diff --git a/README.md b/README.md index 20b4d35..ca6d138 100644 --- a/README.md +++ b/README.md @@ -65,6 +65,30 @@ func main() { ## 版本历史 +### v1.2 — 真实模型接入与 Thinking 死循环修复 + +#### 变更 + +- **接入真实大模型** — 新增 `OpenAIProvider` 和 `ClaudeProvider`，分别基于 OpenAI V3 SDK 和 Anthropic SDK 连接 Deepseek API，替换原有的 Mock Provider +- **Deepseek thinking 模式适配** — 修正 `reasoning_content` 字段丢失导致的 API 400 错误，在 `schema.Message` 中新增字段并在请求中回传 +- **Thinking 死循环修复** — 过渡指令只在首轮插入，防止模型每轮重复调用工具 +- **最大轮数保护** — 新增 `maxTurns = 10` 上限，防止意外死循环 +- **调试输出** — 新增 `dumpMessages` 和 `dumpTools`，每轮打印上下文消息和工具列表 + +#### 踩坑记录 + +| 问题 | 原因 | 解决 | +|---|---|---| +| API 400: `reasoning_content must be passed back` | Deepseek thinking 模式返回了 `reasoning_content` 字段，回传请求时未携带 | 从响应 RawJSON 中提取该字段，用 `param.Override` 注入原始 JSON 回传 | +| 模型在 Phase 2 不调用工具 | Phase 1 的思考内容作为 Assistant 消息追加后，模型认为对话已结束 | 在思考后插入一条 User 角色过渡指令："根据推理，使用工具完成任务" | +| 引擎死循环跑到 maxTurns | 过渡指令每轮都插入，模型每轮都被要求调用工具 | 过渡指令只在 `turnCount == 1` 时插入，后续轮次模型根据观察结果自行判断 | + +#### 经验教训 + +1. **上下文即状态** — Agent 的所有行为都由上下文驱动。插入一条消息就能改变模型行为，不需要改代码逻辑 +2. **非标准 API 字段需手动处理** — 大模型厂商的扩展字段（如 `reasoning_content`）不在 SDK 类型中，需要从 RawJSON 手动提取并用注入方式回传 +3. **过渡指令的作用域很重要** — "使用工具"这种指令适合在首轮引导，重复出现会导致模型无法自行判断任务是否完成 + ### v1.1 — 慢思考模式 (Thinking Phase) 将 ReAct 循环从单阶段升级为双阶段架构： diff --git a/cmd/claw/main.go b/cmd/claw/main.go index 598ee48..3b4f3cd 100644 --- a/cmd/claw/main.go +++ b/cmd/claw/main.go @@ -4,71 +4,63 @@ import ( "context" "log" "os" + "fmt" "go-tiny-claw/internal/engine" "go-tiny-claw/internal/schema" + "go-tiny-claw/internal/provider" ) -// 升级版 Mock Provider -type mockProvider struct { - turn int -} - -func (m *mockProvider) Generate(ctx context.Context, msgs []schema.Message, tools []schema.ToolDefinition) (*schema.Message, error) { - // 如果工具列表为空，说明这是引擎发起的 Phase 1: Thinking 阶段 - if len(tools) == 0 { - return &schema.Message{ - Role: schema.RoleAssistant, - Content: "【推理中】目标是检查文件。我不能直接盲猜，我需要先调用 bash 工具执行 ls 命令，看看当前目录下有什么，然后再做定夺。", - }, nil - } - - // 如果工具列表不为空，说明这是 Phase 2: Action 阶段 - m.turn++ - if m.turn == 1 { - // 第一轮 Action：顺着刚才的 Thinking，精准调用工具 - return &schema.Message{ - Role: schema.RoleAssistant, - Content: "我要执行我刚才计划的步骤了。", - ToolCalls: []schema.ToolCall{ - {ID: "call_123", Name: "bash", Arguments: []byte(`{"command": "ls -la"}`)}, - }, - }, nil - } - - // 第二轮 Action：直接总结退出 - return &schema.Message{ - Role: schema.RoleAssistant, - Content: "根据工具返回的结果，我看到了 main.go，任务圆满完成！", - }, nil -} +// 伪造的工具注册表 (用于测试 Provider 的工具提取能力) type mockRegistry struct{} -func (m *mockRegistry) GetAvailableTools() []schema.ToolDefinition { - // 为了让 Phase 2 能检测到工具，这里返回一个伪造的工具定义数组 - return []schema.ToolDefinition{{Name: "bash"}} +func (m *mockRegistry) GetAvailableTools() []schema.ToolDefinition { + return []schema.ToolDefinition{ + { + Name: "get_weather", + Description: "获取指定城市的当前天气情况。", + InputSchema: map[string]interface{}{ + "type": "object", + "properties": map[string]interface{}{ + "city": map[string]interface{}{ + "type": "string", + }, + }, + "required": []string{"city"}, + }, + }, + } } func (m *mockRegistry) Execute(ctx context.Context, call schema.ToolCall) schema.ToolResult { + log.Printf(" -> [Mock 工具执行] 获取 %s 的天气中...\n", call.Name) return schema.ToolResult{ ToolCallID: call.ID, - Output: "-rw-r--r-- 1 user group 234 Oct 24 10:00 main.go\n", + Output: "API 返回：今天是晴天，气温 25 度。", IsError: false, } } func main() { + + fmt.Printf("11111") workDir, _ := os.Getwd() - p := &mockProvider{} - r := &mockRegistry{} + // 1. 初始化真实的 Provider大脑 (指向智谱 GLM-4.5) + // 这里你可以任意切换 NewZhipuClaudeProvider 或 NewZhipuOpenAIProvider，效果完全一致！ + llmProvider := provider.DeepseekOpenAIProvider("deepseek-v4-flash") + // 2. 注入伪造的工具注册表 + registry := &mockRegistry{} - // 实例化引擎，开启 EnableThinking = true - eng := engine.NewAgentEngine(p, r, workDir, true) + // 3. 实例化并运行引擎，开启 EnableThinking = true (开启慢思考阶段！) + eng := engine.NewAgentEngine(llmProvider, registry, workDir, true) - err := eng.Run(context.Background(), "帮我检查当前目录的文件") + // 设定测试任务 + prompt := "我想去北京跑步，帮我查查天气适合吗？" + + err := eng.Run(context.Background(), prompt) if err != nil { - log.Fatalf("引擎崩溃: %v", err) + log.Fatalf("引擎运行崩溃: %v", err) } } \ No newline at end of file diff --git a/go.mod b/go.mod index 6175f1f..55f0347 100644 --- a/go.mod +++ b/go.mod @@ -1,2 +1,21 @@ module go-tiny-claw +go 1.23.0 + +require ( + github.com/anthropics/anthropic-sdk-go v1.43.0 // indirect + github.com/bahlo/generic-list-go v0.2.0 // indirect + github.com/buger/jsonparser v1.1.2 // indirect + github.com/invopop/jsonschema v0.13.0 // indirect + github.com/joho/godotenv v1.5.1 // indirect + github.com/mailru/easyjson v0.7.7 // indirect + github.com/openai/openai-go/v3 v3.35.0 // indirect + github.com/standard-webhooks/standard-webhooks/libraries v0.0.1 // indirect + github.com/tidwall/gjson v1.18.0 // indirect + github.com/tidwall/match v1.1.1 // indirect + github.com/tidwall/pretty v1.2.1 // indirect + github.com/tidwall/sjson v1.2.5 // indirect + github.com/wk8/go-ordered-map/v2 v2.1.8 // indirect + golang.org/x/sync v0.16.0 // indirect + gopkg.in/yaml.v3 v3.0.1 // indirect +) diff --git a/go.sum b/go.sum new file mode 100644 index 0000000..99f4318 --- /dev/null +++ b/go.sum @@ -0,0 +1,34 @@ +github.com/anthropics/anthropic-sdk-go v1.43.0 h1:ShY3C7lafzHP0ze1dCxL3ZFZzvkGfXJN91DfZTG8zLM= +github.com/anthropics/anthropic-sdk-go v1.43.0/go.mod h1:5cEaslQ6A9ajdL5YUvhNW57LKxEz0OAZ7WEzgZWLD7k= +github.com/bahlo/generic-list-go v0.2.0 h1:5sz/EEAK+ls5wF+NeqDpk5+iNdMDXrh3z3nPnH1Wvgk= +github.com/bahlo/generic-list-go v0.2.0/go.mod h1:2KvAjgMlE5NNynlg/5iLrrCCZ2+5xWbdbCW3pNTGyYg= +github.com/buger/jsonparser v1.1.2 h1:frqHqw7otoVbk5M8LlE/L7HTnIq2v9RX6EJ48i9AxJk= +github.com/buger/jsonparser v1.1.2/go.mod h1:6RYKKt7H4d4+iWqouImQ9R2FZql3VbhNgx27UK13J/0= +github.com/invopop/jsonschema v0.13.0 h1:KvpoAJWEjR3uD9Kbm2HWJmqsEaHt8lBUpd0qHcIi21E= +github.com/invopop/jsonschema v0.13.0/go.mod h1:ffZ5Km5SWWRAIN6wbDXItl95euhFz2uON45H2qjYt+0= +github.com/joho/godotenv v1.5.1 h1:7eLL/+HRGLY0ldzfGMeQkb7vMd0as4CfYvUVzLqw0N0= +github.com/joho/godotenv v1.5.1/go.mod h1:f4LDr5Voq0i2e/R5DDNOoa2zzDfwtkZa6DnEwAbqwq4= +github.com/josharian/intern v1.0.0/go.mod h1:5DoeVV0s6jJacbCEi61lwdGj/aVlrQvzHFFd8Hwg//Y= +github.com/mailru/easyjson v0.7.7 h1:UGYAvKxe3sBsEDzO8ZeWOSlIQfWFlxbzLZe7hwFURr0= +github.com/mailru/easyjson v0.7.7/go.mod h1:xzfreul335JAWq5oZzymOObrkdz5UnU4kGfJJLY9Nlc= +github.com/openai/openai-go/v3 v3.35.0 h1:109x3epXMSE423KW2euR506GGFezcEt0s87MoWejpH0= +github.com/openai/openai-go/v3 v3.35.0/go.mod h1:cdufnVK14cWcT9qA1rRtrXx4FTRsgbDPW7Ia7SS5cZo= +github.com/standard-webhooks/standard-webhooks/libraries v0.0.1 h1:uOfcYT+3QungH6tIGSVCR/Y3KJmgJiHcojJbMTPDZAI= +github.com/standard-webhooks/standard-webhooks/libraries v0.0.1/go.mod h1:L1MQhA6x4dn9r007T033lsaZMv9EmBAdXyU/+EF40fo= +github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk= +github.com/tidwall/gjson v1.18.0 h1:FIDeeyB800efLX89e5a8Y0BNH+LOngJyGrIWxG2FKQY= +github.com/tidwall/gjson v1.18.0/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk= +github.com/tidwall/match v1.1.1 h1:+Ho715JplO36QYgwN9PGYNhgZvoUSc9X2c80KVTi+GA= +github.com/tidwall/match v1.1.1/go.mod h1:eRSPERbgtNPcGhD8UCthc6PmLEQXEWd3PRB5JTxsfmM= +github.com/tidwall/pretty v1.2.0/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU= +github.com/tidwall/pretty v1.2.1 h1:qjsOFOWWQl+N3RsoF5/ssm1pHmJJwhjlSbZ51I6wMl4= +github.com/tidwall/pretty v1.2.1/go.mod h1:ITEVvHYasfjBbM0u2Pg8T2nJnzm8xPwvNhhsoaGGjNU= +github.com/tidwall/sjson v1.2.5 h1:kLy8mja+1c9jlljvWTlSazM7cKDRfJuR/bOJhcY5NcY= +github.com/tidwall/sjson v1.2.5/go.mod h1:Fvgq9kS/6ociJEDnK0Fk1cpYF4FIW6ZF7LAe+6jwd28= +github.com/wk8/go-ordered-map/v2 v2.1.8 h1:5h/BUHu93oj4gIdvHHHGsScSTMijfx5PeYkE/fJgbpc= +github.com/wk8/go-ordered-map/v2 v2.1.8/go.mod h1:5nJHM5DyteebpVlHnWMV0rPz6Zp7+xBAnxjb1X5vnTw= +golang.org/x/sync v0.16.0 h1:ycBJEhp9p4vXvUZNszeOq0kGTPghopOL8q0fq3vstxw= +golang.org/x/sync v0.16.0/go.mod h1:1dzgHSNfp02xaA81J2MS99Qcpr2w7fw1gpm99rleRqA= +gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0= +gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA= +gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM= diff --git a/internal/engine/loop.go b/internal/engine/loop.go index db00b45..4185929 100644 --- a/internal/engine/loop.go +++ b/internal/engine/loop.go @@ -29,6 +29,28 @@ func NewAgentEngine(p provider.LLMProvider, r tools.Registry, workDir string, en } } +// dumpMessages 打印当前上下文中的所有消息 (调试用) +func dumpMessages(msgs []schema.Message) { + for i, msg := range msgs { + content := msg.Content + if len(content) > 80 { + content = content[:80] + "..." + } + log.Printf(" [%02d] %-10s | %s", i, msg.Role, content) + } +} + +// dumpTools 打印当前可用的工具列表 +func dumpTools(tools []schema.ToolDefinition) { + if len(tools) == 0 { + log.Printf(" [Tools] (无可用工具)\n") + return + } + for i, t := range tools { + log.Printf(" [Tool %d] %s — %s", i, t.Name, t.Description) + } +} + // internal/engine/loop.go (续) func (e *AgentEngine) Run(ctx context.Context, userPrompt string) error { log.Printf("[Engine] 引擎启动，锁定工作区: %s\n", e.WorkDir) @@ -46,18 +68,25 @@ func (e *AgentEngine) Run(ctx context.Context, userPrompt string) error { } turnCount := 0 + const maxTurns = 10 for { turnCount++ + if turnCount > maxTurns { + log.Printf("[Engine] 已达最大轮数 (%d)，强制终止。\n", maxTurns) + break + } log.Printf("\n========== [Turn %d] 开始 ==========\n", turnCount) + dumpMessages(contextHistory) // 获取当前挂载的所有工具定义 availableTools := e.registry.GetAvailableTools() + dumpTools(availableTools) // ==================================================================== - // Phase 1: 慢思考阶段 (Thinking) - 剥夺工具，强制规划 + // Phase 1: 慢思考阶段 (Thinking) - 仅第一轮执行初始规划 // ==================================================================== - if e.EnableThinking { + if e.EnableThinking && turnCount == 1 { log.Println("[Engine][Phase 1] 剥夺工具访问权，强制进入慢思考与规划阶段...") // 核心机制：传入的 availableTools 为 nil！ @@ -72,6 +101,12 @@ func (e *AgentEngine) Run(ctx context.Context, userPrompt string) error { fmt.Printf("🧠 [内部思考 Trace]: %s\n", thinkResp.Content) contextHistory = append(contextHistory, *thinkResp) } + + // 插入过渡指令：让模型知道现在可以调用工具了 + contextHistory = append(contextHistory, schema.Message{ + Role: schema.RoleUser, + Content: "根据你的推理，现在请使用可用的工具来完成任务。执行具体行动。", + }) } // ==================================================================== @@ -79,7 +114,7 @@ func (e *AgentEngine) Run(ctx context.Context, userPrompt string) error { // ==================================================================== log.Println("[Engine][Phase 2] 恢复工具挂载，等待模型采取行动...") - // 此时的 contextHistory 中已经包含了上一阶段模型自己的 Thinking Trace。 + // 此时的 contextHistory 中已经包含了上一阶段模型自己的 Thinking Trace + 过渡指令。 // 模型会顺着自己的逻辑，结合恢复的 availableTools 发起精准的工具调用。 actionResp, err := e.provider.Generate(ctx, contextHistory, availableTools) if err != nil { diff --git a/internal/provider/claude.go b/internal/provider/claude.go new file mode 100644 index 0000000..9d2e30c --- /dev/null +++ b/internal/provider/claude.go @@ -0,0 +1,144 @@ +// internal/provider/claude.go +package provider + +import ( + "context" + "encoding/json" + "fmt" + // "os" + + "github.com/anthropics/anthropic-sdk-go" + "github.com/anthropics/anthropic-sdk-go/option" + "go-tiny-claw/internal/schema" +) + +type ClaudeProvider struct { + client anthropic.Client + model string +} + +func DeepseekClaudeProvider(model string) *ClaudeProvider { + apiKey := "sk-1f44696abe644bd684f09cc43f12c557" + if apiKey == "" { + panic("请设置 ZHIPU_API_KEY 环境变量") + } + baseURL := "https://api.deepseek.com/anthropic" + return &ClaudeProvider{ + client: anthropic.NewClient(option.WithAPIKey(apiKey), option.WithBaseURL(baseURL)), + model: model, + } +} + +func (p *ClaudeProvider) Generate(ctx context.Context, msgs []schema.Message, availableTools []schema.ToolDefinition) (*schema.Message, error) { + var anthropicMsgs []anthropic.MessageParam + var systemPrompt string + + // 1. 消息翻译 + for _, msg := range msgs { + switch msg.Role { + case schema.RoleSystem: + systemPrompt = msg.Content + case schema.RoleUser: + if msg.ToolCallID != "" { + anthropicMsgs = append(anthropicMsgs, anthropic.NewUserMessage( + anthropic.NewToolResultBlock(msg.ToolCallID, msg.Content, false), + )) + } else { + anthropicMsgs = append(anthropicMsgs, anthropic.NewUserMessage( + anthropic.NewTextBlock(msg.Content), + )) + } + case schema.RoleAssistant: + var blocks []anthropic.ContentBlockParamUnion + if msg.Content != "" { + blocks = append(blocks, anthropic.NewTextBlock(msg.Content)) + } + + // 将历史工具调用转回 Claude 特有的 ToolUseBlockParam + for _, tc := range msg.ToolCalls { + var inputMap map[string]interface{} + _ = json.Unmarshal(tc.Arguments, &inputMap) + blocks = append(blocks, anthropic.ContentBlockParamUnion{ + OfToolUse: &anthropic.ToolUseBlockParam{ + ID: tc.ID, + Name: tc.Name, + Input: inputMap, + }, + }) + } + if len(blocks) > 0 { + anthropicMsgs = append(anthropicMsgs, anthropic.NewAssistantMessage(blocks...)) + } + } + } + + // 2. 工具 Schema 翻译 + var anthropicTools []anthropic.ToolUnionParam + for _, toolDef := range availableTools { + // ToolInputSchemaParam 是结构体，需要通过 Properties 字段精准填充 + var properties map[string]any + var required []string + + if m, ok := toolDef.InputSchema.(map[string]interface{}); ok { + if p, ok := m["properties"].(map[string]interface{}); ok { + properties = p + } + if r, ok := m["required"].([]string); ok { + required = r + } + } + + tp := anthropic.ToolParam{ + Name: toolDef.Name, + Description: anthropic.String(toolDef.Description), + InputSchema: anthropic.ToolInputSchemaParam{ + Properties: properties, + Required: required, + }, + } + anthropicTools = append(anthropicTools, anthropic.ToolUnionParam{OfTool: &tp}) + } + + // 3. 构建请求并发送 + params := anthropic.MessageNewParams{ + Model: anthropic.Model(p.model), + MaxTokens: 4096, + Messages: anthropicMsgs, + } + + if systemPrompt != "" { + params.System = []anthropic.TextBlockParam{ + {Text: systemPrompt}, + } + } + + if len(anthropicTools) > 0 { + params.Tools = anthropicTools + } + + resp, err := p.client.Messages.New(ctx, params) + if err != nil { + return nil, fmt.Errorf("Claude/Zhipu API 请求失败: %w", err) + } + + // 4. 反向解析 + resultMsg := &schema.Message{ + Role: schema.RoleAssistant, + } + + for _, block := range resp.Content { + switch block.Type { + case "text": + resultMsg.Content += block.Text + case "tool_use": + argsBytes, _ := json.Marshal(block.Input) + resultMsg.ToolCalls = append(resultMsg.ToolCalls, schema.ToolCall{ + ID: block.ID, + Name: block.Name, + Arguments: argsBytes, + }) + } + } + + return resultMsg, nil +} \ No newline at end of file diff --git a/internal/provider/openai.go b/internal/provider/openai.go new file mode 100644 index 0000000..ede73e1 --- /dev/null +++ b/internal/provider/openai.go @@ -0,0 +1,193 @@ +// internal/provider/openai.go +package provider + +import ( + "context" + "encoding/json" + "fmt" + // "os" + + "github.com/openai/openai-go/v3" + "github.com/openai/openai-go/v3/option" + "github.com/openai/openai-go/v3/packages/param" + "github.com/openai/openai-go/v3/shared" + "go-tiny-claw/internal/schema" +) + +type OpenAIProvider struct { + client openai.Client // 值类型，非指针 + model string +} + +// NewZhipuOpenAIProvider 构造函数：基于 OpenAI V3 SDK，指向智谱底座 +func DeepseekOpenAIProvider(model string) *OpenAIProvider { + apiKey := "sk-1f44696abe644bd684f09cc43f12c557" + if apiKey == "" { + panic("请设置 ZHIPU_API_KEY 环境变量") + } + // 核心：将官方 SDK 的地址替换为智谱的兼容端点 + baseURL := "https://api.deepseek.com" + + return &OpenAIProvider{ + client: openai.NewClient(option.WithAPIKey(apiKey), option.WithBaseURL(baseURL)), + model: model, + } +} + +func (p *OpenAIProvider) Generate(ctx context.Context, msgs []schema.Message, availableTools []schema.ToolDefinition) (*schema.Message, error) { + var openaiMsgs []openai.ChatCompletionMessageParamUnion + + // 1. 翻译上下文消息 + for _, msg := range msgs { + switch msg.Role { + case schema.RoleSystem: + openaiMsgs = append(openaiMsgs, openai.SystemMessage(msg.Content)) + + case schema.RoleUser: + if msg.ToolCallID != "" { + // 注意：v3 新版参数顺序是 (content, toolCallID) + openaiMsgs = append(openaiMsgs, openai.ToolMessage(msg.Content, msg.ToolCallID)) + } else { + openaiMsgs = append(openaiMsgs, openai.UserMessage(msg.Content)) + } + + case schema.RoleAssistant: + // Deepseek thinking mode: reasoning_content 必须回传 + if msg.ReasoningContent != "" || len(msg.ToolCalls) > 0 { + msgMap := map[string]interface{}{ + "role": "assistant", + "content": msg.Content, + } + if msg.ReasoningContent != "" { + msgMap["reasoning_content"] = msg.ReasoningContent + } + if len(msg.ToolCalls) > 0 { + var toolCalls []map[string]interface{} + for _, tc := range msg.ToolCalls { + toolCalls = append(toolCalls, map[string]interface{}{ + "id": tc.ID, + "type": "function", + "function": map[string]interface{}{ + "name": tc.Name, + "arguments": string(tc.Arguments), + }, + }) + } + msgMap["tool_calls"] = toolCalls + } + rawJSON, _ := json.Marshal(msgMap) + astParam := param.Override[openai.ChatCompletionAssistantMessageParam](json.RawMessage(rawJSON)) + openaiMsgs = append(openaiMsgs, openai.ChatCompletionMessageParamUnion{ + OfAssistant: &astParam, + }) + break + } + + astParam := openai.ChatCompletionAssistantMessageParam{} + + if msg.Content != "" { + astParam.Content = openai.ChatCompletionAssistantMessageParamContentUnion{ + OfString: openai.String(msg.Content), + } + } + + // 【重要】如果历史包含 ToolCalls，必须原样放回，以维系大模型的逻辑链 + if len(msg.ToolCalls) > 0 { + var toolCalls []openai.ChatCompletionMessageToolCallUnionParam + for _, tc := range msg.ToolCalls { + // OfFunction 对应 GetFunction()，字段类型严格要求为指针 + toolCalls = append(toolCalls, openai.ChatCompletionMessageToolCallUnionParam{ + OfFunction: &openai.ChatCompletionMessageFunctionToolCallParam{ + ID: tc.ID, + Type: "function", + Function: openai.ChatCompletionMessageFunctionToolCallFunctionParam{ + Name: tc.Name, + Arguments: string(tc.Arguments), + }, + }, + }) + } + astParam.ToolCalls = toolCalls + } + + openaiMsgs = append(openaiMsgs, openai.ChatCompletionMessageParamUnion{ + OfAssistant: &astParam, + }) + } + } + + // 2. 翻译工具定义 (v3 新 API 特性适配) + var openaiTools []openai.ChatCompletionToolUnionParam + for _, toolDef := range availableTools { + var params shared.FunctionParameters + + // 尝试直接断言，如果不成功则通过 JSON 往返序列化来保证类型匹配 + if m, ok := toolDef.InputSchema.(map[string]interface{}); ok { + params = shared.FunctionParameters(m) + } else { + // fallback：JSON 往返序列化 + b, _ := json.Marshal(toolDef.InputSchema) + _ = json.Unmarshal(b, ¶ms) + } + + openaiTools = append(openaiTools, openai.ChatCompletionFunctionTool( + shared.FunctionDefinitionParam{ + Name: toolDef.Name, + Description: openai.String(toolDef.Description), + Parameters: params, + }, + )) + } + + // 3. 构建请求并发送 + params := openai.ChatCompletionNewParams{ + Model: p.model, + Messages: openaiMsgs, + } + + // 【慢思考机制支撑】仅当 availableTools 存在时才挂载 Tools + if len(openaiTools) > 0 { + params.Tools = openaiTools + } + + resp, err := p.client.Chat.Completions.New(ctx, params) + if err != nil { + return nil, fmt.Errorf("OpenAI/Zhipu API 请求失败: %w", err) + } + if len(resp.Choices) == 0 { + return nil, fmt.Errorf("API 返回了空的 Choices") + } + + // 4. 将 API Response 反向翻译为内部 schema.Message + choice := resp.Choices[0].Message + resultMsg := &schema.Message{ + Role: schema.RoleAssistant, + Content: choice.Content, + } + + // Deepseek thinking mode: 从原始响应中提取 reasoning_content + rawMsg := choice.RawJSON() + if rawMsg != "" { + var rawMap map[string]json.RawMessage + if err := json.Unmarshal([]byte(rawMsg), &rawMap); err == nil { + if rc, ok := rawMap["reasoning_content"]; ok { + var s string + if json.Unmarshal(rc, &s) == nil { + resultMsg.ReasoningContent = s + } + } + } + } + + for _, tc := range choice.ToolCalls { + if tc.Type == "function" { + resultMsg.ToolCalls = append(resultMsg.ToolCalls, schema.ToolCall{ + ID: tc.ID, + Name: tc.Function.Name, + Arguments: []byte(tc.Function.Arguments), // 提取 JSON 字符串字节 + }) + } + } + + return resultMsg, nil +} \ No newline at end of file diff --git a/internal/schema/message.go b/internal/schema/message.go index 610c845..a627b89 100644 --- a/internal/schema/message.go +++ b/internal/schema/message.go @@ -16,6 +16,9 @@ type Message struct { Role Role `json:"role"` Content string `json:"content"` // 存放纯文本内容 + // ReasoningContent Deepseek thinking 模式的内部推理链，回传时必须保留 + ReasoningContent string `json:"reasoning_content,omitempty"` + // 如果模型决定调用工具，此字段将被填充 (支持并行调用多个工具) ToolCalls []ToolCall `json:"tool_calls,omitempty"` diff --git a/main.go b/main.go index 6263e13..440e843 100644 --- a/main.go +++ b/main.go @@ -50,7 +50,7 @@ func main() { p := &mockProvider{} r := &mockRegistry{} - eng := engine.NewAgentEngine(p, r, workDir) + eng := engine.NewAgentEngine(p, r, workDir, true) err := eng.Run(context.Background(), "帮我检查当前目录的文件") if err != nil {