feat(sdk): 多模态贯通插件边界——媒体字段、媒体注入接口与并发修复

记忆系统在核心 1.1.0 支持了二进制多媒体节点,但那条链路只对**内核自己**开放:
插件把 Triple / Doc 交进来,媒体一律无处安放,且**不报错**。本版补上公开接口
侧缺失的表达能力。

## 一、类型与接口(全部新增,无签名变更)

- `Triple` += `SentenceText`、`MediaDigests`
- `Doc` += `MediaDigests`、`Attachments`;新增 `MediaAttachment`
- `TextEvent` += `Attachments`
- `DocMemoryAPI` += `InsertWithMedia`
- `IOInjector` += `InjectInputMedia` / `InjectInputMediaSync` / `InjectInterruptMedia`
- `PluginSDK` 补上一直缺失的 `SetToolBlocks` 包装(接口里有、便捷方法里没有,
  插件只能自己去拿 injector)

`MediaAttachment` 一个类型服务两个方向:给 `Data`+`MIME` 是新内容(内核按字节
去重),只给 `Digest` 是引用已有内容。读路径**只回元数据不回字节**——一次检索
可能命中几十份媒体,把字节全塞回来会撑爆跨进程消息。

媒体注入为什么不能搭 `SetToolBlocks` 的车:那个方法只在工具处理函数内部可用,
且媒体要等**下一条** tool message 才到模型手上。插件主动发起一轮带媒体的对话、
以及中断注入,需要各自的签名,且媒体在**本轮**就随消息发出。

`Triple.MediaDigests` 非空而 `SentenceText` 为空时,内核会用媒体标记本身充当句子
——媒体引用挂在句子上,没有句子就无处挂起。插件只需填 digest,标记由内核拼:
要求调用方知道格式,等于让一个拼写错误静默切断引用绑定而全链路无人报错。

## 二、修掉两处并发竞态

`sdk/stress_test.go` 的 `-race` 实测报 11 处 DATA RACE,收敛到两个字段:

1. **`PluginSDK` 的 API 字段无锁**。写方是内核(加载/重载插件时依次注入
   injector、memory、doc、llm…),读方是插件在 `Start()` 里起的后台 goroutine
   ——轮询、监听、定时器都要拿 injector 往管道注消息。生产表现是插件重载瞬间
   偶发崩溃:读到半个接口值就 nil 解引用。
2. **`autoRestart` 标志无锁**。`SetAutoRestart` 的文档用法本身就是「外部连接建好
   后再决定能否自动重启」,而连接建立通常在后台 goroutine;内核 registry 在另一个
   goroutine 读 `AutoRestart()` 决定崩溃后重启策略。这对读写天然跨 goroutine。

加 `apiMu sync.RWMutex`。关键约定写进注释:**只在持锁期间取字段值,取完立刻
释放再调用**。持锁调用会把 `InjectInputSync`(阻塞到 agent 回复,可达数分钟)
与 `SetIOInjector` 串到一起,让插件重载卡死。

## 三、压测(sdk/stress_test.go,13 例)

SDK 是被多个 goroutine 同时使用的共享对象,单线程单测全绿不代表并发路径成立。
断言的是不变量而非吞吐:

- 媒体注入高并发不丢不串——每次调用带唯一 tag,逐条校验文本与图片 URL 配对。
  「不串」是重点:若实现里出现任何共享中间状态(把 blocks 暂存到字段再读出),
  高并发下会出现 A 的文本配 B 的图,而两者单独看都「成功」了;
- injector 热替换(含替换成 nil,即内核卸载 API 的真实状态);
- stop / onRemove handler 恰好一次——契约是「执行后清空,幂等」,执行两次的后果
  从重复写文件到 close 已关闭 channel 直接 panic;
- `StageContext` 并发读改写无 lost update(媒体链路让 Extra 成为新热点,
  而 map 并发写在 Go 里是直接 fatal,recover 接不住);
- `OwnTools` scope 不跨插件泄漏;
- 媒体类型 JSON 往返字节级一致(9 种长度,含 0/1/2/3 与 base64 分组边界)
  ——`[]byte` 在 JSON 里是 base64,往返不一致意味着图片静默损坏,
  要到 CAS 校验 digest 时才发现,那时已无从追查;
- `omitempty` 真的生效(读路径不能出现 `"data"` 键);
- nil 依赖全部静默降级不 panic。

## 四、工具链同步

- `proc_main.go.tmpl`:`procIO` 三个媒体方法、`procDocMemory.InsertWithMedia`。
  模板不跟上的后果是**每个外部插件都编不过**(接口未实现),是硬失败;
- `proc_runtime_test.go`:方法清单补 `io.injectMedia*` 与 `doc.insertWithMedia`。
  漏接线时插件调 `InjectInputMedia` 会静默无效果——模板不发这个 RPC,内核也就
  收不到,两边都不报错;
- `yaegi/mocksdk`:与公开 SDK 对齐。它此前漂移严重且**没有任何代码对着它编译**,
  所以漂移不会被编译器抓到:`Triple` 用的是 `Predicate`,而公开 SDK 一直叫
  `Relation` —— 插件在 yaegi 调试期写 `Relation:` 报未知字段,写 `Predicate:` 则
  编成 plugin.bin 时报错,两边都不对。
- README 中英双语补媒体接口文档与用法示例。

## 兼容性

存量插件不需要改一行也不需要重编:新增方法由**插件调用、内核实现**,不调就不
受影响。17 个 example 插件源码零改动通过类型检查;用 SDK 0.9.2 编的旧 plugin.bin
在新内核上直接建链通过(握手校验的是 ProtocolVersion=1,不是 SDK 版本)。

媒体接口需要核心 1.1.1+(更早的核心没有对应 RPC,调用返回 unknown method)。
`CoreVersion` 保持 1.0.0:它是「SDK 能在其上运行」的下限,媒体是可选能力。
This commit is contained in:
JianFeeeee
2026-09-06 10:03:36 +08:00
parent e256023399
commit ce5bff9275
9 changed files with 1339 additions and 97 deletions

View File

@ -36,6 +36,7 @@ The SDK instance injected via `Start(sdk *PluginSDK)` provides:
| Settings | `Settings()` | Access settings API |
| Events | `Events()` | Access event subscriber (subscribe-only for external plugins) |
| Inject | `InjectText(source, channel, text)` / `InjectInterruptText(source, channel, text)` / `InjectTextNoMemory(source, channel, text)` | Inject text into the agent pipeline |
| Media inject | `InjectInputMedia(source, channel, text, blocks)` / `InjectInputMediaSync(...)` / `InjectInterruptMedia(...)` | Inject input carrying images/audio (added in 1.1.0) |
| Auto-Restart | `SetAutoRestart(enabled)` / `AutoRestart()` | Control automatic restart on crash |
### Stage Hooks
@ -106,6 +107,32 @@ Type enum values:
| `InjectInterruptText(source, channel, text)` | Inject interrupt text, interrupt current processing, route to specified channel |
| `InjectTextNoMemory(source, channel, text)` | Inject text without memory recording, route to specified channel |
### Multimodal Injection (added in 1.1.0)
| Method | Description |
|--------|-------------|
| `InjectInputMedia(source, channel, text, blocks)` | Inject media-bearing input, asynchronous |
| `InjectInputMediaSync(source, channel, text, blocks)` | Inject media-bearing input and wait for the reply text |
| `InjectInterruptMedia(source, channel, text, blocks)` | Inject a media-bearing interrupt that can preempt current processing |
`blocks` is `[]sdk.ContentBlock`, the same type `SetToolBlocks` takes:
```go
s.InjectInputMedia("myplugin", "webui", "take a look at this", []sdk.ContentBlock{{
Type: "image_url",
ImageURL: &sdk.ImageURL{URL: "data:image/png;base64," + b64, Detail: "auto"},
}})
```
How this differs from `SetToolBlocks`: that one is only callable inside a tool handler and
its media reaches the model with the *next* tool message. These three let a plugin
**initiate a turn that carries media** — the media goes out with this turn's message and is
automatically stored in the media store with a memory reference attached.
`data:` URLs in the blocks are stored and deduplicated by the kernel; `http(s)` URLs are
passed to the model only and never stored (storing them would require the kernel to make
network requests, bringing timeouts, auth and SSRF into scope).
`source` identifies the origin, `channel` specifies the target output channel.
### Triple Extended Fields
@ -115,6 +142,65 @@ The Triple data structure includes additional fields:
- `Confidence` — confidence score (0.0–1.0)
- `SubjectType` — subject type
- `ObjectType` — object type
- `SentenceText` — the original sentence (added in 1.1.0), written to the `sentences` table; media references hang off the sentence
- `MediaDigests` — associated media digests (added in 1.1.0)
### Media in Memory (added in 1.1.0)
Inside plain-text memory, media is represented as a **marker** of the form
`[<mime> <short digest>] <description>`:
```
[image/png a1b2c3d4e5f6] a purple-blue-red three-band chart
```
The description is the durable semantic memory (retrieval uses it); the digest is the key
back to the bytes (reverse lookup uses it). Markers are generated by the kernel — a plugin
never has to assemble one, it just **supplies the digest**.
#### Graph memory
```go
s.Memory().Commit([]sdk.Triple{{
Subject: "palette", Relation: "contains", Object: "three-band",
MediaDigests: []string{"a1b2c3d4e5f6"}, // short digest is fine, the kernel resolves it
}})
```
With no `SentenceText`, the kernel uses the marker itself as the sentence — media must have
a sentence to hang off, otherwise the reference has nowhere to attach.
#### Knowledge base
```go
s.DocMemory().InsertWithMedia(&sdk.Doc{
Title: "illustrated note",
Content: "body",
}, []sdk.MediaAttachment{
{MIME: "image/png", Data: pngBytes, Name: "chart.png"}, // new content, stored and deduped
{Digest: "a1b2c3d4e5f6"}, // reference existing content
})
```
`Insert` keeps its original signature; markers already present in the body are bound as
document-level references too. `Query` fills `MediaDigests` and `Attachments` (mime plus
description, **no bytes** — one query can match dozens of media items). Removing a document
releases its references.
#### Text memory
```go
s.TextMemory().Append(sdk.TextEvent{
Role: "user", Content: "look at this",
Attachments: []sdk.MediaAttachment{{MIME: "image/png", Data: pngBytes}},
})
```
`RecentEvents` decodes markers in the body back into `Attachments`.
The media store can be disabled kernel-side (`core.memory.media.enabled=false`); all of the
above then degrades to plain-text behaviour — no errors, no panics, identical to how it
behaved before this feature shipped.
### ToolDef Field Reference