hms/docs/superpowers/plans/2026-05-18-ai-agent-breakthrough-plan.md

AI Agent 突破口实施计划

For agentic workers: REQUIRED: Use superpowers:subagent-driven-development (if subagents available) or superpowers:executing-plans to implement this plan. Steps use checkbox (- [ ]) syntax for tracking.

Goal: 将 erp-ai 的 AI 客服从简单问答升级为 ReAct Agent，通过 Function Calling 串联后端分析能力，实现多策略主动关怀对话。

Architecture: 在 AiProvider trait 新增 generate_with_tools() 方法，实现 Agent Orchestrator 的 ReAct 循环（最多 5 轮 Tool Call），Tool 通过已有的 HealthDataProvider trait 访问 erp-health 数据。会话管理从本地 Storage 迁移到 DB 持久化。

Tech Stack: Rust (Axum + SeaORM + reqwest)、TypeScript/React (Taro 4.2 小程序 + Ant Design Web)、PostgreSQL、SSE

Spec: docs/superpowers/specs/2026-05-18-ai-agent-breakthrough-design.md

---

Chunk 1: Phase 0 — 基础设施（5-6 天）

目标：Agent 核心循环跑通，能用一个 Tool 完成完整对话

Task 0.1: Agent DTO — ChatMessage / ToolDefinition / ToolCall / AgentGenerateResponse

Files:

• Modify: crates/erp-ai/src/dto/mod.rs:62-77 (GenerateRequest/GenerateResponse 之后)

• Test: crates/erp-ai/src/dto/mod.rs (编译检查即可，纯数据结构)

• Step 1: 在 dto/mod.rs 末尾添加 Agent 相关 DTO

在 GenerateResponse 定义之后，添加：

// === Agent Function Calling DTO ===

/// Agent 对话消息
#[derive(Debug, Clone, Serialize, Deserialize, utoipa::ToSchema)]
pub struct ChatMessage {
    pub role: ChatMessageRole,
    pub content: String,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub tool_calls: Option<Vec<ToolCall>>,
    #[serde(skip_serializing_if = "Option::is_none")]
    pub tool_call_id: Option<String>,
}

#[derive(Debug, Clone, Serialize, Deserialize, utoipa::ToSchema)]
#[serde(rename_all = "lowercase")]
pub enum ChatMessageRole {
    User,
    Assistant,
    Tool,
}

/// Tool 定义（传给 LLM 的 Function Schema）
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ToolDefinition {
    pub name: String,
    pub description: String,
    pub parameters: serde_json::Value,
}

/// LLM 返回的 Tool Call
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct ToolCall {
    pub id: String,
    pub name: String,
    pub arguments: serde_json::Value,
}

/// Agent 专用生成响应
#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AgentGenerateResponse {
    pub content: Option<String>,
    pub tool_calls: Option<Vec<ToolCall>>,
    /// 复用已有的 TokenUsage（dto/mod.rs 中的定义：input/output u32）
    pub usage: Option<crate::dto::TokenUsage>,
}

• Step 2: cargo check 验证编译

Run: cargo check -p erp-ai Expected: 编译通过（新增类型无外部依赖）

• Step 3: Commit

git add crates/erp-ai/src/dto/mod.rs
git commit -m "feat(ai): 添加 Agent Function Calling DTO — ChatMessage/ToolDefinition/ToolCall/AgentGenerateResponse"

---

Task 0.2: AiProvider trait 新增 generate_with_tools 方法

Files:

• Modify: crates/erp-ai/src/provider/mod.rs:1-30

• Test: 编译检查

• Step 1: 在 AiProvider trait 中新增 generate_with_tools 默认方法

在 crates/erp-ai/src/provider/mod.rs 的 trait 定义中，health_check 之后添加：

    /// Agent 专用生成方法 — 支持 Function Calling
    /// 不支持 FC 的 Provider 使用默认实现（返回错误）
    async fn generate_with_tools(
        &self,
        messages: Vec<crate::dto::ChatMessage>,
        tools: Vec<crate::dto::ToolDefinition>,
        system_prompt: &str,
        model: &str,
        temperature: f32,
        max_tokens: u32,
    ) -> crate::error::AiResult<crate::dto::AgentGenerateResponse> {
        Err(crate::error::AiError::UnsupportedOperation(
            "Function Calling not supported by this provider".into(),
        ))
    }

同时在 src/error.rs 中添加 UnsupportedOperation 变体（如果不存在）：

#[error("unsupported operation: {0}")]
UnsupportedOperation(String),

• Step 2: cargo check 验证

Run: cargo check -p erp-ai Expected: 编译通过（默认实现不破坏现有 Provider）

• Step 3: Commit

git add crates/erp-ai/src/provider/mod.rs crates/erp-ai/src/error.rs
git commit -m "feat(ai): AiProvider trait 新增 generate_with_tools 默认方法"

---

Task 0.3: Claude Provider 实现 generate_with_tools

Files:

• Modify: crates/erp-ai/src/provider/claude.rs (添加 tools 字段 + 响应解析)

• Step 1: 扩展 ClaudeRequest 结构体

在 claude.rs 的 ClaudeRequest struct 中添加 tools 和 system 字段（如无 system 字段则添加）：

#[derive(Debug, Serialize)]
#[serde(rename_all = "snake_case")]
pub struct ClaudeTool {
    name: String,
    description: String,
    input_schema: serde_json::Value,
}

#[derive(Debug, Serialize)]
struct ClaudeRequest {
    model: String,
    max_tokens: u32,
    #[serde(skip_serializing_if = "Option::is_none")]
    temperature: Option<f32>,
    system: String,
    messages: Vec<ClaudeMessage>,
    #[serde(skip_serializing_if = "Option::is_none")]
    tools: Option<Vec<ClaudeTool>>,
    stream: bool,
}

#[derive(Debug, Serialize, Deserialize)]
struct ClaudeMessage {
    role: String,
    content: serde_json::Value,  // 改为 Value 以支持 tool_use/tool_result 内容块
}

• Step 2: 实现 generate_with_tools

在 impl AiProvider for ClaudeProvider 中添加：

async fn generate_with_tools(
    &self,
    messages: Vec<crate::dto::ChatMessage>,
    tools: Vec<crate::dto::ToolDefinition>,
    system_prompt: &str,
    model: &str,
    temperature: f32,
    max_tokens: u32,
) -> crate::error::AiResult<crate::dto::AgentGenerateResponse> {
    let claude_messages: Vec<ClaudeMessage> = messages.iter().map(|m| {
        // 根据角色和内容构建 Anthropic 格式消息
        // assistant 带 tool_calls 时构造 tool_use content blocks
        // tool 角色时构造 tool_result content block
        // ... 完整转换逻辑
    }).collect();

    let claude_tools: Vec<ClaudeTool> = tools.iter().map(|t| ClaudeTool {
        name: t.name.clone(),
        description: t.description.clone(),
        input_schema: t.parameters.clone(),
    }).collect();

    let req = ClaudeRequest {
        model: model.to_string(),
        max_tokens,
        temperature: Some(temperature),
        system: system_prompt.to_string(),
        messages: claude_messages,
        tools: Some(claude_tools),
        stream: false,
    };

    let resp = self.client.post(&self.api_url)
        .header("x-api-key", &self.api_key)
        .header("anthropic-version", "2023-06-01")
        .json(&req)
        .send().await
        .map_err(|e| AiError::ProviderError(e.to_string()))?;

    let parsed: serde_json::Value = resp.json().await
        .map_err(|e| AiError::ProviderError(e.to_string()))?;

    // 解析 content blocks — 区分 text 和 tool_use
    let mut content_text = None;
    let mut tool_calls = None;

    if let Some(blocks) = parsed["content"].as_array() {
        for block in blocks {
            match block["type"].as_str() {
                Some("text") => {
                    content_text = block["text"].as_str().map(|s| s.to_string());
                }
                Some("tool_use") => {
                    let tc = ToolCall {
                        id: block["id"].as_str().unwrap_or_default().to_string(),
                        name: block["name"].as_str().unwrap_or_default().to_string(),
                        arguments: block["input"].clone(),
                    };
                    tool_calls.get_or_insert_with(Vec::new).push(tc);
                }
                _ => {}
            }
        }
    }

    let usage = parsed["usage"].as_object().map(|u| crate::dto::TokenUsage {
        input: u["input_tokens"].as_u64().unwrap_or(0) as u32,
        output: u["output_tokens"].as_u64().unwrap_or(0) as u32,
    });

    Ok(AgentGenerateResponse { content: content_text, tool_calls, usage })
}

• Step 3: cargo check + cargo test -p erp-ai

Run: cargo check -p erp-ai && cargo test -p erp-ai Expected: 编译通过，现有测试不受影响

• Step 4: Commit

git add crates/erp-ai/src/provider/claude.rs
git commit -m "feat(ai): Claude Provider 实现 generate_with_tools — tool_use/tool_result 解析"

---

Task 0.4: OpenAI Provider 实现 generate_with_tools

Files:

• Modify: crates/erp-ai/src/provider/openai.rs

• Step 1: 扩展 ChatRequest 和 ChatMessageResp

在 openai.rs 中：

#[derive(Debug, Serialize)]
struct ChatTool {
    r#type: String,  // "function"
    function: ChatFunction,
}

#[derive(Debug, Serialize)]
struct ChatFunction {
    name: String,
    description: String,
    parameters: serde_json::Value,
}

#[derive(Debug, Serialize)]
struct ChatRequest {
    model: String,
    max_tokens: u32,
    #[serde(skip_serializing_if = "Option::is_none")]
    temperature: Option<f32>,
    messages: Vec<OpenAiMessage>,
    #[serde(skip_serializing_if = "Option::is_none")]
    tools: Option<Vec<ChatTool>>,
    stream: bool,
}

#[derive(Debug, Serialize, Deserialize)]
struct OpenAiMessage {
    role: String,
    #[serde(skip_serializing_if = "Option::is_none")]
    content: Option<String>,
    #[serde(skip_serializing_if = "Option::is_none")]
    tool_calls: Option<Vec<OpenAiToolCall>>,
    #[serde(skip_serializing_if = "Option::is_none")]
    tool_call_id: Option<String>,
}

#[derive(Debug, Serialize, Deserialize)]
struct OpenAiToolCall {
    id: String,
    r#type: String,
    function: OpenAiFunction,
}

#[derive(Debug, Serialize, Deserialize)]
struct OpenAiFunction {
    name: String,
    arguments: String,
}

• Step 2: 实现 generate_with_tools

在 impl AiProvider for OpenAiProvider 中，转换消息格式（user→user, assistant+tool_calls→assistant, tool→tool），发送请求，解析 choices[0].message.tool_calls。

• Step 3: cargo check + cargo test -p erp-ai

• Step 4: Commit

git add crates/erp-ai/src/provider/openai.rs
git commit -m "feat(ai): OpenAI Provider 实现 generate_with_tools — function calling 支持"

---

Task 0.5: Ollama Provider 降级处理

Files:

• Modify: crates/erp-ai/src/provider/ollama.rs

• Step 1: Ollama 使用默认的 generate_with_tools（返回 UnsupportedOperation）

Ollama 的 Function Calling 支持不稳定，Phase 0 不实现。依赖 trait 默认方法即可。

如果需要显式声明不支持（更好的错误信息），在 impl AiProvider for OllamaProvider 中添加：

async fn generate_with_tools(
    &self,
    _messages: Vec<crate::dto::ChatMessage>,
    _tools: Vec<crate::dto::ToolDefinition>,
    _system_prompt: &str,
    _model: &str,
    _temperature: f32,
    _max_tokens: u32,
) -> crate::error::AiResult<crate::dto::AgentGenerateResponse> {
    Err(crate::error::AiError::UnsupportedOperation(
        "Ollama does not support Function Calling. Use Claude or OpenAI provider for Agent features.".into(),
    ))
}

• Step 2: cargo check -p erp-ai

• Step 3: Commit

git add crates/erp-ai/src/provider/ollama.rs
git commit -m "feat(ai): Ollama Provider 声明不支持 Function Calling"

---

Task 0.6: HealthDataProvider 扩展 — 新增 appointments 和 medication 方法

Files:

• Modify: crates/erp-core/src/health_provider.rs:10-42 (trait 定义)

• Create: crates/erp-core/src/health_provider.rs (新增 DTO: AppointmentSummaryDto, MedicationSummaryDto)

• Modify: crates/erp-health/src/health_provider_impl.rs (实现新方法)

• Test: cargo test -p erp-health

• Step 1: 在 trait 中新增两个方法 + DTO

在 health_provider.rs 的 trait 定义末尾添加：

    /// 获取患者即将到来的预约
    async fn get_upcoming_appointments(
        &self,
        tenant_id: Uuid,
        patient_id: Uuid,
    ) -> AppResult<Vec<AppointmentSummaryDto>>;

    /// 获取患者当前用药列表
    async fn get_medication_list(
        &self,
        tenant_id: Uuid,
        patient_id: Uuid,
    ) -> AppResult<Vec<MedicationSummaryDto>>;

新增 DTO：

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct AppointmentSummaryDto {
    pub id: Uuid,
    pub department: String,
    pub doctor_name: String,
    pub scheduled_at: chrono::DateTime<chrono::Utc>,
    pub status: String,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
pub struct MedicationSummaryDto {
    pub name: String,
    pub dosage: String,
    pub frequency: String,
}

• Step 2: 在 erp-health 实现新方法

在 health_provider_impl.rs 的 impl HealthDataProvider for HealthDataProviderImpl 中，基于现有的 appointment_service 和 medication_record_service 实现查询，返回脱敏后的 DTO。

• Step 3: cargo check + cargo test -p erp-health

Run: cargo check -p erp-health && cargo test -p erp-health Expected: 编译通过

• Step 4: Commit

git add crates/erp-core/src/health_provider.rs crates/erp-health/src/health_provider_impl.rs
git commit -m "feat(core): HealthDataProvider 新增 get_upcoming_appointments + get_medication_list"

---

Task 0.7: 数据库迁移 — 会话/消息/日志/用户画像 4 张表

Files:

• Create: crates/erp-server/migration/src/m20260518_000148_create_ai_chat_tables.rs

• Modify: crates/erp-server/migration/src/lib.rs (注册新迁移)

• Step 1: 创建迁移文件

参考现有迁移文件格式（如 m20260516_000147），创建包含 4 张表的迁移：

• ai_chat_sessions — 会话表（含 tenant_id, user_id, patient_id, title, status, metadata + 标准字段）

• ai_chat_messages — 消息表（含 session_id FK, role, content, tool_calls JSONB, tool_call_id, token_count + 标准字段）

• ai_tool_call_logs — 日志表（append-only：tenant_id, session_id, message_id, tool_name, parameters, result_summary, execution_ms, success, created_at, created_by）

• ai_user_profiles — 用户画像表（tenant_id, user_id UNIQUE, preferences JSONB, health_interests TEXT[], frequent_topics TEXT[], personality_summary, last_updated_at + 标准字段，省略 created_by/updated_by 由 Agent 自动维护）

• Step 2: 在 lib.rs 注册迁移

在 migration/src/lib.rs 的 MigratorTrait 列表中添加新迁移。

• Step 3: cargo check -p erp-server

• Step 4: 启动后端验证迁移执行

Run: cd crates/erp-server && cargo run Expected: 日志显示迁移 000148 执行成功

• Step 5: Commit

git add crates/erp-server/migration/src/m20260518_000148_create_ai_chat_tables.rs crates/erp-server/migration/src/lib.rs
git commit -m "feat(db): 迁移 000148 — AI 聊天会话/消息/工具日志/用户画像 4 张表"

---

Task 0.8: AgentTool trait + ToolRegistry + ToolContext + DisplayHint

Files:

• Create: crates/erp-ai/src/agent/mod.rs (模块入口)

• Create: crates/erp-ai/src/agent/tool.rs (AgentTool trait + ToolContext + ToolResult + DisplayHint)

• Create: crates/erp-ai/src/agent/registry.rs (ToolRegistry)

• Test: crates/erp-ai/src/agent/tool_test.rs (单元测试)

• Step 1: 创建 agent 模块骨架

crates/erp-ai/src/agent/mod.rs:

pub mod tool;
pub mod registry;
pub mod orchestrator;

pub use tool::{AgentTool, ToolContext, ToolResult, DisplayHint};
pub use registry::ToolRegistry;
pub use orchestrator::AgentOrchestrator;

• Step 2: 实现 AgentTool trait + ToolContext + ToolResult + DisplayHint

crates/erp-ai/src/agent/tool.rs:

use async_trait::async_trait;
use chrono::{DateTime, Utc};
use erp_core::health_provider::HealthDataProvider;
use sea_orm::DatabaseConnection;
use serde::{Deserialize, Serialize};
use uuid::Uuid;

#[async_trait]
pub trait AgentTool: Send + Sync {
    fn name(&self) -> &str;
    fn description(&self) -> &str;
    fn parameters_schema(&self) -> serde_json::Value;
    async fn execute(&self, ctx: &ToolContext, params: serde_json::Value) -> ToolResult;
}

pub struct ToolContext {
    pub tenant_id: Uuid,
    pub user_id: Uuid,
    pub patient_id: Option<Uuid>,
    pub db: DatabaseConnection,
    pub health_provider: std::sync::Arc<dyn HealthDataProvider>,
}

pub struct ToolResult {
    pub output: String,
    pub display_hint: Option<DisplayHint>,
}

#[derive(Debug, Clone, Serialize, Deserialize)]
#[serde(tag = "type", rename_all = "snake_case")]
pub enum DisplayHint {
    VitalCard {
        indicator_type: String,
        values: Vec<(String, f64)>,
        unit: String,
    },
    LabReportCard {
        report_date: String,
        abnormal_count: usize,
    },
    ActionConfirm {
        action_type: String,
        summary: String,
        confirm_payload: serde_json::Value,
    },
    RiskAlert {
        level: String,
        message: String,
    },
    Text,
}

• Step 3: 实现 ToolRegistry

crates/erp-ai/src/agent/registry.rs:

use std::collections::HashMap;
use std::sync::Arc;
use super::tool::AgentTool;

pub struct ToolRegistry {
    tools: HashMap<String, Arc<dyn AgentTool>>,
}

impl ToolRegistry {
    pub fn new() -> Self {
        Self { tools: HashMap::new() }
    }

    pub fn register(&mut self, tool: Arc<dyn AgentTool>) {
        self.tools.insert(tool.name().to_string(), tool);
    }

    pub fn get(&self, name: &str) -> Option<&Arc<dyn AgentTool>> {
        self.tools.get(name)
    }

    pub fn all_tools(&self) -> Vec<&Arc<dyn AgentTool>> {
        self.tools.values().collect()
    }

    /// 生成传给 LLM 的 ToolDefinition 列表
    pub fn tool_definitions(&self) -> Vec<crate::dto::ToolDefinition> {
        self.tools.values().map(|t| crate::dto::ToolDefinition {
            name: t.name().to_string(),
            description: t.description().to_string(),
            parameters: t.parameters_schema(),
        }).collect()
    }
}

• Step 4: 在 lib.rs 注册 agent 模块

在 crates/erp-ai/src/lib.rs 添加 pub mod agent;

• Step 5: cargo check -p erp-ai

• Step 6: Commit

git add crates/erp-ai/src/agent/ crates/erp-ai/src/lib.rs
git commit -m "feat(ai): AgentTool trait + ToolRegistry + ToolContext + DisplayHint"

---

Task 0.9: AgentOrchestrator — ReAct 循环

Files:

• Create: crates/erp-ai/src/agent/orchestrator.rs

• Test: crates/erp-ai/src/agent/orchordinator_test.rs

• Step 1: 实现 AgentOrchestrator

crates/erp-ai/src/agent/orchestrator.rs:

use crate::agent::registry::ToolRegistry;
use crate::agent::tool::{AgentTool, ToolContext, ToolResult};
use crate::dto::{AgentGenerateResponse, ChatMessage, ChatMessageRole, ToolCall};
use crate::error::AiResult;
use crate::provider::AiProvider;
use std::sync::Arc;

pub struct AgentOrchestrator {
    provider: Arc<dyn AiProvider>,
    tool_registry: Arc<ToolRegistry>,
    max_iterations: usize,  // 默认 5
}

impl AgentOrchestrator {
    pub fn new(provider: Arc<dyn AiProvider>, tool_registry: Arc<ToolRegistry>) -> Self {
        Self { provider, tool_registry, max_iterations: 5 }
    }

    /// 执行 Agent ReAct 循环
    pub async fn run(
        &self,
        system_prompt: &str,
        messages: &mut Vec<ChatMessage>,
        ctx: &ToolContext,
    ) -> AiResult<AgentRunResult> {
        let tools = self.tool_registry.tool_definitions();
        let mut iterations = 0;
        let mut total_input_tokens = 0u32;
        let mut total_output_tokens = 0u32;

        loop {
            iterations += 1;

            let response = self.provider.generate_with_tools(
                messages.clone(),
                tools.clone(),
                system_prompt,
                "auto",  // 模型由 Provider 内部决定
                0.7,
                2048,
            ).await?;

            if let Some(ref usage) = response.usage {
                total_input_tokens += usage.input_tokens;
                total_output_tokens += usage.output_tokens;
            }

            // 如果没有 tool_calls，Agent 给出最终回复
            let tool_calls = match response.tool_calls {
                Some(tc) if !tc.is_empty() => tc,
                _ => {
                    return Ok(AgentRunResult {
                        reply: response.content.unwrap_or_default(),
                        total_input_tokens,
                        total_output_tokens,
                        iterations,
                    });
                }
            };

            // 达到上限：强制结束
            if iterations >= self.max_iterations {
                // 追加 User 角色指令让 LLM 基于已有信息生成最终回复
                messages.push(ChatMessage {
                    role: ChatMessageRole::User,
                    content: "（系统提示：已收集足够信息，请直接总结回复用户，不要再调用工具）".to_string(),
                    tool_calls: None,
                    tool_call_id: None,
                });
                continue;
            }

            // 将 assistant 的 tool_calls 加入消息历史
            messages.push(ChatMessage {
                role: ChatMessageRole::Assistant,
                content: response.content.unwrap_or_default(),
                tool_calls: Some(tool_calls.clone()),
                tool_call_id: None,
            });

            // 执行每个 Tool Call
            for tc in &tool_calls {
                let tool_result = match self.tool_registry.get(&tc.name) {
                    Some(tool) => {
                        match tool.execute(ctx, tc.arguments.clone()).await {
                            Ok(result) => result.output,
                            Err(e) => format!("Tool '{}' 执行失败: {}", tc.name, e),
                        }
                    }
                    None => format!("未知 Tool: {}", tc.name),
                };

                messages.push(ChatMessage {
                    role: ChatMessageRole::Tool,
                    content: tool_result,
                    tool_calls: None,
                    tool_call_id: Some(tc.id.clone()),
                });
            }
        }
    }
}

pub struct AgentRunResult {
    pub reply: String,
    pub total_input_tokens: u32,
    pub total_output_tokens: u32,
    pub iterations: usize,
}

• Step 2: cargo check -p erp-ai

• Step 3: Commit

git add crates/erp-ai/src/agent/orchestrator.rs
git commit -m "feat(ai): AgentOrchestrator — ReAct 循环（最多 5 轮 Tool Call + 强制终止）"

---

Task 0.10: 实现 query_patient_vitals Tool — 端到端验证

Files:

• Create: crates/erp-ai/src/agent/tools/mod.rs

• Create: crates/erp-ai/src/agent/tools/query_vitals.rs

• Modify: crates/erp-ai/src/agent/mod.rs (注册 tools 子模块)

• Step 1: 创建 tools 子模块

crates/erp-ai/src/agent/tools/mod.rs:

pub mod query_vitals;

crates/erp-ai/src/agent/tools/query_vitals.rs:

use async_trait::async_trait;
use crate::agent::tool::{AgentTool, ToolContext, ToolResult, DisplayHint};
use serde::{Deserialize, Serialize};
use erp_core::health_provider::TimeRange;
use chrono::Utc;

pub struct QueryPatientVitalsTool;

#[async_trait]
impl AgentTool for QueryPatientVitalsTool {
    fn name(&self) -> &str { "query_patient_vitals" }

    fn description(&self) -> &str {
        "查询患者最近的体征数据（血压、血糖、心率等）。需要提供患者 ID 和天数范围（默认 7 天）。"
    }

    fn parameters_schema(&self) -> serde_json::Value {
        serde_json::json!({
            "type": "object",
            "properties": {
                "days": {
                    "type": "integer",
                    "description": "查询最近多少天的数据，默认 7 天"
                }
            }
        })
    }

    async fn execute(&self, ctx: &ToolContext, params: serde_json::Value) -> ToolResult {
        let patient_id = match ctx.patient_id {
            Some(id) => id,
            None => return ToolResult {
                output: "未关联患者档案，无法查询体征数据".to_string(),
                display_hint: None,
            },
        };

        let days = params["days"].as_i64().unwrap_or(7);
        let now = Utc::now();
        let start = now - chrono::Duration::days(days);

        let range = TimeRange { start, end: now };
        let metrics = vec![
            "blood_pressure_systolic".into(),
            "blood_pressure_diastolic".into(),
            "heart_rate".into(),
            "blood_glucose".into(),
        ];

        match ctx.health_provider.get_vital_signs(ctx.tenant_id, patient_id, &metrics, &range).await {
            Ok(vitals) => {
                if vitals.is_empty() {
                    return ToolResult {
                        output: "该时间段内无体征数据".to_string(),
                        display_hint: None,
                    };
                }

                let mut output = String::from("最近体征数据：\n");
                for v in &vitals {
                    output.push_str(&format!("- {}: ", v.metric));
                    let values_str: Vec<String> = v.values.iter()
                        .take(10)
                        .map(|(date, val)| format!("{}={}", date, val))
                        .collect();
                    output.push_str(&values_str.join(", "));
                    output.push_str(&format!(" ({})\n", v.unit));
                }

                ToolResult {
                    output,
                    display_hint: Some(DisplayHint::VitalCard {
                        indicator_type: vitals[0].metric.clone(),
                        values: vitals[0].values.iter().take(10)
                            .map(|(d, v)| (d.clone(), *v))
                            .collect(),
                        unit: vitals[0].unit.clone(),
                    }),
                }
            }
            Err(e) => ToolResult {
                output: format!("查询体征数据失败: {}", e),
                display_hint: None,
            },
        }
    }
}

• Step 2: 更新 agent/mod.rs 注册 tools 子模块

添加 pub mod tools; 并在 pub use 中导出。

• Step 3: cargo check -p erp-ai

• Step 4: Commit

git add crates/erp-ai/src/agent/tools/ crates/erp-ai/src/agent/mod.rs
git commit -m "feat(ai): 实现 query_patient_vitals Tool — 首个端到端 Agent Tool"

---

Task 0.11: 改造 chat_handler — 接入 AgentOrchestrator

Files:

• Modify: crates/erp-ai/src/handler/chat_handler.rs (替换原有简单逻辑)

• Modify: crates/erp-ai/src/state.rs (添加 ToolRegistry 字段)

• Modify: crates/erp-ai/src/module.rs (注册新权限码 + 初始化 ToolRegistry)

• Step 1: 在 AiState 中添加 ToolRegistry

state.rs 新增字段：

pub tool_registry: Arc<crate::agent::ToolRegistry>,

• Step 2: 在 module.rs 中初始化 ToolRegistry 并注入 AiState

在模块初始化时：

let mut tool_registry = ToolRegistry::new();
tool_registry.register(Arc::new(QueryPatientVitalsTool));
// 后续 Phase 1 添加更多 Tool

• Step 3: 重写 chat_handler 使用 AgentOrchestrator

替换原有的 chat() 函数核心逻辑：

1. 从请求中获取 session_id（或创建新会话）
2. 从 DB 加载会话历史消息
3. 将用户消息保存到 DB
4. 构建 ToolContext（从 AiState 获取 health_provider, db）
5. 构建 system prompt（多策略，Phase 1 完善）
6. 创建 AgentOrchestrator 并调用 run()
7. 将 Agent 回复保存到 DB
8. 返回 ChatResponse

注意：Phase 0 先用简化版 session 管理（直接传 session_id 参数），完整的 Session CRUD API 留到 Phase 2。

路由说明：Phase 0 复用现有 POST /ai/chat 路由（module.rs:361 已注册），改造 handler 内部逻辑。Phase 2 会变更为 Spec 定义的 /api/v1/ai/chat/sessions/{id}/messages。

模型选择：Phase 0 硬编码 "auto" 由 Provider 内部决定模型。后续可通过 AiState.provider_registry 动态选择。

• Step 4: 在 module.rs 注册新权限码

// 现有权限码补充
("ai.chat.session.list", "查看 AI 会话列表"),
("ai.chat.session.manage", "创建/关闭 AI 会话"),
("ai.chat.session.history", "查看 AI 会话消息历史"),

• Step 5: cargo check + cargo test --workspace

• Step 6: 功能验证 — 启动后端，用 Postman 测试

cd crates/erp-server && cargo run

Postman 发送 POST /api/v1/ai/chat：

{
  "message": "我最近血压怎么样",
  "history": []
}

Expected: Agent 返回包含血压数据的自然语言回复。

• Step 7: Commit

git add crates/erp-ai/src/handler/chat_handler.rs crates/erp-ai/src/state.rs crates/erp-ai/src/module.rs
git commit -m "feat(ai): 改造 chat_handler 接入 AgentOrchestrator — ReAct Agent 首次跑通"

---

Task 0.12: Phase 0 集成测试

Files:

• Create: crates/erp-server/tests/integration/ai_agent_test.rs

• Step 1: 编写集成测试

测试场景：

1. 发送简单问候 → Agent 直接回复（无 Tool Call）
2. 发送体征查询 → Agent 调用 query_patient_vitals Tool → 回复包含数据
3. 达到 5 轮上限 → Agent 正常结束回复
4. 无关联患者 → Tool 返回提示信息

注意：集成测试需要 mock LLM Provider（避免真实 API 调用），可创建 MockProvider 实现 AiProvider trait。

• Step 2: cargo test --workspace

• Step 3: Commit

git add crates/erp-server/tests/integration/ai_agent_test.rs
git commit -m "test(ai): Phase 0 集成测试 — Agent 循环 + Tool 执行 + 降级场景"

---

Phase 0 完成标准

• cargo check 全 workspace 通过
• cargo test --workspace 全部通过
• Postman 调用 /api/v1/ai/chat，Agent 能查到患者体征数据并自然回复
• 代码已提交并推送

---