Compare commits
16
Commits
9e4f50c955
...
9e476e79bb
| Author | SHA1 | Date | |
|---|---|---|---|
 徐涛
|
9e476e79bb | ||
|
徐涛
|
3bd135ec98 | ||
|
徐涛
|
76f3235ed7 | ||
|
徐涛
|
6315f2d008 | ||
|
徐涛
|
fba78f5f33 | ||
|
徐涛
|
3c1a3ee62e | ||
|
徐涛
|
8dcd1f2482 | ||
|
徐涛
|
602bd43fce | ||
|
徐涛
|
a52588cd0f | ||
|
徐涛
|
4fdc62754c | ||
|
徐涛
|
9f5e8702a2 | ||
|
徐涛
|
2d0d5c1592 | ||
|
徐涛
|
7b2d2db322 | ||
|
徐涛
|
c084c57e2c | ||
|
徐涛
|
c2c0d498ee | ||
|
徐涛
|
e54edbc037 |
@@ -0,0 +1,91 @@
|
||||
# Changelog
|
||||
|
||||
本项目所有重要变更均记录于此文件。格式参考 [Keep a Changelog](https://keepachangelog.com/zh-CN/1.1.0/)。
|
||||
|
||||
## [0.1.0] - 2026-07-04
|
||||
|
||||
首个公开版本。涵盖 Phase 0-4c 的全部核心能力、Provider IR 重构、LlmCycle 简化，以及面向用户的 7 个离线示例。
|
||||
|
||||
### Added
|
||||
|
||||
**LLM 调用周期（Phase 0）**
|
||||
- 统一 IR 类型层：`Message`、`ContentBlock`、`MessageRequest`、`MessageResponse`、`ToolDefinition`、`StopReason`
|
||||
- `LlmProvider` trait + 4 个 Provider 实现：OpenAI Chat Completions、Anthropic Messages、DeepSeek（OpenAI 兼容）、Qwen（OpenAI 兼容）
|
||||
- `ProviderRegistry`：多 Provider 注册与发现
|
||||
- `LlmCycle`：重试策略 + 用量追踪 + 自动 tool 循环 + Auto-compaction
|
||||
- `HookExecutor`：OnTurnStart / OnTurnEnd / OnPlanStepComplete 等生命周期钩子
|
||||
- `StreamEvents`：AssistantTextDelta / ToolExecutionStarted / MessageComplete 等流式事件
|
||||
|
||||
**提示词工程（Phase 1）**
|
||||
- `PromptTemplate`：变量插值 + 条件渲染
|
||||
- `PromptComposer`：system/user/assistant/tool_result 消息链组合
|
||||
|
||||
**工具系统（Phase 2）**
|
||||
- `ToolRegistry`：注册、发现、并行调用、超时控制
|
||||
- `BaseTool` trait（含 `ToolContext` 执行上下文）
|
||||
- `McpClient`：stdio transport（StreamableHttp 已预留接口）
|
||||
- `PermissionChecker`：白名单 / 黑名单 / 自定义权限
|
||||
|
||||
**记忆系统（Phase 3）**
|
||||
- `MemoryStore` trait + `InMemoryStore` 默认实现
|
||||
- `ConversationMemory`：sliding window / 全量模式，集成 `llm::compact`
|
||||
- `KnowledgeStore`：知识页面存储
|
||||
- `MemoryRetriever`：TextOverlap Dice 系数评分
|
||||
- `EvictionPolicy`：None / Ttl / Capacity 三种淘汰策略
|
||||
|
||||
**Agent 运行时（Phase 4a/b/c）**
|
||||
- `Agent` trait：name / system_prompt / tool_definitions
|
||||
- `AgentSession` + `RuntimeBundle` + `AgentBuilder`：链式构造 + 依赖注入
|
||||
- `AgentError`：Llm / Tool / Memory / HookBlocked / LimitExceeded / Config / PlanParse / Other 8 个变体
|
||||
- `TaskAgent` trait + `JsonPlanParser`：自主执行 + 外部驱动
|
||||
- `Plan` / `Step` / `StepStatus`：纯数据结构
|
||||
- `SessionMemory`：基于 `MemoryStore` 的会话级 key-value 桥接
|
||||
|
||||
**面向用户的资产**
|
||||
- `agcore::llm::mock::MockProvider`：公开的 Mock Provider，支持 `chat` + `chat_stream`，无需 API key 即可运行示例
|
||||
- 7 个离线可运行示例（`cargo run --example ...`）：
|
||||
- `prompt_composer` — 模板变量插值 + 消息链构建
|
||||
- `custom_tool` — 工具注册 + 权限检查
|
||||
- `agent_session_demo` — AgentBuilder → AgentSession → submit_turn 完整链路
|
||||
- `task_agent_demo` — JsonPlanParser → Plan → Step 状态机
|
||||
- `conversation_memory_demo` — SessionMemory 命名空间隔离
|
||||
- `knowledge_search_demo` — KnowledgeStore + 停用词过滤
|
||||
- `streaming_events_demo` — 流式事件消费
|
||||
|
||||
### Changed
|
||||
|
||||
- **`LlmCycle` 切换到 IR 消息类型**：内部消息从 `OpenaiChatMessage` 切到统一 `Message` 类型，移除 Phase 0 的桥接转换层（方案：`docs/10c-phase2-llm-cycle-simplify.md`）
|
||||
- **Provider IR 重构**：`LlmProvider` trait 签名同步切到 `MessageRequest` / `MessageResponse`；新增 Anthropic / DeepSeek / Qwen Provider（方案：`docs/10-llm-provider-refinement.md`、`docs/10b-phase1-provider-adaptation.md`）
|
||||
- **`AgentBuilder` 扩展 `session_memory_backend`**：Phase 4c 接入会话级记忆
|
||||
- **错误消息面向最终用户友好化**：`AgentError` / `LlmError` / `ToolError` / `MemoryError` / `PromptError` 全部改写为带可操作建议的友好消息
|
||||
- **`composer.rs` IR 迁移**：`PromptComposer` 内部从 `OpenaiChatMessage` 切到 `Message` / `ContentBlock`，与新类型系统保持一致
|
||||
- **`knowledge.rs` 锁修复**：`std::sync::Mutex` → `tokio::sync::Mutex`，避免 MutexGuard 跨 `.await` 持有
|
||||
|
||||
### Deprecated
|
||||
|
||||
- `ChatResponse` — 自 0.1.0 起标记为 deprecated，请改用 `MessageResponse`
|
||||
- `ToolDefinition` — 自 0.1.0 起标记为 deprecated（仍直接对应 OpenAI wire-format）；v0.2 将引入 IR 工具类型
|
||||
|
||||
### Fixed
|
||||
|
||||
- 修复 Provider IR 重构后 `session.rs` / `cycle.rs` 测试模块的导入缺失回归
|
||||
- 修复 `knowledge.rs::search()` 中 MutexGuard 跨 `.await` 持有的潜在阻塞
|
||||
- 清零全部 clippy 警告（`#[allow(dead_code)]` 已在最低必要范围使用）
|
||||
|
||||
### Removed
|
||||
|
||||
- 移除 Phase 0 的 `OpenaiChatMessage ↔ Message` 桥接层（LlmCycle 简化后已无外部调用方）
|
||||
|
||||
---
|
||||
|
||||
## 版本基线
|
||||
|
||||
| 指标 | 数值 |
|
||||
|------|------|
|
||||
| `cargo build --all-targets` | ✅ 通过 |
|
||||
| `cargo test --all-targets` | ✅ 182 passed / 0 failed |
|
||||
| `cargo clippy --all-targets -- -D warnings` | ✅ 0 警告 |
|
||||
| 离线示例 | ✅ 7 个全部 `cargo run` 退出码 0 |
|
||||
| 许可证 | Apache-2.0 |
|
||||
|
||||
[0.1.0]: https://git.archgrid.xyz/xt/agcore/releases/tag/v0.1.0
|
||||
@@ -0,0 +1,201 @@
|
||||
Apache License
|
||||
Version 2.0, January 2004
|
||||
http://www.apache.org/licenses/
|
||||
|
||||
TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
|
||||
|
||||
1. Definitions.
|
||||
|
||||
"License" shall mean the terms and conditions for use, reproduction,
|
||||
and distribution as defined by Sections 1 through 9 of this document.
|
||||
|
||||
"Licensor" shall mean the copyright owner or entity authorized by
|
||||
the copyright owner that is granting the License.
|
||||
|
||||
"Legal Entity" shall mean the union of the acting entity and all
|
||||
other entities that control, are controlled by, or are under common
|
||||
control with that entity. For the purposes of this definition,
|
||||
"control" means (i) the power, direct or indirect, to cause the
|
||||
direction or management of such entity, whether by contract or
|
||||
otherwise, or (ii) ownership of fifty percent (50%) or more of the
|
||||
outstanding shares, or (iii) beneficial ownership of such entity.
|
||||
|
||||
"You" (or "Your") shall mean an individual or Legal Entity
|
||||
exercising permissions granted by this License.
|
||||
|
||||
"Source" form shall mean the preferred form for making modifications,
|
||||
including but not limited to software source code, documentation
|
||||
source, and configuration files.
|
||||
|
||||
"Object" form shall mean any form resulting from mechanical
|
||||
transformation or translation of a Source form, including but
|
||||
not limited to compiled object code, generated documentation,
|
||||
and conversions to other media types.
|
||||
|
||||
"Work" shall mean the work of authorship, whether in Source or
|
||||
Object form, made available under the License, as indicated by a
|
||||
copyright notice that is included in or attached to the work
|
||||
(an example is provided in the Appendix below).
|
||||
|
||||
"Derivative Works" shall mean any work, whether in Source or Object
|
||||
form, that is based on (or derived from) the Work and for which the
|
||||
editorial revisions, annotations, elaborations, or other modifications
|
||||
represent, as a whole, an original work of authorship. For the purposes
|
||||
of this License, Derivative Works shall not include works that remain
|
||||
separable from, or merely link (or bind by name) to the interfaces of,
|
||||
the Work and Derivative Works thereof.
|
||||
|
||||
"Contribution" shall mean any work of authorship, including
|
||||
the original version of the Work and any modifications or additions
|
||||
to that Work or Derivative Works thereof, that is intentionally
|
||||
submitted to Licensor for inclusion in the Work by the copyright owner
|
||||
or by an individual or Legal Entity authorized to submit on behalf of
|
||||
the copyright owner. For the purposes of this definition, "submitted"
|
||||
means any form of electronic, verbal, or written communication sent
|
||||
to the Licensor or its representatives, including but not limited to
|
||||
communication on electronic mailing lists, source code control systems,
|
||||
and issue tracking systems that are managed by, or on behalf of, the
|
||||
Licensor for the purpose of discussing and improving the Work, but
|
||||
excluding communication that is conspicuously marked or otherwise
|
||||
designated in writing by the copyright owner as "Not a Contribution."
|
||||
|
||||
"Contributor" shall mean Licensor and any individual or Legal Entity
|
||||
on behalf of whom a Contribution has been received by Licensor and
|
||||
subsequently incorporated within the Work.
|
||||
|
||||
2. Grant of Copyright License. Subject to the terms and conditions of
|
||||
this License, each Contributor hereby grants to You a perpetual,
|
||||
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
||||
copyright license to reproduce, prepare Derivative Works of,
|
||||
publicly display, publicly perform, sublicense, and distribute the
|
||||
Work and such Derivative Works in Source or Object form.
|
||||
|
||||
3. Grant of Patent License. Subject to the terms and conditions of
|
||||
this License, each Contributor hereby grants to You a perpetual,
|
||||
worldwide, non-exclusive, no-charge, royalty-free, irrevocable
|
||||
(except as stated in this section) patent license to make, have made,
|
||||
use, offer to sell, sell, import, and otherwise transfer the Work,
|
||||
where such license applies only to those patent claims licensable
|
||||
by such Contributor that are necessarily infringed by their
|
||||
Contribution(s) alone or by combination of their Contribution(s)
|
||||
with the Work to which such Contribution(s) was submitted. If You
|
||||
institute patent litigation against any entity (including a
|
||||
cross-claim or counterclaim in a lawsuit) alleging that the Work
|
||||
or a Contribution incorporated within the Work constitutes direct
|
||||
or contributory patent infringement, then any patent licenses
|
||||
granted to You under this License for that Work shall terminate
|
||||
as of the date such litigation is filed.
|
||||
|
||||
4. Redistribution. You may reproduce and distribute copies of the
|
||||
Work or Derivative Works thereof in any medium, with or without
|
||||
modifications, and in Source or Object form, provided that You
|
||||
meet the following conditions:
|
||||
|
||||
(a) You must give any other recipients of the Work or
|
||||
Derivative Works a copy of this License; and
|
||||
|
||||
(b) You must cause any modified files to carry prominent notices
|
||||
stating that You changed the files; and
|
||||
|
||||
(c) You must retain, in the Source form of any Derivative Works
|
||||
that You distribute, all copyright, patent, trademark, and
|
||||
attribution notices from the Source form of the Work,
|
||||
excluding those notices that do not pertain to any part of
|
||||
the Derivative Works; and
|
||||
|
||||
(d) If the Work includes a "NOTICE" text file as part of its
|
||||
distribution, then any Derivative Works that You distribute must
|
||||
include a readable copy of the attribution notices contained
|
||||
within such NOTICE file, excluding those notices that do not
|
||||
pertain to any part of the Derivative Works, in at least one
|
||||
of the following places: within a NOTICE text file distributed
|
||||
as part of the Derivative Works; within the Source form or
|
||||
documentation, if provided along with the Derivative Works; or,
|
||||
within a display generated by the Derivative Works, if and
|
||||
wherever such third-party notices normally appear. The contents
|
||||
of the NOTICE file are for informational purposes only and
|
||||
do not modify the License. You may add Your own attribution
|
||||
notices within Derivative Works that You distribute, alongside
|
||||
or as an addendum to the NOTICE text from the Work, provided
|
||||
that such additional attribution notices cannot be construed
|
||||
as modifying the License.
|
||||
|
||||
You may add Your own copyright statement to Your modifications and
|
||||
may provide additional or different license terms and conditions
|
||||
for use, reproduction, or distribution of Your modifications, or
|
||||
for any such Derivative Works as a whole, provided Your use,
|
||||
reproduction, and distribution of the Work otherwise complies with
|
||||
the conditions stated in this License.
|
||||
|
||||
5. Submission of Contributions. Unless You explicitly state otherwise,
|
||||
any Contribution intentionally submitted for inclusion in the Work
|
||||
by You to the Licensor shall be under the terms and conditions of
|
||||
this License, without any additional terms or conditions.
|
||||
Notwithstanding the above, nothing herein shall supersede or modify
|
||||
the terms of any separate license agreement you may have executed
|
||||
with Licensor regarding such Contributions.
|
||||
|
||||
6. Trademarks. This License does not grant permission to use the trade
|
||||
names, trademarks, service marks, or product names of the Licensor,
|
||||
except as required for reasonable and customary use in describing the
|
||||
origin of the Work and reproducing the content of the NOTICE file.
|
||||
|
||||
7. Disclaimer of Warranty. Unless required by applicable law or
|
||||
agreed to in writing, Licensor provides the Work (and each
|
||||
Contributor provides its Contributions) on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
|
||||
implied, including, without limitation, any warranties or conditions
|
||||
of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
|
||||
PARTICULAR PURPOSE. You are solely responsible for determining the
|
||||
appropriateness of using or redistributing the Work and assume any
|
||||
risks associated with Your exercise of permissions under this License.
|
||||
|
||||
8. Limitation of Liability. In no event and under no legal theory,
|
||||
whether in tort (including negligence), contract, or otherwise,
|
||||
unless required by applicable law (such as deliberate and grossly
|
||||
negligent acts) or agreed to in writing, shall any Contributor be
|
||||
liable to You for damages, including any direct, indirect, special,
|
||||
incidental, or consequential damages of any character arising as a
|
||||
result of this License or out of the use or inability to use the
|
||||
Work (including but not limited to damages for loss of goodwill,
|
||||
work stoppage, computer failure or malfunction, or any and all
|
||||
other commercial damages or losses), even if such Contributor
|
||||
has been advised of the possibility of such damages.
|
||||
|
||||
9. Accepting Warranty or Additional Liability. While redistributing
|
||||
the Work or Derivative Works thereof, You may choose to offer,
|
||||
and charge a fee for, acceptance of support, warranty, indemnity,
|
||||
or other liability obligations and/or rights consistent with this
|
||||
License. However, in accepting such obligations, You may act only
|
||||
on Your own behalf and on Your sole responsibility, not on behalf
|
||||
of any other Contributor, and only if You agree to indemnify,
|
||||
defend, and hold each Contributor harmless for any liability
|
||||
incurred by, or claims asserted against, such Contributor by reason
|
||||
of your accepting any such warranty or additional liability.
|
||||
|
||||
END OF TERMS AND CONDITIONS
|
||||
|
||||
APPENDIX: How to apply the Apache License to your work.
|
||||
|
||||
To apply the Apache License to your work, attach the following
|
||||
boilerplate notice, with the fields enclosed by brackets "[]"
|
||||
replaced with your own identifying information. (Don't include
|
||||
the brackets!) The text should be enclosed in the appropriate
|
||||
comment syntax for the file format. We also recommend that a
|
||||
file or class name and description of purpose be included on the
|
||||
same "printed page" as the copyright notice for easier
|
||||
identification within third-party archives.
|
||||
|
||||
Copyright 2026 AG Core Contributors
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
|
||||
Unless required by applicable law or agreed to in writing, software
|
||||
distributed under the License is distributed on an "AS IS" BASIS,
|
||||
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
See the License for the specific language governing permissions and
|
||||
limitations under the License.
|
||||
@@ -1,3 +1,245 @@
|
||||
# AG Core
|
||||
|
||||
AG Core是一个用于提供构建智能体的底层工具箱，其中提供的功能主要包括基础的大模型调用的完整周期，提示词的组合与优化，记忆检索与管理、MCP与工具的调用等。
|
||||
> AG Core 是一个用于构建 AI 智能体（Agent）的 Rust 工具箱，提供 LLM 调用、提示词工程、工具系统、记忆检索、Agent 运行时等核心能力，所有模块通过 trait 抽象、可插拔、可离线测试。
|
||||
|
||||
## 这是什么？
|
||||
|
||||
AG Core 不是 Agent 产品，而是 Agent 的**底层依赖库**：上层应用（TUI / Bot 网关 / 业务服务）基于 AG Core 装配出符合自身场景的 Agent。设计原则：
|
||||
|
||||
- **模块化** —— 五大功能领域（LLM / Prompt / Tool / Memory / Agent）独立 crate-internal 模块，通过 trait 解耦
|
||||
- **可插拔** —— LLM Provider、记忆存储、工具注册全部通过 trait 抽象，可替换为自有实现
|
||||
- **可离线** —— 公开 `MockProvider`，所有示例与测试无需 API key 即可运行
|
||||
- **异步优先** —— 所有 IO API 均为 `async`，基于 tokio 运行时
|
||||
|
||||
适合用来：
|
||||
|
||||
- 搭建支持多轮对话 + 工具调用的智能体服务
|
||||
- 接入多家 LLM（OpenAI / Anthropic / DeepSeek / Qwen 等）
|
||||
- 在 Agent 中组合提示词模板、知识检索、MCP 工具
|
||||
- 在 Rust 中复用一套"业务无关"的 Agent 基础设施
|
||||
|
||||
## 快速上手
|
||||
|
||||
5 分钟上手，无需 API key（使用 `MockProvider` 预设响应）。
|
||||
|
||||
**1. 添加依赖**
|
||||
|
||||
```toml
|
||||
[dependencies]
|
||||
agcore = "0.1"
|
||||
tokio = { version = "1", features = ["macros", "rt-multi-thread"] }
|
||||
```
|
||||
|
||||
**2. 第一个 Agent**
|
||||
|
||||
```rust,no_run
|
||||
use std::sync::Arc;
|
||||
|
||||
use agcore::agent::{Agent, AgentBuilder, AgentSession};
|
||||
use agcore::llm::hooks::HookExecutor;
|
||||
use agcore::llm::mock::MockProvider;
|
||||
use agcore::llm::provider::LlmProvider;
|
||||
use agcore::llm::types::message::{ContentBlock, Message};
|
||||
use agcore::llm::types::response_v2::{MessageResponse, StopReason};
|
||||
use agcore::llm::types::Usage;
|
||||
use agcore::tools::ToolRegistry;
|
||||
|
||||
struct Greeter;
|
||||
|
||||
impl Agent for Greeter {
|
||||
fn name(&self) -> &str { "greeter" }
|
||||
fn system_prompt(&self) -> Option<&str> { Some("你是一个中文助手，回答简短。") }
|
||||
}
|
||||
|
||||
fn text_response(text: &str) -> MessageResponse {
|
||||
MessageResponse {
|
||||
id: "r".into(),
|
||||
model: "mock".into(),
|
||||
message: Message::Assistant {
|
||||
content: vec![ContentBlock::Text { text: text.into() }],
|
||||
},
|
||||
usage: Usage::from_input_output(2, text.chars().count() as u32),
|
||||
stop_reason: StopReason::Stop,
|
||||
extra: Default::default(),
|
||||
}
|
||||
}
|
||||
|
||||
#[tokio::main]
|
||||
async fn main() {
|
||||
// 1. MockProvider：预设响应，无须 API key
|
||||
let provider: Arc<dyn LlmProvider> =
|
||||
Arc::new(MockProvider::new(vec![text_response("你好，我是 agcore。")]));
|
||||
|
||||
// 2. 装配 RuntimeBundle（必填：provider / tool_registry / hook_executor）
|
||||
let bundle = Arc::new(
|
||||
AgentBuilder::new()
|
||||
.provider(provider)
|
||||
.tool_registry(Arc::new(ToolRegistry::new()))
|
||||
.hook_executor(Arc::new(HookExecutor::new()))
|
||||
.build()
|
||||
.expect("装配 RuntimeBundle 失败"),
|
||||
);
|
||||
|
||||
// 3. 创建会话并提交首轮
|
||||
let agent: Arc<dyn Agent> = Arc::new(Greeter);
|
||||
let mut session = AgentSession::new(agent, "demo", bundle);
|
||||
let resp = session.submit_turn("你好").await.expect("submit_turn 失败");
|
||||
println!("LLM: {}", resp.text());
|
||||
}
|
||||
```
|
||||
|
||||
跑起来：
|
||||
|
||||
```bash
|
||||
cargo run
|
||||
# LLM: 你好，我是 agcore。
|
||||
```
|
||||
|
||||
接真实 Provider（如 OpenAI）只需把 `MockProvider` 换成：
|
||||
|
||||
```rust,no_run
|
||||
use agcore::llm::provider::{create_provider, ProviderConfig, ProviderType};
|
||||
|
||||
let provider = create_provider(
|
||||
ProviderType::OpenaiChat,
|
||||
ProviderConfig {
|
||||
base_url: "https://api.openai.com/v1".into(),
|
||||
api_key: std::env::var("OPENAI_API_KEY").expect("未设置 OPENAI_API_KEY"),
|
||||
model: "gpt-4o-mini".into(),
|
||||
},
|
||||
).expect("创建 Provider 失败");
|
||||
```
|
||||
|
||||
更多端到端示例见 [`examples/`](./examples/) 目录（共 7 个，全部可 `cargo run --example <name>`）：
|
||||
|
||||
| 示例 | 说明 |
|
||||
|------|------|
|
||||
| `agent_session_demo` | Agent + 会话 + SessionMemory 完整链路（MockProvider 离线） |
|
||||
| `custom_tool` | 自定义工具注册、单次 / 并行调用、权限检查 |
|
||||
| `prompt_composer` | 提示词模板与组合器（纯离线） |
|
||||
| `task_agent_demo` | Plan 解析、Step 状态机、错误路径 |
|
||||
| `conversation_memory_demo` | 对话记忆滑动窗口与隔离 |
|
||||
| `knowledge_search_demo` | 知识页面关键词检索 |
|
||||
| `streaming_events_demo` | LLM 流式响应事件消费（含错误路径） |
|
||||
|
||||
## 核心模块
|
||||
|
||||
| 模块 | 一句话说明 |
|
||||
|------|----------|
|
||||
| `agcore::llm` | LLM 调用周期（`LlmProvider` trait + `LlmCycle` 重试/用量 + 流式事件 + auto-compaction + Hook + 公开 `MockProvider`） |
|
||||
| `agcore::prompt` | 提示词工程（`PromptTemplate` 变量插值 + `PromptTemplateRegistry` + `PromptComposer` 多角色消息构造 + `validate_messages`） |
|
||||
| `agcore::tools` | 工具系统（`BaseTool` trait + `ToolRegistry` 注册/调用 + `PermissionChecker` 黑白名单 + MCP stdio 客户端） |
|
||||
| `agcore::memory` | 记忆系统（`MemoryStore` trait + `InMemoryStore` 默认实现 + `ConversationMemory` 滑动窗口 + `KnowledgeStore` + `MemoryRetriever`） |
|
||||
| `agcore::agent` | Agent 运行时（`Agent` trait 角色定义 + `AgentBuilder` + `RuntimeBundle` 依赖注入 + `AgentSession` 会话 + `SessionMemory` + `Plan`/`Step` 任务编排） |
|
||||
|
||||
## 架构关系图
|
||||
|
||||
```
|
||||
┌─────────────────────────────────────────────────────────────┐
|
||||
│ 应用层 (TUI / Bot 网关 / 业务服务) │
|
||||
└───────────────────────────┬─────────────────────────────────┘
|
||||
│ 使用
|
||||
┌───────────────────────────▼─────────────────────────────────┐
|
||||
│ Agent Runtime (agcore::agent) │
|
||||
│ Agent / AgentBuilder / RuntimeBundle / AgentSession / │
|
||||
│ SessionMemory / Plan / Step │
|
||||
└─────┬───────────────┬───────────────┬───────────────┬───────┘
|
||||
│ │ │ │
|
||||
┌─────▼─────┐ ┌──────▼──────┐ ┌──────▼──────┐ ┌──────▼──────┐
|
||||
│ LLM │ │ Prompt │ │ Tool │ │ Memory │
|
||||
│ agcore:: │ │ agcore:: │ │ agcore:: │ │ agcore:: │
|
||||
│ llm │ │ prompt │ │ tools │ │ memory │
|
||||
└─────┬─────┘ └─────────────┘ └─────┬───────┘ └──────┬──────┘
|
||||
│ │ │
|
||||
└──────────────┬────────────────┘ │
|
||||
▼ │
|
||||
┌─────────────────┐ │
|
||||
│ Mock Provider │◄──────────────────────┘
|
||||
│ 公开 API │ 离线测试 / 示例
|
||||
└─────────────────┘
|
||||
```
|
||||
|
||||
## 模块依赖关系
|
||||
|
||||
```mermaid
|
||||
graph BT
|
||||
LLM["<b>llm</b><br/>Provider / Cycle /<br/>Hooks / Stream /<br/>Compact / Mock"]:::core
|
||||
Prompt["<b>prompt</b><br/>Template / Composer"]:::core
|
||||
Tool["<b>tools</b><br/>BaseTool / Registry /<br/>Permission / MCP"]:::core
|
||||
Memory["<b>memory</b><br/>Store / Conversation /<br/>Knowledge / Retriever"]:::core
|
||||
Agent["<b>agent</b><br/>Agent / Builder /<br/>Session / Plan"]:::core
|
||||
|
||||
Prompt --> LLM
|
||||
Tool --> LLM
|
||||
Memory --> LLM
|
||||
Agent --> LLM
|
||||
Agent --> Tool
|
||||
Agent --> Memory
|
||||
|
||||
classDef core fill:#60a5fa,stroke:#2563eb,color:#fff
|
||||
```
|
||||
|
||||
**依赖规则**：
|
||||
|
||||
- `llm` 是叶子，被其他四个模块使用
|
||||
- `prompt` / `tools` / `memory` 互相不依赖，可独立使用
|
||||
- `agent` 编译期依赖 `llm` / `tools` / `memory`，与 `prompt` 无直接编译依赖（system prompt 以 `&str` 形式传入）
|
||||
- 上层应用只应依赖 `agent` + 必要的子模块，不应跨层直接 `use`
|
||||
|
||||
## 环境变量
|
||||
|
||||
> AG Core 库本身不读环境变量。下表是 [`examples/simple_visit.rs`](./examples/simple_visit.rs) 这一真实调用示例所使用的环境变量，以及通常推荐的取值。
|
||||
|
||||
| 变量 | 必填 | 推荐值 | 说明 |
|
||||
|------|------|-------|------|
|
||||
| `OPENAI_API_KEY` | 用 OpenAI 时 | — | OpenAI / 兼容 Provider（DeepSeek / Qwen）的 API key |
|
||||
| `OPENAI_BASE_URL` | 是 | `https://api.openai.com/v1` | OpenAI 兼容端点 base URL（DeepSeek/Qwen 用对应地址） |
|
||||
| `OPENAI_MODEL` | 是 | `gpt-4o-mini` | OpenAI 模型名 |
|
||||
| `ANTHROPIC_API_KEY` | 用 Anthropic 时 | — | Anthropic Claude API key |
|
||||
| `ANTHROPIC_BASE_URL` | 是 | `https://api.anthropic.com` | Anthropic 兼容端点 base URL |
|
||||
| `ANTHROPIC_MODEL` | 是 | `claude-3-5-sonnet-latest` | Claude 模型名 |
|
||||
| `PROVIDER` | 否 | `openai` | Provider 类型：`openai` / `openai-response` / `anthropic` / `deepseek` / `qwen` |
|
||||
| `RUST_LOG` | 否 | `agcore=info` | tracing 日志级别（其他 crate 可加 `=debug`） |
|
||||
|
||||
示例（运行 `examples/simple_visit.rs` 真实调用 OpenAI）：
|
||||
|
||||
```bash
|
||||
export OPENAI_API_KEY=sk-...
|
||||
export OPENAI_BASE_URL=https://api.openai.com/v1
|
||||
export OPENAI_MODEL=gpt-4o-mini
|
||||
export PROVIDER=openai
|
||||
export RUST_LOG=agcore=debug
|
||||
cargo run --example simple_visit
|
||||
```
|
||||
|
||||
## 参考项目
|
||||
|
||||
AG Core 在 Phase 4 设计阶段调研了 4 个 2026 年公开的 AI Agent 项目（详见 [`docs/note-agent-harness-references.md`](./docs/note-agent-harness-references.md)）：
|
||||
|
||||
| 项目 | 类型 | 语言 | 借鉴点 |
|
||||
|------|------|------|-------|
|
||||
| [OpenClaw](https://github.com/openclaw/openclaw) | 消息网关 | TypeScript | 多渠道适配模式 |
|
||||
| [Hermes Agent](https://github.com/NousResearch/hermes-agent) | 自主学习智能体 | Python | 实体与会话解耦 |
|
||||
| [OpenHuman](https://github.com/tinyhumansai/openhuman) | 桌面助手 | Rust + Tauri | 记忆树与 Token 压缩 |
|
||||
| [OpenHarness](https://github.com/HKUDS/OpenHarness) | Agent Harness 框架 | Python | 显式依赖注入容器 + 三级权限 |
|
||||
|
||||
AG Core 不"抄代码"，只参考架构模式。当前实现已经采纳：
|
||||
|
||||
- **OpenHarness 风格** —— 显式 `RuntimeBundle` 依赖注入容器
|
||||
- **Hermes 风格** —— `Agent` trait（角色）与 `AgentSession`（会话）解耦
|
||||
|
||||
后续 v0.2+ 计划借鉴 OpenHuman 的 Memory Tree / TokenJuice 模式。
|
||||
|
||||
## 许可证
|
||||
|
||||
本项目基于 **Apache License 2.0** 授权。完整文本见 [`LICENSE`](./LICENSE)。
|
||||
|
||||
```
|
||||
Copyright 2026 AG Core Contributors
|
||||
|
||||
Licensed under the Apache License, Version 2.0 (the "License");
|
||||
you may not use this file except in compliance with the License.
|
||||
You may obtain a copy of the License at
|
||||
|
||||
http://www.apache.org/licenses/LICENSE-2.0
|
||||
```
|
||||
@@ -0,0 +1,340 @@
|
||||
# v0.1 发布实施计划
|
||||
|
||||
> 状态：待实施
|
||||
> 关联文档：`docs/roadmap.md`、`docs/8-examples-plan.md`、`docs/10c-phase2-llm-cycle-simplify.md`
|
||||
|
||||
## 背景与目标
|
||||
|
||||
AG Core 已完成 Phase 0-4c 全部 7 个 Phase 的核心功能，以及 Provider IR 重构（新类型系统 + Anthropic/DeepSeek/Qwen Provider）+ LlmCycle 简化（IR 消息类型切换 + 桥接层移除）。
|
||||
|
||||
**当前基线**：
|
||||
- `cargo build` — ✅ 通过（5 个 warning）
|
||||
- `cargo test` — ❌ 编译失败（4 个 error：session.rs/cycle.rs 测试模块中 `ContentBlock`/`ProviderCapabilities`/`ProviderFeatures` 导入缺失，均为 Provider IR 重构后未同步的回归）
|
||||
- LlmCycle 简化合并前全量测试为 177 通过，合并后尚未跑通过过全量测试
|
||||
|
||||
v0.1 发布的阻塞项不是"功能不足"，而是"已有功能不能被用户快速看见和使用"。本计划聚焦于：
|
||||
|
||||
1. **扫清技术债** — composer.rs 迁移、锁修复、clippy 清零
|
||||
2. **提供可离线运行的示例** — 让用户 5 分钟内上手
|
||||
3. **补充面向用户的文档** — README、MockProvider、错误消息友好化
|
||||
4. **完成发布前准备** — CI 验证、Roadmap 同步、版本标记
|
||||
|
||||
---
|
||||
|
||||
## 总体时间线
|
||||
|
||||
预计 8-10 个工作日（2 周），分为两条并行线：
|
||||
|
||||
```
|
||||
Week 1 ────┬── 测试修复 + 技术债扫清
|
||||
│ ├ T0 测试编译回归修复（~0.5d）← ⚠️ 任何改动前的前置步骤
|
||||
│ ├ T1.1 composer.rs IR 迁移（~1d）
|
||||
│ ├ T1.2 knowledge.rs 锁修复（~0.5d）
|
||||
│ ├ T1.3 旧类型废弃标记（~0.5d）
|
||||
│ └ T1.4 clippy 清零（~0.5d）
|
||||
│
|
||||
├── 示例 + 开发者体验
|
||||
│ ├ T2.0 MockProvider 公开化（~0.5d）
|
||||
│ ├ T2.1 4个🥇示例（~2-3d，可并行）
|
||||
│ └ T3.1 README 初稿（~1d，并行）
|
||||
│
|
||||
Week 2 ────┬── 示例继续
|
||||
│ ├ T2.2 3个🥈示例（~2d）
|
||||
│ └ T3.2 错误消息 review（~0.5d）
|
||||
│
|
||||
└── 发布准备
|
||||
├ T3.3 README 定稿（~1d）
|
||||
├ T4.1 CI 示例验证（~0.5d）
|
||||
├ T4.2 Roadmap 更新（~0.5d）
|
||||
├ T4.3 CHANGELOG 初始化（~0.5d）
|
||||
└ T4.4 v0.1 tag（~0.5d）
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 实施步骤
|
||||
|
||||
### Phase A — 测试修复 + 技术债扫清（Week 1 前半）
|
||||
|
||||
> **重要**：Task A0 是后续所有任务的前提——不修复测试编译，所有 Task 的验收条件（`cargo test` 全通过）都不可执行。
|
||||
|
||||
#### Task A0：修复测试编译回归（~0.5d）
|
||||
|
||||
**问题**：Provider IR 重构后，`ContentBlock`/`ProviderCapabilities`/`ProviderFeatures` 被移到新位置，但 2 个测试模块的导入未同步，导致 `cargo test` 编译失败。
|
||||
|
||||
**涉及文件**：
|
||||
- `src/agent/session.rs` — 测试模块缺少 `ContentBlock` 导入
|
||||
- `src/llm/cycle.rs` — 测试模块缺少 `ProviderCapabilities`/`ProviderFeatures` 导入
|
||||
|
||||
**修复方案**：在相应测试模块内补全 `use` 语句（每处加一行即可）。
|
||||
|
||||
**额外收益**：修复后 `convert.rs` 中 2 处 `irrefutable if let` 警告也一并修正（改为 `let`），减少 clippy 基数。
|
||||
|
||||
**前置依赖**：无
|
||||
|
||||
**验收条件**：
|
||||
- `cargo test` 全量编译通过
|
||||
- `cargo test` 全量运行通过
|
||||
|
||||
---
|
||||
|
||||
#### Task A1：`composer.rs` IR 迁移（~1d）
|
||||
|
||||
**前置依赖**：Task A0（否则无法通过 `cargo test` 验证）
|
||||
|
||||
**涉及文件**：
|
||||
- `src/prompt/composer.rs` — 主要改动
|
||||
- `src/prompt/mod.rs` — 类型重导出
|
||||
|
||||
**改动范围**：约 60 处 `OpenaiChatMessage` → `Message`，`ContentField`/`OpenaiContentPart` → `ContentBlock`
|
||||
|
||||
**具体清单**：
|
||||
1. `PromptComposer` 内部 `messages` 字段类型 `Vec<OpenaiChatMessage>` → `Vec<Message>`
|
||||
2. 所有 `.system_text()`/`.user_text()`/`.assistant_text()`/`.developer_text()`/`.tool_result()` 构造方法 — 新 `Message` 已有同名方法，直接替换调用
|
||||
3. 所有 `xx_content()`/`xx_contents()` 方法 — `ContentBlock` 替代 `ContentField`/`OpenaiContentPart`
|
||||
4. `set_message_name()` — 新 `Message` 是扁平 enum，无 `name` 字段，移除或忽略
|
||||
5. `build()` 返回类型 `Vec<OpenaiChatMessage>` → `Vec<Message>`
|
||||
6. `validate_messages()` 消息校验函数 — `OpenaiChatMessage::Tool { .. }` → `Message::ToolResult { .. }`，`tool_calls` 从 `Assistant` 变体的 `ContentBlock::ToolUse` 中提取
|
||||
|
||||
**测试**：内联测试中 `OpenaiChatMessage::Tool { .. }` pattern match 需同步更新
|
||||
|
||||
**验收条件**：
|
||||
- `cargo build` 无错误
|
||||
- `cargo test` 中 prompt 模块测试全通过
|
||||
- 无新增 clippy 警告
|
||||
|
||||
---
|
||||
|
||||
#### Task A2：`knowledge.rs` 锁修复（~0.5d）
|
||||
|
||||
**涉及文件**：`src/memory/knowledge.rs`
|
||||
|
||||
**问题**：`std::sync::Mutex` guard 在 `search()` 方法中跨 `.await` 持有，可能在高并发下阻塞 tokio 工作线程
|
||||
|
||||
**修复方案**：`std::sync::Mutex<Vec<PageIndexEntry>>` → `tokio::sync::Mutex<Vec<PageIndexEntry>>`
|
||||
|
||||
**影响范围**：5 处 `.lock().unwrap()` 调用（`rebuild_index`、`add_page`、`delete_page`、`search`、`get_index`）
|
||||
|
||||
**验收条件**：
|
||||
- `cargo build` 无错误
|
||||
- `cargo test` 中 memory 模块测试全通过
|
||||
- clippy 不再报 `await_holding_lock` 警告
|
||||
|
||||
---
|
||||
|
||||
#### Task A3：旧类型公开 API 废弃标记（~0.5d）
|
||||
|
||||
**涉及文件**：`src/llm/types/mod.rs`
|
||||
|
||||
**改动**：
|
||||
- `ChatResponse` struct 加 `#[deprecated(since = "0.1.0", note = "请改用 MessageResponse")]`
|
||||
- `ToolDefinition` 类型别名加 `#[deprecated]`（如果仍公开）
|
||||
- 保留结构体定义（OpenAI `chat_inner()` 内部转换层仍然在用），不删除
|
||||
|
||||
**验收条件**：
|
||||
- `cargo build` 产生废弃警告（期望行为）但不产生编译错误
|
||||
- 外部调用方能看到有用提示
|
||||
|
||||
---
|
||||
|
||||
#### Task A4：clippy 警告清零（~0.5d）
|
||||
|
||||
当前剩余 8 个：
|
||||
|
||||
| # | 警告类型 | 文件 | 修复方式 |
|
||||
|---|---------|------|---------|
|
||||
| 1 | irrefutable `if let` | — | 改为直接 `let` |
|
||||
| 2 | `with_plan_step_index` 未用 | `src/llm/hooks.rs` | 加 `#[allow(dead_code)]` |
|
||||
| 3-5 | 字段未读（api_key, stop_sequence, next_block_index） | `anthropic.rs`, `response_v2.rs` | 加 `#[allow(dead_code)]` |
|
||||
| 6 | 手动前缀剥离 | — | 用 `strip_prefix()` |
|
||||
| 7 | MutexGuard 跨 await | knowledge.rs | Task A2 修复后自动消失 |
|
||||
| 8 | `if` 相同分支 | retriever.rs | 合并条件 |
|
||||
|
||||
**验收条件**：`cargo clippy --lib -p agcore` 0 警告
|
||||
|
||||
---
|
||||
|
||||
### Phase B — 示例实现（Week 1 后半 ~ Week 2 前半）
|
||||
|
||||
#### Task B0：MockProvider 公开化（~0.5d）
|
||||
|
||||
**涉及文件**：
|
||||
- `src/llm/mock.rs` — 新建，公开 `MockProvider` struct
|
||||
- `src/llm/mod.rs` — 加 `pub mod mock;`
|
||||
- `src/agent/session.rs` — 测试中 `MockProvider` 改为引用 `crate::llm::mock::MockProvider`
|
||||
|
||||
**MockProvider 接口**：
|
||||
|
||||
```rust
|
||||
pub struct MockProvider { .. }
|
||||
impl MockProvider {
|
||||
pub fn new(responses: Vec<MessageResponse>) -> Self;
|
||||
pub fn chat(&self, request: MessageRequest) -> Result<MessageResponse, LlmError>;
|
||||
pub fn chat_stream(&self, request: MessageRequest)
|
||||
-> Result<Pin<Box<dyn Stream<Item = Result<StreamEvent, LlmError>> + Send>>, LlmError>;
|
||||
}
|
||||
```
|
||||
|
||||
**决策**：同时实现 `chat_stream`，从 `MessageResponse` 拆解为 `StreamEvent::ContentBlockStart + ContentBlockDelta * N + MessageComplete` 序列，使流式示例也能不依赖 API key。
|
||||
|
||||
**验收条件**：
|
||||
- 公开 `MockProvider` 可被外部 crate 引用
|
||||
- `cargo test` 中所有 session 测试通过
|
||||
- `agent_session_demo.rs` 示例可引用 `MockProvider` 并运行
|
||||
|
||||
---
|
||||
|
||||
#### Task B1：🥇 示例 × 4（~2-3d，可并行）
|
||||
|
||||
4 个示例相互无依赖，按技术债消除进度安排：
|
||||
|
||||
| # | 示例 | 代码量 | 前置依赖 | 说明 |
|
||||
|---|------|--------|---------|------|
|
||||
| B1a | `agent_session_demo.rs` | ~100 行 | Task B0 | AgentBuilder → AgentSession → submit_turn → SessionMemory 完整链路 |
|
||||
| B1b | `custom_tool.rs` | ~80 行 | 无 | 实现模拟工具 → ToolRegistry 注册 → invoke/invoke_all → PermissionChecker |
|
||||
| B1c | `prompt_composer.rs` | ~60 行 | Task A1 | 模板变量插值 → PromptComposer 构建消息链 → 断言验证，纯离线 |
|
||||
| B1d | `task_agent_demo.rs` | ~70 行 | 无 | 构造 JSON → JsonPlanParser → Plan → Step 状态机 → Hook 事件 |
|
||||
|
||||
每个示例的详细设计见 `docs/8-examples-plan.md`，按该文档直接实现。
|
||||
|
||||
**验收条件**：
|
||||
- `cargo run --example prompt_composer` → 成功退出
|
||||
- `cargo run --example custom_tool` → 成功退出
|
||||
- `cargo run --example agent_session_demo` → 成功退出
|
||||
- `cargo run --example task_agent_demo` → 成功退出
|
||||
|
||||
---
|
||||
|
||||
#### Task B2：🥈 示例 × 3（~2d）
|
||||
|
||||
| # | 示例 | 代码量 | 前置依赖 |
|
||||
|---|------|--------|---------|
|
||||
| B2a | `conversation_memory_demo.rs` | ~70 行 | 无 |
|
||||
| B2b | `knowledge_search_demo.rs` | ~60 行 | 无 |
|
||||
| B2c | `streaming_events_demo.rs` | ~80 行 | Task B0（MockProvider 需支持 `chat_stream`） |
|
||||
|
||||
**验收条件**：额外 3 个示例均可 `cargo run` 成功退出
|
||||
|
||||
---
|
||||
|
||||
### Phase C — 开发者体验 + 文档（Week 2 后半，与 Phase B 后段并行）
|
||||
|
||||
#### Task C1：README 完整版（~1.5d）
|
||||
|
||||
**涉及文件**：`README.md`
|
||||
|
||||
**内容结构**：
|
||||
|
||||
```
|
||||
# AG Core
|
||||
|
||||
## 这是什么？ ← 一句话定位
|
||||
## 快速上手 ← cargo add + MockProvider 示例
|
||||
## 核心模块 ← 7 个模块一句话说明
|
||||
## 架构关系图 ← ASCII 或 Mermaid
|
||||
## 模块依赖关系 ← 依赖关系图 + 说明
|
||||
## 环境变量 ← 当前支持的环境变量表
|
||||
## 参考项目 ← OpenClaw / Hermes / OpenHuman / OpenHarness
|
||||
## 许可证 ← MIT / Apache-2.0
|
||||
```
|
||||
|
||||
**要求**：快速上手代码段在提交前必须真实编译通过。
|
||||
|
||||
---
|
||||
|
||||
#### Task C2：错误消息友好化 review（~0.5d）
|
||||
|
||||
**涉及文件**：
|
||||
- `src/agent/error.rs` — AgentError 消息
|
||||
- `src/llm/error.rs` — LlmError 消息
|
||||
- `src/tools/error.rs` — ToolError 消息
|
||||
- `src/memory/error.rs` — MemoryError 消息
|
||||
- `src/prompt/error.rs` — PromptError 消息
|
||||
|
||||
**检查标准**：
|
||||
- 用户能理解"哪里错了"
|
||||
- 有可操作的建议（"请检查 API key"、"请配置环境变量 LLM_API_KEY"）
|
||||
- 无英文残留的工程师视角消息
|
||||
|
||||
---
|
||||
|
||||
### Phase D — 发布准备（Week 2 末）
|
||||
|
||||
#### Task D1：CI 示例验证集成（~0.5d）
|
||||
|
||||
- 确认 `cargo build` 无新增警告
|
||||
- 确认 `cargo test` 全部通过（退出码 0）
|
||||
- 确认 `cargo test --examples` 全部通过
|
||||
- 确认 `cargo clippy --lib -p agcore` 0 警告
|
||||
|
||||
#### Task D2：Roadmap 更新（~0.5d）
|
||||
|
||||
更新 `docs/roadmap.md`：
|
||||
- Phase 2 状态更新为 ✅ 全部交付物已完成（Provider IR 重构 + LlmCycle 简化）
|
||||
- 测试计数更新为 D1 确认的最终实际数值
|
||||
- 移除 clippy 警告计数
|
||||
- 添加 v0.1 发布里程碑
|
||||
|
||||
#### Task D3：依赖裁剪（~0.5d，非阻塞，可跳过）
|
||||
|
||||
`tokio = { features = ["full"] }` → 按需 feature（`rt`、`macros`、`sync`、`time`、`net`）
|
||||
|
||||
**理由**：减少编译时间 + 减少安全面。时间不够可延后到 v0.2。
|
||||
|
||||
#### Task D4：CHANGELOG 初始化（~0.5d）
|
||||
|
||||
**涉及文件**：`CHANGELOG.md`
|
||||
|
||||
**内容要求**：
|
||||
- 版本号：`v0.1.0`
|
||||
- 发布日期：标记当天
|
||||
- 变更摘要：Phase 0-4c（LLM 调用周期、提示词工程、工具系统、记忆系统、Agent 运行时、任务执行、会话级记忆）+ Provider IR 重构（统一类型系统 + Anthropic/DeepSeek/Qwen 适配）+ LlmCycle 简化
|
||||
- 格式参考 [Keep a Changelog](https://keepachangelog.com/) 规范
|
||||
|
||||
**验收条件**：
|
||||
- `CHANGELOG.md` 文件存在，内容与当前版本一致
|
||||
- 文件随 tag commit 一起提交
|
||||
|
||||
---
|
||||
|
||||
#### Task D5：版本标记（~0.5d）
|
||||
|
||||
```bash
|
||||
git tag v0.1.0 && git push --tags
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 并行机会
|
||||
|
||||
| 并行组 | 包含任务 | 说明 |
|
||||
|--------|---------|------|
|
||||
| 组 0 | A0 | 单独执行，后续所有 Task 的前提 |
|
||||
| 组 1 | A1 + A2 + A3 | A0 完成后可并行修 |
|
||||
| 组 2 | B1a + B1b + B1d | 三个示例无前置依赖（B1a 需等 B0，B1c 需等 A1） |
|
||||
| 组 3 | C1 + C2 + D2 | README / 错误消息 / Roadmap 更新，纯文档工作 |
|
||||
| 组 4 | D1 + D3 | CI 验证 + 依赖裁剪，可并行 |
|
||||
|
||||
---
|
||||
|
||||
## 风险与应对
|
||||
|
||||
| 风险 | 影响 | 概率 | 应对 |
|
||||
|------|------|------|------|
|
||||
| A0 修复后仍有未发现的测试回归 | 🔴 后续 Task 验收不可信 | 低 | `cargo test` 全量通过后才启动 A1 |
|
||||
| composer.rs 迁移发现未预见的 API 依赖 | 🔴 A1 延期 | 低 | 保持每次 commit 可编译；分 2 次提交（先私有不影响构建，再改返回类型） |
|
||||
| MockProvider stream 实现比预期复杂 | 🟡 B2c 延期 | 中 | 先简化实现（只支持基本 text delta），复杂 case 留给 v0.2 |
|
||||
| 示例与库 API 不同步 | 🟡 持续风险 | 中 | 所有示例纳入 `cargo test --examples` 作为 CI gate |
|
||||
| clippy `#[allow(dead_code)]` 积累过多 | 🟢 低 | 高 | v0.2 发布前专门 review 一次允许列表 |
|
||||
| tokio feature 裁剪导致编译失败 | 🟡 D3 延期 | 低 | 裁剪前确认现有 feature 覆盖所有 `tokio::` 调用点 |
|
||||
|
||||
---
|
||||
|
||||
## 验收标准
|
||||
|
||||
1. **代码质量**：`cargo build` 0 错误，`cargo clippy --lib -p agcore` 0 警告
|
||||
2. **测试覆盖**：`cargo test` 全量编译通过且全部运行通过，`cargo test --examples` 全通过
|
||||
3. **示例可用**：4 个🥇 + 3 个🥈共 7 个示例均可离线 `cargo run` 成功退出
|
||||
4. **文档完整**：README 包含快速上手 + 架构概览，错误消息全部友好化
|
||||
5. **版本标记**：`git tag v0.1.0`
|
||||
6. **Roadmap 同步**：`docs/roadmap.md` 测试计数和状态与代码一致
|
||||
@@ -0,0 +1,522 @@
|
||||
# Phase 5：热身准备 — 实施方案
|
||||
|
||||
## 1. 背景与目标
|
||||
|
||||
Phase 5 是 v0.2.0 发布周期的**热身准备阶段**，包含三个互不依赖的 Step，为后续 Phase 6-12 的端到端集成提供基础设施。
|
||||
|
||||
**核心目标**：
|
||||
- 为 Phase 8（端到端示例）提供零 API key 的运行路径（Ollama）
|
||||
- 为公共枚举的向后兼容性加上编译期护栏（`#[non_exhaustive]`）
|
||||
- 为 Provider 构造提供统一的超时与重试配置入口（`ProviderConfig` 扩展）
|
||||
|
||||
三个 Step 之间**无依赖关系**，但出于实现效率考虑，按 **5.2 → 5.3 → 5.1** 顺序执行。理由：5.2 先新增 `Ollama` 枚举变体，5.3 再加 `#[non_exhaustive]`，避免枚举标记后添加变体需要在外部 crate 加 `_ =>` 兜底分支的困扰。
|
||||
|
||||
## 2. 需求分析
|
||||
|
||||
### Step 5.2 — Ollama Provider
|
||||
|
||||
| 维度 | 内容 |
|
||||
|------|------|
|
||||
| **需求** | 新增 `OllamaProvider`，newtype 包装 `GenericOpenaiProvider`，默认连接本地 Ollama 实例 |
|
||||
| **优先级** | P0 — 为 Phase 8 端到端示例提供无需 API key 的运行路径 |
|
||||
| **预期交付物** | `src/llm/provider/ollama.rs` 新建文件；`ProviderType` 新增 `Ollama` 变体 |
|
||||
| **代码量** | ~55 行 |
|
||||
|
||||
### Step 5.3 — `#[non_exhaustive]` 前置标记
|
||||
|
||||
| 维度 | 内容 |
|
||||
|------|------|
|
||||
| **需求** | 为 4 个公共枚举添加 `#[non_exhaustive]` 属性，避免后续新增变体时破坏下游 match |
|
||||
| **优先级** | P1 — 编译期兼容性保障 |
|
||||
| **预期交付物** | 修改 4 个枚举定义，各加一行属性 |
|
||||
| **代码量** | ~4 行 |
|
||||
|
||||
### Step 5.1 — ProviderConfig 扩展
|
||||
|
||||
| 维度 | 内容 |
|
||||
|------|------|
|
||||
| **需求** | `ProviderConfig` 新增 `timeout_secs` 和 `max_retries` 字段；实现 `Default`、`from_env()` 构造；timeout 传导到各 Provider HTTP Client |
|
||||
| **优先级** | P0 — 与 Roadmap 一致，Phase 8（MVP 出口）依赖 from_env |
|
||||
| **预期交付物** | `ProviderConfig` 扩展；`create_provider()` 超时注入；`from_env()` + 单元测试 |
|
||||
| **代码量** | ~60 行 + 测试 |
|
||||
|
||||
## 3. 方案设计
|
||||
|
||||
### 3.1 Step 5.2 — Ollama Provider（先执行）
|
||||
|
||||
#### 改动文件清单
|
||||
|
||||
| 文件 | 操作 | 说明 |
|
||||
|------|------|------|
|
||||
| `src/llm/provider/ollama.rs` | **新建** | OllamaProvider newtype 包装 |
|
||||
| `src/llm/provider.rs` | 修改 | `ProviderType` 新增 `Ollama` 变体；`FromStr` 加解析；`create_provider()` 加分支 |
|
||||
| `src/llm/provider/mod.rs` 或其他模块注册文件 | 修改（如需要） | 注册 `pub mod ollama` |
|
||||
|
||||
#### 关键代码
|
||||
|
||||
**`src/llm/provider/ollama.rs`**（新建）：
|
||||
|
||||
```rust
|
||||
//! Ollama Provider —— OpenAI-compatible 协议的 newtype 包装，零 API key。
|
||||
//!
|
||||
//! 默认 base_url = `http://localhost:11434/v1`，空 api_key 也可工作。
|
||||
//! 实现方式同 DeepSeekProvider / QwenProvider，共享 GenericOpenaiProvider 的 HTTP/SSE/转换逻辑。
|
||||
|
||||
use reqwest::Client;
|
||||
use std::pin::Pin;
|
||||
|
||||
use async_trait::async_trait;
|
||||
use futures_core::Stream;
|
||||
|
||||
use super::openai::GenericOpenaiProvider;
|
||||
use super::{LlmProvider, ProviderCapabilities};
|
||||
use crate::llm::error::LlmError;
|
||||
use crate::llm::types::request_v2::MessageRequest;
|
||||
use crate::llm::types::response_v2::{MessageResponse, StreamEvent};
|
||||
|
||||
pub struct OllamaProvider(pub GenericOpenaiProvider);
|
||||
|
||||
impl OllamaProvider {
|
||||
pub fn new(base_url: String, api_key: String, model: String) -> Self {
|
||||
let url = if base_url.is_empty() {
|
||||
"http://localhost:11434/v1".to_string()
|
||||
} else {
|
||||
base_url
|
||||
};
|
||||
Self(GenericOpenaiProvider::new_with_name(
|
||||
url,
|
||||
api_key,
|
||||
model,
|
||||
"ollama",
|
||||
))
|
||||
}
|
||||
|
||||
/// 替换默认 HTTP Client（用于 timeout 注入等场景）。
|
||||
/// 与 `OpenaiChatProvider::with_client` 和 `DeepSeekProvider::with_client` 一致。
|
||||
pub fn with_client(self, client: Client) -> Self {
|
||||
Self(self.0.with_client(client))
|
||||
}
|
||||
}
|
||||
|
||||
#[async_trait]
|
||||
impl LlmProvider for OllamaProvider {
|
||||
async fn chat(&self, request: MessageRequest) -> Result<MessageResponse, LlmError> {
|
||||
self.0.chat(request).await
|
||||
}
|
||||
|
||||
async fn chat_stream(
|
||||
&self,
|
||||
request: MessageRequest,
|
||||
) -> Result<Pin<Box<dyn Stream<Item = Result<StreamEvent, LlmError>> + Send>>, LlmError> {
|
||||
self.0.chat_stream(request).await
|
||||
}
|
||||
|
||||
fn capabilities(&self) -> ProviderCapabilities {
|
||||
let mut caps = self.0.capabilities();
|
||||
caps.provider_name = "ollama";
|
||||
caps
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
**`src/llm/provider.rs`** 的修改：
|
||||
|
||||
```rust
|
||||
// ProviderType 新增变体
|
||||
#[derive(Debug, Clone, Copy, PartialEq, Eq)]
|
||||
pub enum ProviderType {
|
||||
OpenaiChat,
|
||||
OpenaiResponse,
|
||||
Anthropic,
|
||||
DeepSeek,
|
||||
Qwen,
|
||||
/// Ollama（本地），默认 base_url = `http://localhost:11434/v1`。
|
||||
Ollama,
|
||||
}
|
||||
|
||||
// FromStr 加解析
|
||||
fn from_str(s: &str) -> Result<Self, Self::Err> {
|
||||
match s.to_lowercase().as_str() {
|
||||
// ... 已有条目 ...
|
||||
"ollama" => Ok(ProviderType::Ollama),
|
||||
_ => Err(format!("未知的 Provider 类型: {s}")),
|
||||
}
|
||||
}
|
||||
|
||||
// create_provider() 加分支
|
||||
// Step 5.2 阶段仅展示基本构造。Step 5.1（ProviderConfig 扩展）
|
||||
// 执行到此分支时，将同步补充 with_client 链式调用注入 timeout：
|
||||
//
|
||||
// let client = Client::builder()
|
||||
// .timeout(Duration::from_secs(config.timeout_secs))
|
||||
// .build()?;
|
||||
// Ok(Box::new(
|
||||
// ollama::OllamaProvider::new(config.base_url, config.api_key, config.model)
|
||||
// .with_client(client),
|
||||
// ))
|
||||
ProviderType::Ollama => Ok(Box::new(ollama::OllamaProvider::new(
|
||||
config.base_url,
|
||||
config.api_key,
|
||||
config.model,
|
||||
))),
|
||||
```
|
||||
|
||||
#### 集成方式
|
||||
|
||||
OllamaProvider 的 newtype 包装模式与 `DeepSeekProvider`、`QwenProvider` 完全一致，`LlmProvider` trait 委托给 `self.0`。`capabilities().provider_name` 返回 `"ollama"`。
|
||||
|
||||
### 3.2 Step 5.3 — `#[non_exhaustive]` 前置标记
|
||||
|
||||
#### 改动文件清单
|
||||
|
||||
| 文件 | 行号 | 枚举 | 操作 |
|
||||
|------|------|------|------|
|
||||
| `src/llm/provider.rs` | ~21 | `ProviderType` | 加 `#[non_exhaustive]` |
|
||||
| `src/llm/types/response_v2.rs` | ~22 | `StopReason` | 加 `#[non_exhaustive]` |
|
||||
| `src/llm/types/shared.rs` | ~16 | `FinishReason` | 加 `#[non_exhaustive]` |
|
||||
| `src/memory/store.rs` | ~35 | `EvictionPolicy` | 加 `#[non_exhaustive]` |
|
||||
|
||||
**排除清单**：`SlotMode`。
|
||||
|
||||
**决策理由**：`SlotMode` 枚举在 Phase 10（`src/llm/context.rs`）中才实际定义，Phase 5 尚不存在此类型。`#[non_exhaustive]` 无法标注不存在的枚举，因此排除标注。Roadmap（v0.2.0 §Phase 5 Step 5.3）列出的 `SlotMode`(预置) 推迟到 Phase 10 实现时一并添加。
|
||||
|
||||
#### 关键代码
|
||||
|
||||
每个枚举在 `derive` 上方或下方加一行属性：
|
||||
|
||||
```rust
|
||||
#[derive(Debug, Clone, Copy, PartialEq, Eq, Serialize, Deserialize)]
|
||||
#[non_exhaustive]
|
||||
pub enum ProviderType {
|
||||
// ...
|
||||
}
|
||||
```
|
||||
|
||||
#### 影响分析
|
||||
|
||||
- `#[non_exhaustive]` 是纯编译期属性，不影响运行时行为
|
||||
- 同一 crate 内的 exhaustive match 不受影响（同 crate 可穷举）
|
||||
- 下游 crate 的 match 必须加 `_ =>` 兜底分支，这是期望行为——确保未来新增变体时不会 silent break
|
||||
- **单向门**：此步骤一旦通过 `v0.2.0` 发布到公共 API 后，**不可回退**。回退意味着移除 `#[non_exhaustive]`，可能破坏已添加 `_ =>` 的下游代码。因此必须在发布前完成并确认所有枚举变体正确
|
||||
|
||||
### 3.3 Step 5.1 — ProviderConfig 扩展（最后执行）
|
||||
|
||||
#### 改动文件清单
|
||||
|
||||
| 文件 | 操作 | 说明 |
|
||||
|------|------|------|
|
||||
| `src/llm/provider.rs` | 修改 | `ProviderConfig` 加字段；加 `impl Default`；加 `from_env()`；`create_provider` 注入 timeout |
|
||||
| `src/llm/provider/openai.rs` | 修改 | `GenericOpenaiProvider` 新增 `timeout_secs` 字段；`new_with_name` 接受 timeout 参数；`map_reqwest_error` 参数化 |
|
||||
| `src/llm/provider/anthropic.rs` | 修改 | 新增 `timeout_secs` 字段；`new()` 接受 timeout 参数；`map_reqwest_error` 参数化 |
|
||||
| `src/llm/provider/anthropic.rs` | 修改 | 新增 `with_timeout()` 方法（返回 `Result<Self, LlmError>`） |
|
||||
| `src/llm/provider/openai_compat.rs` | 修改 | `DeepSeekProvider` 和 `QwenProvider` 新增公开 `with_client()` 方法 |
|
||||
| `src/llm/provider/ollama.rs` | 修改 | `OllamaProvider` 新增公开 `with_client()` 方法 |
|
||||
| `Cargo.toml` | 修改 | 加 `temp_env` dev-dependency |
|
||||
| 测试文件（`provider.rs` 内联或独立） | 新增 | `from_env` 单元测试 + timeout 传导集成测试 |
|
||||
|
||||
#### 数据结构
|
||||
|
||||
```rust
|
||||
/// Provider 构造参数 —— 通用 base_url + api_key + model + timeout/retry 配置。
|
||||
pub struct ProviderConfig {
|
||||
pub base_url: String,
|
||||
pub api_key: String,
|
||||
pub model: String,
|
||||
/// 请求超时秒数（默认 30）。应用于 Provider 的 HTTP Client 级别。
|
||||
pub timeout_secs: u64,
|
||||
/// 最大重试次数（默认 3）。当前此字段仅由 `from_env()` 采集，
|
||||
/// 实际重试逻辑由 `CycleConfig.retry.max_retries` 控制。
|
||||
/// 未来可合并到统一的 retry 配置。
|
||||
pub max_retries: u32,
|
||||
}
|
||||
|
||||
impl Default for ProviderConfig {
|
||||
fn default() -> Self {
|
||||
Self {
|
||||
base_url: String::new(),
|
||||
api_key: String::new(),
|
||||
model: String::new(),
|
||||
timeout_secs: 30,
|
||||
max_retries: 3,
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
impl ProviderConfig {
|
||||
/// 从环境变量构造 ProviderConfig。
|
||||
///
|
||||
/// 必填变量：
|
||||
/// - `{prefix}_BASE_URL`
|
||||
/// - `{prefix}_API_KEY`
|
||||
/// - `{prefix}_MODEL`
|
||||
///
|
||||
/// 可选变量（有默认值）：
|
||||
/// - `{prefix}_TIMEOUT_SECS`（默认 30）
|
||||
/// - `{prefix}_MAX_RETRIES`（默认 3）
|
||||
pub fn from_env(prefix: &str) -> Result<Self, String> {
|
||||
let base_url = std::env::var(format!("{prefix}_BASE_URL"))
|
||||
.map_err(|_| format!("{prefix}_BASE_URL 环境变量未设置"))?;
|
||||
let api_key = std::env::var(format!("{prefix}_API_KEY"))
|
||||
.map_err(|_| format!("{prefix}_API_KEY 环境变量未设置"))?;
|
||||
let model = std::env::var(format!("{prefix}_MODEL"))
|
||||
.map_err(|_| format!("{prefix}_MODEL 环境变量未设置"))?;
|
||||
let timeout_secs = match std::env::var(format!("{prefix}_TIMEOUT_SECS")) {
|
||||
Ok(v) => v.parse().unwrap_or_else(|_| {
|
||||
tracing::warn!("{prefix}_TIMEOUT_SECS='{v}' 解析失败，使用默认值 30");
|
||||
30
|
||||
}),
|
||||
Err(_) => 30,
|
||||
};
|
||||
let max_retries = match std::env::var(format!("{prefix}_MAX_RETRIES")) {
|
||||
Ok(v) => v.parse().unwrap_or_else(|_| {
|
||||
tracing::warn!("{prefix}_MAX_RETRIES='{v}' 解析失败，使用默认值 3");
|
||||
3
|
||||
}),
|
||||
Err(_) => 3,
|
||||
};
|
||||
|
||||
// ponytail: max_retries 当前仅采集，不传入 Provider。
|
||||
// 实际重试由 CycleConfig.retry.max_retries 控制。
|
||||
// 此 warn 在应用启动时通常只触发一次，多次调用 from_env 时
|
||||
// 重复输出的风险低。如有噪声，可改用 std::sync::Once 控制。
|
||||
if max_retries != 3 {
|
||||
tracing::warn!(
|
||||
"ProviderConfig.max_retries={} 已采集但当前未生效；\
|
||||
重试次数由 CycleConfig.retry.max_retries 控制",
|
||||
max_retries,
|
||||
);
|
||||
}
|
||||
|
||||
Ok(Self {
|
||||
base_url,
|
||||
api_key,
|
||||
model,
|
||||
timeout_secs,
|
||||
max_retries,
|
||||
})
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
#### Timeout 传导模式
|
||||
|
||||
在 `create_provider()` 中，对基于 `GenericOpenaiProvider` 的 Provider（OpenAI / DeepSeek / Qwen / Ollama），通过同一模式注入 timeout：构造带 timeout 的 `Client` 后调用 `with_client(client)`。
|
||||
|
||||
所有 OpenAI-compatible 分支新增的 `with_client()` 公开方法：
|
||||
|
||||
| Provider | 方法 | 位置 |
|
||||
|----------|------|------|
|
||||
| `OpenaiChatProvider` | 已有 `with_client(Client) -> Self` | `openai.rs` |
|
||||
| `DeepSeekProvider` | 新增 `with_client(Client) -> Self` | `openai_compat.rs` |
|
||||
| `QwenProvider` | 新增 `with_client(Client) -> Self` | `openai_compat.rs` |
|
||||
| `OllamaProvider` | 新增 `with_client(Client) -> Self` | `ollama.rs`（新建文件） |
|
||||
|
||||
**关于 `new_with_client` 的说明**：`DeepSeekProvider` 和 `QwenProvider` 当前已有测试用的 `new_with_client(base_url, api_key, model, client)` 方法（通过 `inner.http_client = client` 直接写字段）。新增 `with_client` 后，`new_with_client` 应重构为 `Self::new(base_url, api_key, model).with_client(client)` 代理，统一走公开 API 路径。
|
||||
|
||||
代码示例（以 DeepSeek 为例，OpenAI/Qwen/Ollama 模式完全一致）：
|
||||
|
||||
```rust
|
||||
ProviderType::DeepSeek => {
|
||||
let client = Client::builder()
|
||||
.timeout(Duration::from_secs(config.timeout_secs))
|
||||
.build()
|
||||
.map_err(|e| LlmError::Other(format!("创建 HTTP 客户端失败: {e}")))?;
|
||||
Ok(Box::new(
|
||||
openai_compat::DeepSeekProvider::new(
|
||||
config.base_url,
|
||||
config.api_key,
|
||||
config.model,
|
||||
)
|
||||
.with_client(client),
|
||||
))
|
||||
}
|
||||
```
|
||||
|
||||
Anthropic 由于需要保留 `default_headers`，使用独立的 `with_timeout` 模式：
|
||||
|
||||
AnthropicProvider 新增 `with_timeout` 方法：
|
||||
|
||||
```rust
|
||||
impl AnthropicProvider {
|
||||
/// 替换默认 HTTP Client 的超时配置。
|
||||
///
|
||||
/// ⚠️ 副作用：此方法**完全重建** `http_client`，调用后原有通过 `with_client`
|
||||
/// 注入的 Client 将被替换。headers 逻辑与 `new()` 中的构造保持一致。
|
||||
pub fn with_timeout(mut self, secs: u64) -> Result<Self, LlmError> {
|
||||
// ponytail: 重建 http_client 时保留已有默认 headers（x-api-key / anthropic-version）。
|
||||
// 如后续 AnthropicProvider 的 headers 变为动态，此方法需同步更新。
|
||||
let key_header = HeaderValue::from_str(&self.api_key)
|
||||
.map_err(|_| LlmError::Other("Anthropic API key 包含无效的 HTTP 头部字符".into()))?;
|
||||
let version_header = HeaderValue::from_static("2023-06-01");
|
||||
|
||||
self.http_client = Client::builder()
|
||||
.timeout(Duration::from_secs(secs))
|
||||
.default_headers({
|
||||
let mut headers = HeaderMap::new();
|
||||
headers.insert("x-api-key", key_header);
|
||||
headers.insert("anthropic-version", version_header);
|
||||
headers
|
||||
})
|
||||
.build()
|
||||
.map_err(|e| LlmError::Other(format!("创建 Anthropic HTTP 客户端失败: {e}")))?;
|
||||
Ok(self)
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
#### `map_reqwest_error` 中的硬编码超时修复
|
||||
|
||||
`openai.rs` 和 `anthropic.rs` 中的 `map_reqwest_error` 辅助函数当前在超时错误中返回硬编码的 `Duration::from_secs(120)`：
|
||||
|
||||
```rust
|
||||
// 现状 —— 硬编码 120s，与可配置 timeout 脱节
|
||||
LlmError::Timeout { duration: Duration::from_secs(120) }
|
||||
```
|
||||
|
||||
**修复方式**：采用**方案 A**——在 Provider struct 中存储 `timeout_secs` 字段，`map_reqwest_error` 读取该字段的值而非硬编码 120s。
|
||||
|
||||
```rust
|
||||
// 修复后 —— 参数化，从 Provider 存储的 timeout_secs 读取
|
||||
// GenericOpenaiProvider 新增 timeout_secs 字段：
|
||||
pub struct GenericOpenaiProvider {
|
||||
http_client: Client,
|
||||
base_url: String,
|
||||
api_key: String,
|
||||
model: String,
|
||||
provider_name: &'static str,
|
||||
extra_headers: Vec<(String, String)>,
|
||||
timeout_secs: u64, // ← 新增，由 new_with_name 的参数传入
|
||||
}
|
||||
|
||||
// map_reqwest_error 使用 self.timeout_secs 而非硬编码 120：
|
||||
LlmError::Timeout { duration: Duration::from_secs(self.timeout_secs) }
|
||||
```
|
||||
|
||||
**方案 B（从 reqwest::Client 提取 timeout）已被否决**：`reqwest::Client` 不提供 timeout getter，无法从已构造的 client 中反向读取超时配置。
|
||||
|
||||
如果漏掉此修复，用户设置 `AG_LLM_TIMEOUT_SECS=60` 后超时，错误消息仍显示 "LLM 请求超时（120s）"，与实际配置不符。
|
||||
|
||||
---
|
||||
|
||||
#### max_retries 说明
|
||||
|
||||
`ProviderConfig.max_retries` 当前仅由 `from_env()` 采集存储，**实际重试操作由 `CycleConfig.retry.max_retries` 控制**。两者之间的关系通过文档注释声明：
|
||||
|
||||
```rust
|
||||
/// 最大重试次数（默认 3）。当前此字段仅由 `from_env()` 采集，
|
||||
/// 实际重试逻辑由 `CycleConfig.retry.max_retries` 控制。
|
||||
/// 未来 Phase 6+ 可统一合并此字段到 CycleConfig。
|
||||
```
|
||||
|
||||
#### 测试设计
|
||||
|
||||
使用 `temp_env` 在单元测试中隔离环境变量：
|
||||
|
||||
```rust
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
|
||||
#[test]
|
||||
fn provider_config_from_env_requires_all_vars() {
|
||||
// 未设置任何变量时应返回 Err
|
||||
let result = ProviderConfig::from_env("TEST_PROVIDER");
|
||||
assert!(result.is_err());
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn provider_config_from_env_uses_defaults() {
|
||||
temp_env::with_vars([
|
||||
("TEST_PROVIDER_BASE_URL", Some("http://localhost:11434/v1")),
|
||||
("TEST_PROVIDER_API_KEY", Some("")),
|
||||
("TEST_PROVIDER_MODEL", Some("llama3")),
|
||||
], || {
|
||||
let config = ProviderConfig::from_env("TEST_PROVIDER").unwrap();
|
||||
assert_eq!(config.timeout_secs, 30);
|
||||
assert_eq!(config.max_retries, 3);
|
||||
});
|
||||
}
|
||||
|
||||
#[test]
|
||||
fn provider_config_from_env_reads_custom_timeout() {
|
||||
temp_env::with_vars([
|
||||
("TEST_PROVIDER_BASE_URL", Some("http://x")),
|
||||
("TEST_PROVIDER_API_KEY", Some("k")),
|
||||
("TEST_PROVIDER_MODEL", Some("m")),
|
||||
("TEST_PROVIDER_TIMEOUT_SECS", Some("60")),
|
||||
("TEST_PROVIDER_MAX_RETRIES", Some("5")),
|
||||
], || {
|
||||
let config = ProviderConfig::from_env("TEST_PROVIDER").unwrap();
|
||||
assert_eq!(config.timeout_secs, 60);
|
||||
assert_eq!(config.max_retries, 5);
|
||||
});
|
||||
}
|
||||
}
|
||||
```
|
||||
|
||||
## 4. 实现计划
|
||||
|
||||
### Step 5.2 — Ollama Provider（~55 行）
|
||||
|
||||
| 步骤 | 操作 | 验证 |
|
||||
|------|------|------|
|
||||
| 1 | 创建 `src/llm/provider/ollama.rs`，实现 `OllamaProvider` newtype | 编译通过 |
|
||||
| 2 | 在 `provider.rs` 注册 `pub mod ollama` | 编译通过 |
|
||||
| 3 | `ProviderType` 新增 `Ollama` 变体 | 编译通过 |
|
||||
| 4 | `FromStr` 加 `"ollama"` 解析 | 编译通过 |
|
||||
| 5 | `create_provider()` 加 `Ollama =>` 分支 | 编译通过 |
|
||||
| 6 | 运行 `cargo build` | 无错误 |
|
||||
|
||||
### Step 5.3 — `#[non_exhaustive]` 前置标记（~4 行）
|
||||
|
||||
| 步骤 | 操作 | 验证 |
|
||||
|------|------|------|
|
||||
| 1 | `ProviderType`（`provider.rs`）加 `#[non_exhaustive]` | 编译通过 |
|
||||
| 2 | `StopReason`（`response_v2.rs`）加 `#[non_exhaustive]` | 编译通过 |
|
||||
| 3 | `FinishReason`（`shared.rs`）加 `#[non_exhaustive]` | 编译通过 |
|
||||
| 4 | `EvictionPolicy`（`memory/store.rs`）加 `#[non_exhaustive]` | 编译通过 |
|
||||
| 5 | 运行 `cargo build --all-targets` | 无 warning |
|
||||
|
||||
### Step 5.1 — ProviderConfig 扩展（~60 行 + 测试）
|
||||
|
||||
| 步骤 | 操作 | 验证 |
|
||||
|------|------|------|
|
||||
| 1 | `ProviderConfig` 加 `timeout_secs` / `max_retries` 字段 | 编译通过 |
|
||||
| 2 | 实现 `impl Default for ProviderConfig` | 编译通过 |
|
||||
| 3 | 实现 `ProviderConfig::from_env()` | 编译通过 |
|
||||
| 4 | `GenericOpenaiProvider` 和 `AnthropicProvider` 新增 `timeout_secs` 字段，`new_with_name`／`new()` 接受 timeout 参数 | 编译通过 |
|
||||
| 5 | `map_reqwest_error` 在各 Provider 中改为从 `self.timeout_secs` 读取，移除硬编码 120s | 编译通过 |
|
||||
| 6 | `create_provider()` 中各分支注入 timeout（OpenAI-compatible 用 `Client::builder().timeout()` + `with_client`；Anthropic 用 `with_timeout()`） | 编译通过 |
|
||||
| 7 | `DeepSeekProvider`／`QwenProvider` 的 `new_with_client` 重构为 `Self::new(...).with_client(client)` 代理 | 测试通过 |
|
||||
| 8 | `Cargo.toml` 添加 `temp_env` dev-dependency | `cargo build` 通过 |
|
||||
| 8 | 添加 `from_env` 单元测试 + timeout 传导集成测试 | `cargo test` 通过 |
|
||||
| 9 | 完整验证 | 见第 6 节 |
|
||||
|
||||
## 5. 风险评估
|
||||
|
||||
| 风险 | 影响 | 概率 | 缓解措施 |
|
||||
|------|------|------|----------|
|
||||
| `create_provider()` 中 `Client::builder().build()` 返回 `Result`，当前代码使用 `.expect()`，改为 `map_err` 转为 `LlmError` 后需确保所有分支正确转换 | 编译期强制处理，遗漏分支直接报错 | 低 | `create_provider` 返回 `Result<Box<dyn LlmProvider>, LlmError>`，`map_err` 天然适配。新增的 timeout 注入路径逐一检查 |
|
||||
| `AnthropicProvider` 的 `default_headers` 在 `with_timeout` 中重建时与 `new()` 中的 headers 不一致 | Anthropic 认证失败 | 低 | `with_timeout` 方法复制 `new()` 中的 headers 构造逻辑。通过已有测试验证认证通过 |
|
||||
| Ollama 实际运行时行为差异：版本兼容性、API 路径、模型名等 | 运行时才能发现 | 中 | Phase 5 仅做类型级验证（`cargo build`），Phase 8 端到端测试时通过 Ollama mock 或真实实例验证 |
|
||||
| `max_retries` 存储了却未实际使用，造成困惑 | 开发者误以为已生效 | 中 | 通过文档注释明确声明 `max_retries` 当前仅采集，实际重试由 `CycleConfig.retry.max_retries` 控制 |
|
||||
| `temp_env` 测试在多线程并发测试中互相污染环境变量 | 偶发测试失败 | 中（Rust 默认单线程测试用 `--test-threads=1` 可避免） | 将 `from_env` 测试控制在同一测试文件，避免并行执行。必要时在 CI 中确保 `--test-threads=1` |
|
||||
|
||||
## 6. 验收标准
|
||||
|
||||
以下条件**全部满足**方可认为 Phase 5 完成：
|
||||
|
||||
- [ ] `cargo build --all-targets` 通过，无错误
|
||||
- [ ] `cargo test --all-targets` 通过，新增测试覆盖 `from_env` 的必填/选填/默认值场景
|
||||
- [ ] `cargo clippy --all-targets -- -D warnings` 通过，无任何 warning
|
||||
- [ ] `cargo doc --no-deps -D warnings` 通过，所有公共 API 有文档注释（`///`）
|
||||
- [ ] 新增文件：1（`ollama.rs`）
|
||||
- [ ] 修改文件：9（`provider.rs`、`openai.rs`、`anthropic.rs`、`openai_compat.rs`、`response_v2.rs`、`shared.rs`、`store.rs`、`Cargo.toml`、测试文件）
|
||||
- [ ] 净代码增量：~160 行
|
||||
- [ ] `ProviderType` 新增 `Ollama` 变体，`"ollama"` 字符串可解析
|
||||
- [ ] 4 个公共枚举带有 `#[non_exhaustive]` 属性
|
||||
- [ ] `ProviderConfig` 可从环境变量构造（`from_env()`），含默认值
|
||||
- [ ] timeout 值已传导到 `create_provider()` 中各 Provider 的 HTTP Client 配置
|
||||
- [ ] timeout 传导验证通过至少一个端到端 wiremock 集成测试（模拟 HTTP 服务在超时后返回 408，验证 Provider 返回 `LlmError::Timeout`）
|
||||
- [ ] `DeepSeekProvider`、`QwenProvider`、`OllamaProvider` 均有公开 `with_client()` 方法，可在 `create_provider` 中注入 timeout Client
|
||||
- [ ] `map_reqwest_error` 中不再硬编码 `Duration::from_secs(120)`，改为参数化读取
|
||||
@@ -0,0 +1,344 @@
|
||||
# 笔记：opencode 子代理调度、分发与合并及工作流推进
|
||||
|
||||
> 基于 `/Users/midnite/Samples/opencode` 源码调研，2026-07-04
|
||||
|
||||
---
|
||||
|
||||
## 一、整体架构
|
||||
|
||||
```
|
||||
LLM（主 Agent）
|
||||
│
|
||||
├── 调用 Task tool（tool call）
|
||||
│ ↓
|
||||
│ TaskTool.execute() ← packages/opencode/src/tool/task.ts
|
||||
│ │
|
||||
│ ├── agent.get() ← 查找 Agent 定义（agent.ts）
|
||||
│ ├── deriveSubagentPermission() ← 权限合并（subagent-permissions.ts）
|
||||
│ ├── sessions.create() ← 创建子 session
|
||||
│ │
|
||||
│ ├── [前台] background.wait() + background.waitForPromotion() race
|
||||
│ │ ↓ 完成
|
||||
│ │ renderOutput() → XML <task> 标签返回
|
||||
│ │
|
||||
│ └── [后台] background.start() → notify() 异步注入结果
|
||||
│
|
||||
└── 会话循环（runLoop） ← prompt.ts
|
||||
│
|
||||
├── 检测 subtask type part → handleSubtask()
|
||||
├── 检测 compaction → compaction.process()
|
||||
└── 正常流程 → LLM.stream() → processor.handleEvent()
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 二、子代理调度（Dispatch）
|
||||
|
||||
### 2.1 三种触发入口
|
||||
|
||||
| 入口 | 触发方式 | 调用链路 |
|
||||
|------|---------|---------|
|
||||
| A — LLM 自主 | LLM 调用 `task` tool | 系统提示词中注入了 Task tool 描述 + `describeTask()` 输出子代理列表 → LLM 决策 |
|
||||
| B — `subtask` part | 消息中有 `type: "subtask"` 的 part | `handleSubtask()` 直接执行 TaskTool，不走 LLM |
|
||||
| C — `agent` part | 消息中有 `type: "agent"` 的 part | 转为"调用 task tool 带 subagent: XXX"的提示词，引导 LLM |
|
||||
|
||||
### 2.2 TaskTool.execute() 完整流程（task.ts）
|
||||
|
||||
```
|
||||
execute(params, ctx):
|
||||
1. background 开关检查（需 experimental flag）
|
||||
2. ctx.ask() 权限询问
|
||||
3. agent.get(subagent_type) 查找子代理定义
|
||||
4. task_id 存在 → sessions.get(task_id) 恢复已有子 session
|
||||
task_id 不存在 → sessions.create() 创建新子 session
|
||||
5. deriveSubagentSessionPermission() 合并权限
|
||||
6. 添加默认 deny 规则（todowrite / task）
|
||||
7. 确定 model（继承或子代理自定义）
|
||||
8. 执行 runTask() → ops.resolvePromptParts() + ops.prompt()
|
||||
9. 结果格式化为 XML ← renderOutput()
|
||||
```
|
||||
|
||||
### 2.3 关键：子 session 创建（task.ts lines 121-158）
|
||||
|
||||
```typescript
|
||||
// 权限继承
|
||||
const childPermission = deriveSubagentSessionPermission({
|
||||
parentSessionPermission: parent.permission ?? [],
|
||||
subagent: next,
|
||||
})
|
||||
|
||||
// 默认 deny 规则
|
||||
const childToolDenies = [
|
||||
// 子代理自己的 permission 没允许 todowrite → 默认 deny
|
||||
...(next.permission.some(r => r.permission === "todowrite") ? []
|
||||
: [{ permission: "todowrite", pattern: "*", action: "deny" }]),
|
||||
// 子代理自己的 permission 没允许 task → 默认 deny（防嵌套）
|
||||
...(next.permission.some(r => r.permission === "task") ? []
|
||||
: [{ permission: "task", pattern: "*", action: "deny" }]),
|
||||
// 主 agent 专有工具也不给子代理
|
||||
...(cfg.experimental?.primary_tools?.map(p => ({ permission: p, ... })) ?? []),
|
||||
]
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 三、通信格式：Tool Call / Tool Result
|
||||
|
||||
### 3.1 父→子：Task tool 参数
|
||||
|
||||
```
|
||||
{
|
||||
subagent_type: "explore" | "general" | ...,
|
||||
description: "简短描述（3-5词）",
|
||||
prompt: "子代理的完整任务描述",
|
||||
task_id?: "恢复已有子 session 时使用",
|
||||
command?: "触发该调用的 CLI 命令（可选）",
|
||||
background?: true // 后台模式（需 experimental flag）
|
||||
}
|
||||
```
|
||||
|
||||
### 3.2 子→父：XML 包装的纯文本（renderOutput）
|
||||
|
||||
```xml
|
||||
<task id="ses_xxxxx" state="completed">
|
||||
<summary>任务简述</summary>
|
||||
<task_result>
|
||||
子 agent 输出的完整文本内容...
|
||||
</task_result>
|
||||
</task>
|
||||
```
|
||||
|
||||
错误时：
|
||||
|
||||
```xml
|
||||
<task id="ses_xxxxx" state="error">
|
||||
<summary>任务失败</summary>
|
||||
<task_error>
|
||||
Error: 具体错误信息...
|
||||
</task_error>
|
||||
</task>
|
||||
```
|
||||
|
||||
### 3.3 传递给 LLM 的方式
|
||||
|
||||
**前台模式**：
|
||||
```
|
||||
TaskTool.execute() 返回 { output: "<task>...</task>" }
|
||||
↓
|
||||
AI SDK 将其转为 tool result，存入数据库 tool part
|
||||
↓
|
||||
下一轮 LLM 调用时，tool result 作为消息历史的一部分传入
|
||||
↓
|
||||
LLM 看到 XML，自行解析使用
|
||||
```
|
||||
|
||||
**后台模式**：
|
||||
```
|
||||
TaskTool.execute() 立即返回 <task state="running">...
|
||||
↓
|
||||
子 agent 完成后 → background.wait() 触发 → inject()
|
||||
↓
|
||||
向父 session 注入合成 text part（synthetic: true）
|
||||
携带 <task state="completed">... 结果
|
||||
↓
|
||||
父 LLM 在下一轮循环中看到该消息
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 四、分发与合并（Distribution & Merge）
|
||||
|
||||
### 4.1 并行分发
|
||||
|
||||
- **无专用分发层**。依赖 LLM 在单条消息中发出多个 tool call
|
||||
- `task.txt` 引导 LLM：*"Launch multiple agents concurrently whenever possible"*
|
||||
- 底层通过 Effect.ts 的 `Effect.forkIn(scope, { startImmediately: true })` 实现同一消息内多 tool call 并发
|
||||
- **子 agent 之间完全隔离**，无直接通信
|
||||
|
||||
### 4.2 结果合并
|
||||
|
||||
**无专用合并逻辑。** 合并完全通过 LLM 的上下文理解完成：
|
||||
|
||||
- 前台：tool result 自然进入消息历史，LLM 下一轮读取
|
||||
- CLI 命令：额外注入 "Summarize the task tool output above and continue with your task." 引导 LLM 总结
|
||||
- LLM 自主调用：无额外引导，LLM 自行决定如何使用
|
||||
|
||||
### 4.3 前台/后台切换机制（task.ts lines 303-333）
|
||||
|
||||
```typescript
|
||||
// 前台执行
|
||||
return yield* Effect.raceFirst(
|
||||
background.wait({ id: nextSession.id }), // 等完成
|
||||
background.waitForPromotion(nextSession.id), // 等 promote 到后台
|
||||
)
|
||||
```
|
||||
|
||||
当用户将前台任务 promote 到后台时，`waitForPromotion` 先返回（标记 `metadata.background = true`），TaskTool 转而返回后台模式的输出。
|
||||
|
||||
### 4.4 后台作业引擎（core/background-job.ts）
|
||||
|
||||
纯内存、非持久化注册表。使用 Effect.ts 的 `SynchronizedRef` 做并发控制。
|
||||
|
||||
| 操作 | 行为 |
|
||||
|------|------|
|
||||
| `start()` | 创建 job，fork run effect，返回 info |
|
||||
| `extend()` | 追加顺序执行的 run（通过 `Deferred` 链式等待前一个完成） |
|
||||
| `wait()` | `Deferred.await(done)`，可选 timeout |
|
||||
| `waitForPromotion()` | 等待 `promoted` Deferred 或检测 `background` 标记 |
|
||||
| `promote()` | 标记 `background = true`，触发 `onPromote` callback |
|
||||
| `cancel()` | 设置 `cancelled`，close scope（中断所有子 fork） |
|
||||
|
||||
---
|
||||
|
||||
## 五、工作流推进（Workflow Progression）
|
||||
|
||||
### 5.1 核心循环（prompt.ts → runLoop）
|
||||
|
||||
```
|
||||
runLoop(sessionID):
|
||||
while true:
|
||||
1. MessageV2.filterCompactedEffect() 获取消息
|
||||
2. MessageV2.latest() 取最近 user/assistant/tasks
|
||||
3. 检查 finish 状态
|
||||
- 不是 tool-calls 且有 finish → break（退出循环）
|
||||
4. 取 tasks（subtask / compaction 队列）
|
||||
- subtask → handleSubtask() → continue
|
||||
- compaction → compaction.process() → continue/break
|
||||
5. 检查 overflow → 自动创建 compaction task → continue
|
||||
6. 构建 assistant message
|
||||
7. SessionProcessor.create() 创建 handle
|
||||
8. SessionTools.resolve() 解析所有工具
|
||||
9. 构建 system prompt（环境信息 + skills + MCP + instructions）
|
||||
10. handle.process() — 启动 LLM stream
|
||||
11. 检查 result:
|
||||
- "compact" → 返回给外层触发 compaction
|
||||
- "stop" → break
|
||||
- "continue" → 继续循环
|
||||
```
|
||||
|
||||
### 5.2 SessionProcessor 事件处理（processor.ts）
|
||||
|
||||
| Stream 事件 | 处理逻辑 |
|
||||
|------------|---------|
|
||||
| `reasoning-start/delta/end` | 创建 reasoning part → 增量追加 → 最终持久化 |
|
||||
| `tool-input-start/delta/end` | 创建/更新 tool part（pending 状态） |
|
||||
| `tool-call` | 标记 running → 设置 input → **doom loop 检测** |
|
||||
| `tool-result` | `completeToolCall()` → 持久化结果 + 附件 |
|
||||
| `tool-error` | `failToolCall()` → 标记错误 |
|
||||
| `provider-error` | 抛出异常 → 触发重试 |
|
||||
| `text-start/delta/end` | 流式文本 → `updatePartDelta()` **增量持久化** |
|
||||
| `step-start` | 创建快照（snapshot） |
|
||||
| `step-finish` | 生成 patch diff → 更新 usage/tokens → **overflow 检测** → 触发 summary |
|
||||
| `finish` | stream 结束 |
|
||||
|
||||
### 5.3 Doom Loop 检测（processor.ts lines 351-377）
|
||||
|
||||
连续 3 次**完全相同的 tool call**（相同名称 + 相同输入）触发权限询问：
|
||||
|
||||
```typescript
|
||||
const recentParts = parts.slice(-DOOM_LOOP_THRESHOLD) // DOOM_LOOP_THRESHOLD = 3
|
||||
if (recentParts.length === DOOM_LOOP_THRESHOLD &&
|
||||
recentParts.every(part =>
|
||||
part.type === "tool" &&
|
||||
part.tool === value.name &&
|
||||
part.state.status !== "pending" &&
|
||||
JSON.stringify(part.state.input) === JSON.stringify(input)
|
||||
)) {
|
||||
yield* permission.ask({ permission: "doom_loop", ... })
|
||||
}
|
||||
```
|
||||
|
||||
### 5.4 Compaction 工作流
|
||||
|
||||
两种触发方式：
|
||||
|
||||
| 触发条件 | 行为 |
|
||||
|---------|------|
|
||||
| step-finish 检测到 `isOverflow()` + `auto: true` | 创建 compaction task → 下一轮循环执行 → 压缩后 continue |
|
||||
| step-finish 检测到 `isOverflow()` + `auto: false` | 标记 `assistantMessage.error` → idle 等待用户干预 |
|
||||
|
||||
Compaction 使用专门的 `compaction` agent（hidden, mode=primary, `*=deny`）执行。
|
||||
压缩后的消息标记 `compacted: true`，后续通过 `MessageV2.filterCompactedEffect()` 过滤。
|
||||
|
||||
### 5.5 重试机制（processor.ts lines 658-672）
|
||||
|
||||
```typescript
|
||||
Effect.retry(
|
||||
SessionRetry.policy({
|
||||
provider: input.model.providerID,
|
||||
parse, // 错误解析（区分可重试/不可重试）
|
||||
set: (info) => status.set(sessionID, { type: "retry", ... }),
|
||||
}),
|
||||
)
|
||||
```
|
||||
|
||||
遇 provider 错误自动重试，LLM stream 完成后 `Effect.ensuring(cleanup)` 保证资源释放。
|
||||
|
||||
---
|
||||
|
||||
## 六、六种内置 Agent
|
||||
|
||||
| 名称 | Mode | Hidden | 用途 | 核心权限特征 |
|
||||
|------|------|--------|------|-------------|
|
||||
| `build` | primary | 否 | 默认 agent，全部工具 | question/plan_enter=allow |
|
||||
| `plan` | primary | 否 | 计划模式，禁用编辑 | edit=deny(除 plans), task(general)=deny |
|
||||
| `general` | subagent | 否 | 通用子代理 | todowrite=deny（默认禁止改 todo） |
|
||||
| `explore` | subagent | 否 | 只读代码探索 | `*=deny`，仅 read/grep/glob/bash/webfetch/websearch |
|
||||
| `compaction` | primary | 是 | 会话压缩（自动） | `*=deny` |
|
||||
| `title` | primary | 是 | 生成会话标题 | `*=deny`（step=1 时异步 fork） |
|
||||
| `summary` | primary | 是 | 生成消息摘要 | `*=deny`（每个 step-finish 时异步 fork） |
|
||||
|
||||
用户可通过 `config.agent` 自定义 agent（支持 `mode: "all"`），也可通过 `agent.generate` 让 LLM 辅助生成。
|
||||
|
||||
---
|
||||
|
||||
## 七、权限模型总结
|
||||
|
||||
```
|
||||
父 session permission
|
||||
│
|
||||
├── 仅继承 deny 规则 + external_directory 规则 ← subagent-permissions.ts
|
||||
│ （父 agent 的 allow 规则不传播到子代理）
|
||||
│
|
||||
├── 子代理自身 permission（来自 agent 定义）
|
||||
│
|
||||
├── 默认 deny：
|
||||
│ - todowrite（除非子代理明确允许）
|
||||
│ - task（除非子代理明确允许，默认防嵌套）
|
||||
│
|
||||
└── 主 agent 专有工具 deny（来自 config.experimental.primary_tools）
|
||||
```
|
||||
|
||||
子代理的 session 权限 = **父 deny + 父 external_directory + 自身 permission - 默认 deny - primary_tools deny**。
|
||||
|
||||
---
|
||||
|
||||
## 八、关键设计决策
|
||||
|
||||
| 决策 | 意图 | 效果/局限 |
|
||||
|------|------|----------|
|
||||
| 结果以 XML 纯文本嵌入上下文 | 简单、LLM 可直接理解 | LLM 自行解析 XML；大结果可能被截断 |
|
||||
| 无专用 merge 逻辑 | 简洁，不引入额外抽象 | 依赖 LLM 的理解能力处理返回结果 |
|
||||
| 默认禁止子代理嵌套 task | 防止无限递归 | 限制了多级分解场景 |
|
||||
| 同一消息多 tool call 并发 | 利用 LLM 并行能力 | 子 agent 隔离，无法协作 |
|
||||
| Effect.ts 贯穿全程 | 类型安全、结构化并发 | 学习曲线陡峭 |
|
||||
| session 作为隔离边界 | 天然权限/消息隔离 | 每个子 session 独立数据库记录，开销较大 |
|
||||
| 后台引擎纯内存 | 有意识取舍（注释说明） | 进程重启丢失状态 |
|
||||
|
||||
---
|
||||
|
||||
## 九、参考源码路径
|
||||
|
||||
| 文件 | 角色 |
|
||||
|------|------|
|
||||
| `packages/opencode/src/tool/task.ts` | Task tool 核心实现（调度入口） |
|
||||
| `packages/opencode/src/tool/task.txt` | Task tool 的 LLM 使用说明 |
|
||||
| `packages/opencode/src/agent/agent.ts` | Agent 定义注册中心 |
|
||||
| `packages/opencode/src/agent/subagent-permissions.ts` | 子代理权限推导 |
|
||||
| `packages/opencode/src/tool/registry.ts` | 工具注册 + `describeTask()` 列出可用子代理 |
|
||||
| `packages/opencode/src/session/prompt.ts` | 会话循环 + `handleSubtask()` + 提示词构建 |
|
||||
| `packages/opencode/src/session/processor.ts` | LLM stream 事件处理器 |
|
||||
| `packages/opencode/src/session/tools.ts` | Tool ↔ AI SDK 桥接 |
|
||||
| `packages/opencode/src/session/system.ts` | 系统提示词生成（含 Task tool 说明） |
|
||||
| `packages/opencode/src/background/job.ts` | 后台作业包装层 |
|
||||
| `packages/core/src/background-job.ts` | 后台作业核心引擎（内存注册表） |
|
||||
+313
-60
@@ -1,13 +1,13 @@
|
||||
# AG Core Roadmap
|
||||
|
||||
> 定稿日期：2026-05-11
|
||||
> 最后更新：2026-06-11（Phase 4c 编码实施完成）
|
||||
> 最后更新：2026-07-04
|
||||
|
||||
## 愿景
|
||||
|
||||
AG Core 定位为构建 AI 智能体的底层工具箱，通过模块化、可插拔的架构，提供大模型调用、提示词工程、工具系统、记忆检索四大核心能力，支持快速组合出符合业务需求的智能体应用。
|
||||
|
||||
**当前状态**：Phase 0 基础设施已全部完成，Phase 1 提示词工程已全部完成，Phase 2 工具系统已全部完成，Phase 3 记忆系统已全部完成，Phase 4a 核心胶水层已全部完成，Phase 4b 任务执行已全部完成，Phase 4c 会话级记忆已全部完成（116 个测试通过，0 警告）。
|
||||
**当前状态**：v0.1.0 已发布（2026-07-04）。Phase 0-4c 全部完成，Provider IR 重构 + LlmCycle 简化 + 7 个离线示例已交付。v0.2.0 已细分为 8 个增量 Phase（Phase 5-12），本周开发启动。
|
||||
|
||||
---
|
||||
|
||||
@@ -240,95 +240,323 @@ graph BT
|
||||
|
||||
---
|
||||
|
||||
## 扩展计划（v0.2+）
|
||||
## v0.2.0 — 生产就绪（Production-Ready Core）
|
||||
|
||||
> 以下功能在已完成的 phase 中已实现基础能力或在 Phase 4 阶段明确了边界，后续可按维度增量扩展。
|
||||
> 设计参考：见 `docs/note-agent-harness-references.md`（OpenClaw / Hermes / OpenHuman / OpenHarness 横向对比）。
|
||||
> OpenCode 借鉴：见 `docs/note-opencode-agent-switching.md`（Agent 切换 + System Prompt 拼接机制）。
|
||||
**目标**：解决 Rust Agent 工具箱从"能跑"到"能被人依赖"的鸿沟。持久化、配置层、上下文管理三大块补齐后，开发者可在 30 分钟内写出生产可用的 Agent 服务。
|
||||
|
||||
### 已有扩展项（沿用）
|
||||
**总体规模**：8 个增量 Phase（Phase 5-12），17 个可验证 Step。
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| Prompt Optimizer | `prompt` | 提示词自动优化 | P3 | 待实现 |
|
||||
| 流式接口优化 | `llm/stream` | 流式响应解析与事件化 | P0 | ✅ 已完成基础实现 |
|
||||
### 功能清单
|
||||
|
||||
### v0.2+ 新增扩展项
|
||||
#### P0 — 必须交付
|
||||
|
||||
> 以下为基于 Phase 4 设计讨论确定的 v0.2+ 候选扩展方向，按维度分组。
|
||||
> 标注为"v0.2 待评估"表示在 Phase 4 完成后再决定是否启动。
|
||||
| # | 功能 | 模块 | 方案要点 |
|
||||
|---|------|------|---------|
|
||||
| 1 | SqliteStore | `memory` | `rusqlite` + `bundled` feature，`MemoryStore` 的 SQLite 实现，进程重启数据不丢 |
|
||||
| 2 | ProviderConfig 扩展 + `from_env()` | `llm` | 补全 `timeout_secs` / `max_retries` 字段；`AG_LLM_*` 环境变量辅助函数 |
|
||||
| 3 | ToolDefinition IR 正式化 | `tools` | 移除 deprecated OpenAI wire 格式，替换为自定义 `ToolDef` 结构体 |
|
||||
| 4 | API 稳定性管理 | `*` | 公开枚举加 `#[non_exhaustive]`；CHANGELOG 记录 Breaking Changes；废弃 API 用 `#[deprecated]` 标记 |
|
||||
| 5 | Quick Start + 端到端示例 | `examples/` | 30 行 `main.rs` 快速开始；一个"SQLite 持久化 + Provider + 工具调用 + 多轮对话"的可运行示例（`cargo run --example`） |
|
||||
|
||||
#### Multi-Agent / 协同
|
||||
#### P1 — 重要但不阻塞
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| Multi-Agent 协同（Swarm） | `agent` | 子 Agent 委派、并行子任务、结果聚合 | P2 | v0.2 待评估 |
|
||||
| # | 功能 | 模块 | 方案要点 |
|
||||
|---|------|------|---------|
|
||||
| 6 | Ollama Provider | `llm/provider` | OpenAI Compat，本地 LLM 支持，实现量极小 |
|
||||
| 7 | VectorRetriever trait | `memory` | 语义检索 trait 抽象（`index` / `search`），不绑定后端实现 |
|
||||
| 8 | 流式 `submit_turn_stream` | `agent` | `AgentSession` 新增 `submit_turn_stream()`，返回 `Stream<Item = StreamEvent>` |
|
||||
| 9 | 测试补强 | `*` | wiremock Provider roundtrip 测试；多线程并发写入 MemoryStore 测试 |
|
||||
|
||||
#### 技能（Skills）
|
||||
#### P2 — 有时间再做
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| Markdown 技能按需加载 | `agent` / `prompt` | 兼容 `SKILL.md` 格式（Hermes / OpenHarness 风格），按 prompt 上下文动态加载 | P2 | v0.2 待评估 |
|
||||
| # | 功能 | 模块 | 备注 |
|
||||
|---|------|------|------|
|
||||
| 10 | MCP StreamableHttp | `tools` | 当前仅预留枚举变体 |
|
||||
| 11 | Gemini Provider | `llm/provider` | 协议差异大，实现成本较高 |
|
||||
| 12 | 文件系统 MemoryStore 后端 | `memory` | JSON/JSONL 轻量持久化 |
|
||||
|
||||
#### 记忆（Memory）
|
||||
### ContextSlot 上下文管理
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| 多通道检索（hybrid） | `memory/retriever` | 在 TextOverlap 之上叠加向量检索通道 | P2 | v0.2 待评估 |
|
||||
| KnowledgeGraph 深度记忆 | `memory` | 实体-关系图、`note-knowledge-graph-design.md` 已记录设计 | P3 | v0.2 待评估 |
|
||||
| TokenJuice 智能压缩 | `memory` / `llm/compact` | 借鉴 OpenHuman TokenJuice，对工具结果做语义压缩而非字节截断 | P3 | v0.2 待评估 |
|
||||
**模块归属**：`src/llm/context.rs`（与 `compact.rs` 同级）
|
||||
|
||||
#### 交互层（TUI / Gateway）
|
||||
**核心概念**：`ContextSlot` 是一段带策略配置的消息列表，以 `slot_id` 为 namespace 独立持久化到 `MemoryStore`。支持三种模式、三种来源和派生关联（记录 `parent_id`）。
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| TUI / 多平台 Gateway | 应用层 | OpenClaw / Hermes 风格的消息平台桥接（Feishu / Telegram / Discord 等） | P3 | v0.2+ 应用层 |
|
||||
**核心类型**：
|
||||
|
||||
#### 训练基础设施
|
||||
```rust
|
||||
pub struct ContextSlot { id, session_id, config, messages, store }
|
||||
pub struct SlotConfig { mode: SlotMode, source: SlotSource, budget, compact }
|
||||
pub enum SlotMode {
|
||||
Full, // 完整对话历史
|
||||
Focused(FocusedConfig), // 聚焦：保持 LLM 注意力
|
||||
Readonly, // 只读参考上下文
|
||||
}
|
||||
pub struct FocusedConfig { keep_system, recent_turns, inject_summary }
|
||||
pub enum SlotSource {
|
||||
New, // 全新空槽，独立持久化
|
||||
Derived { parent_id, strategy: DeriveStrategy }, // 从父 slot 派生
|
||||
Static(Vec<Message>), // 预置消息，不持久化
|
||||
}
|
||||
pub enum DeriveStrategy { Full, Focused(FocusedConfig) }
|
||||
pub struct ContextBudget { system, history, tools, tool_results, reserve }
|
||||
```
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| RL 轨迹导出 | `agent` | ShareGPT 格式轨迹、Atropos 集成（Hermes 风格） | P3 | v0.3+ 探索 |
|
||||
**持久化 Key 命名**：
|
||||
- `slot_msg:{session_id}:{slot_id}:{index}` → 消息内容
|
||||
- `slot_meta:{session_id}:{slot_id}` → `SlotMeta`（含 `parent_id`）
|
||||
- `slot_rel:{session_id}:{child_id}:parent` → `"{parent_id}"`
|
||||
|
||||
#### 安全治理
|
||||
**`AgentSession` 扩展**：
|
||||
- `create_slot(id, config)` — 创建新 slot
|
||||
- `switch_slot(id)` — 切换当前 slot
|
||||
- `list_slots()` — 列出所有 slot
|
||||
- `derive_slot(id, parent_id, strategy)` — 从父 slot 派生
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| Human-in-the-loop 审批 | `agent` / `tools/permission` | 高危工具执行前的异步审批回调（OpenHarness `permission_prompt` 模式） | P2 | v0.2 待评估 |
|
||||
**与 `ConversationMemory` 的关系**：保留不废除。`ConversationMemory` 继续服务传统对话场景。
|
||||
|
||||
#### 流式 / 实时
|
||||
**v0.2 不做**：
|
||||
- ❌ `slot.fork()` / `merge()` — 分支方法推迟到 v0.3+
|
||||
- ❌ `inject_summary` 自动生成 — v0.2 仅消费端（从 `SessionMemory` 读取），生成在 v0.3+
|
||||
- ❌ 血缘关系图遍历 — 只存 `parent_id`，不做查询
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| 流式 `submit_turn` | `agent/session` | Phase 4 v1 只暴露非流式 `submit_turn()`；v0.2 包装 `LlmCycle::submit_stream` 暴露流式入口 | P2 | v0.2 待评估 |
|
||||
**依赖**：Phase 0（MemoryStore trait）、Phase 3（MemoryStore 持久化）
|
||||
**优先级**：P1
|
||||
|
||||
#### Agent 切换 / Prompt 动态（OpenCode 借鉴）
|
||||
---
|
||||
|
||||
| 扩展项 | 所在模块 | 说明 | 优先级 | 状态 |
|
||||
|-------|---------|------|--------|------|
|
||||
| Agent 身份切换（角色轮换） | `agent` | 借鉴 OpenCode Tab 键切换 build/plan：同一 `AgentSession` 持有可热替换的 `Agent` 引用，切换时不重置消息历史，在末尾追加 `synthetic: true` 的状态变更消息。详见 `docs/note-opencode-agent-switching.md` §4 | P2 | v0.2 待评估 |
|
||||
| System Prompt 多层动态拼接 | `agent/session` | 借鉴 OpenCode `request.ts:58-66`：拆分 `base_prompt + agent_prompt + env_context` 三层，`AgentSession::submit_turn` 每轮重算（不缓存），便于按 agent 类型动态切换 | P2 | v0.2 待评估 |
|
||||
| **多 Context 切换** | `agent` | **Phase 4c 的 SessionMemory 数据结构已预留信息桥接通道，v0.2+ 在其上包装 `ContextManager` 实现完整的多 context 切换：创建/销毁/切换 context、通过 SessionMemory 桥接关键信息。详见 `docs/note-context-switch-design.md`** | P2 | v0.2 待评估 |
|
||||
### v0.2.0 实施计划 — 8 个增量 Phase
|
||||
|
||||
> **编号说明**：Phase 5-12 接续 v0.1 的 Phase 0-4c，按开发顺序排列。
|
||||
|
||||
#### Phase 5: 热身准备（Warmup）
|
||||
|
||||
**目标**：快速交付三个互不依赖的独立改动，建立交付节奏。
|
||||
|
||||
| Step | 内容 | 文件范围 | 验证标准 |
|
||||
|------|------|---------|---------|
|
||||
| **5.1** | `ProviderConfig` 扩展：补 `timeout_secs`(def=30) + `max_retries`(def=3)；新增 `ProviderConfig::from_env(prefix)` | `llm/provider.rs` + 各 Provider `new()` 构造函数 | `cargo test` + `from_env()` 单元测试 |
|
||||
| **5.2** | `OllamaProvider`：基于 `GenericOpenaiProvider` 包装，改 base_url 为 `http://localhost:11434`；`ProviderType` 新增 `Ollama` | `llm/provider/provider.rs` + `llm/provider/ollama.rs`（新增） | `cargo build` — 纯类型级验证 |
|
||||
| **5.3** | 公开枚举 `#[non_exhaustive]` 前置标记：`ProviderType` / `StopReason` / `FinishReason` / `EvictionPolicy` / `SlotMode`(预置) | 各枚举定义处 | 编译通过 + `cargo clippy` 0 警告 |
|
||||
|
||||
**依赖**：无（三个 Step 互不冲突）
|
||||
**优先级**：P0（5.1）+ P1（5.2）+ P0 前置（5.3）
|
||||
**为何独立成 Phase**：三个改动零文件重叠，可以并行推进。它们是后续所有 Phase 的"门把手"——先做完热身再进入核心工作。
|
||||
|
||||
---
|
||||
|
||||
#### Phase 6: ToolDefinition IR 正式化
|
||||
|
||||
**目标**：引入 `ToolDef` 新类型，替换已标记 `#[deprecated]` 的 `ToolDefinition`（`OpenaiToolDefinition` 别名）。
|
||||
|
||||
**这是 v0.2 技术风险最高的 Phase**，影响 4 个模块约 8 个文件。通过 5 个 Step 逐文件切割确保每步可编译。
|
||||
|
||||
| Step | 内容 | 验证标准 |
|
||||
|------|------|---------|
|
||||
| **6.1** | `types/tool.rs` 新增 `ToolDef` 结构体 + `From<ToolDef> for OpenaiToolDefinition` + 反向 `From` | 单元测试 roundtrip |
|
||||
| **6.2** | `types/mod.rs` 切别名 `pub type ToolDefinition = ToolDef`；`MessageRequest.tools` 改 `Vec<ToolDef>` | `cargo build` 编译断点 |
|
||||
| **6.3** | `cycle.rs` 4 个方法签名 + `registry.rs` `definitions()` 签名更新 | `cargo build` |
|
||||
| **6.4** | Provider 适配层（openai.rs / anthropic.rs / openai_compat.rs）：`build_request()` 内做 `ToolDef → wire-format` 转换 | `cargo test` 每个 provider 测试 |
|
||||
| **6.5** | 所有测试/示例中 `ToolDefinition` → `ToolDef` 修复；移除旧 `#[deprecated]` alias | `cargo test --all-targets` 全绿 |
|
||||
|
||||
**边界切割技巧**：
|
||||
- Step 6.1 → 6.2 之间是安全 checkpoint：新类型存在但旧代码照常编译
|
||||
- Provider 层不改序列化逻辑，只加一层 `From` 转换
|
||||
- 当前代码中 `ToolDefinition` 已是 `#[deprecated(since = "0.1.0")]`，用户已有迁移预期
|
||||
|
||||
**依赖**：无（仅与 Phase 5.3 有枚举兼容关系）
|
||||
**优先级**：P0
|
||||
|
||||
---
|
||||
|
||||
#### Phase 7: SqliteStore 持久化
|
||||
|
||||
**目标**：实现 `MemoryStore` 的 SQLite 后端，进程重启数据不丢。
|
||||
|
||||
**与 Phase 6 无耦合，可重叠开发。**
|
||||
|
||||
| Step | 内容 | 文件 | 验证标准 |
|
||||
|------|------|-----|---------|
|
||||
| **7.1** | 新增 `memory/store/sqlite.rs`：`Mutex<Connection>` + `spawn_blocking`，实现 `save/get/delete/list` + prefix 过滤 | `memory/store/sqlite.rs` + `Cargo.toml`（add `rusqlite`） | 单元测试 CRUD + prefix 查询 |
|
||||
| **7.2** | WAL 模式 + 并发安全 + 集成测试（`tokio::spawn` 10 个并发 task） | `sqlite.rs` 扩展 | 并发写入 100 轮无 race |
|
||||
|
||||
**设计决策**：
|
||||
- 用 `Mutex<Connection>` 而非连接池（ponytail：一个连接够用就不加 r2d2）
|
||||
- WAL 模式：`PRAGMA journal_mode=WAL` 解决读写锁
|
||||
|
||||
**依赖**：`MemoryStore` trait（v0.1 Phase 3 已就绪）
|
||||
**优先级**：P0
|
||||
|
||||
---
|
||||
|
||||
#### Phase 8: MVP 集成出口（v0.2.0-rc.1 候选）
|
||||
|
||||
**目标**：P0 五项全部交付。开发者 clone 仓库后 10 分钟跑起持久化 Agent。
|
||||
|
||||
| Step | 内容 | 验证标准 |
|
||||
|------|------|---------|
|
||||
| **8.1** | API 稳定性扫尾：`#[deprecated]` 整理 + CHANGELOG v0.2 + 公开类型回顾 | 人工 review + `cargo doc` 无 warning |
|
||||
| **8.2** | Quick Start 示例（30 行 `main.rs`）：MockProvider + EchoTool + 一次 `submit_turn` | `cargo run --example quick_start` exit 0 |
|
||||
| **8.3** | 端到端示例：SqliteStore + Ollama/OpenAI(from_env) + 自定义 Tool + 多轮对话 | `cargo run --example end_to_end`（Mock fallback，无需 API key）|
|
||||
|
||||
**Phase 8 完成后可打 `v0.2.0-rc.1` 标签**。
|
||||
|
||||
**依赖**：Phase 5（ProviderConfig from_env）+ Phase 6（ToolDef）+ Phase 7（SqliteStore）
|
||||
**优先级**：P0
|
||||
|
||||
---
|
||||
|
||||
#### Phase 9: 流式体验增强
|
||||
|
||||
**目标**：Agent 会话支持流式输出，开发者看到实时 token。
|
||||
|
||||
| Step | 内容 | 文件 | 验证标准 |
|
||||
|------|------|-----|---------|
|
||||
| **9.1** | `AgentSession::submit_turn_stream(user_input) -> impl Stream<Item=StreamEvent>` | `agent/session.rs` | 单元测试验证流事件序列：`TextDelta → ... → MessageComplete` |
|
||||
|
||||
**注意**：tool 自动循环时流中插入 `ToolExecutionStarted` 事件，用户端 UI 显示"正在调用工具..."。
|
||||
|
||||
**依赖**：Phase 6（ToolDef）+ `LlmProvider.chat_stream`（v0.1 已有）
|
||||
**优先级**：P1
|
||||
|
||||
---
|
||||
|
||||
#### Phase 10: ContextSlot 上下文管理
|
||||
|
||||
**目标**：支持多上下文分区管理，Agent 可在不同 slot 之间切换。
|
||||
|
||||
| Step | 内容 | 验证标准 |
|
||||
|------|------|---------|
|
||||
| **10.1** | `src/llm/context.rs`：`ContextSlot` + `SlotConfig` / `SlotMode` / `SlotSource` / `ContextBudget` 核心类型 | `cargo build` |
|
||||
| **10.2** | ContextSlot 持久化：基于 `MemoryStore` trait（不绑定 SqliteStore）实现 save/load/list + slot 命名空间 key 策略 | 单元测试：slot 创建/写入/读取/隔离（不串数据） |
|
||||
| **10.3** | `AgentSession` 扩展：`create_slot` / `switch_slot` / `list_slots` / `derive_slot` + `AgentBuilder` 默认创建 `"default"` slot | 集成测试 + 新示例 `context_slot_demo` |
|
||||
|
||||
**如何保证简单场景无感**：`AgentBuilder::build()` 内部检查，如果用户没手动 `create_slot`，自动创建 `"default"` slot → `submit_turn` 默认写到 default slot。
|
||||
|
||||
**依赖**：Phase 7（SqliteStore 作为推荐持久化后端；`MemoryStore` trait 即可）
|
||||
**优先级**：P1
|
||||
|
||||
---
|
||||
|
||||
#### Phase 11: 测试与检索补强
|
||||
|
||||
**目标**：补全测试覆盖 + 语义检索抽象。
|
||||
|
||||
| Step | 内容 | 验证标准 |
|
||||
|------|------|---------|
|
||||
| **11.1** | `VectorRetriever` trait：`index(id, embeddings)` + `search(query, k)` | 编译 + mock 测试 |
|
||||
| **11.2** | wiremock Provider roundtrip 测试：模拟 OpenAI/Anthropic HTTP 端点 | `cargo test` 新增 10+ roundtrip 测试 |
|
||||
| **11.3** | 并发测试补强：InMemoryStore + SqliteStore 多线程写入验证 | 跑 100 轮无 race |
|
||||
|
||||
**依赖**：无（可随时做）
|
||||
**优先级**：P1
|
||||
|
||||
---
|
||||
|
||||
#### Phase 12: P2 锦上添花（可选）
|
||||
|
||||
**目标**：时间允许时按优先级交付。
|
||||
|
||||
| 优先级 | 功能 | 实现量估计 | 备注 |
|
||||
|--------|------|-----------|------|
|
||||
| **12.1** | 文件系统 MemoryStore（JSON/JSONL） | ~80 行 | 最简单，适合练手 |
|
||||
| **12.2** | MCP StreamableHttp 传输 | ~150 行 | 协议还在演进 |
|
||||
| **12.3** | Gemini Provider | ~300 行 | 协议差异大，建议推迟到 v0.3 |
|
||||
|
||||
**依赖**：无（独立交付）
|
||||
|
||||
---
|
||||
|
||||
### v0.2.0 Phase 依赖关系图
|
||||
|
||||
```mermaid
|
||||
graph BT
|
||||
P5["Phase 5<br/>热身准备"]:::warmup
|
||||
P6["Phase 6<br/>ToolDefinition IR"]:::core
|
||||
P7["Phase 7<br/>SqliteStore"]:::core
|
||||
P8["Phase 8<br/>MVP 出口 (rc.1)"]:::mvp
|
||||
P9["Phase 9<br/>流式体验增强"]:::p1
|
||||
P10["Phase 10<br/>ContextSlot"]:::p1
|
||||
P11["Phase 11<br/>测试与检索"]:::p1
|
||||
P12["Phase 12<br/>P2 锦上添花"]:::p2
|
||||
|
||||
P8 --> P5
|
||||
P8 --> P6
|
||||
P8 --> P7
|
||||
|
||||
P9 --> P6
|
||||
|
||||
P10 --> P7
|
||||
P10 --> P8
|
||||
|
||||
P11 -.-> P7
|
||||
|
||||
classDef warmup fill:#e2e8f0,stroke:#94a3b8
|
||||
classDef core fill:#fbbf24,stroke:#d97706
|
||||
classDef mvp fill:#4ade80,stroke:#16a34a
|
||||
classDef p1 fill:#93c5fd,stroke:#2563eb
|
||||
classDef p2 fill:#c4b5fd,stroke:#7c3aed
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 关键里程碑
|
||||
|
||||
| 里程碑 | Phase 完成条件 | 可验证指标 |
|
||||
|--------|---------------|-----------|
|
||||
| **M1** | Phase 5 | 热身三项完成：`from_env()` 可用 / Ollama 类型存在 / `#[non_exhaustive]` 就位 |
|
||||
| **M2** | Phase 6 | `ToolDef` 全量切换，`cargo test --all-targets` 全绿 |
|
||||
| **M3** | Phase 7 | SqliteStore CRUD + 并发测试通过，进程重启数据不丢 |
|
||||
| **M4** | **Phase 8 (rc.1)** | P0 五项全部交付，`cargo run --example quick_start` 跑通 |
|
||||
| **M5** | Phase 9 | `submit_turn_stream` 流式事件序列验证通过 |
|
||||
| **M6** | Phase 10 | ContextSlot 创建/切换/派生集成测试通过 |
|
||||
| **M7** | Phase 11 | wiremock + 并发测试补强，测试总量 200+ |
|
||||
| **M8** | Phase 12（可选） | P2 功能按需交付 |
|
||||
|
||||
---
|
||||
|
||||
## v0.3+ 展望
|
||||
|
||||
### 已规划的功能
|
||||
|
||||
| 功能 | 说明 | 预计版本 |
|
||||
|------|------|---------|
|
||||
| ContextSlot 分支（fork/merge） | 在决策点 fork 出子上下文，分支独立演进，可合并/丢弃 | v0.3 |
|
||||
| 摘要自动生成 | Hook 驱动，`OnTurnEnd` 自动将对话摘要写入 `SessionMemory`，`inject_summary` 消费端已在 v0.2 就绪 | v0.3 |
|
||||
| 知识图谱 | 实体-关系图，`docs/note-knowledge-graph-design.md` 已记录设计 | v0.3+ |
|
||||
| Multi-Agent 协同（Swarm） | 子 Agent 委派、并行子任务、结果聚合 | v0.4+ |
|
||||
| 精确 tokenizer 计数 | 绑定具体模型的 tokenizer 计数，替代当前的字符估算 | v0.3+ |
|
||||
| 血缘关系图遍历 | 以 `parent_id` 为基础，提供 slot 血缘链查询 | v0.3+ |
|
||||
| Markdown 技能按需加载 | 兼容 `SKILL.md` 格式，按 prompt 上下文动态加载 | v0.3+ |
|
||||
| TokenJuice 语义压缩 | 对工具结果做语义压缩而非字节截断 | v0.3+ |
|
||||
| Human-in-the-loop 审批 | 高危工具执行前的异步审批回调 | v0.3+ |
|
||||
| RL 轨迹导出 | ShareGPT 格式轨迹、Atropos 集成 | v0.4+ |
|
||||
|
||||
### 明确不做（agcore 范围外）
|
||||
|
||||
| 功能 | 原因 |
|
||||
|------|------|
|
||||
| TUI / 多平台 Gateway | 应用层职责（Feishu / Telegram / Discord 桥接） |
|
||||
| 配置自动加载（config/figment） | 配置来源策略应由上游应用决定，agcore 不定义配置格式 |
|
||||
| 提示词自动优化 | 属于智能层，不应内建于 core 库 |
|
||||
|
||||
---
|
||||
|
||||
## 风险与建议
|
||||
|
||||
1. **Phase 0 已完成**：LLM 调用周期基础设施已全部实现，可以支撑后续模块开发
|
||||
2. **并行可能性**：Phase 0 和 Phase 1 可并行开展（无相互依赖），可加速早期交付
|
||||
3. **MCP 协议复杂性**：MCP 涉及协议握手、session 管理、长期连接，建议预留充足时间调研协议细节
|
||||
4. **Scope 蔓延风险**：当前 specs 只有 1 份文档，建议每个模块上线前都产出对应 spec，避免边实现边设计
|
||||
5. **Phase 4 抽象化边界**：AG Core 定位为"支持库"而非"Agent 产品"，Phase 4（4a/4b/4c）需严格控制范围——只暴露 trait + 最小 reference impl，业务循环（多轮 turn 编排、对话记忆自动回写、Task 拆解策略）留给上层应用。`SessionMemory`（Phase 4c）提供信息桥接通道但不实现 context 切换逻辑。多 context 切换管理延后至 v0.2+。详细设计决策见 `docs/7-agent-runtime.md`
|
||||
6. **参考项目语言差异**：OpenClaw / Hermes / OpenHarness 均为 Python/TypeScript 实现，OpenHuman 虽是 Rust + Tauri 但定位是桌面应用。借鉴时**只取架构模式**，不照搬具体实现（如 Pydantic 工具校验、SQLite Memory Tree、Node+Python 双进程等）
|
||||
1. **持久化依赖**：`rusqlite` + `bundled` 零外部依赖编译，但 SQLite 不适配所有场景（分布式/高并发写）。`MemoryStore` trait 的抽象层允许下游自行实现 Redis / PostgreSQL 后端
|
||||
2. **ContextSlot 心智负担**：`ContextSlot` 引入了一等抽象的复杂度。建议通过 `AgentBuilder` 默认创建 `"default"` slot，让简单场景无感使用
|
||||
3. **向量检索生态**：`VectorRetriever` trait-only 不绑定实现，需社区贡献或用户自行适配 pgvector / qdrant / lancedb
|
||||
4. **Scope 蔓延**：agcore 定位为"支持库"而非"Agent 产品"，始终以 trait + reference impl 为边界，业务循环留给上层
|
||||
5. **API 稳定性**：v0.2 引入 `#[non_exhaustive]` 和 `#[deprecated]` 机制，但不承诺 SemVer 稳定——仍在快速迭代期
|
||||
|
||||
---
|
||||
|
||||
## 下一步行动
|
||||
|
||||
1. **Phase 4c 已完成**：Phase 4a + 4b + 4c 已交付（116 测试通过，0 clippy 警告）。可启动 v0.2+ 扩展评估（如多 Context 切换、Multi-Agent 协同等）
|
||||
2. **Context 切换备忘**：`docs/note-context-switch-design.md` 记录了多 context 切换方案讨论，作为 v0.2+ 扩展项的输入
|
||||
3. **参考项目调研沉淀**：已完成 OpenClaw / Hermes / OpenHuman / OpenHarness 横向调研，结果沉淀至 `docs/note-agent-harness-references.md`，作为 v0.2+ 扩展项的输入
|
||||
4. **Phase 3 备用设计就绪**：`docs/note-knowledge-graph-design.md` 记录了 KnowledgeGraph、高级评分、RecallBased 淘汰等设计，v0.2+ 记忆扩展可直接参考
|
||||
1. **Phase 5 启动**：ProviderConfig from_env + Ollama Provider + #[non_exhaustive] 前置，三个 Step 并行推进
|
||||
2. **Phase 6 方案准备**：ToolDef 结构体定义 + 兼容转换，出实施笔记（实施时直接走代码评审）
|
||||
3. **示例先行**：每完成一个 Phase 立即更新对应示例，验证通过后再合入
|
||||
4. **里程碑追踪**：以 Phase 8（MVP 出口）为 v0.2.0-rc.1 节点，逐 Phase 验收
|
||||
|
||||
**已完成 / 进行中阶段**：
|
||||
- ✅ Phase 0 Foundation — 全部交付物已完成
|
||||
@@ -338,3 +566,28 @@ graph BT
|
||||
- ✅ Phase 4a Core Glue — 全部交付物已完成
|
||||
- ✅ Phase 4b Task Execution — 全部交付物已完成
|
||||
- ✅ Phase 4c Session Memory — 全部交付物已完成
|
||||
- ✅ Provider IR 重构 — 统一类型系统 + OpenAI/Anthropic/DeepSeek/Qwen 适配
|
||||
- ✅ LlmCycle 简化 — IR 消息类型切换 + Phase 0 桥接层移除
|
||||
- ✅ v0.1 Release — 技术债扫清、MockProvider 公开化、7 个离线示例、README + 错误消息友好化、CHANGELOG 初始化
|
||||
- 📋 **v0.2 规划细化完成** — 8 个增量 Phase（Phase 5-12），17 个可验证 Step，覆盖 P0-P2 全部 12 项功能 + ContextSlot
|
||||
|
||||
---
|
||||
|
||||
## v0.1 发布里程碑（2026-07-04）
|
||||
|
||||
**质量基线**：
|
||||
|
||||
| 指标 | 数值 |
|
||||
|------|------|
|
||||
| `cargo build --all-targets` | ✅ 通过 |
|
||||
| `cargo test --all-targets` | ✅ **182 passed / 0 failed** |
|
||||
| `cargo clippy --all-targets -- -D warnings` | ✅ 0 警告 |
|
||||
| 离线示例（`cargo run --example`） | ✅ 7 个全部 exit 0 |
|
||||
|
||||
**关键交付**：
|
||||
1. **Provider IR 重构** — 统一 `Message` / `ContentBlock` / `MessageRequest` / `MessageResponse` 类型层；4 个 Provider 适配（OpenAI Chat / Anthropic Messages / DeepSeek / Qwen）；`LlmProvider` trait 签名同步切换
|
||||
2. **LlmCycle 简化** — `LlmCycle` 内部消息类型切到 IR 层；移除 Phase 0 的 `OpenaiChatMessage ↔ Message` 桥接；测试从 116 → 182（含 provider 测试）
|
||||
3. **`MockProvider` 公开化** — `agcore::llm::mock::MockProvider` 支持 `chat` + `chat_stream`，无需 API key 即可运行示例
|
||||
4. **7 个离线示例** — `prompt_composer` / `custom_tool` / `agent_session_demo` / `task_agent_demo` / `conversation_memory_demo` / `knowledge_search_demo` / `streaming_events_demo`
|
||||
5. **错误消息友好化** — `AgentError` / `LlmError` / `ToolError` / `MemoryError` / `PromptError` 全部面向最终用户改写（给出可操作的建议）
|
||||
6. **文档完整** — README 完整版（快速上手 + 架构图 + 环境变量）、Apache-2.0 LICENSE
|
||||
|
||||
@@ -0,0 +1,130 @@
|
||||
//! agent_session_demo —— Agent 装配 + 会话链路 + SessionMemory 桥接。
|
||||
//!
|
||||
//! 演示：
|
||||
//! 1. 实现 `Agent` trait（定义角色 + system prompt）
|
||||
//! 2. 用 `MockProvider` 预设响应（离线可跑）
|
||||
//! 3. `AgentBuilder` 装配 `RuntimeBundle`
|
||||
//! 4. `AgentSession::submit_turn` 跑多轮对话
|
||||
//! 5. `SessionMemory` 读写 + snapshot 输出
|
||||
//! 6. 跨 session 数据隔离验证
|
||||
//!
|
||||
//! 运行：`cargo run --example agent_session_demo`
|
||||
|
||||
use std::sync::Arc;
|
||||
|
||||
use agcore::agent::{Agent, AgentBuilder, AgentSession};
|
||||
use agcore::llm::hooks::HookExecutor;
|
||||
use agcore::llm::mock::MockProvider;
|
||||
use agcore::llm::types::message::{ContentBlock, Message};
|
||||
use agcore::llm::types::response_v2::{MessageResponse, StopReason};
|
||||
use agcore::llm::types::Usage;
|
||||
use agcore::tools::ToolRegistry;
|
||||
|
||||
/// 计算器角色 Agent。
|
||||
struct CalculatorAgent;
|
||||
|
||||
impl Agent for CalculatorAgent {
|
||||
fn name(&self) -> &str {
|
||||
"calculator"
|
||||
}
|
||||
fn system_prompt(&self) -> Option<&str> {
|
||||
Some("你是一个简洁的计算器助手，每轮回答一句话。")
|
||||
}
|
||||
}
|
||||
|
||||
/// 构造预设的纯文本 Assistant 响应。
|
||||
fn assistant_text(text: &str) -> MessageResponse {
|
||||
MessageResponse {
|
||||
id: String::new(),
|
||||
model: String::new(),
|
||||
message: Message::Assistant {
|
||||
content: vec![ContentBlock::Text { text: text.into() }],
|
||||
},
|
||||
usage: Usage::from_input_output(8, 4),
|
||||
stop_reason: StopReason::Stop,
|
||||
extra: Default::default(),
|
||||
}
|
||||
}
|
||||
|
||||
#[tokio::main]
|
||||
async fn main() {
|
||||
// 1. MockProvider：预设三轮响应（无须 API key 即可离线运行）
|
||||
let provider = Arc::new(MockProvider::new(vec![
|
||||
assistant_text("1 + 1 = 2"),
|
||||
assistant_text("2 + 2 = 4"),
|
||||
assistant_text("会话即将结束。"),
|
||||
]));
|
||||
|
||||
// 2. AgentBuilder 装配 RuntimeBundle（必填：provider / tool_registry / hook_executor）
|
||||
let bundle = Arc::new(
|
||||
AgentBuilder::new()
|
||||
.provider(provider)
|
||||
.tool_registry(Arc::new(ToolRegistry::new()))
|
||||
.hook_executor(Arc::new(HookExecutor::new()))
|
||||
.build()
|
||||
.expect("RuntimeBundle 装配失败"),
|
||||
);
|
||||
|
||||
// 3. 创建会话
|
||||
let agent: Arc<dyn Agent> = Arc::new(CalculatorAgent);
|
||||
let mut session = AgentSession::new(agent, "demo-session", bundle.clone());
|
||||
assert_eq!(session.turn_index(), 0);
|
||||
|
||||
// 4. 提交第一轮
|
||||
println!("=== 提交第 1 轮 ===");
|
||||
let resp = session.submit_turn("1+1=?").await.expect("submit_turn 失败");
|
||||
println!("LLM: {}", resp.text());
|
||||
session
|
||||
.set_session_data("last_q", "1+1=?")
|
||||
.await
|
||||
.expect("set_session_data 失败");
|
||||
session
|
||||
.set_session_data("last_a", resp.text())
|
||||
.await
|
||||
.expect("set_session_data 失败");
|
||||
assert_eq!(session.turn_index(), 1);
|
||||
|
||||
// 5. 提交第二轮
|
||||
println!("\n=== 提交第 2 轮 ===");
|
||||
let resp = session.submit_turn("再加一次 2+2=?").await.unwrap();
|
||||
println!("LLM: {}", resp.text());
|
||||
assert_eq!(session.turn_index(), 2);
|
||||
|
||||
// 6. 验证 SessionMemory 读取
|
||||
println!("\n=== Session Memory 读取 ===");
|
||||
println!(
|
||||
"last_q = {:?}",
|
||||
session.get_session_data("last_q").await.unwrap()
|
||||
);
|
||||
println!(
|
||||
"last_a = {:?}",
|
||||
session.get_session_data("last_a").await.unwrap()
|
||||
);
|
||||
|
||||
// 7. Snapshot 格式化输出
|
||||
println!("\n=== Session Memory Snapshot ===");
|
||||
println!("{}", session.session_memory().snapshot().await.unwrap());
|
||||
|
||||
// 8. 跨 session 数据隔离验证
|
||||
println!("=== 数据隔离验证 ===");
|
||||
let other = AgentSession::new(
|
||||
Arc::new(CalculatorAgent),
|
||||
"other-session",
|
||||
bundle,
|
||||
);
|
||||
assert!(
|
||||
other.get_session_data("last_q").await.unwrap().is_none(),
|
||||
"新会话不应看到旧 session 的 last_q"
|
||||
);
|
||||
println!("新会话 last_q = None ✓");
|
||||
|
||||
// 9. 用量累计验证
|
||||
println!("\n=== 用量累计 ===");
|
||||
let total = session.usage().total();
|
||||
println!(
|
||||
"prompt={}, completion={}, total={}",
|
||||
total.prompt_tokens, total.completion_tokens, total.total_tokens
|
||||
);
|
||||
|
||||
println!("\n✓ agent_session_demo 完成");
|
||||
}
|
||||
@@ -0,0 +1,143 @@
|
||||
//! conversation_memory_demo —— 对话记忆滑动窗口与隔离。
|
||||
//!
|
||||
//! 演示：
|
||||
//! 1. `ConversationMemoryConfig` 构造（SlidingWindow / Full 策略）
|
||||
//! 2. `add_message` 写入多角色消息（Message IR）
|
||||
//! 3. 滑动窗口自动淘汰旧消息
|
||||
//! 4. Full 策略保留全部
|
||||
//! 5. `get_history` / `len` / `clear`
|
||||
//! 6. 跨 session 数据隔离（共用 MemoryStore）
|
||||
//!
|
||||
//! 运行：`cargo run --example conversation_memory_demo`
|
||||
|
||||
use std::sync::Arc;
|
||||
|
||||
use agcore::llm::types::message::Message;
|
||||
use agcore::memory::{ConversationMemory, ConversationMemoryConfig, InMemoryStore, MemoryStrategy};
|
||||
|
||||
fn message_text(msg: &Message) -> &str {
|
||||
match msg {
|
||||
Message::User { content }
|
||||
| Message::System { content }
|
||||
| Message::Assistant { content }
|
||||
| Message::ToolResult { content, .. } => content
|
||||
.iter()
|
||||
.filter_map(|b| match b {
|
||||
agcore::llm::types::message::ContentBlock::Text { text } => Some(text.as_str()),
|
||||
_ => None,
|
||||
})
|
||||
.next()
|
||||
.unwrap_or(""),
|
||||
Message::UserImage { .. } => "[image]",
|
||||
}
|
||||
}
|
||||
|
||||
#[tokio::main]
|
||||
async fn main() {
|
||||
// 1. 滑动窗口策略：写入 5 条但只保留最近 3 条
|
||||
println!("=== SlidingWindow 策略（max_turns=3）===");
|
||||
let store = Arc::new(InMemoryStore::new());
|
||||
let config = ConversationMemoryConfig {
|
||||
strategy: MemoryStrategy::SlidingWindow,
|
||||
max_turns: 3,
|
||||
compact_config: None,
|
||||
};
|
||||
let mut memory = ConversationMemory::new(store, "session-1", config);
|
||||
|
||||
for i in 0..5 {
|
||||
memory
|
||||
.add_message(Message::user_text(format!("消息 {i}")))
|
||||
.await
|
||||
.expect("写入失败");
|
||||
}
|
||||
println!("写入 5 条 → len = {} (期望 3)", memory.len());
|
||||
let history = memory.get_history();
|
||||
for (i, msg) in history.iter().enumerate() {
|
||||
println!(" [{}] {}", i, message_text(msg));
|
||||
}
|
||||
assert_eq!(memory.len(), 3);
|
||||
assert_eq!(message_text(&history[0]), "消息 2", "最旧应是消息 2");
|
||||
assert_eq!(message_text(&history[2]), "消息 4", "最新应是消息 4");
|
||||
|
||||
// 2. Full 策略：保留全部
|
||||
println!("\n=== Full 策略（max_turns=3）===");
|
||||
let store2 = Arc::new(InMemoryStore::new());
|
||||
let config2 = ConversationMemoryConfig {
|
||||
strategy: MemoryStrategy::Full,
|
||||
max_turns: 3,
|
||||
compact_config: None,
|
||||
};
|
||||
let mut memory2 = ConversationMemory::new(store2, "session-2", config2);
|
||||
for i in 0..5 {
|
||||
memory2
|
||||
.add_message(Message::user_text(format!("Full {i}")))
|
||||
.await
|
||||
.unwrap();
|
||||
}
|
||||
println!("写入 5 条 → len = {} (期望 5)", memory2.len());
|
||||
assert_eq!(memory2.len(), 5);
|
||||
|
||||
// 3. 多角色混合 + clear
|
||||
println!("\n=== 多角色写入 + clear ===");
|
||||
let store3 = Arc::new(InMemoryStore::new());
|
||||
let mut memory3 = ConversationMemory::new(
|
||||
store3,
|
||||
"session-3",
|
||||
ConversationMemoryConfig::default(),
|
||||
);
|
||||
memory3
|
||||
.add_message(Message::user_text("你好"))
|
||||
.await
|
||||
.unwrap();
|
||||
memory3
|
||||
.add_message(Message::assistant("你好！有什么可以帮你的吗？"))
|
||||
.await
|
||||
.unwrap();
|
||||
memory3
|
||||
.add_message(Message::user_text("今天天气怎么样？"))
|
||||
.await
|
||||
.unwrap();
|
||||
memory3
|
||||
.add_message(Message::assistant("我无法查询实时天气，但你可以查看天气应用。"))
|
||||
.await
|
||||
.unwrap();
|
||||
println!(
|
||||
"写入 4 条多角色消息 → len = {}, 最后一条: {:?}",
|
||||
memory3.len(),
|
||||
message_text(memory3.get_history().last().unwrap())
|
||||
);
|
||||
assert_eq!(memory3.len(), 4);
|
||||
|
||||
memory3.clear().await.unwrap();
|
||||
println!(
|
||||
"clear 后 → len = {}, is_empty = {}",
|
||||
memory3.len(),
|
||||
memory3.is_empty()
|
||||
);
|
||||
assert!(memory3.is_empty());
|
||||
|
||||
// 4. Session 隔离
|
||||
println!("\n=== Session 隔离（共用 InMemoryStore）===");
|
||||
let store4 = Arc::new(InMemoryStore::new());
|
||||
let mut a = ConversationMemory::new(
|
||||
store4.clone(),
|
||||
"s-a",
|
||||
ConversationMemoryConfig::default(),
|
||||
);
|
||||
let mut b = ConversationMemory::new(
|
||||
store4.clone(),
|
||||
"s-b",
|
||||
ConversationMemoryConfig::default(),
|
||||
);
|
||||
a.add_message(Message::user_text("A 的消息")).await.unwrap();
|
||||
b.add_message(Message::user_text("B 的消息")).await.unwrap();
|
||||
println!(
|
||||
"A.len = {}, B.len = {} (期望 1/1，互不污染)",
|
||||
a.len(),
|
||||
b.len()
|
||||
);
|
||||
assert_eq!(a.len(), 1);
|
||||
assert_eq!(b.len(), 1);
|
||||
|
||||
println!("\n✓ conversation_memory_demo 完成");
|
||||
}
|
||||
@@ -0,0 +1,169 @@
|
||||
//! custom_tool —— 自定义工具注册、单次 / 并行调用、权限检查。
|
||||
//!
|
||||
//! 演示：
|
||||
//! 1. 实现 `BaseTool` trait（WeatherTool + DeleteFileTool）
|
||||
//! 2. 注册到 `ToolRegistry`
|
||||
//! 3. 单次 `invoke`（含 tool_call_id 关联）
|
||||
//! 4. 并行 `invoke_all`
|
||||
//! 5. 调用未注册工具 → `ToolError::NotFound`
|
||||
//! 6. `PermissionChecker` 黑名单阻断 `DeleteFileTool`
|
||||
//!
|
||||
//! 运行：`cargo run --example custom_tool`
|
||||
|
||||
use std::sync::Arc;
|
||||
|
||||
use agcore::tools::{
|
||||
BaseTool, Permission, PermissionChecker, PermissionConfig, ToolContext, ToolError, ToolRef,
|
||||
ToolRegistry,
|
||||
};
|
||||
use async_trait::async_trait;
|
||||
use serde_json::{json, Value};
|
||||
|
||||
/// 天气查询工具 —— 模拟根据城市返回天气数据。
|
||||
struct WeatherTool;
|
||||
|
||||
#[async_trait]
|
||||
impl BaseTool for WeatherTool {
|
||||
fn name(&self) -> &str {
|
||||
"get_weather"
|
||||
}
|
||||
fn description(&self) -> &str {
|
||||
"查询指定城市的天气"
|
||||
}
|
||||
fn parameters(&self) -> Value {
|
||||
json!({
|
||||
"type": "object",
|
||||
"properties": {
|
||||
"city": { "type": "string", "description": "城市名" }
|
||||
},
|
||||
"required": ["city"]
|
||||
})
|
||||
}
|
||||
fn required_permissions(&self) -> Vec<Permission> {
|
||||
vec![Permission::Network]
|
||||
}
|
||||
async fn execute(
|
||||
&self,
|
||||
args: Value,
|
||||
_ctx: &ToolContext<'_>,
|
||||
) -> Result<Value, ToolError> {
|
||||
let city = args["city"].as_str().unwrap_or("未知");
|
||||
// 模拟查询：根据城市名给出不同温度
|
||||
let (temperature, condition) = match city {
|
||||
"北京" => (22_i32, "晴"),
|
||||
"上海" => (25, "多云"),
|
||||
"广州" => (28, "雷阵雨"),
|
||||
_ => (20, "晴"),
|
||||
};
|
||||
Ok(json!({
|
||||
"city": city,
|
||||
"temperature": temperature,
|
||||
"condition": condition
|
||||
}))
|
||||
}
|
||||
}
|
||||
|
||||
/// 删除文件工具 —— 用于演示权限黑名单阻断。
|
||||
struct DeleteFileTool;
|
||||
|
||||
#[async_trait]
|
||||
impl BaseTool for DeleteFileTool {
|
||||
fn name(&self) -> &str {
|
||||
"delete_file"
|
||||
}
|
||||
fn description(&self) -> &str {
|
||||
"删除指定路径的文件"
|
||||
}
|
||||
fn parameters(&self) -> Value {
|
||||
json!({
|
||||
"type": "object",
|
||||
"properties": { "path": { "type": "string" } },
|
||||
"required": ["path"]
|
||||
})
|
||||
}
|
||||
fn required_permissions(&self) -> Vec<Permission> {
|
||||
vec![Permission::Delete]
|
||||
}
|
||||
async fn execute(
|
||||
&self,
|
||||
_args: Value,
|
||||
_ctx: &ToolContext<'_>,
|
||||
) -> Result<Value, ToolError> {
|
||||
Ok(json!({"deleted": true}))
|
||||
}
|
||||
}
|
||||
|
||||
#[tokio::main]
|
||||
async fn main() {
|
||||
// 1. 注册工具
|
||||
let mut registry = ToolRegistry::new();
|
||||
let weather: ToolRef = Arc::new(WeatherTool);
|
||||
let deleter: ToolRef = Arc::new(DeleteFileTool);
|
||||
registry.register(weather).expect("注册 get_weather 失败");
|
||||
registry.register(deleter).expect("注册 delete_file 失败");
|
||||
println!("=== 已注册工具 ===");
|
||||
println!("{:?}", registry.list_tools());
|
||||
|
||||
// 2. 单次 invoke：tool_call_id 用于回传时与原始调用关联
|
||||
println!("\n=== 单次 invoke ===");
|
||||
let result = registry
|
||||
.invoke("call_1", "get_weather", json!({"city": "北京"}))
|
||||
.await
|
||||
.expect("invoke 失败");
|
||||
println!("tool_call_id: {}", result.tool_call_id);
|
||||
println!("tool_name: {}", result.tool_name);
|
||||
println!("output: {}", result.output.unwrap());
|
||||
|
||||
// 3. 并行 invoke_all：三个并行天气查询
|
||||
println!("\n=== 并行 invoke_all（30s 超时）===");
|
||||
let calls = vec![
|
||||
("c1".into(), "get_weather".into(), json!({"city": "北京"})),
|
||||
("c2".into(), "get_weather".into(), json!({"city": "上海"})),
|
||||
("c3".into(), "get_weather".into(), json!({"city": "广州"})),
|
||||
];
|
||||
let results = registry.invoke_all(calls, 30).await;
|
||||
assert_eq!(results.len(), 3);
|
||||
for r in &results {
|
||||
let output = r.output.as_ref().unwrap();
|
||||
println!("[{}] {}", r.tool_call_id, output);
|
||||
}
|
||||
|
||||
// 4. 调用未注册工具
|
||||
println!("\n=== 未注册工具 ===");
|
||||
let err = registry
|
||||
.invoke("c_x", "nope_tool", json!({}))
|
||||
.await
|
||||
.unwrap_err();
|
||||
println!("错误: {err}");
|
||||
|
||||
// 5. 权限检查：默认 PermissionConfig 黑名单含 Delete
|
||||
println!("\n=== 权限检查（默认 PermissionConfig，denied = [Delete, Shell]）===");
|
||||
let mut registry_with_checker = ToolRegistry::new().with_permission_checker(PermissionChecker::new(
|
||||
PermissionConfig::default(),
|
||||
));
|
||||
registry_with_checker
|
||||
.register(Arc::new(WeatherTool) as ToolRef)
|
||||
.unwrap();
|
||||
registry_with_checker
|
||||
.register(Arc::new(DeleteFileTool) as ToolRef)
|
||||
.unwrap();
|
||||
|
||||
// get_weather 声明 Network → 在 allowed 列表 → 通过
|
||||
let r = registry_with_checker
|
||||
.invoke("c1", "get_weather", json!({"city": "北京"}))
|
||||
.await
|
||||
.unwrap();
|
||||
println!(
|
||||
"get_weather 权限检查: {}",
|
||||
if r.output.is_ok() { "通过 ✓" } else { "阻断 ✗" }
|
||||
);
|
||||
|
||||
// delete_file 声明 Delete → 在 denied 列表 → 阻断
|
||||
let err = registry_with_checker
|
||||
.invoke("c2", "delete_file", json!({"path": "/tmp/x"}))
|
||||
.await
|
||||
.unwrap_err();
|
||||
println!("delete_file 权限检查: 阻断 ✗ ({err})");
|
||||
|
||||
println!("\n✓ custom_tool 完成");
|
||||
}
|
||||
@@ -0,0 +1,149 @@
|
||||
//! knowledge_search_demo —— 知识页面存储与关键词检索。
|
||||
//!
|
||||
//! 演示：
|
||||
//! 1. `KnowledgeStore` 存储多个 `KnowledgePage`
|
||||
//! 2. `MemoryRetriever` 按关键词检索 + TextOverlap (Dice) 评分
|
||||
//! 3. 评分 [0.0, 1.0] 范围校验
|
||||
//! 4. `RetrieverConfig::min_score` 阈值过滤
|
||||
//! 5. `RetrieverConfig::max_results` 截断
|
||||
//! 6. 空 query 返回空结果
|
||||
//!
|
||||
//! 运行：`cargo run --example knowledge_search_demo`
|
||||
|
||||
use std::sync::Arc;
|
||||
|
||||
use agcore::memory::{
|
||||
InMemoryStore, KnowledgePage, KnowledgeStore, MemoryRetriever, MemoryStore, RetrieverConfig,
|
||||
};
|
||||
use time::OffsetDateTime;
|
||||
|
||||
fn make_page(id: &str, title: &str, content: &str) -> KnowledgePage {
|
||||
let now = OffsetDateTime::now_utc();
|
||||
KnowledgePage {
|
||||
id: id.to_string(),
|
||||
title: title.to_string(),
|
||||
summary: content.chars().take(30).collect(),
|
||||
content: content.to_string(),
|
||||
tags: Vec::new(),
|
||||
references: Vec::new(),
|
||||
created_at: now,
|
||||
updated_at: now,
|
||||
}
|
||||
}
|
||||
|
||||
#[tokio::main]
|
||||
async fn main() {
|
||||
// 1. 创建知识库 + 批量存储页面
|
||||
let store: Arc<dyn MemoryStore> = Arc::new(InMemoryStore::new());
|
||||
let ks = KnowledgeStore::new(store);
|
||||
|
||||
let pages = vec![
|
||||
make_page("rust-1", "Rust 入门", "Rust 是一门系统级编程语言，注重安全性与并发。"),
|
||||
make_page("python-1", "Python 简介", "Python 是一门动态类型的高级编程语言。"),
|
||||
make_page(
|
||||
"langgraph-1",
|
||||
"LangGraph 框架",
|
||||
"LangGraph 是 LangChain 的状态图扩展，用于构建多步 Agent。",
|
||||
),
|
||||
make_page("rust-async", "Rust 异步编程", "Rust 异步基于 tokio 与 futures 抽象。"),
|
||||
];
|
||||
for p in &pages {
|
||||
ks.add_page(p.clone()).await.expect("保存页面失败");
|
||||
}
|
||||
println!("=== 已存储 {} 个知识页面 ===", pages.len());
|
||||
let index = ks.get_index().await;
|
||||
for entry in &index {
|
||||
println!(" - {} ({})", entry.title, entry.id);
|
||||
}
|
||||
|
||||
// 2. 关键词检索 —— 期望命中 Rust 相关页面
|
||||
println!("\n=== 关键词检索：'Rust 异步' ===");
|
||||
let retriever = MemoryRetriever::new(ks, RetrieverConfig::default());
|
||||
let result = retriever.retrieve("Rust 异步").await.unwrap();
|
||||
println!("query: {}", result.query);
|
||||
for item in &result.items {
|
||||
println!(
|
||||
" 命中: {} (score={:.3})",
|
||||
item.page.title, item.score
|
||||
);
|
||||
assert!(
|
||||
(0.0..=1.0).contains(&item.score),
|
||||
"score 应在 [0, 1] 区间"
|
||||
);
|
||||
}
|
||||
assert!(!result.items.is_empty(), "应至少命中一个页面");
|
||||
|
||||
// 3. min_score 阈值过滤
|
||||
println!("\n=== min_score=0.5 阈值过滤（无关 query）===");
|
||||
let store2: Arc<dyn MemoryStore> = Arc::new(InMemoryStore::new());
|
||||
let ks2 = KnowledgeStore::new(store2);
|
||||
ks2.add_page(make_page("rust-1", "Rust 入门", "Rust 入门内容。"))
|
||||
.await
|
||||
.unwrap();
|
||||
let cfg = RetrieverConfig {
|
||||
max_results: 20,
|
||||
min_score: 0.5,
|
||||
};
|
||||
let retriever2 = MemoryRetriever::new(ks2, cfg);
|
||||
let result = retriever2
|
||||
.retrieve("完全不相关的火锅配方")
|
||||
.await
|
||||
.unwrap();
|
||||
println!(
|
||||
"无关 query → items.len = {} (期望 0)",
|
||||
result.items.len()
|
||||
);
|
||||
assert!(result.items.is_empty());
|
||||
|
||||
// 4. max_results 截断
|
||||
println!("\n=== max_results=2 截断 ===");
|
||||
let store3: Arc<dyn MemoryStore> = Arc::new(InMemoryStore::new());
|
||||
let ks3 = KnowledgeStore::new(store3);
|
||||
for i in 0..5 {
|
||||
ks3.add_page(make_page(
|
||||
&format!("rust-{i}"),
|
||||
"Rust 主题",
|
||||
&format!("第 {i} 篇关于 Rust 的内容"),
|
||||
))
|
||||
.await
|
||||
.unwrap();
|
||||
}
|
||||
let cfg = RetrieverConfig {
|
||||
max_results: 2,
|
||||
min_score: 0.0,
|
||||
};
|
||||
let retriever3 = MemoryRetriever::new(ks3, cfg);
|
||||
let result = retriever3.retrieve("Rust").await.unwrap();
|
||||
println!(
|
||||
"5 个相关页面 → 返回 items.len = {} (期望 2)",
|
||||
result.items.len()
|
||||
);
|
||||
assert_eq!(result.items.len(), 2);
|
||||
|
||||
// 5. 空 query
|
||||
println!("\n=== 空 query ===");
|
||||
let empty = retriever3.retrieve("").await.unwrap();
|
||||
println!("空 query → items.len = {}", empty.items.len());
|
||||
assert!(empty.items.is_empty());
|
||||
|
||||
// 6. 停用词过滤：`extract_keywords` 在检索前过滤单字符词与停用词
|
||||
println!("\n=== 停用词过滤 ===");
|
||||
let mixed = retriever3.retrieve("the Rust is").await.unwrap();
|
||||
println!(
|
||||
"query='the Rust is' → 命中 {} 个 (停用词 'the'/'is' 被过滤，仅 'rust' 进入搜索)",
|
||||
mixed.items.len()
|
||||
);
|
||||
assert!(
|
||||
!mixed.items.is_empty(),
|
||||
"非停用词 'rust' 应命中页面（即使 query 中含停用词）"
|
||||
);
|
||||
|
||||
let only_stop = retriever3.retrieve("the is are").await.unwrap();
|
||||
println!(
|
||||
"纯停用词 query='the is are' → 命中 {} 个 (期望 0，所有 token 均被过滤)",
|
||||
only_stop.items.len()
|
||||
);
|
||||
assert!(only_stop.items.is_empty(), "纯停用词 query 必须返回空结果");
|
||||
|
||||
println!("\n✓ knowledge_search_demo 完成");
|
||||
}
|
||||
@@ -0,0 +1,108 @@
|
||||
//! prompt_composer —— 提示词模板与组合器离线示例。
|
||||
//!
|
||||
//! 演示：
|
||||
//! 1. `PromptTemplate::compile` + `render` 变量插值（`{{var}}` 语法）
|
||||
//! 2. 缺失变量返回 `PromptError`
|
||||
//! 3. `PromptTemplateRegistry` 注册 + 按名渲染
|
||||
//! 4. `PromptComposer` 构造多角色消息序列
|
||||
//! 5. `validate_messages` 校验消息序列合法性
|
||||
//!
|
||||
//! 运行：`cargo run --example prompt_composer`
|
||||
|
||||
use agcore::llm::types::message::{ContentBlock, Message};
|
||||
use agcore::prompt::{
|
||||
validate_messages, PromptComposer, PromptTemplate, PromptTemplateRegistry, TemplateContext,
|
||||
};
|
||||
|
||||
fn message_text(msg: &Message) -> String {
|
||||
match msg {
|
||||
Message::System { content }
|
||||
| Message::User { content }
|
||||
| Message::Assistant { content }
|
||||
| Message::ToolResult { content, .. } => content
|
||||
.iter()
|
||||
.filter_map(|b| match b {
|
||||
ContentBlock::Text { text } => Some(text.as_str()),
|
||||
_ => None,
|
||||
})
|
||||
.collect(),
|
||||
Message::UserImage { .. } => "[image]".into(),
|
||||
}
|
||||
}
|
||||
|
||||
fn main() {
|
||||
// 1. PromptTemplate::compile + render —— 直接构造模板
|
||||
println!("=== PromptTemplate::compile + render ===");
|
||||
let tpl = PromptTemplate::compile(
|
||||
"今日 {{location}} 天气：{{condition}}，温度 {{temperature}}",
|
||||
)
|
||||
.expect("编译失败");
|
||||
let mut ctx = TemplateContext::new();
|
||||
ctx.insert("location", "北京");
|
||||
ctx.insert("condition", "晴");
|
||||
ctx.insert("temperature", "25°C");
|
||||
let rendered = tpl.render(&ctx).expect("渲染失败");
|
||||
println!("渲染结果: {rendered}");
|
||||
|
||||
// 2. 缺失变量 → PromptError
|
||||
println!("\n=== 缺失变量 ===");
|
||||
match tpl.render(&TemplateContext::new()) {
|
||||
Ok(_) => println!("意外成功"),
|
||||
Err(e) => println!("按预期报错: {e}"),
|
||||
}
|
||||
|
||||
// 3. PromptTemplateRegistry —— 按名注册 + 渲染（支持 #if 条件）
|
||||
println!("\n=== PromptTemplateRegistry ===");
|
||||
let mut registry = PromptTemplateRegistry::new();
|
||||
registry
|
||||
.register("weather", "今日 {{location}}：{{condition}}")
|
||||
.expect("注册失败");
|
||||
registry
|
||||
.register("greet", "你好 {{name}}！{{#if formal}} 见到您很荣幸。{{/if}}")
|
||||
.expect("注册失败");
|
||||
|
||||
let mut ctx = TemplateContext::new();
|
||||
ctx.insert("name", "Alice");
|
||||
println!(
|
||||
"greet (formal=false): {}",
|
||||
registry.render("greet", &ctx).unwrap()
|
||||
);
|
||||
ctx.insert("formal", true);
|
||||
println!(
|
||||
"greet (formal=true): {}",
|
||||
registry.render("greet", &ctx).unwrap()
|
||||
);
|
||||
|
||||
// 4. PromptComposer —— 构造多角色消息序列
|
||||
println!("\n=== PromptComposer ===");
|
||||
let messages = PromptComposer::new()
|
||||
.system("你是一个天气助手")
|
||||
.user("今天天气怎么样？")
|
||||
.assistant("请告诉我城市名。")
|
||||
.user(rendered)
|
||||
.build();
|
||||
println!("消息数: {}", messages.len());
|
||||
for (i, m) in messages.iter().enumerate() {
|
||||
let role = match m {
|
||||
Message::System { .. } => "system",
|
||||
Message::User { .. } | Message::UserImage { .. } => "user",
|
||||
Message::Assistant { .. } => "assistant",
|
||||
Message::ToolResult { .. } => "tool",
|
||||
};
|
||||
println!("[{i}] {role}: {}", message_text(m));
|
||||
}
|
||||
|
||||
// 5. validate_messages —— 消息序列合法性校验
|
||||
println!("\n=== validate_messages ===");
|
||||
match validate_messages(&messages) {
|
||||
Ok(()) => println!("消息序列合法 ✓"),
|
||||
Err(e) => println!("消息序列非法: {e}"),
|
||||
}
|
||||
let empty: Vec<Message> = Vec::new();
|
||||
match validate_messages(&empty) {
|
||||
Ok(()) => println!("空消息合法"),
|
||||
Err(e) => println!("空消息按预期报错: {e}"),
|
||||
}
|
||||
|
||||
println!("\n✓ prompt_composer 完成");
|
||||
}
|
||||
@@ -0,0 +1,117 @@
|
||||
//! streaming_events_demo —— LLM 流式响应事件流消费（含错误路径）。
|
||||
//!
|
||||
//! 演示：
|
||||
//! 1. `MockProvider::chat_stream` 输出标准 `StreamEvent` 流（离线可跑）
|
||||
//! 2. `LlmCycle::submit_stream` 消费流
|
||||
//! 3. match 各类 `StreamEvent`：MessageStart / ContentBlockStart / TextDelta /
|
||||
//! ContentBlockEnd / CostUpdate / MessageComplete
|
||||
//! 4. 实时累计文本与解析事件计数
|
||||
//! 5. 从 `MessageComplete` 拿到完整 `MessageResponse`
|
||||
//! 6. **错误路径**：队列耗尽时 `chat_stream` 返回 `Err`，`submit_stream` 同样
|
||||
//! 返回 `Err`（错误不进流，直接 fail-fast）
|
||||
//!
|
||||
//! 运行：`cargo run --example streaming_events_demo`
|
||||
|
||||
use std::sync::Arc;
|
||||
|
||||
use agcore::llm::cycle::{CycleConfig, LlmCycle};
|
||||
use agcore::llm::mock::MockProvider;
|
||||
use agcore::llm::provider::LlmProvider;
|
||||
use agcore::llm::types::message::{ContentBlock, Message};
|
||||
use agcore::llm::types::response_v2::{MessageResponse, StopReason, StreamEvent};
|
||||
use agcore::llm::types::Usage;
|
||||
use futures_util::StreamExt;
|
||||
|
||||
/// 构造预设的纯文本响应。
|
||||
fn text_response(id: &str, text: &str) -> MessageResponse {
|
||||
MessageResponse {
|
||||
id: id.to_string(),
|
||||
model: "mock".into(),
|
||||
message: Message::Assistant {
|
||||
content: vec![ContentBlock::Text { text: text.into() }],
|
||||
},
|
||||
usage: Usage::from_input_output(3, text.chars().count() as u32),
|
||||
stop_reason: StopReason::Stop,
|
||||
extra: Default::default(),
|
||||
}
|
||||
}
|
||||
|
||||
/// 消费一个流直到终止事件，并打印事件轨迹。
|
||||
async fn drain_stream(stream: &mut (impl futures_util::Stream<Item = StreamEvent> + Unpin)) {
|
||||
while let Some(event) = stream.next().await {
|
||||
match event {
|
||||
StreamEvent::MessageStart { id, model } => {
|
||||
println!("[MessageStart] id={id}, model={model}");
|
||||
}
|
||||
StreamEvent::ContentBlockStart { index, .. } => {
|
||||
println!("[ContentBlockStart] index={index}");
|
||||
}
|
||||
StreamEvent::TextDelta { text } => {
|
||||
print!("[TextDelta] {text}");
|
||||
use std::io::Write;
|
||||
std::io::stdout().flush().ok();
|
||||
}
|
||||
StreamEvent::ContentBlockEnd { index } => {
|
||||
println!("\n[ContentBlockEnd] index={index}");
|
||||
}
|
||||
StreamEvent::CostUpdate { usage } => {
|
||||
let p = usage.prompt_tokens.unwrap_or(0);
|
||||
let c = usage.completion_tokens.unwrap_or(0);
|
||||
println!("[CostUpdate] prompt={p}, completion={c}");
|
||||
}
|
||||
StreamEvent::MessageComplete { full_response } => {
|
||||
println!(
|
||||
"[MessageComplete] stop_reason={:?}",
|
||||
full_response.stop_reason
|
||||
);
|
||||
break;
|
||||
}
|
||||
other => {
|
||||
// 未匹配的事件变体（ThinkingDelta / RefusalDelta / ToolCall* 等）
|
||||
// 在 MockProvider 当前实现下不可达；保留打印以便未来 Provider 扩展时易调试。
|
||||
eprintln!("[unhandled event] {other:?}");
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#[tokio::main]
|
||||
async fn main() {
|
||||
// 1. 准备 Mock provider（仅含一轮文本响应，第二次调用会触发错误）
|
||||
let provider = Arc::new(MockProvider::new(vec![text_response(
|
||||
"resp-001",
|
||||
"今天天气晴朗，适合户外活动。",
|
||||
)]));
|
||||
let dyn_provider: Arc<dyn LlmProvider> = provider.clone();
|
||||
|
||||
// ===== 阶段 1：正常流式响应 =====
|
||||
println!("=== 阶段 1：正常流式响应 ===");
|
||||
let mut cycle = LlmCycle::new_with_arc(dyn_provider.clone(), CycleConfig::default());
|
||||
let mut stream = cycle
|
||||
.submit_stream("讲个笑话".to_string(), vec![])
|
||||
.await
|
||||
.expect("阶段 1 submit_stream 应成功");
|
||||
drain_stream(&mut stream).await;
|
||||
|
||||
// ===== 阶段 2：错误路径（队列耗尽）=====
|
||||
// 队列中已无响应 → `chat_stream` 返回 `Err(LlmError::Other)` →
|
||||
// `submit_stream` 用 `?` 立即传播（错误不进流，fail-fast）。
|
||||
// 上层 Agent 通过 `match` 或 `?` 处理 `AgentError::Llm(_)`。
|
||||
println!("\n=== 阶段 2：错误路径（队列耗尽）===");
|
||||
let mut cycle = LlmCycle::new_with_arc(dyn_provider, CycleConfig::default());
|
||||
let result = cycle
|
||||
.submit_stream("第二次提问".to_string(), vec![])
|
||||
.await;
|
||||
match result {
|
||||
Ok(_) => panic!("阶段 2 必须失败（队列耗尽）"),
|
||||
Err(e) => {
|
||||
eprintln!("[expected error] {e}");
|
||||
assert!(
|
||||
e.to_string().contains("预设响应已用完"),
|
||||
"应包含 MockProvider 的队列耗尽提示"
|
||||
);
|
||||
}
|
||||
}
|
||||
|
||||
println!("\n✓ streaming_events_demo 完成");
|
||||
}
|
||||
@@ -0,0 +1,113 @@
|
||||
//! task_agent_demo —— Plan 解析、Step 状态机、错误路径。
|
||||
//!
|
||||
//! 演示：
|
||||
//! 1. `JsonPlanParser::parse` 解析合法 JSON 输入
|
||||
//! 2. `Plan` / `Step` 数据结构遍历
|
||||
//! 3. `StepStatus` 状态机：Pending → Running → Completed / Failed / Skipped
|
||||
//! 4. `is_pending()` / `is_terminal()` 语义
|
||||
//! 5. 错误路径：非法 JSON / 空 steps / 缺字段 → `AgentError::PlanParse`
|
||||
//!
|
||||
//! 运行：`cargo run --example task_agent_demo`
|
||||
//!
|
||||
//! ## 已知技术债（v0.2 迁移指南）
|
||||
//!
|
||||
//! 本示例使用 `#[deprecated]` 标记的旧 wire-format 类型：
|
||||
//! - `ChatResponse`、`OpenaiChatMessage`、`FinishReason` —— `OpenaiChatProvider::chat_inner()`
|
||||
//! 内部转换层仍在使用（参见 `docs/10a-phase0-types-and-trait.md` §2.5.1），
|
||||
//! 故结构体定义保留。
|
||||
//! - `StepStatus::Completed(ChatResponse)` —— 因为 `Step` 的"已完成"变体需携带
|
||||
//! provider 响应，目前沿用旧的 `ChatResponse`。
|
||||
//!
|
||||
//! **触发迁移的条件**：v0.2 引入 IR 层的 `StepResult` / 切换为 `MessageResponse`。
|
||||
//! **迁移路径**：将本文件 `ChatResponse`/`OpenaiChatMessage`/`FinishReason` 替换为
|
||||
//! `MessageResponse`/`Message`/`StopReason`，移除顶部 `#![allow(deprecated)]`。
|
||||
//! 上层应用代码（`TaskAgent` 消费者）也可同步迁移。
|
||||
|
||||
#![allow(deprecated)]
|
||||
|
||||
use agcore::agent::{AgentError, JsonPlanParser, PlanParser, Step, StepStatus};
|
||||
use agcore::llm::types::openai_message::OpenaiChatMessage;
|
||||
use agcore::llm::types::shared::FinishReason;
|
||||
use agcore::llm::types::{ChatResponse, Usage};
|
||||
|
||||
#[tokio::main]
|
||||
async fn main() {
|
||||
// 1. JsonPlanParser 解析合法 JSON
|
||||
println!("=== JsonPlanParser 解析合法 JSON ===");
|
||||
let parser = JsonPlanParser;
|
||||
let input = r#"{
|
||||
"steps": [
|
||||
{"description": "查询北京天气"},
|
||||
{"description": "根据天气计算穿衣建议"},
|
||||
{"description": "生成最终回复"}
|
||||
]
|
||||
}"#;
|
||||
let plan = parser
|
||||
.parse(input, "为用户生成今日出行建议")
|
||||
.await
|
||||
.expect("合法 JSON 应解析成功");
|
||||
println!("Plan goal: {}", plan.goal);
|
||||
println!("Plan id: {}", plan.id);
|
||||
println!("Steps: {}", plan.steps.len());
|
||||
for (i, step) in plan.steps.iter().enumerate() {
|
||||
println!(
|
||||
" [{}] {} (pending={}, terminal={})",
|
||||
i,
|
||||
step.description,
|
||||
step.status.is_pending(),
|
||||
step.status.is_terminal()
|
||||
);
|
||||
}
|
||||
assert!(plan.steps.iter().all(|s| s.status.is_pending()));
|
||||
|
||||
// 2. Step 状态机：Pending → Running → Completed
|
||||
println!("\n=== Step 状态机：Pending → Running → Completed ===");
|
||||
let mut step = plan.steps.into_iter().next().expect("应有 step");
|
||||
println!("初始: pending={}", step.status.is_pending());
|
||||
assert!(step.status.is_pending());
|
||||
|
||||
step.status = StepStatus::Running;
|
||||
println!("Running: pending={}, terminal={}", step.status.is_pending(), step.status.is_terminal());
|
||||
|
||||
step.status = StepStatus::Completed(ChatResponse {
|
||||
message: OpenaiChatMessage::assistant_text("天气：晴，22°C"),
|
||||
usage: Usage::from_input_output(5, 10),
|
||||
stop_reason: Some(FinishReason::Stop),
|
||||
});
|
||||
println!("Completed: pending={}, terminal={}", step.status.is_pending(), step.status.is_terminal());
|
||||
assert!(step.status.is_terminal());
|
||||
|
||||
// 3. 失败路径
|
||||
println!("\n=== Step 状态机：失败路径 ===");
|
||||
let mut fail_step = Step::new(0, "调用天气 API");
|
||||
fail_step.status = StepStatus::Failed(AgentError::Other("API 不可用".into()));
|
||||
println!("Failed: pending={}, terminal={}", fail_step.status.is_pending(), fail_step.status.is_terminal());
|
||||
assert!(fail_step.status.is_terminal());
|
||||
|
||||
// 4. 跳过路径
|
||||
let mut skipped = Step::new(0, "可选步骤");
|
||||
skipped.status = StepStatus::Skipped;
|
||||
println!("Skipped: terminal={}", skipped.status.is_terminal());
|
||||
assert!(skipped.status.is_terminal());
|
||||
|
||||
// 5. 错误路径 1：非法 JSON
|
||||
println!("\n=== 错误路径：非法 JSON ===");
|
||||
let err = parser.parse("not json", "goal").await.unwrap_err();
|
||||
println!("错误: {err}");
|
||||
assert!(matches!(err, AgentError::PlanParse(_)));
|
||||
|
||||
// 6. 错误路径 2：空 steps
|
||||
println!("\n=== 错误路径：空 steps ===");
|
||||
let err = parser.parse(r#"{"steps": []}"#, "goal").await.unwrap_err();
|
||||
println!("错误: {err}");
|
||||
assert!(matches!(err, AgentError::PlanParse(_)));
|
||||
|
||||
// 7. 错误路径 3：缺 description 字段
|
||||
println!("\n=== 错误路径：缺 description 字段 ===");
|
||||
let bad_input = r#"{"steps": [{"description": "ok"}, {"no_desc": true}]}"#;
|
||||
let err = parser.parse(bad_input, "goal").await.unwrap_err();
|
||||
println!("错误: {err}");
|
||||
assert!(matches!(err, AgentError::PlanParse(_)));
|
||||
|
||||
println!("\n✓ task_agent_demo 完成");
|
||||
}
|
||||
@@ -7,12 +7,14 @@
|
||||
//! - **不绑定业务循环**：`submit_turn` 在 `AgentSession` 上，不在 trait 上
|
||||
|
||||
use crate::agent::runtime::RuntimeBundle;
|
||||
#[allow(deprecated)]
|
||||
use crate::llm::types::ToolDefinition;
|
||||
|
||||
/// Agent 角色抽象。
|
||||
///
|
||||
/// 实现此 trait 即可接入 Agent Runtime。典型实现是 struct 持有静态配置（name、system prompt 模板），
|
||||
/// 也可以是基于配置动态生成的轻量实现。
|
||||
#[allow(deprecated)]
|
||||
pub trait Agent: Send + Sync {
|
||||
/// 角色名（用于日志、调试、UI 展示）。
|
||||
fn name(&self) -> &str;
|
||||
|
||||
@@ -89,17 +89,18 @@ impl AgentBuilder {
|
||||
/// 构造 `RuntimeBundle`，校验必填字段。
|
||||
///
|
||||
/// **错误**：`provider` / `tool_registry` / `hook_executor` 任一缺失则返回
|
||||
/// `AgentError::Config("missing <field>")`，不 panic。
|
||||
/// `AgentError::Config(...)`，提示调用 `.provider(...)` / `.tool_registry(...)` /
|
||||
/// `.hook_executor(...)` 补齐。不 panic。
|
||||
pub fn build(self) -> Result<RuntimeBundle, AgentError> {
|
||||
let provider = self
|
||||
.provider
|
||||
.ok_or_else(|| AgentError::Config("missing provider".into()))?;
|
||||
.ok_or_else(|| AgentError::Config("缺少 LLM provider，请先调用 .provider(...)".into()))?;
|
||||
let tool_registry = self
|
||||
.tool_registry
|
||||
.ok_or_else(|| AgentError::Config("missing tool_registry".into()))?;
|
||||
.ok_or_else(|| AgentError::Config("缺少 tool_registry，请先调用 .tool_registry(...)（即使是空 ToolRegistry 也需要传入）".into()))?;
|
||||
let hook_executor = self
|
||||
.hook_executor
|
||||
.ok_or_else(|| AgentError::Config("missing hook_executor".into()))?;
|
||||
.ok_or_else(|| AgentError::Config("缺少 hook_executor，请先调用 .hook_executor(...)（空 HookExecutor 也可）".into()))?;
|
||||
|
||||
let config = self.config.unwrap_or_default();
|
||||
|
||||
|
||||
+1
-1
@@ -129,7 +129,7 @@ mod tests {
|
||||
|
||||
#[test]
|
||||
fn config_not_recoverable() {
|
||||
assert!(!AgentError::Config("missing provider".into()).is_recoverable());
|
||||
assert!(!AgentError::Config("缺少 provider".into()).is_recoverable());
|
||||
}
|
||||
|
||||
#[test]
|
||||
|
||||
+4
-46
@@ -15,8 +15,7 @@ use crate::agent::runtime::RuntimeBundle;
|
||||
use crate::agent::session_memory::SessionMemory;
|
||||
use crate::llm::cycle::{CostTracker, CycleConfig, LlmCycle};
|
||||
use crate::llm::hooks::{HookContext, HookEvent};
|
||||
use crate::llm::types::message::{ContentBlock, Message};
|
||||
use crate::llm::types::openai_message::OpenaiChatMessage;
|
||||
use crate::llm::types::message::Message;
|
||||
use crate::llm::types::response_v2::MessageResponse;
|
||||
use crate::memory::store::InMemoryStore;
|
||||
|
||||
@@ -176,15 +175,12 @@ impl AgentSession {
|
||||
mod tests {
|
||||
use super::*;
|
||||
use crate::agent::builder::AgentBuilder;
|
||||
use crate::llm::error::LlmError;
|
||||
use crate::llm::hooks::{Hook, HookContext, HookExecutor, HookResult};
|
||||
use crate::llm::provider::{LlmProvider, ProviderCapabilities, ProviderFeatures};
|
||||
use crate::llm::types::request_v2::MessageRequest;
|
||||
use crate::llm::types::response_v2::{MessageResponse, StopReason, StreamEvent};
|
||||
use crate::llm::mock::MockProvider;
|
||||
use crate::llm::types::message::ContentBlock;
|
||||
use crate::llm::types::response_v2::{MessageResponse, StopReason};
|
||||
use crate::tools::ToolRegistry;
|
||||
use async_trait::async_trait;
|
||||
use futures_core::Stream;
|
||||
use std::pin::Pin;
|
||||
use std::sync::atomic::{AtomicU32, Ordering};
|
||||
|
||||
/// 计数 hook —— 每被调用一次 +1。
|
||||
@@ -208,44 +204,6 @@ mod tests {
|
||||
}
|
||||
}
|
||||
|
||||
/// MockProvider：按调用顺序返回预设响应。
|
||||
struct MockProvider {
|
||||
responses: std::sync::Mutex<Vec<MessageResponse>>,
|
||||
}
|
||||
|
||||
impl MockProvider {
|
||||
fn new(responses: Vec<MessageResponse>) -> Self {
|
||||
Self {
|
||||
responses: std::sync::Mutex::new(responses),
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#[async_trait]
|
||||
impl LlmProvider for MockProvider {
|
||||
async fn chat(&self, _request: MessageRequest) -> Result<MessageResponse, LlmError> {
|
||||
let mut responses = self.responses.lock().unwrap();
|
||||
if responses.is_empty() {
|
||||
return Err(LlmError::Other("no more mock responses".into()));
|
||||
}
|
||||
Ok(responses.remove(0))
|
||||
}
|
||||
async fn chat_stream(
|
||||
&self,
|
||||
_request: MessageRequest,
|
||||
) -> Result<Pin<Box<dyn Stream<Item = Result<StreamEvent, LlmError>> + Send>>, LlmError>
|
||||
{
|
||||
unimplemented!()
|
||||
}
|
||||
fn capabilities(&self) -> ProviderCapabilities {
|
||||
ProviderCapabilities {
|
||||
provider_name: "mock",
|
||||
supported_models: None,
|
||||
features: ProviderFeatures::default(),
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
struct StubAgent {
|
||||
name: String,
|
||||
prompt: Option<String>,
|
||||
|
||||
@@ -10,6 +10,7 @@
|
||||
//! - 重试由上层新建 `Plan` 实现，`TaskAgent` 不做自动重试
|
||||
|
||||
use crate::agent::error::AgentError;
|
||||
#[allow(deprecated)]
|
||||
use crate::llm::types::ChatResponse;
|
||||
|
||||
use async_trait::async_trait;
|
||||
@@ -55,6 +56,7 @@ impl Step {
|
||||
/// 均未派生 `Clone`（保留原始错误信息，传递所有权而非克隆）。如需复制 `Plan`，
|
||||
/// 只能 clone 处于 `Pending` / `Running` / `Completed` / `Skipped` 状态的步骤。
|
||||
#[derive(Debug)]
|
||||
#[allow(deprecated)]
|
||||
pub enum StepStatus {
|
||||
/// 初始状态 —— 等待执行。
|
||||
Pending,
|
||||
|
||||
@@ -5,6 +5,7 @@ pub mod convert;
|
||||
pub mod cycle;
|
||||
pub mod error;
|
||||
pub mod hooks;
|
||||
pub mod mock;
|
||||
pub mod provider;
|
||||
pub mod stream;
|
||||
pub mod types;
|
||||
|
||||
+15
-21
@@ -40,15 +40,14 @@ pub fn from_openai(msg: &OpenaiChatMessage) -> Message {
|
||||
let mut blocks = content_to_blocks(content);
|
||||
if let Some(calls) = tool_calls {
|
||||
for call in calls {
|
||||
if let OpenaiToolCall::Function { id, function } = call {
|
||||
let input: Value =
|
||||
serde_json::from_str(&function.arguments).unwrap_or(Value::Null);
|
||||
blocks.push(ContentBlock::ToolUse {
|
||||
id: id.clone(),
|
||||
name: function.name.clone(),
|
||||
input,
|
||||
});
|
||||
}
|
||||
let OpenaiToolCall::Function { id, function } = call;
|
||||
let input: Value =
|
||||
serde_json::from_str(&function.arguments).unwrap_or(Value::Null);
|
||||
blocks.push(ContentBlock::ToolUse {
|
||||
id: id.clone(),
|
||||
name: function.name.clone(),
|
||||
input,
|
||||
});
|
||||
}
|
||||
}
|
||||
Message::Assistant { content: blocks }
|
||||
@@ -178,8 +177,8 @@ pub fn content_to_blocks(field: &ContentField) -> Vec<ContentBlock> {
|
||||
// ponytail: 简化处理 —— URL 直接通过，data URI 拆出
|
||||
// data:<mime>;base64,<b64> → ImageSource { data: b64, mime, is_url: false }。
|
||||
let url = &image_url.url;
|
||||
if let Some(rest) = url.strip_prefix("data:") {
|
||||
if let Some((mime, b64)) = rest.split_once(";base64,") {
|
||||
if let Some(rest) = url.strip_prefix("data:")
|
||||
&& let Some((mime, b64)) = rest.split_once(";base64,") {
|
||||
return Some(ContentBlock::Image {
|
||||
source: crate::llm::types::message::ImageSource {
|
||||
data: b64.to_string(),
|
||||
@@ -188,7 +187,6 @@ pub fn content_to_blocks(field: &ContentField) -> Vec<ContentBlock> {
|
||||
},
|
||||
});
|
||||
}
|
||||
}
|
||||
Some(ContentBlock::Image {
|
||||
source: crate::llm::types::message::ImageSource {
|
||||
data: url.clone(),
|
||||
@@ -255,8 +253,7 @@ pub fn blocks_to_content(blocks: &[ContentBlock]) -> ContentField {
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
use crate::llm::types::message::ImageSource;
|
||||
use crate::llm::types::shared::{AudioFormat, ImageDetail};
|
||||
use crate::llm::types::shared::ImageDetail;
|
||||
use serde_json::json;
|
||||
|
||||
#[test]
|
||||
@@ -365,13 +362,10 @@ mod tests {
|
||||
assert!(matches!(content, ContentField::String(_)));
|
||||
let calls = tool_calls.expect("tool_calls");
|
||||
assert_eq!(calls.len(), 1);
|
||||
if let OpenaiToolCall::Function { id, function } = &calls[0] {
|
||||
assert_eq!(id, "call_1");
|
||||
assert_eq!(function.name, "search");
|
||||
assert!(function.arguments.contains("rust"));
|
||||
} else {
|
||||
panic!("expected Function variant");
|
||||
}
|
||||
let OpenaiToolCall::Function { id, function } = &calls[0];
|
||||
assert_eq!(id, "call_1");
|
||||
assert_eq!(function.name, "search");
|
||||
assert!(function.arguments.contains("rust"));
|
||||
}
|
||||
_ => panic!("expected Assistant"),
|
||||
}
|
||||
|
||||
+10
-8
@@ -16,11 +16,12 @@ use crate::llm::compact::{should_compact, microcompact, CompactConfig, CompactSt
|
||||
use crate::llm::cycle::retry::should_retry;
|
||||
use crate::llm::error::LlmError;
|
||||
use crate::llm::hooks::{HookContext, HookExecutor};
|
||||
use crate::llm::provider::{LlmProvider, ProviderCapabilities, ProviderFeatures};
|
||||
use crate::llm::provider::LlmProvider;
|
||||
use crate::llm::stream::StreamEvent;
|
||||
use crate::llm::types::message::{ContentBlock, Message};
|
||||
use crate::llm::types::request_v2::MessageRequest;
|
||||
use crate::llm::types::response_v2::{MessageResponse, StopReason};
|
||||
#[allow(deprecated)]
|
||||
use crate::llm::types::{ToolChoice, ToolDefinition};
|
||||
|
||||
/// LLM 调用周期配置。
|
||||
@@ -79,6 +80,7 @@ pub struct LlmCycle {
|
||||
compact_state: CompactState,
|
||||
}
|
||||
|
||||
#[allow(deprecated)]
|
||||
impl LlmCycle {
|
||||
/// 创建一个新的 LlmCycle（持有 `Box<dyn LlmProvider>` 的独占所有权）。
|
||||
///
|
||||
@@ -175,13 +177,10 @@ impl LlmCycle {
|
||||
messages: Vec<Message>,
|
||||
tools: Vec<ToolDefinition>,
|
||||
) -> Result<MessageResponse, LlmError> {
|
||||
let ir_messages: Vec<Message> = messages;
|
||||
let ir_tools: Vec<ToolDefinition> = tools.clone();
|
||||
|
||||
let request = MessageRequest {
|
||||
model: self.config.model.clone(),
|
||||
messages: ir_messages,
|
||||
tools: ir_tools,
|
||||
messages,
|
||||
tools,
|
||||
tool_choice: ToolChoice::Auto,
|
||||
max_tokens: self.config.max_tokens,
|
||||
temperature: self.config.temperature,
|
||||
@@ -672,6 +671,7 @@ fn truncate_tool_result(s: &str, max_bytes: usize) -> String {
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
use crate::llm::provider::{ProviderCapabilities, ProviderFeatures};
|
||||
use crate::tools::{BaseTool, ToolRegistry};
|
||||
use async_trait::async_trait;
|
||||
use futures_core::Stream;
|
||||
@@ -858,8 +858,10 @@ mod tests {
|
||||
|
||||
#[tokio::test]
|
||||
async fn test_submit_with_tools_max_turns_exceeded() {
|
||||
let mut config = CycleConfig::default();
|
||||
config.max_tool_turns = Some(2);
|
||||
let config = CycleConfig {
|
||||
max_tool_turns: Some(2),
|
||||
..Default::default()
|
||||
};
|
||||
let responses = vec![
|
||||
assistant_tool_call_response(vec![("c1", "add", r#"{"a":1,"b":1}"#)]),
|
||||
assistant_tool_call_response(vec![("c2", "add", r#"{"a":1,"b":1}"#)]),
|
||||
|
||||
+15
-13
@@ -5,33 +5,35 @@ use std::time::Duration;
|
||||
/// 错误按可重试性分为两类：
|
||||
/// - **可重试**：`RateLimit`、`Timeout`、状态码 >= 500
|
||||
/// - **不可重试**：`Authentication`、`ContextLength`、状态码 4xx（除 429）
|
||||
///
|
||||
/// 错误消息面向最终用户（中文），并尽量附带可操作的修复建议（如检查 API key、减少上下文）。
|
||||
#[derive(thiserror::Error, Debug)]
|
||||
pub enum LlmError {
|
||||
/// API 认证失败（如 API key 无效）。
|
||||
#[error("认证失败: {0}")]
|
||||
/// API 认证失败（API key 无效、过期或权限不足）。
|
||||
#[error("LLM 认证失败: {0}。请检查环境变量中的 API key（如 OPENAI_API_KEY / ANTHROPIC_API_KEY）是否正确")]
|
||||
Authentication(String),
|
||||
|
||||
/// 请求被限流，可选地附带重试等待时间。
|
||||
#[error("限流(retry_after={retry_after:?})")]
|
||||
/// 请求被限流，可选地附带重试等待时间。可重试。
|
||||
#[error("LLM 限流（服务方），建议等待 {retry_after:?} 后重试")]
|
||||
RateLimit { retry_after: Option<Duration> },
|
||||
|
||||
/// HTTP 请求失败，包含状态码和响应体。
|
||||
#[error("请求失败(status={status}): {body}")]
|
||||
/// HTTP 请求失败（网络错误或非 2xx 状态码），包含状态码与响应体。
|
||||
#[error("LLM 请求失败（HTTP {status}）: {body}。请检查 Provider 端点地址（base_url）和网络连通性")]
|
||||
Request { status: u16, body: String },
|
||||
|
||||
/// 请求超时。
|
||||
#[error("请求超时(duration={duration:?})")]
|
||||
/// 请求超时。可重试。
|
||||
#[error("LLM 请求超时（{duration:?}）。请检查网络连接，或调大 LlmCycle 超时配置")]
|
||||
Timeout { duration: Duration },
|
||||
|
||||
/// 流式响应处理错误（预留）。
|
||||
#[error("流式响应错误: {0}")]
|
||||
/// 流式响应处理错误（SSE 解析失败、流中断等）。可重试。
|
||||
#[error("LLM 流式响应错误: {0}。可重试或改用非流式接口")]
|
||||
Stream(String),
|
||||
|
||||
/// 上下文长度超限。
|
||||
#[error("上下文超限(actual={actual}, limit={limit})")]
|
||||
/// 上下文长度超出模型窗口限制。
|
||||
#[error("LLM 上下文超限：当前 {actual} tokens > 模型上限 {limit} tokens。请减少消息历史、缩短 prompt，或启用 auto-compaction（llm::compact）")]
|
||||
ContextLength { actual: u32, limit: u32 },
|
||||
|
||||
/// 其他未分类的 LLM 调用失败。
|
||||
#[error("LLM 调用失败: {0}")]
|
||||
Other(String),
|
||||
}
|
||||
}
|
||||
@@ -75,6 +75,7 @@ impl<'a> HookContext<'a> {
|
||||
}
|
||||
|
||||
/// 设置 plan step 序号（仅 OnPlanStepComplete 使用，Phase 4b 新增）。
|
||||
#[allow(dead_code)]
|
||||
pub(crate) fn with_plan_step_index(mut self, plan_step_index: usize) -> Self {
|
||||
self.plan_step_index = Some(plan_step_index);
|
||||
self
|
||||
|
||||
+305
@@ -0,0 +1,305 @@
|
||||
//! 公开的 [`MockProvider`] —— 在示例与测试中按顺序返回预设响应。
|
||||
//!
|
||||
//! 提供与真实 LLM Provider 一致的 `chat` / `chat_stream` 行为，
|
||||
//! 但不发起任何网络请求，因此适合做离线示例、回归测试和 CI gate。
|
||||
//!
|
||||
//! # 用法
|
||||
//!
|
||||
//! ```no_run
|
||||
//! use std::sync::Arc;
|
||||
//! use agcore::llm::mock::MockProvider;
|
||||
//! use agcore::llm::provider::LlmProvider;
|
||||
//! use agcore::llm::types::message::{ContentBlock, Message};
|
||||
//! use agcore::llm::types::response_v2::{MessageResponse, StopReason};
|
||||
//! use agcore::llm::types::Usage;
|
||||
//!
|
||||
//! let response = MessageResponse {
|
||||
//! id: "resp-1".into(),
|
||||
//! model: "mock".into(),
|
||||
//! message: Message::Assistant {
|
||||
//! content: vec![ContentBlock::Text { text: "hi".into() }],
|
||||
//! },
|
||||
//! usage: Usage::from_input_output(2, 1),
|
||||
//! stop_reason: StopReason::Stop,
|
||||
//! extra: Default::default(),
|
||||
//! };
|
||||
//!
|
||||
//! let provider: Arc<dyn LlmProvider> = Arc::new(MockProvider::new(vec![response]));
|
||||
//! ```
|
||||
//!
|
||||
//! # 流式行为
|
||||
//!
|
||||
//! `chat_stream` 把当前响应的 `ContentBlock` 序列拆解为标准流事件：
|
||||
//!
|
||||
//! - `Text` → `ContentBlockStart(Text)` + `TextDelta` + `ContentBlockEnd`
|
||||
//! - `Thinking` → `ContentBlockStart(Thinking)` + `ThinkingDelta` + `ContentBlockEnd`
|
||||
//! - `ToolUse` → `ContentBlockStart(ToolUse)` + `ToolCallArgumentsDelta` + `ToolCallEnd`
|
||||
//! - 其他变体（Image/Audio/File/Extension）—— 不在流路径中模拟，跳过
|
||||
|
||||
use std::pin::Pin;
|
||||
use std::sync::Mutex;
|
||||
|
||||
use async_stream::stream;
|
||||
use futures_core::Stream;
|
||||
|
||||
use crate::llm::error::LlmError;
|
||||
use crate::llm::provider::{LlmProvider, ProviderCapabilities, ProviderFeatures};
|
||||
use crate::llm::types::message::{ContentBlock, ContentBlockType, Message};
|
||||
use crate::llm::types::request_v2::MessageRequest;
|
||||
use crate::llm::types::response_v2::{MessageResponse, PartialUsage, StreamEvent};
|
||||
|
||||
/// 按调用顺序返回预设响应的 [`LlmProvider`]。
|
||||
///
|
||||
/// 内部用 `Mutex<Vec<MessageResponse>>` 存储队列；`chat` / `chat_stream` 均弹出队首响应。
|
||||
/// 当队列耗尽时返回 `LlmError::Other("MockProvider: 预设响应已用完")`。
|
||||
pub struct MockProvider {
|
||||
responses: Mutex<Vec<MessageResponse>>,
|
||||
}
|
||||
|
||||
impl MockProvider {
|
||||
/// 用预设响应列表创建。
|
||||
pub fn new(responses: Vec<MessageResponse>) -> Self {
|
||||
Self {
|
||||
responses: Mutex::new(responses),
|
||||
}
|
||||
}
|
||||
|
||||
/// 创建空队列的 Provider（后续可用 `extend` 注入响应）。
|
||||
pub fn empty() -> Self {
|
||||
Self::new(Vec::new())
|
||||
}
|
||||
|
||||
/// 追加更多响应到队列末尾。
|
||||
pub fn extend(&self, responses: Vec<MessageResponse>) {
|
||||
self.responses.lock().unwrap().extend(responses);
|
||||
}
|
||||
|
||||
/// 队列中剩余响应数量。
|
||||
pub fn remaining(&self) -> usize {
|
||||
self.responses.lock().unwrap().len()
|
||||
}
|
||||
|
||||
fn pop(&self) -> Result<MessageResponse, LlmError> {
|
||||
let mut guard = self.responses.lock().unwrap();
|
||||
if guard.is_empty() {
|
||||
return Err(LlmError::Other("MockProvider: 预设响应已用完".into()));
|
||||
}
|
||||
Ok(guard.remove(0))
|
||||
}
|
||||
}
|
||||
|
||||
#[async_trait::async_trait]
|
||||
impl LlmProvider for MockProvider {
|
||||
async fn chat(&self, _request: MessageRequest) -> Result<MessageResponse, LlmError> {
|
||||
self.pop()
|
||||
}
|
||||
|
||||
async fn chat_stream(
|
||||
&self,
|
||||
_request: MessageRequest,
|
||||
) -> Result<Pin<Box<dyn Stream<Item = Result<StreamEvent, LlmError>> + Send>>, LlmError>
|
||||
{
|
||||
let response = self.pop()?;
|
||||
// 提前 clone 出在 stream 闭包中需要的字段；最后 yield 时 move response。
|
||||
let id = response.id.clone();
|
||||
let model = response.model.clone();
|
||||
let usage = response.usage;
|
||||
let blocks: Vec<ContentBlock> = match &response.message {
|
||||
Message::Assistant { content } => content.clone(),
|
||||
_ => Vec::new(),
|
||||
};
|
||||
|
||||
let stream = stream! {
|
||||
yield Ok(StreamEvent::MessageStart {
|
||||
id: id.clone(),
|
||||
model: model.clone(),
|
||||
});
|
||||
|
||||
for (i, block) in blocks.iter().enumerate() {
|
||||
let index = i as u32;
|
||||
match block {
|
||||
ContentBlock::Text { text } => {
|
||||
yield Ok(StreamEvent::ContentBlockStart {
|
||||
index,
|
||||
block_type: ContentBlockType::Text,
|
||||
});
|
||||
if !text.is_empty() {
|
||||
yield Ok(StreamEvent::TextDelta { text: text.clone() });
|
||||
}
|
||||
yield Ok(StreamEvent::ContentBlockEnd { index });
|
||||
}
|
||||
ContentBlock::Thinking { text, .. } => {
|
||||
yield Ok(StreamEvent::ContentBlockStart {
|
||||
index,
|
||||
block_type: ContentBlockType::Thinking,
|
||||
});
|
||||
if !text.is_empty() {
|
||||
yield Ok(StreamEvent::ThinkingDelta { text: text.clone() });
|
||||
}
|
||||
yield Ok(StreamEvent::ContentBlockEnd { index });
|
||||
}
|
||||
ContentBlock::ToolUse { id: tu_id, name, input } => {
|
||||
yield Ok(StreamEvent::ContentBlockStart {
|
||||
index,
|
||||
block_type: ContentBlockType::ToolUse {
|
||||
id: tu_id.clone(),
|
||||
name: name.clone(),
|
||||
},
|
||||
});
|
||||
yield Ok(StreamEvent::ToolCallArgumentsDelta {
|
||||
index,
|
||||
arguments: serde_json::to_string(input)
|
||||
.unwrap_or_else(|_| "{}".into()),
|
||||
});
|
||||
yield Ok(StreamEvent::ToolCallEnd { index });
|
||||
}
|
||||
// 其他变体（Image / Audio / File / ToolResult / Extension / Refusal）
|
||||
// 在 mock 流路径下跳过：示例仅需 text/thinking/tool 三种主流路径。
|
||||
_ => {}
|
||||
}
|
||||
}
|
||||
|
||||
yield Ok(StreamEvent::CostUpdate {
|
||||
usage: PartialUsage {
|
||||
prompt_tokens: Some(usage.prompt_tokens),
|
||||
completion_tokens: Some(usage.completion_tokens),
|
||||
total_tokens: Some(usage.total_tokens),
|
||||
..Default::default()
|
||||
},
|
||||
});
|
||||
|
||||
yield Ok(StreamEvent::MessageComplete { full_response: response });
|
||||
};
|
||||
|
||||
Ok(Box::pin(stream))
|
||||
}
|
||||
|
||||
fn capabilities(&self) -> ProviderCapabilities {
|
||||
ProviderCapabilities {
|
||||
provider_name: "mock",
|
||||
supported_models: None,
|
||||
features: ProviderFeatures::default(),
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
#[cfg(test)]
|
||||
mod tests {
|
||||
use super::*;
|
||||
use crate::llm::types::{StopReason, Usage};
|
||||
use futures_util::StreamExt;
|
||||
use serde_json::json;
|
||||
|
||||
fn text_response(text: &str) -> MessageResponse {
|
||||
MessageResponse {
|
||||
id: "r".into(),
|
||||
model: "mock".into(),
|
||||
message: Message::Assistant {
|
||||
content: vec![ContentBlock::Text { text: text.into() }],
|
||||
},
|
||||
usage: Usage::from_input_output(3, 7),
|
||||
stop_reason: StopReason::Stop,
|
||||
extra: Default::default(),
|
||||
}
|
||||
}
|
||||
|
||||
#[tokio::test]
|
||||
async fn chat_returns_queued_response() {
|
||||
let provider = MockProvider::new(vec![text_response("hello")]);
|
||||
let resp = provider
|
||||
.chat(MessageRequest::default())
|
||||
.await
|
||||
.unwrap();
|
||||
assert_eq!(resp.text(), "hello");
|
||||
assert_eq!(provider.remaining(), 0);
|
||||
}
|
||||
|
||||
#[tokio::test]
|
||||
async fn chat_errors_when_empty() {
|
||||
let provider = MockProvider::empty();
|
||||
let err = provider.chat(MessageRequest::default()).await.unwrap_err();
|
||||
assert!(matches!(err, LlmError::Other(_)));
|
||||
}
|
||||
|
||||
#[tokio::test]
|
||||
async fn extend_appends_responses() {
|
||||
let provider = MockProvider::new(vec![text_response("a")]);
|
||||
provider.extend(vec![text_response("b"), text_response("c")]);
|
||||
assert_eq!(provider.remaining(), 3);
|
||||
}
|
||||
|
||||
#[tokio::test]
|
||||
async fn chat_stream_emits_text_delta_sequence() {
|
||||
let provider = MockProvider::new(vec![text_response("hi")]);
|
||||
let mut stream = provider.chat_stream(MessageRequest::default()).await.unwrap();
|
||||
|
||||
let mut seen_start = false;
|
||||
let mut seen_block_start = false;
|
||||
let mut seen_delta = false;
|
||||
let mut seen_block_end = false;
|
||||
let mut seen_complete = false;
|
||||
|
||||
while let Some(event) = stream.next().await {
|
||||
match event.unwrap() {
|
||||
StreamEvent::MessageStart { .. } => seen_start = true,
|
||||
StreamEvent::ContentBlockStart {
|
||||
block_type: ContentBlockType::Text,
|
||||
..
|
||||
} => seen_block_start = true,
|
||||
StreamEvent::TextDelta { text } => {
|
||||
assert_eq!(text, "hi");
|
||||
seen_delta = true;
|
||||
}
|
||||
StreamEvent::ContentBlockEnd { .. } => seen_block_end = true,
|
||||
StreamEvent::MessageComplete { .. } => {
|
||||
seen_complete = true;
|
||||
break;
|
||||
}
|
||||
_ => {}
|
||||
}
|
||||
}
|
||||
|
||||
assert!(seen_start);
|
||||
assert!(seen_block_start);
|
||||
assert!(seen_delta);
|
||||
assert!(seen_block_end);
|
||||
assert!(seen_complete);
|
||||
}
|
||||
|
||||
#[tokio::test]
|
||||
async fn chat_stream_emits_tool_call_arguments() {
|
||||
let response = MessageResponse {
|
||||
id: "r".into(),
|
||||
model: "mock".into(),
|
||||
message: Message::Assistant {
|
||||
content: vec![ContentBlock::ToolUse {
|
||||
id: "call_1".into(),
|
||||
name: "search".into(),
|
||||
input: json!({"q": "rust"}),
|
||||
}],
|
||||
},
|
||||
usage: Usage::from_input_output(2, 5),
|
||||
stop_reason: StopReason::ToolUse,
|
||||
extra: Default::default(),
|
||||
};
|
||||
let provider = MockProvider::new(vec![response]);
|
||||
let mut stream = provider.chat_stream(MessageRequest::default()).await.unwrap();
|
||||
|
||||
let mut saw_tool_args = false;
|
||||
let mut saw_tool_end = false;
|
||||
while let Some(event) = stream.next().await {
|
||||
match event.unwrap() {
|
||||
StreamEvent::ToolCallArgumentsDelta { arguments, .. } => {
|
||||
let v: serde_json::Value = serde_json::from_str(&arguments).unwrap();
|
||||
assert_eq!(v, json!({"q": "rust"}));
|
||||
saw_tool_args = true;
|
||||
}
|
||||
StreamEvent::ToolCallEnd { .. } => saw_tool_end = true,
|
||||
StreamEvent::MessageComplete { .. } => break,
|
||||
_ => {}
|
||||
}
|
||||
}
|
||||
assert!(saw_tool_args);
|
||||
assert!(saw_tool_end);
|
||||
}
|
||||
}
|
||||
@@ -21,7 +21,7 @@ use tracing::{debug, error, info, warn};
|
||||
use super::{LlmProvider, ProviderCapabilities, ProviderFeatures};
|
||||
use crate::llm::error::LlmError;
|
||||
use crate::llm::types::message::{ContentBlock, ContentBlockType, Message};
|
||||
use crate::llm::types::request_v2::{MessageRequest, ThinkingConfig};
|
||||
use crate::llm::types::request_v2::MessageRequest;
|
||||
use crate::llm::types::response_v2::{
|
||||
MessageResponse, PartialMessageResponse, PartialUsage, StopReason, StreamEvent,
|
||||
};
|
||||
@@ -36,6 +36,7 @@ const DEFAULT_MAX_TOKENS: u32 = 4096;
|
||||
pub struct AnthropicProvider {
|
||||
http_client: Client,
|
||||
base_url: String,
|
||||
#[allow(dead_code)]
|
||||
api_key: String,
|
||||
model: String,
|
||||
}
|
||||
@@ -193,7 +194,7 @@ impl AnthropicProvider {
|
||||
.json(&body)
|
||||
.send()
|
||||
.await
|
||||
.map_err(|e| Self::map_reqwest_error(e))?;
|
||||
.map_err(Self::map_reqwest_error)?;
|
||||
|
||||
let status = response.status();
|
||||
if !status.is_success() {
|
||||
@@ -229,7 +230,7 @@ impl AnthropicProvider {
|
||||
.json(&body)
|
||||
.send()
|
||||
.await
|
||||
.map_err(|e| Self::map_reqwest_error(e))?;
|
||||
.map_err(Self::map_reqwest_error)?;
|
||||
|
||||
let status = response.status();
|
||||
if !status.is_success() {
|
||||
@@ -586,6 +587,7 @@ enum AnthropicDelta {
|
||||
struct AnthropicMessageDeltaInner {
|
||||
stop_reason: Option<String>,
|
||||
#[serde(default)]
|
||||
#[allow(dead_code)]
|
||||
stop_sequence: Option<String>,
|
||||
}
|
||||
|
||||
@@ -599,6 +601,7 @@ pub struct AnthropicSseStream {
|
||||
chunks: Pin<Box<dyn Stream<Item = Result<Bytes, LlmError>> + Send>>,
|
||||
buffer: String,
|
||||
partial: PartialMessageResponse,
|
||||
#[allow(dead_code)]
|
||||
next_block_index: u32,
|
||||
saw_terminal: bool,
|
||||
}
|
||||
|
||||
@@ -22,7 +22,7 @@ use serde_json::Value;
|
||||
use tracing::{debug, error, info};
|
||||
|
||||
use super::{LlmProvider, ProviderCapabilities, ProviderFeatures};
|
||||
use crate::llm::convert::{blocks_to_content, content_to_blocks, from_openai, to_openai};
|
||||
use crate::llm::convert::{from_openai, to_openai};
|
||||
use crate::llm::error::LlmError;
|
||||
use crate::llm::types::message::{ContentBlock, ContentBlockType, Message};
|
||||
use crate::llm::types::openai_message::{ContentField, OpenaiChatMessage};
|
||||
@@ -720,8 +720,8 @@ impl Stream for ChunkToEventStream {
|
||||
}
|
||||
let data = if let Some(p) = trimmed.strip_prefix("data: ") {
|
||||
p
|
||||
} else if trimmed.starts_with("data:") {
|
||||
&trimmed[5..]
|
||||
} else if let Some(p) = trimmed.strip_prefix("data:") {
|
||||
p
|
||||
} else {
|
||||
continue;
|
||||
};
|
||||
|
||||
+1
-1
@@ -23,7 +23,7 @@ use crate::llm::types::old_stream::LegacyStreamEvent;
|
||||
use crate::llm::types::response_v2::MessageResponse;
|
||||
use crate::llm::types::response_v2::StopReason;
|
||||
use crate::llm::types::usage::Usage;
|
||||
use crate::llm::types::{FinishReason, OpenaiChatChunk, OpenaiToolCall};
|
||||
use crate::llm::types::{OpenaiChatChunk, OpenaiToolCall};
|
||||
|
||||
// 唯一的对外 `StreamEvent` 定义（高精度 IR 事件，来自 `response_v2`）。
|
||||
//
|
||||
|
||||
@@ -40,6 +40,7 @@ pub use usage::{CompletionTokensDetails, CostTracker, PromptTokensDetails, Usage
|
||||
// Phase 1 起移除 `ChatRequest` 别名 —— 新代码统一使用 `MessageRequest`（v2 IR）。
|
||||
// `ChatResponse` 结构体仍存在，作为 OpenAI `chat_inner()` 内部 wire-format 转换目标。
|
||||
/// 旧 wire-format 响应结构（保留用于 OpenAI 内部转换层）。
|
||||
#[deprecated(since = "0.1.0", note = "请改用 MessageResponse")]
|
||||
#[derive(Debug, Clone)]
|
||||
pub struct ChatResponse {
|
||||
pub message: OpenaiChatMessage,
|
||||
@@ -47,6 +48,7 @@ pub struct ChatResponse {
|
||||
pub stop_reason: Option<FinishReason>,
|
||||
}
|
||||
|
||||
#[allow(deprecated)]
|
||||
impl From<OpenaiChatResponse> for ChatResponse {
|
||||
fn from(response: OpenaiChatResponse) -> Self {
|
||||
let message = response
|
||||
@@ -63,6 +65,7 @@ impl From<OpenaiChatResponse> for ChatResponse {
|
||||
}
|
||||
}
|
||||
|
||||
#[allow(deprecated)]
|
||||
impl From<ChatResponse> for OpenaiChatChunk {
|
||||
fn from(response: ChatResponse) -> Self {
|
||||
let delta = Delta::from(response.message.clone());
|
||||
@@ -95,4 +98,5 @@ impl From<ChatResponse> for OpenaiChatChunk {
|
||||
}
|
||||
|
||||
/// 工具定义别名（无新类型冲突，保留）。
|
||||
#[deprecated(since = "0.1.0", note = "ToolDefinition 仍直接对应 OpenAI wire-format；未来 v0.2 引入 IR 工具类型后会再次更新")]
|
||||
pub type ToolDefinition = OpenaiToolDefinition;
|
||||
|
||||
@@ -12,7 +12,9 @@ pub struct StreamOptions {
|
||||
}
|
||||
|
||||
#[derive(Debug, Clone)]
|
||||
#[derive(Default)]
|
||||
pub enum ToolChoice {
|
||||
#[default]
|
||||
None,
|
||||
Auto,
|
||||
Required,
|
||||
@@ -20,12 +22,6 @@ pub enum ToolChoice {
|
||||
AllowedTools { tool_names: Vec<String> },
|
||||
}
|
||||
|
||||
impl Default for ToolChoice {
|
||||
fn default() -> Self {
|
||||
// Default 行为由调用方显式选择；此处选择不暴露任何工具 —— 不暴露比错暴露安全。
|
||||
ToolChoice::None
|
||||
}
|
||||
}
|
||||
|
||||
impl Serialize for ToolChoice {
|
||||
fn serialize<S>(&self, serializer: S) -> Result<S::Ok, S::Error>
|
||||
|
||||
@@ -174,8 +174,8 @@ impl ConversationMemory {
|
||||
}
|
||||
}
|
||||
|
||||
if let Some(ref compact_config) = self.config.compact_config {
|
||||
if should_compact(&self.messages, compact_config, &self.compact_state) {
|
||||
if let Some(ref compact_config) = self.config.compact_config
|
||||
&& should_compact(&self.messages, compact_config, &self.compact_state) {
|
||||
let keep_recent = compact_config.keep_recent;
|
||||
let freed = microcompact(&mut self.messages, keep_recent);
|
||||
if freed > 0 {
|
||||
@@ -184,7 +184,6 @@ impl ConversationMemory {
|
||||
let _ = self.compact_state.record_failure();
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
}
|
||||
|
||||
@@ -237,7 +236,7 @@ mod tests {
|
||||
};
|
||||
let mut conv = ConversationMemory::new(store, "s1", config);
|
||||
for i in 0..5 {
|
||||
conv.add_message(Message::user_text(&format!("msg-{i}")))
|
||||
conv.add_message(Message::user_text(format!("msg-{i}")))
|
||||
.await
|
||||
.unwrap();
|
||||
}
|
||||
@@ -254,7 +253,7 @@ mod tests {
|
||||
};
|
||||
let mut conv = ConversationMemory::new(store, "s1", config);
|
||||
for i in 0..5 {
|
||||
conv.add_message(Message::user_text(&format!("msg-{i}")))
|
||||
conv.add_message(Message::user_text(format!("msg-{i}")))
|
||||
.await
|
||||
.unwrap();
|
||||
}
|
||||
|
||||
+13
-6
@@ -3,21 +3,28 @@
|
||||
use thiserror::Error;
|
||||
|
||||
/// 记忆系统错误枚举。
|
||||
///
|
||||
/// 错误消息面向最终用户（中文），并尽量附带可操作的修复建议（如检查环境变量、重试）。
|
||||
#[derive(Debug, Error)]
|
||||
pub enum MemoryError {
|
||||
#[error("Item not found: {0}")]
|
||||
/// 按 ID 未找到指定记忆条目。可重试——通常是 namespace 拼写错误或条目已被淘汰。
|
||||
#[error("未找到记忆条目 '{0}'，请检查 ID 或 namespace 是否正确")]
|
||||
NotFound(String),
|
||||
|
||||
#[error("Storage error: {0}")]
|
||||
/// 底层存储失败（磁盘满、连接断开等）。一般不可恢复，建议上层记录并告警。
|
||||
#[error("存储失败: {0}，请检查后端存储是否可用（磁盘 / 网络 / 权限）")]
|
||||
Storage(String),
|
||||
|
||||
#[error("Serialization error: {0}")]
|
||||
/// 序列化 / 反序列化失败（数据结构变更、JSON 字段缺失等）。一般不可恢复。
|
||||
#[error("序列化失败: {0}，请检查数据结构兼容性或升级记忆格式")]
|
||||
Serialization(String),
|
||||
|
||||
#[error("Invalid input: {0}")]
|
||||
/// 调用方传入的参数不合法（空 ID、负数容量等）。可重试——修正参数即可。
|
||||
#[error("参数无效: {0}")]
|
||||
InvalidInput(String),
|
||||
|
||||
#[error("Retrieval error: {0}")]
|
||||
/// 检索过程出错（索引重建失败、评分异常等）。可重试。
|
||||
#[error("检索失败: {0}，可重试或调整 query")]
|
||||
RetrievalError(String),
|
||||
}
|
||||
|
||||
@@ -26,4 +33,4 @@ impl MemoryError {
|
||||
pub fn is_recoverable(&self) -> bool {
|
||||
matches!(self, Self::NotFound(_) | Self::RetrievalError(_))
|
||||
}
|
||||
}
|
||||
}
|
||||
+28
-21
@@ -19,7 +19,7 @@ pub const KNOWLEDGE_PREFIX: &str = "knowledge_";
|
||||
/// 同时维护一个 `Vec<PageIndexEntry>` 索引以加速列表遍历。
|
||||
pub struct KnowledgeStore {
|
||||
store: Arc<dyn MemoryStore>,
|
||||
index: std::sync::Mutex<Vec<PageIndexEntry>>,
|
||||
index: tokio::sync::Mutex<Vec<PageIndexEntry>>,
|
||||
}
|
||||
|
||||
impl KnowledgeStore {
|
||||
@@ -27,7 +27,7 @@ impl KnowledgeStore {
|
||||
pub fn new(store: Arc<dyn MemoryStore>) -> Self {
|
||||
Self {
|
||||
store,
|
||||
index: std::sync::Mutex::new(Vec::new()),
|
||||
index: tokio::sync::Mutex::new(Vec::new()),
|
||||
}
|
||||
}
|
||||
|
||||
@@ -40,7 +40,7 @@ impl KnowledgeStore {
|
||||
..Default::default()
|
||||
})
|
||||
.await?;
|
||||
let mut index = self.index.lock().unwrap();
|
||||
let mut index = self.index.lock().await;
|
||||
index.clear();
|
||||
for item in items {
|
||||
let page: KnowledgePage = serde_json::from_str(&item.content)
|
||||
@@ -66,7 +66,7 @@ impl KnowledgeStore {
|
||||
created_at: now,
|
||||
};
|
||||
self.store.save(item).await?;
|
||||
let mut index = self.index.lock().unwrap();
|
||||
let mut index = self.index.lock().await;
|
||||
// 替换或追加
|
||||
if let Some(existing) = index.iter_mut().find(|e| e.id == page.id) {
|
||||
*existing = PageIndexEntry::from(&page);
|
||||
@@ -106,7 +106,7 @@ impl KnowledgeStore {
|
||||
pub async fn delete_page(&self, id: &str) -> Result<(), MemoryError> {
|
||||
let full_id = format!("{KNOWLEDGE_PREFIX}{id}");
|
||||
self.store.delete(&full_id).await?;
|
||||
let mut index = self.index.lock().unwrap();
|
||||
let mut index = self.index.lock().await;
|
||||
index.retain(|e| e.id != id);
|
||||
Ok(())
|
||||
}
|
||||
@@ -120,24 +120,31 @@ impl KnowledgeStore {
|
||||
return Ok(Vec::new());
|
||||
}
|
||||
let needle = query.to_lowercase();
|
||||
// 锁内仅 clone 匹配的 entry id，避免异步 get_page() 持有 index 锁。
|
||||
let ids: Vec<String> = {
|
||||
let index = self.index.lock().await;
|
||||
index
|
||||
.iter()
|
||||
.filter(|entry| {
|
||||
entry.title.to_lowercase().contains(&needle)
|
||||
|| entry.summary.to_lowercase().contains(&needle)
|
||||
|| entry.tags.iter().any(|t| t.to_lowercase().contains(&needle))
|
||||
})
|
||||
.map(|entry| entry.id.clone())
|
||||
.collect()
|
||||
};
|
||||
let mut results = Vec::new();
|
||||
let index = self.index.lock().unwrap();
|
||||
for entry in index.iter() {
|
||||
if entry.title.to_lowercase().contains(&needle)
|
||||
|| entry.summary.to_lowercase().contains(&needle)
|
||||
|| entry.tags.iter().any(|t| t.to_lowercase().contains(&needle))
|
||||
{
|
||||
if let Some(page) = self.get_page(&entry.id).await? {
|
||||
results.push(page);
|
||||
}
|
||||
for id in &ids {
|
||||
if let Some(page) = self.get_page(id).await? {
|
||||
results.push(page);
|
||||
}
|
||||
}
|
||||
Ok(results)
|
||||
}
|
||||
|
||||
/// 获取内容目录（所有页面的轻量级索引条目）。
|
||||
pub fn get_index(&self) -> Vec<PageIndexEntry> {
|
||||
self.index.lock().unwrap().clone()
|
||||
pub async fn get_index(&self) -> Vec<PageIndexEntry> {
|
||||
self.index.lock().await.clone()
|
||||
}
|
||||
}
|
||||
|
||||
@@ -224,7 +231,7 @@ mod tests {
|
||||
let ks = KnowledgeStore::new(store);
|
||||
ks.add_page(make_page("p1", "A", &[])).await.unwrap();
|
||||
ks.add_page(make_page("p2", "B", &[])).await.unwrap();
|
||||
let index = ks.get_index();
|
||||
let index = ks.get_index().await;
|
||||
assert_eq!(index.len(), 2);
|
||||
}
|
||||
|
||||
@@ -235,13 +242,13 @@ mod tests {
|
||||
// 添加页面
|
||||
ks.add_page(make_page("p1", "A", &[])).await.unwrap();
|
||||
ks.add_page(make_page("p2", "B", &[])).await.unwrap();
|
||||
assert_eq!(ks.get_index().len(), 2);
|
||||
assert_eq!(ks.get_index().await.len(), 2);
|
||||
|
||||
// 模拟 index 漂移：清空后重建
|
||||
ks.index.lock().unwrap().clear();
|
||||
assert_eq!(ks.get_index().len(), 0);
|
||||
ks.index.lock().await.clear();
|
||||
assert_eq!(ks.get_index().await.len(), 0);
|
||||
|
||||
ks.rebuild_index().await.unwrap();
|
||||
assert_eq!(ks.get_index().len(), 2);
|
||||
assert_eq!(ks.get_index().await.len(), 2);
|
||||
}
|
||||
}
|
||||
|
||||
@@ -113,9 +113,7 @@ fn extract_keywords(query: &str, stop_words: &HashSet<String>) -> Vec<String> {
|
||||
.split(|c: char| !c.is_alphanumeric())
|
||||
.filter_map(|s| {
|
||||
let lower = s.to_lowercase();
|
||||
if lower.is_empty() || lower.chars().count() < 2 {
|
||||
None
|
||||
} else if stop_words.contains(&lower) {
|
||||
if lower.is_empty() || lower.chars().count() < 2 || stop_words.contains(&lower) {
|
||||
None
|
||||
} else {
|
||||
Some(lower)
|
||||
|
||||
+1
-1
@@ -4,4 +4,4 @@ pub mod composer;
|
||||
|
||||
pub use error::PromptError;
|
||||
pub use template::{PromptTemplate, PromptTemplateRegistry, TemplateContext, TemplateValue};
|
||||
pub use composer::PromptComposer;
|
||||
pub use composer::{validate_messages, PromptComposer};
|
||||
|
||||
+63
-128
@@ -1,12 +1,11 @@
|
||||
use crate::llm::types::openai_message::{ContentField, OpenaiChatMessage, OpenaiContentPart};
|
||||
use crate::llm::types::request::OpenaiChatRequest;
|
||||
use crate::llm::types::message::{ContentBlock, Message};
|
||||
use crate::prompt::error::PromptError;
|
||||
use crate::prompt::template::{PromptTemplate, TemplateContext};
|
||||
|
||||
/// 提示词组合器——构建多角色消息序列。
|
||||
#[derive(Default)]
|
||||
pub struct PromptComposer {
|
||||
messages: Vec<OpenaiChatMessage>,
|
||||
messages: Vec<Message>,
|
||||
}
|
||||
|
||||
impl PromptComposer {
|
||||
@@ -16,7 +15,7 @@ impl PromptComposer {
|
||||
}
|
||||
|
||||
/// 从已有的消息列表初始化。
|
||||
pub fn from_messages(messages: Vec<OpenaiChatMessage>) -> Self {
|
||||
pub fn from_messages(messages: Vec<Message>) -> Self {
|
||||
Self { messages }
|
||||
}
|
||||
|
||||
@@ -24,34 +23,32 @@ impl PromptComposer {
|
||||
|
||||
/// 添加一条纯文本 system 消息。
|
||||
pub fn system(mut self, text: impl Into<String>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::system_text(text.into()));
|
||||
self.push_message(Message::system(text.into()));
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条纯文本 user 消息。
|
||||
pub fn user(mut self, text: impl Into<String>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::user_text(text.into()));
|
||||
self.push_message(Message::user_text(text.into()));
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条纯文本 assistant 消息。
|
||||
pub fn assistant(mut self, text: impl Into<String>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::assistant_text(text.into()));
|
||||
self.push_message(Message::assistant(text.into()));
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条纯文本 developer 消息（o1 系列模型使用）。
|
||||
/// IR 层统一映射为 `Message::System`，由 Provider 在发送时按目标模型决定 `role`。
|
||||
pub fn developer(mut self, text: impl Into<String>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::developer_text(text.into()));
|
||||
self.push_message(Message::system(text.into()));
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条 Tool 消息（工具执行结果回传）。
|
||||
pub fn tool(mut self, tool_call_id: impl Into<String>, content: impl Into<String>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::tool_result(
|
||||
tool_call_id.into(),
|
||||
content.into(),
|
||||
));
|
||||
self.push_message(Message::tool_result(tool_call_id.into(), content.into(), false));
|
||||
self
|
||||
}
|
||||
|
||||
@@ -64,7 +61,7 @@ impl PromptComposer {
|
||||
ctx: &TemplateContext,
|
||||
) -> Result<Self, PromptError> {
|
||||
let text = template.render(ctx)?;
|
||||
self.push_message(OpenaiChatMessage::user_text(text));
|
||||
self.push_message(Message::user_text(text));
|
||||
Ok(self)
|
||||
}
|
||||
|
||||
@@ -75,7 +72,7 @@ impl PromptComposer {
|
||||
ctx: &TemplateContext,
|
||||
) -> Result<Self, PromptError> {
|
||||
let text = template.render(ctx)?;
|
||||
self.push_message(OpenaiChatMessage::system_text(text));
|
||||
self.push_message(Message::system(text));
|
||||
Ok(self)
|
||||
}
|
||||
|
||||
@@ -86,7 +83,7 @@ impl PromptComposer {
|
||||
ctx: &TemplateContext,
|
||||
) -> Result<Self, PromptError> {
|
||||
let text = template.render(ctx)?;
|
||||
self.push_message(OpenaiChatMessage::assistant_text(text));
|
||||
self.push_message(Message::assistant(text));
|
||||
Ok(self)
|
||||
}
|
||||
|
||||
@@ -97,150 +94,98 @@ impl PromptComposer {
|
||||
ctx: &TemplateContext,
|
||||
) -> Result<Self, PromptError> {
|
||||
let text = template.render(ctx)?;
|
||||
self.push_message(OpenaiChatMessage::developer_text(text));
|
||||
self.push_message(Message::system(text));
|
||||
Ok(self)
|
||||
}
|
||||
|
||||
// ===== 多模态 ContentPart =====
|
||||
// ===== 多模态 ContentBlock =====
|
||||
|
||||
/// 添加一条含指定 ContentPart 的 system 消息。
|
||||
pub fn system_content(mut self, part: OpenaiContentPart) -> Self {
|
||||
self.push_message(OpenaiChatMessage::System {
|
||||
content: ContentField::Array(vec![part]),
|
||||
name: None,
|
||||
/// 添加一条含指定 ContentBlock 的 system 消息。
|
||||
pub fn system_content(mut self, block: ContentBlock) -> Self {
|
||||
self.push_message(Message::System {
|
||||
content: vec![block],
|
||||
});
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条含指定 ContentPart 的 user 消息。
|
||||
pub fn user_content(mut self, part: OpenaiContentPart) -> Self {
|
||||
self.push_message(OpenaiChatMessage::User {
|
||||
content: ContentField::Array(vec![part]),
|
||||
name: None,
|
||||
/// 添加一条含指定 ContentBlock 的 user 消息。
|
||||
pub fn user_content(mut self, block: ContentBlock) -> Self {
|
||||
self.push_message(Message::User {
|
||||
content: vec![block],
|
||||
});
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条含指定 ContentPart 的 assistant 消息。
|
||||
pub fn assistant_content(mut self, part: OpenaiContentPart) -> Self {
|
||||
self.push_message(OpenaiChatMessage::Assistant {
|
||||
content: ContentField::Array(vec![part]),
|
||||
refusal: None,
|
||||
name: None,
|
||||
tool_calls: None,
|
||||
/// 添加一条含指定 ContentBlock 的 assistant 消息。
|
||||
pub fn assistant_content(mut self, block: ContentBlock) -> Self {
|
||||
self.push_message(Message::Assistant {
|
||||
content: vec![block],
|
||||
});
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条含指定 ContentPart 的 developer 消息。
|
||||
pub fn developer_content(mut self, part: OpenaiContentPart) -> Self {
|
||||
self.push_message(OpenaiChatMessage::Developer {
|
||||
content: ContentField::Array(vec![part]),
|
||||
name: None,
|
||||
/// 添加一条含指定 ContentBlock 的 developer 消息。
|
||||
pub fn developer_content(mut self, block: ContentBlock) -> Self {
|
||||
self.push_message(Message::System {
|
||||
content: vec![block],
|
||||
});
|
||||
self
|
||||
}
|
||||
|
||||
/// 添加一条含指定 ContentPart 的 Tool 消息。
|
||||
/// 添加一条含指定 ContentBlock 的 Tool 消息。
|
||||
pub fn tool_content(
|
||||
mut self,
|
||||
tool_call_id: impl Into<String>,
|
||||
part: OpenaiContentPart,
|
||||
block: ContentBlock,
|
||||
) -> Self {
|
||||
self.push_message(OpenaiChatMessage::Tool {
|
||||
content: ContentField::Array(vec![part]),
|
||||
self.push_message(Message::ToolResult {
|
||||
tool_call_id: tool_call_id.into(),
|
||||
content: vec![block],
|
||||
is_error: false,
|
||||
});
|
||||
self
|
||||
}
|
||||
|
||||
/// 批量添加 ContentPart 作为 user 消息。
|
||||
pub fn user_contents(mut self, parts: Vec<OpenaiContentPart>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::User {
|
||||
content: ContentField::Array(parts),
|
||||
name: None,
|
||||
});
|
||||
/// 批量添加 ContentBlock 作为 user 消息。
|
||||
pub fn user_contents(mut self, blocks: Vec<ContentBlock>) -> Self {
|
||||
self.push_message(Message::User { content: blocks });
|
||||
self
|
||||
}
|
||||
|
||||
/// 批量添加 ContentPart 作为 system 消息。
|
||||
pub fn system_contents(mut self, parts: Vec<OpenaiContentPart>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::System {
|
||||
content: ContentField::Array(parts),
|
||||
name: None,
|
||||
});
|
||||
/// 批量添加 ContentBlock 作为 system 消息。
|
||||
pub fn system_contents(mut self, blocks: Vec<ContentBlock>) -> Self {
|
||||
self.push_message(Message::System { content: blocks });
|
||||
self
|
||||
}
|
||||
|
||||
/// 批量添加 ContentPart 作为 assistant 消息。
|
||||
pub fn assistant_contents(mut self, parts: Vec<OpenaiContentPart>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::Assistant {
|
||||
content: ContentField::Array(parts),
|
||||
refusal: None,
|
||||
name: None,
|
||||
tool_calls: None,
|
||||
});
|
||||
/// 批量添加 ContentBlock 作为 assistant 消息。
|
||||
pub fn assistant_contents(mut self, blocks: Vec<ContentBlock>) -> Self {
|
||||
self.push_message(Message::Assistant { content: blocks });
|
||||
self
|
||||
}
|
||||
|
||||
/// 批量添加 ContentPart 作为 developer 消息。
|
||||
pub fn developer_contents(mut self, parts: Vec<OpenaiContentPart>) -> Self {
|
||||
self.push_message(OpenaiChatMessage::Developer {
|
||||
content: ContentField::Array(parts),
|
||||
name: None,
|
||||
});
|
||||
self
|
||||
}
|
||||
|
||||
// ===== 角色标识 =====
|
||||
|
||||
/// 为上一条添加的消息设置 `name` 字段。
|
||||
pub fn with_name(mut self, name: impl Into<String>) -> Self {
|
||||
let name = name.into();
|
||||
if let Some(msg) = self.messages.last_mut() {
|
||||
set_message_name(msg, name);
|
||||
}
|
||||
/// 批量添加 ContentBlock 作为 developer 消息。
|
||||
pub fn developer_contents(mut self, blocks: Vec<ContentBlock>) -> Self {
|
||||
self.push_message(Message::System { content: blocks });
|
||||
self
|
||||
}
|
||||
|
||||
// ===== 构建 =====
|
||||
|
||||
/// 构建最终的消息列表。
|
||||
pub fn build(self) -> Vec<OpenaiChatMessage> {
|
||||
pub fn build(self) -> Vec<Message> {
|
||||
self.messages
|
||||
}
|
||||
|
||||
/// 构建并直接创建 ChatRequest（需搭配 model 参数）。
|
||||
/// 返回的 `OpenaiChatRequest` 中 `tools`、`temperature`、`max_tokens` 等字段均为 `None`，
|
||||
/// 可通过结构体更新语法补全：`OpenaiChatRequest { tools: Some(...), ..req }`。
|
||||
pub fn build_request(self, model: impl Into<String>) -> OpenaiChatRequest {
|
||||
OpenaiChatRequest {
|
||||
model: model.into(),
|
||||
messages: self.messages,
|
||||
..Default::default()
|
||||
}
|
||||
}
|
||||
|
||||
// ===== 内部方法 =====
|
||||
|
||||
fn push_message(&mut self, msg: OpenaiChatMessage) {
|
||||
fn push_message(&mut self, msg: Message) {
|
||||
self.messages.push(msg);
|
||||
}
|
||||
}
|
||||
|
||||
fn set_message_name(msg: &mut OpenaiChatMessage, name: String) {
|
||||
match msg {
|
||||
OpenaiChatMessage::Developer { name: n, .. } => *n = Some(name),
|
||||
OpenaiChatMessage::System { name: n, .. } => *n = Some(name),
|
||||
OpenaiChatMessage::User { name: n, .. } => *n = Some(name),
|
||||
OpenaiChatMessage::Assistant { name: n, .. } => *n = Some(name),
|
||||
OpenaiChatMessage::Tool { .. } => {}
|
||||
OpenaiChatMessage::Function { .. } => {}
|
||||
}
|
||||
}
|
||||
|
||||
/// 验证消息序列是否符合 OpenAI API 要求。
|
||||
pub fn validate_messages(messages: &[OpenaiChatMessage]) -> Result<(), PromptError> {
|
||||
/// 验证消息序列是否符合 LLM API 要求（Tool 消息必须紧跟含 tool_calls 的 Assistant）。
|
||||
pub fn validate_messages(messages: &[Message]) -> Result<(), PromptError> {
|
||||
if messages.is_empty() {
|
||||
return Err(PromptError::InvalidSequence(
|
||||
"消息列表不能为空".to_string(),
|
||||
@@ -251,10 +196,7 @@ pub fn validate_messages(messages: &[OpenaiChatMessage]) -> Result<(), PromptErr
|
||||
|
||||
for (i, msg) in messages.iter().enumerate() {
|
||||
match msg {
|
||||
OpenaiChatMessage::Tool {
|
||||
tool_call_id,
|
||||
..
|
||||
} => {
|
||||
Message::ToolResult { tool_call_id, .. } => {
|
||||
if last_tool_call_ids.is_empty() {
|
||||
return Err(PromptError::InvalidSequence(format!(
|
||||
"消息[{i}] Tool 消息前必须有 Assistant 消息且含 tool_calls"
|
||||
@@ -267,21 +209,14 @@ pub fn validate_messages(messages: &[OpenaiChatMessage]) -> Result<(), PromptErr
|
||||
)));
|
||||
}
|
||||
}
|
||||
OpenaiChatMessage::Assistant {
|
||||
tool_calls: Some(calls),
|
||||
..
|
||||
} => {
|
||||
Message::Assistant { content } => {
|
||||
last_tool_call_ids.clear();
|
||||
for call in calls {
|
||||
let crate::llm::types::OpenaiToolCall::Function { id, .. } = call;
|
||||
last_tool_call_ids.push(id.clone());
|
||||
for block in content {
|
||||
if let ContentBlock::ToolUse { id, .. } = block {
|
||||
last_tool_call_ids.push(id.clone());
|
||||
}
|
||||
}
|
||||
}
|
||||
OpenaiChatMessage::Assistant {
|
||||
tool_calls: None, ..
|
||||
} => {
|
||||
last_tool_call_ids.clear();
|
||||
}
|
||||
_ => {
|
||||
last_tool_call_ids.clear();
|
||||
}
|
||||
@@ -318,18 +253,18 @@ mod tests {
|
||||
|
||||
assert_eq!(msgs.len(), 4);
|
||||
match &msgs[3] {
|
||||
OpenaiChatMessage::Tool {
|
||||
Message::ToolResult {
|
||||
tool_call_id,
|
||||
content,
|
||||
..
|
||||
} => {
|
||||
assert_eq!(tool_call_id, "call_123");
|
||||
match content {
|
||||
ContentField::String(s) => assert_eq!(s, "Sunny, 25°C"),
|
||||
_ => {}
|
||||
match &content[0] {
|
||||
ContentBlock::Text { text } => assert_eq!(text, "Sunny, 25°C"),
|
||||
_ => panic!("Expected Text block"),
|
||||
}
|
||||
}
|
||||
_ => panic!("Expected Tool message"),
|
||||
_ => panic!("Expected ToolResult message"),
|
||||
}
|
||||
}
|
||||
|
||||
@@ -345,7 +280,7 @@ mod tests {
|
||||
|
||||
#[test]
|
||||
fn test_validate_messages_empty() {
|
||||
let msgs: Vec<OpenaiChatMessage> = vec![];
|
||||
let msgs: Vec<Message> = vec![];
|
||||
assert!(validate_messages(&msgs).is_err());
|
||||
}
|
||||
|
||||
|
||||
+7
-7
@@ -2,27 +2,27 @@ use thiserror::Error;
|
||||
|
||||
#[derive(Error, Debug)]
|
||||
pub enum PromptError {
|
||||
#[error("模板解析错误: {0}")]
|
||||
#[error("模板解析错误: {0}。请检查模板语法（{{var}} / {{#if}} / {{#each}}）")]
|
||||
Parse(String),
|
||||
|
||||
#[error("渲染错误: 变量 '{0}' 未找到")]
|
||||
#[error("渲染错误: 变量 '{0}' 未找到。请在 TemplateContext 中插入该变量")]
|
||||
VariableNotFound(String),
|
||||
|
||||
#[error("渲染错误: 引用的子模板 '{0}' 未注册")]
|
||||
#[error("渲染错误: 引用的子模板 '{0}' 未注册。请先用 PromptTemplateRegistry::register 注册该子模板")]
|
||||
PartialNotFound(String),
|
||||
|
||||
#[error("渲染错误: '{0}' 不是数组，无法遍历")]
|
||||
#[error("渲染错误: '{0}' 不是数组，无法遍历。请确认传入的是数组或先判空")]
|
||||
NotAnArray(String),
|
||||
|
||||
#[error("渲染递归超过最大深度限制 ({0})")]
|
||||
#[error("渲染递归超过最大深度限制 ({0})。请检查是否有循环 include，或调大 MAX_DEPTH")]
|
||||
MaxDepthReached(u8),
|
||||
|
||||
#[error("渲染错误: {0}")]
|
||||
Render(String),
|
||||
|
||||
#[error("消息序列校验失败: {0}")]
|
||||
#[error("消息序列校验失败: {0}。请检查消息角色顺序（例如 tool 必须在 assistant tool_call 之后）")]
|
||||
InvalidSequence(String),
|
||||
|
||||
#[error("文件读取错误: {0}")]
|
||||
#[error("文件读取错误: {0}。请检查模板文件路径与权限")]
|
||||
Io(#[from] std::io::Error),
|
||||
}
|
||||
|
||||
+13
-13
@@ -5,44 +5,44 @@ use std::sync::Arc;
|
||||
/// 工具调用过程中可能发生的所有错误。
|
||||
#[derive(thiserror::Error, Debug, Clone)]
|
||||
pub enum ToolError {
|
||||
/// 工具未注册。
|
||||
#[error("工具 '{0}' 未注册")]
|
||||
/// 工具未注册。不可恢复——需调用方先 `registry.register(...)`。
|
||||
#[error("工具 '{0}' 未注册。请先用 ToolRegistry::register(...) 注册该工具，或检查 LLM 输出的工具名拼写")]
|
||||
NotFound(String),
|
||||
|
||||
/// 工具执行失败（可恢复——文本回传 LLM）。
|
||||
#[error("工具 '{0}' 执行失败: {1}")]
|
||||
/// 工具执行失败（可恢复——文本回传 LLM 由其决定重试或放弃）。
|
||||
#[error("工具 '{0}' 执行失败: {1}。可让 LLM 调整参数后重试，或切换到备选工具")]
|
||||
ExecutionFailed(String, String),
|
||||
|
||||
/// 工具参数无效（可恢复——文本回传 LLM）。
|
||||
#[error("工具 '{0}' 参数无效: {1}")]
|
||||
#[error("工具 '{0}' 参数无效: {1}。请检查 LLM 输出的参数是否符合 BaseTool::parameters() 声明的 JSON Schema")]
|
||||
InvalidArguments(String, String),
|
||||
|
||||
/// 权限被拒绝（不可恢复——终止循环）。
|
||||
#[error("权限被拒绝: 工具 '{0}' 需要 {1} 权限")]
|
||||
#[error("权限被拒绝: 工具 '{0}' 需要 {1} 权限。请在 PermissionConfig 中显式允许，或人工确认后绕过")]
|
||||
PermissionDenied(String, String),
|
||||
|
||||
/// MCP 协议错误（不可恢复）。
|
||||
#[error("MCP 协议错误: {0}")]
|
||||
#[error("MCP 协议错误: {0}。请检查 MCP server 版本与本库兼容，或查看 server 日志")]
|
||||
McpError(String),
|
||||
|
||||
/// MCP 未初始化（不可恢复）。
|
||||
#[error("MCP 未初始化: {0}")]
|
||||
#[error("MCP 未初始化: {0}。请先调用 McpClient::initialize() 完成握手")]
|
||||
McpNotInitialized(String),
|
||||
|
||||
/// MCP 超时（不可恢复）。
|
||||
#[error("MCP 超时: {0}")]
|
||||
#[error("MCP 超时: {0}。请检查 MCP server 是否存活，或调大超时配置")]
|
||||
McpTimeout(String),
|
||||
|
||||
/// IO 错误（不可恢复）。
|
||||
#[error("IO 错误: {0}")]
|
||||
#[error("IO 错误: {0}。请检查文件路径、权限或磁盘空间")]
|
||||
Io(Arc<std::io::Error>),
|
||||
|
||||
/// 取消。
|
||||
/// 工具执行被取消。
|
||||
#[error("工具执行已取消: {0}")]
|
||||
Cancelled(String),
|
||||
|
||||
/// 其他未分类错误。
|
||||
#[error("其他错误: {0}")]
|
||||
#[error("工具调用错误: {0}")]
|
||||
Other(String),
|
||||
}
|
||||
|
||||
@@ -105,7 +105,7 @@ mod tests {
|
||||
|
||||
#[test]
|
||||
fn test_io_is_not_recoverable() {
|
||||
let io_err = std::io::Error::new(std::io::ErrorKind::Other, "disk");
|
||||
let io_err = std::io::Error::other("disk");
|
||||
let err = ToolError::from(io_err);
|
||||
assert!(!err.is_recoverable());
|
||||
}
|
||||
|
||||
@@ -20,6 +20,7 @@ use tokio::io::{AsyncBufReadExt, AsyncWriteExt, BufReader};
|
||||
use tokio::process::{Child, ChildStdin, ChildStdout, Command};
|
||||
use tokio::sync::{oneshot, Mutex};
|
||||
|
||||
#[allow(deprecated)]
|
||||
use crate::llm::types::ToolDefinition;
|
||||
use crate::tools::base::{BaseTool, ToolContext, ToolRef};
|
||||
use crate::tools::error::ToolError;
|
||||
@@ -135,6 +136,7 @@ impl std::fmt::Debug for McpClient {
|
||||
}
|
||||
}
|
||||
|
||||
#[allow(deprecated)]
|
||||
impl McpClient {
|
||||
/// 创建一个 MCP 客户端。
|
||||
pub fn new(server_name: impl Into<String>, transport: McpTransport) -> Self {
|
||||
@@ -540,6 +542,7 @@ enum McpClientHandle {
|
||||
}
|
||||
|
||||
#[async_trait]
|
||||
#[allow(deprecated)]
|
||||
impl BaseTool for McpToolAdapter {
|
||||
fn name(&self) -> &str {
|
||||
&self.name
|
||||
|
||||
@@ -7,6 +7,7 @@ use std::time::Duration;
|
||||
use futures::future::join_all;
|
||||
use serde_json::Value;
|
||||
|
||||
#[allow(deprecated)]
|
||||
use crate::llm::types::ToolDefinition;
|
||||
use crate::tools::base::{ToolContext, ToolRef};
|
||||
use crate::tools::error::ToolError;
|
||||
@@ -70,6 +71,7 @@ impl std::fmt::Debug for ToolRegistry {
|
||||
}
|
||||
}
|
||||
|
||||
#[allow(deprecated)]
|
||||
impl ToolRegistry {
|
||||
/// 创建一个新的工具注册表。
|
||||
pub fn new() -> Self {
|
||||
|
||||
Reference in New Issue
Block a user