refactor(stream): move stream contracts out of adapters

docs/architecture/core-decomposition.md

@@ -226,15 +226,15 @@ flowchart TB
 
 ### 7.3 适配层（Adapters）
 
-适配层负责协议、transport、外部 provider 和宿主通信转换，物理位置是 `src/crates/adapters`。其中 `ai-adapters` 负责 AI provider 请求/响应与 stream 协议，`api-layer` 负责产品宿主共用的后端 API adapter，`transport` 负责事件投递和 host transport adapter，`webdriver` 负责 WebDriver 协议和浏览器自动化 adapter。适配层不拥有产品能力选择，也不承载可复用 OS service 实现。
+适配层负责协议、transport、外部 provider 和宿主通信转换，物理位置是 `src/crates/adapters`。其中 `ai-adapters` 负责 AI provider 请求/响应映射和 provider stream 协议解析，解析结果应转换为 execution 层拥有的统一 stream 契约；`api-layer` 负责产品宿主共用的后端 API adapter，`transport` 负责事件投递和 host transport adapter，`webdriver` 负责 WebDriver 协议和浏览器自动化 adapter。适配层不拥有产品能力选择，也不承载可复用 OS service 实现。
 
 ### 7.4 服务实现层（Services）
 
 服务实现层负责接触本地系统和 runtime infrastructure 的可复用具体实现，物理位置是 `src/crates/services`。其中 `services-core` 承载轻量 service primitive，`services-integrations` 承载 MCP、Git、remote、file watch 和产品领域 port 的具体实现，`terminal` 承载 PTY、shell integration 和 terminal session infrastructure。服务实现层可以实现 `contracts`、`execution` 或 `product-domains` 定义的 port，但不选择产品 profile，也不直接暴露 UI/协议入口。
 
 ### 7.5 执行原语层（Execution Primitives）
 
-执行原语层提供 provider-neutral 的 runtime building blocks，物理位置是 `src/crates/execution`。`agent-runtime`、`agent-stream`、`harness`、`runtime-services`、`tool-contracts`、`tool-provider-groups` 和 `tool-execution` 分别定义 agent loop facts、stream normalization、workflow descriptor、typed service bundle、tool manifest / permission / result policy、tool group facts 和低层 tool execution helper。当前 Cargo package / lib 名保持兼容，但物理目录按职责命名。它们只能依赖稳定契约或明确的 provider-neutral DTO，不直接创建 Tauri handle、filesystem manager、Git provider、MCP client、AI client 或 host process。
+执行原语层提供 provider-neutral 的 runtime building blocks，物理位置是 `src/crates/execution`。`agent-runtime`、`agent-stream`、`harness`、`runtime-services`、`tool-contracts`、`tool-provider-groups` 和 `tool-execution` 分别定义 agent loop facts、统一 stream DTO / tool-call 累积 / replay 契约、workflow descriptor、typed service bundle、tool manifest / permission / result policy、tool group facts 和低层 tool execution helper。当前 Cargo package / lib 名保持兼容，但物理目录按职责命名。它们只能依赖稳定契约或明确的 provider-neutral DTO，不直接创建 Tauri handle、filesystem manager、Git provider、MCP client、AI client 或 host process。
 
 ### 7.6 稳定契约与产品领域层（Stable Contracts and Product Domains）
 

docs/plans/core-decomposition-completed.md

@@ -43,6 +43,7 @@
 - `src/crates` 已按六层物理布局整理：`interfaces/`、`assembly/`、`adapters/`、`services/`、`execution/`、`contracts/`。
 - 旧 `surfaces` 和 `providers` 目标层级已被移除：协议入口归入 `interfaces`，协议/transport/provider 转换归入 `adapters`，OS/runtime infrastructure 具体实现归入 `services`。
 - execution 下 tool 相关目录已按职责命名：`tool-contracts`、`tool-provider-groups`、`tool-execution`。Cargo package / lib 名保持兼容。
+- `agent-stream` 已成为统一 stream DTO、tool-call 累积和 replay 契约 owner；provider stream 解析测试归属 `ai-adapters`。
 - AGENTS、README、DeepReview path classifier、core boundary rules 和 Cargo workspace path 已同步到当前分层。
 
 ## 2. 已建立的保护

docs/plans/core-decomposition-plan.md

@@ -23,6 +23,7 @@
 - Desktop / CLI / ACP 仍通过 `bitfun-core/product-full` 获取完整能力；Server / Web / Mobile Web 不直接依赖 core，但尚未完成按交付形态裁剪最小 feature / dependency。
 - `runtime-services` 已有 typed builder、capability availability 和 core product runtime provider adapter，但不少 concrete provider 仍在 core 创建或持有。
 - `tool-contracts` 已承接 provider-neutral tool manifest、admission、catalog、result policy 等纯策略；`tool-execution` 只承接部分低层 IO/search helper；Bash、terminal lifecycle、indexed search、remote shell、permission wait、checkpoint orchestration 和完整 execution pipeline 仍未完全迁移。
+- `agent-stream` 已承接统一 stream DTO、tool-call 累积和 replay 契约；provider SSE / 响应解析测试归属 `ai-adapters`。
 - `harness` 当前主要承接 descriptor / route plan / registry contract；Deep Review、DeepResearch、MiniApp 的 concrete workflow execution 仍在 core 或产品路径。
 - `product-domains` 已承接 MiniApp / function-agent 的部分纯领域逻辑；worker、host side effect、AI acquisition、marker IO 等 concrete path 仍未完成 owner 迁移。
 

src/crates/adapters/ai-adapters/AGENTS.md

@@ -3,8 +3,8 @@
 Scope: this guide applies to `src/crates/adapters/ai-adapters`.
 
 `bitfun-ai-adapters` owns provider-specific request/response mapping and stream
-normalization. Keep provider quirks here instead of leaking them into core tool
-contracts or product runtime logic.
+protocol parsing. Keep provider quirks here, then convert stream chunks into the
+provider-neutral contracts owned by `bitfun-agent-stream`.
 
 ## Guardrails
 
@@ -16,6 +16,8 @@ contracts or product runtime logic.
   count.
 - Do not change shared stream or usage semantics without updating the focused
   adapter tests and downstream usage expectations.
+- Do not move provider-neutral stream DTOs, replay policy, or tool-call
+  accumulation ownership back into this crate.
 
 ## Verification
 

src/crates/adapters/ai-adapters/Cargo.toml

@@ -11,6 +11,7 @@ crate-type = ["rlib"]
 
 [dependencies]
 anyhow = { workspace = true }
+bitfun-agent-stream = { path = "../../execution/agent-stream" }
 bitfun-core-types = { path = "../../contracts/core-types" }
 chrono = { workspace = true }
 eventsource-stream = { workspace = true }
@@ -22,3 +23,9 @@ serde_json = { workspace = true }
 tokio = { workspace = true }
 tokio-stream = { workspace = true }
 urlencoding = { workspace = true }
+
+[dev-dependencies]
+async-trait = { workspace = true }
+axum = { workspace = true }
+bitfun-events = { path = "../../contracts/events" }
+tokio-util = { workspace = true }

src/crates/adapters/ai-adapters/README.md

@@ -6,8 +6,7 @@ This crate owns the portable AI integration layer:
 
 - provider request building
 - provider-specific message conversion
-- SSE / stream parsing
-- streamed tool-call aggregation
+- SSE / stream parsing into provider-neutral stream contracts
 - shared AI-facing transport types
 - provider model discovery
 - connection health checks
@@ -26,8 +25,8 @@ and re-exports this crate where convenient.
 
 - `client`: shared HTTP transport, retries, aggregation, health checks
 - `providers`: OpenAI / Anthropic / Gemini request and discovery adapters
-- `stream`: provider SSE parsing into unified streaming events
-- `tool_call_accumulator`: reconstruct structured tool calls from streamed deltas
+- `stream`: provider SSE parsing into unified streaming events from `bitfun-agent-stream`
+- `tool_call_accumulator`: compatibility re-export; canonical implementation lives in `bitfun-agent-stream`
 - `types`: portable request/response/config/message types
 
 ## Design Rule

src/crates/adapters/ai-adapters/src/stream/types/unified.rs

@@ -1,81 +1 @@
-use serde::{Deserialize, Serialize};
-use serde_json::Value;
-use std::borrow::Cow;
-use std::fmt;
-
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct UnifiedToolCall {
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub tool_call_index: Option<usize>,
-    pub id: Option<String>,
-    pub name: Option<String>,
-    pub arguments: Option<String>,
-    #[serde(default)]
-    pub arguments_is_snapshot: bool,
-}
-
-/// Unified AI response format
-#[derive(Clone, Serialize, Deserialize, Default)]
-pub struct UnifiedResponse {
-    pub text: Option<String>,
-    pub reasoning_content: Option<String>,
-    /// Signature for Anthropic extended thinking (returned in multi-turn conversations)
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub thinking_signature: Option<String>,
-    pub tool_call: Option<UnifiedToolCall>,
-    pub usage: Option<UnifiedTokenUsage>,
-    pub finish_reason: Option<String>,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub provider_metadata: Option<Value>,
-}
-
-impl fmt::Debug for UnifiedResponse {
-    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> fmt::Result {
-        let reasoning_summary = self.reasoning_content.as_ref().map(|s| {
-            if s.len() > 100 {
-                let end = s
-                    .char_indices()
-                    .take_while(|(i, _)| *i < 100)
-                    .last()
-                    .map(|(i, c)| i + c.len_utf8())
-                    .unwrap_or(0);
-                // Guard against multi-byte chars pushing end past the string length
-                let end = end.min(s.len());
-                Cow::Owned(format!("{}... ({} bytes)", &s[..end], s.len()))
-            } else {
-                Cow::Borrowed(s.as_str())
-            }
-        });
-        f.debug_struct("UnifiedResponse")
-            .field("text", &self.text)
-            .field("reasoning_content", &reasoning_summary)
-            .field("thinking_signature", &"<omitted>")
-            .field("tool_call", &self.tool_call)
-            .field("usage", &self.usage)
-            .field("finish_reason", &self.finish_reason)
-            .field("provider_metadata", &"<omitted>")
-            .finish()
-    }
-}
-
-/// Unified token usage statistics
-#[derive(Debug, Clone, Serialize, Deserialize)]
-pub struct UnifiedTokenUsage {
-    pub prompt_token_count: u32,
-    pub candidates_token_count: u32,
-    pub total_token_count: u32,
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub reasoning_token_count: Option<u32>,
-    /// Cache READ tokens (i.e., served from cache this call). Universal across
-    /// providers: OpenAI `cached_tokens`, DeepSeek `prompt_cache_hit_tokens`,
-    /// Anthropic `cache_read_input_tokens`, Gemini `cachedContentTokenCount`.
-    /// Hit rate consumers must use this as numerator and `prompt_token_count`
-    /// as denominator.
-    #[serde(skip_serializing_if = "Option::is_none")]
-    pub cached_content_token_count: Option<u32>,
-    /// Cache WRITE tokens (only Anthropic reports this per-token; others either
-    /// have no creation concept or bill creation by storage time). Disjoint from
-    /// `cached_content_token_count`. Do NOT include in hit-rate numerator.
-    #[serde(skip_serializing_if = "Option::is_none", default)]
-    pub cache_creation_token_count: Option<u32>,
-}
+pub use bitfun_agent_stream::{UnifiedResponse, UnifiedTokenUsage, UnifiedToolCall};