同源
- 逐平台 `--no-bundle --target <os/arch>`(bundle 会连 windows 一起编)
- 判成功同时看**退出码 + 产物存在**
- 全部产物齐了才 `sha256sum`(边打边算会漏掉后生成的包)
- 有任一失败即整体失败,不生成 SHA256SUMS
## 6. 版本
SDK 仍为 1.2.0(main 是下一个未发布中版本);1.2.0 条目补记本次新增接口,
并注明新标志位需核心 1.2.0+(旧核心会忽略这些字段,不报错但不生效)。
This commit is contained in:
144
sdk/plugin.go
144
sdk/plugin.go
@ -35,12 +35,63 @@ const (
|
||||
StageAfterOutput Stage = "after_output"
|
||||
)
|
||||
|
||||
// 上下文策略:决定一次工具调用/输入/注入是否依据其内容裁剪上下文。
|
||||
//
|
||||
// 默认(空串或 ContextPolicyNone)**不裁剪**:裁剪会归档丢弃低相关事件,
|
||||
// 必须由工具/通道/注入点显式声明才发生——否则一个只想往上下文里塞内容的
|
||||
// 插件会在背后把别人的内容挤掉,且看不出是谁干的。
|
||||
const (
|
||||
ContextPolicyNone = "none"
|
||||
ContextPolicyPrune = "prune"
|
||||
)
|
||||
|
||||
// ValidContextPolicy 校验策略取值;空串等价于 ContextPolicyNone。
|
||||
func ValidContextPolicy(policy string) bool {
|
||||
switch policy {
|
||||
case "", ContextPolicyNone, ContextPolicyPrune:
|
||||
return true
|
||||
}
|
||||
return false
|
||||
}
|
||||
|
||||
// InjectOptions 声明一次注入行为在记忆层与上下文层的表现。
|
||||
//
|
||||
// 零值 = 记入记忆 + 不裁剪上下文,与历史行为(三参数注入方法)完全一致,
|
||||
// 因此调用方只有在确实需要改变行为时才需要填它。
|
||||
//
|
||||
// 为什么注入也要这两个标志:注入的内容来源千差万别——轮询到的频道消息
|
||||
// 属于真实对话(该记),而“任务还在跑”“连接已重连”这类提醒不该污染记忆,
|
||||
// 也不该把上下文按它的内容裁一遍。按调用点声明比按通道一刀切准确。
|
||||
//
|
||||
// NoMemory: 此次注入不参与记忆计算(向量化/关键词提取/蒸馏),原文仍留在上下文
|
||||
// ContextPolicy: 此次注入后是否依据(清洗后的)内容裁剪上下文;默认不裁剪。
|
||||
//
|
||||
// 中断注入也允许声明 prune——它同样会携带内容进入上下文。
|
||||
//
|
||||
// CleanerName: 此次注入的内容用哪个**已注册的通道 cleaner** 清洗。
|
||||
//
|
||||
// 空串 = 按注入的 source 查通道定义(既有行为)。
|
||||
// 为什么要能显式指定:注入的 source 未必是注册过的输入通道名,
|
||||
// 而注入内容往往带 ANSI/JSON 包装,需要清洗后才是有效内容;
|
||||
// 不指定就只能退到「按 source 查不到就不清洗」。
|
||||
type InjectOptions struct {
|
||||
NoMemory bool
|
||||
ContextPolicy string
|
||||
CleanerName string
|
||||
}
|
||||
|
||||
// ChannelDef 描述通道在记忆计算层的行为,与 ToolDef.NoMemory/Cleaner 语义一致。
|
||||
// NoMemory: 此通道输入/输出不参与记忆计算(向量化/关键词提取/蒸馏),但原文保留在上下文中
|
||||
// Cleaner: 计算层过滤函数,不改原文;仅在向量化/jieba/蒸馏/存档提取关键词时调用
|
||||
// ContextPolicy: 此通道的输入到达后是否据此裁剪上下文,默认 none(不裁剪)
|
||||
//
|
||||
// JSON tag 是必需的:通道定义要跨进程传给内核,而 Cleaner 是函数(必须忽略)。
|
||||
// 没有 tag 时既无法整体 marshal(func 不支持),又会诱使调用方手写字段白名单——
|
||||
// 那样新增字段会被静默丢掉。
|
||||
type ChannelDef struct {
|
||||
NoMemory bool
|
||||
Cleaner func(string) string
|
||||
NoMemory bool `json:"no_memory,omitempty"`
|
||||
Cleaner func(string) string `json:"-"`
|
||||
ContextPolicy string `json:"context_policy,omitempty"`
|
||||
}
|
||||
|
||||
// StageContext provides context for stage handlers.
|
||||
@ -106,7 +157,7 @@ type ToolDef struct {
|
||||
Parameters map[string]interface{} `json:"parameters"`
|
||||
NoMemory bool `json:"no_memory,omitempty"` // 此工具输出不参与记忆计算,但原文保留
|
||||
Cleaner func(string) string `json:"-"` // 计算层过滤函数,不改原文;仅在向量化/jieba/蒸馏时调用
|
||||
ContextPolicy string `json:"context_policy,omitempty"` // 工具上下文策略:"none"(默认) / "prune"
|
||||
ContextPolicy string `json:"context_policy,omitempty"` // 上下文策略:""(默认,不裁剪) / ContextPolicyNone / ContextPolicyPrune
|
||||
}
|
||||
|
||||
// IOInjector provides methods for injecting input and interrupts into the agent pipeline.
|
||||
@ -125,6 +176,17 @@ type IOInjector interface {
|
||||
InjectInputMedia(source, channel, text string, blocks []ContentBlock)
|
||||
InjectInputMediaSync(source, channel, text string, blocks []ContentBlock) string
|
||||
InjectInterruptMedia(source, channel, text string, blocks []ContentBlock)
|
||||
|
||||
// 以下 Opts 变体让调用点在**这一次注入**上声明记忆与裁剪行为。
|
||||
//
|
||||
// 上面那些不带 opts 的方法等价于传零值 InjectOptions(记入记忆 + 不裁剪),
|
||||
// 保留它们是为了不破坏已有插件;新代码应当用 Opts 变体把意图写清楚。
|
||||
InjectTextOpts(source, channel, text string, opts InjectOptions)
|
||||
InjectInterruptTextOpts(source, channel, text string, opts InjectOptions)
|
||||
InjectInputSyncOpts(source, channel, text string, opts InjectOptions) string
|
||||
InjectInputMediaOpts(source, channel, text string, blocks []ContentBlock, opts InjectOptions)
|
||||
InjectInputMediaSyncOpts(source, channel, text string, blocks []ContentBlock, opts InjectOptions) string
|
||||
InjectInterruptMediaOpts(source, channel, text string, blocks []ContentBlock, opts InjectOptions)
|
||||
}
|
||||
|
||||
// EventType identifies the kind of system event.
|
||||
@ -505,54 +567,94 @@ func (s *PluginSDK) injector() IOInjector {
|
||||
}
|
||||
|
||||
// InjectInterruptText injects a text interrupt that can preempt current LLM processing.
|
||||
// 等价于 InjectInterruptTextOpts(..., InjectOptions{}):记入记忆、不裁剪。
|
||||
func (s *PluginSDK) InjectInterruptText(source, channel, text string) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectInterruptText(source, channel, text)
|
||||
}
|
||||
s.InjectInterruptTextOpts(source, channel, text, InjectOptions{})
|
||||
}
|
||||
|
||||
// InjectText injects a text message into the agent pipeline.
|
||||
// 等价于 InjectTextOpts(..., InjectOptions{}):记入记忆、不裁剪。
|
||||
func (s *PluginSDK) InjectText(source, channel, text string) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectText(source, channel, text)
|
||||
}
|
||||
s.InjectTextOpts(source, channel, text, InjectOptions{})
|
||||
}
|
||||
|
||||
// InjectTextNoMemory injects a text message without generating memory.
|
||||
// 等价于 InjectTextOpts(..., InjectOptions{NoMemory: true})。
|
||||
func (s *PluginSDK) InjectTextNoMemory(source, channel, text string) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectTextNoMemory(source, channel, text)
|
||||
}
|
||||
s.InjectTextOpts(source, channel, text, InjectOptions{NoMemory: true})
|
||||
}
|
||||
|
||||
// InjectInputSync injects a text message and synchronously waits for the agent reply,
|
||||
// returning the reply text (empty string if none). Replies must be dispatched back
|
||||
// to the source channel by the caller.
|
||||
func (s *PluginSDK) InjectInputSync(source, channel, text string) string {
|
||||
io := s.injector()
|
||||
if io == nil {
|
||||
return ""
|
||||
}
|
||||
return io.InjectInputSync(source, channel, text)
|
||||
return s.InjectInputSyncOpts(source, channel, text, InjectOptions{})
|
||||
}
|
||||
|
||||
// InjectInputMedia 注入带媒体内容块(image_url/audio_url)的输入。
|
||||
// blocks 会落进媒体存储被记忆引用捕获,同时作为当前轮 content 数组
|
||||
// 发给 LLM,让模型在「本轮」就看到图/听到音频——区别于 SetToolBlocks
|
||||
// 的「下一轮 tool message」语义。
|
||||
// 等价于 InjectInputMediaOpts(..., InjectOptions{})。
|
||||
func (s *PluginSDK) InjectInputMedia(source, channel, text string, blocks []ContentBlock) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectInputMedia(source, channel, text, blocks)
|
||||
}
|
||||
s.InjectInputMediaOpts(source, channel, text, blocks, InjectOptions{})
|
||||
}
|
||||
|
||||
// InjectInputMediaSync 注入带媒体内容块的输入并同步等待 agent 回复。
|
||||
// 等价于 InjectInputMediaSyncOpts(..., InjectOptions{})。
|
||||
func (s *PluginSDK) InjectInputMediaSync(source, channel, text string, blocks []ContentBlock) string {
|
||||
return s.InjectInputMediaSyncOpts(source, channel, text, blocks, InjectOptions{})
|
||||
}
|
||||
|
||||
// ---- 带 InjectOptions 的注入(声明记忆/裁剪行为)----
|
||||
|
||||
// InjectTextOpts 注入文本到 agent,并在这一次注入上声明记忆与裁剪行为。
|
||||
func (s *PluginSDK) InjectTextOpts(source, channel, text string, opts InjectOptions) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectTextOpts(source, channel, text, opts)
|
||||
}
|
||||
}
|
||||
|
||||
// InjectInterruptTextOpts 注入可抢占当前处理的中断文本。
|
||||
//
|
||||
// 中断也允许声明 ContextPolicyPrune:中断同样携带内容进入上下文,
|
||||
// 是否需要据此裁剪由调用方决定(默认不裁剪)。
|
||||
func (s *PluginSDK) InjectInterruptTextOpts(source, channel, text string, opts InjectOptions) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectInterruptTextOpts(source, channel, text, opts)
|
||||
}
|
||||
}
|
||||
|
||||
// InjectInputSyncOpts 注入输入并同步等待回复,同时在这次注入上声明记忆/裁剪行为。
|
||||
func (s *PluginSDK) InjectInputSyncOpts(source, channel, text string, opts InjectOptions) string {
|
||||
io := s.injector()
|
||||
if io == nil {
|
||||
return ""
|
||||
}
|
||||
return io.InjectInputMediaSync(source, channel, text, blocks)
|
||||
return io.InjectInputSyncOpts(source, channel, text, opts)
|
||||
}
|
||||
|
||||
// InjectInputMediaOpts 注入带媒体块的输入,并声明记忆/裁剪行为。
|
||||
func (s *PluginSDK) InjectInputMediaOpts(source, channel, text string, blocks []ContentBlock, opts InjectOptions) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectInputMediaOpts(source, channel, text, 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 (s *PluginSDK) InjectInputMediaSyncOpts(source, channel, text string, blocks []ContentBlock, opts InjectOptions) string {
|
||||
io := s.injector()
|
||||
if io == nil {
|
||||
return ""
|
||||
}
|
||||
return io.InjectInputMediaSyncOpts(source, channel, text, blocks, opts)
|
||||
}
|
||||
|
||||
// InjectInterruptMediaOpts 注入带媒体块的中断,并声明记忆/裁剪行为。
|
||||
func (s *PluginSDK) InjectInterruptMediaOpts(source, channel, text string, blocks []ContentBlock, opts InjectOptions) {
|
||||
if io := s.injector(); io != nil {
|
||||
io.InjectInterruptMediaOpts(source, channel, text, blocks, opts)
|
||||
}
|
||||
}
|
||||
|
||||
// InjectInterruptMedia 注入带媒体内容块的中断,可抢占当前 LLM 处理。
|
||||
|
||||
@ -43,6 +43,7 @@ type injectCall struct {
|
||||
channel string
|
||||
text string
|
||||
blocks []ContentBlock
|
||||
opts InjectOptions // 调用点声明的记忆/裁剪行为
|
||||
}
|
||||
|
||||
func (r *recordingInjector) record(c injectCall) {
|
||||
@ -93,6 +94,41 @@ func (r *recordingInjector) InjectInterruptMedia(s, c, t string, b []ContentBloc
|
||||
r.record(injectCall{kind: "interruptMedia", source: s, channel: c, text: t, blocks: b})
|
||||
}
|
||||
|
||||
// ---- 带 InjectOptions 的注入:记录 opts 以便测试断言标志位确实传到了内核 ----
|
||||
|
||||
func (r *recordingInjector) InjectTextOpts(s, c, t string, o InjectOptions) {
|
||||
r.nText.Add(1)
|
||||
r.record(injectCall{kind: "textOpts", source: s, channel: c, text: t, opts: o})
|
||||
}
|
||||
|
||||
func (r *recordingInjector) InjectInterruptTextOpts(s, c, t string, o InjectOptions) {
|
||||
r.nInterrupt.Add(1)
|
||||
r.record(injectCall{kind: "interruptTextOpts", source: s, channel: c, text: t, opts: o})
|
||||
}
|
||||
|
||||
func (r *recordingInjector) InjectInputSyncOpts(s, c, t string, o InjectOptions) string {
|
||||
r.nSync.Add(1)
|
||||
r.record(injectCall{kind: "syncOpts", source: s, channel: c, text: t, opts: o})
|
||||
return "reply:" + t
|
||||
}
|
||||
|
||||
func (r *recordingInjector) InjectInputMediaOpts(s, c, t string, b []ContentBlock, o InjectOptions) {
|
||||
r.nMedia.Add(1)
|
||||
r.record(injectCall{kind: "mediaOpts", source: s, channel: c, text: t, blocks: b, opts: o})
|
||||
}
|
||||
|
||||
func (r *recordingInjector) InjectInputMediaSyncOpts(s, c, t string, b []ContentBlock, o InjectOptions) string {
|
||||
r.nMedia.Add(1)
|
||||
r.nSync.Add(1)
|
||||
r.record(injectCall{kind: "mediaSyncOpts", source: s, channel: c, text: t, blocks: b, opts: o})
|
||||
return "reply:" + t
|
||||
}
|
||||
|
||||
func (r *recordingInjector) InjectInterruptMediaOpts(s, c, t string, b []ContentBlock, o InjectOptions) {
|
||||
r.nMedia.Add(1)
|
||||
r.record(injectCall{kind: "interruptMediaOpts", source: s, channel: c, text: t, blocks: b, opts: o})
|
||||
}
|
||||
|
||||
func (r *recordingInjector) snapshot() []injectCall {
|
||||
r.mu.Lock()
|
||||
defer r.mu.Unlock()
|
||||
|
||||
Reference in New Issue
Block a user