From be751e76f69a22b123768abdc76803e01a443d1b Mon Sep 17 00:00:00 2001
From: Jerry Yan <792602257@qq.com>
Date: Mon, 4 May 2026 09:40:41 +0800
Subject: [PATCH] feat(infra/llm): OpenAI client + base_url override +
 reasoning_content compat

---
 go.mod                            |   3 +
 go.sum                            |   6 +
 internal/infra/llm/openai.go      | 201 ++++++++++++++++++++++++++++++
 internal/infra/llm/openai_test.go |  96 ++++++++++++++
 4 files changed, 306 insertions(+)
 create mode 100644 internal/infra/llm/openai.go
 create mode 100644 internal/infra/llm/openai_test.go

diff --git a/go.mod b/go.mod
index 04c18d1..8669704 100644
--- a/go.mod
+++ b/go.mod
@@ -30,6 +30,7 @@ require (
 	github.com/cpuguy83/dockercfg v0.3.2 // indirect
 	github.com/davecgh/go-spew v1.1.2-0.20180830191138-d8f796af33cc // indirect
 	github.com/distribution/reference v0.6.0 // indirect
+	github.com/dlclark/regexp2 v1.10.0 // indirect
 	github.com/docker/go-connections v0.6.0 // indirect
 	github.com/docker/go-units v0.5.0 // indirect
 	github.com/ebitengine/purego v0.10.0 // indirect
@@ -58,9 +59,11 @@ require (
 	github.com/moby/sys/user v0.4.0 // indirect
 	github.com/moby/sys/userns v0.1.0 // indirect
 	github.com/moby/term v0.5.2 // indirect
+	github.com/openai/openai-go v1.12.0 // indirect
 	github.com/opencontainers/go-digest v1.0.0 // indirect
 	github.com/opencontainers/image-spec v1.1.1 // indirect
 	github.com/pelletier/go-toml/v2 v2.2.4 // indirect
+	github.com/pkoukk/tiktoken-go v0.1.8 // indirect
 	github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 // indirect
 	github.com/power-devops/perfstat v0.0.0-20240221224432-82ca36839d55 // indirect
 	github.com/sagikazarmark/locafero v0.11.0 // indirect
diff --git a/go.sum b/go.sum
index da4326a..dfc6d97 100644
--- a/go.sum
+++ b/go.sum
@@ -37,6 +37,8 @@ github.com/dhui/dktest v0.4.6 h1:+DPKyScKSEp3VLtbMDHcUq6V5Lm5zfZZVb0Sk7Ahom4=
 github.com/dhui/dktest v0.4.6/go.mod h1:JHTSYDtKkvFNFHJKqCzVzqXecyv+tKt8EzceOmQOgbU=
 github.com/distribution/reference v0.6.0 h1:0IXCQ5g4/QMHHkarYzh5l+u8T3t73zM5QvfrDyIgxBk=
 github.com/distribution/reference v0.6.0/go.mod h1:BbU0aIcezP1/5jX/8MP0YiH4SdvB5Y4f/wlDRiLyi3E=
+github.com/dlclark/regexp2 v1.10.0 h1:+/GIL799phkJqYW+3YbOd8LCcbHzT0Pbo8zl70MHsq0=
+github.com/dlclark/regexp2 v1.10.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/docker/docker v28.3.3+incompatible h1:Dypm25kh4rmk49v1eiVbsAtpAsYURjYkaKubwuBdxEI=
 github.com/docker/docker v28.3.3+incompatible/go.mod h1:eEKB0N0r5NX/I1kEveEz05bcu8tLC/8azJZsviup8Sk=
 github.com/docker/go-connections v0.6.0 h1:LlMG9azAe1TqfR7sO+NJttz1gy6KO7VJBh+pMmjSD94=
@@ -122,6 +124,8 @@ github.com/morikuni/aec v1.0.0 h1:nP9CBfwrvYnBRgY6qfDQkygYDmYwOilePFkwzv4dU8A=
 github.com/morikuni/aec v1.0.0/go.mod h1:BbKIizmSmc5MMPqRYbxO4ZU0S0+P200+tUnFx7PXmsc=
 github.com/oklog/ulid/v2 v2.1.1 h1:suPZ4ARWLOJLegGFiZZ1dFAkqzhMjL3J1TzI+5wHz8s=
 github.com/oklog/ulid/v2 v2.1.1/go.mod h1:rcEKHmBBKfef9DhnvX7y1HZBYxjXb0cP5ExxNsTT1QQ=
+github.com/openai/openai-go v1.12.0 h1:NBQCnXzqOTv5wsgNC36PrFEiskGfO5wccfCWDo9S1U0=
+github.com/openai/openai-go v1.12.0/go.mod h1:g461MYGXEXBVdV5SaR/5tNzNbSfwTBBefwc+LlDCK0Y=
 github.com/opencontainers/go-digest v1.0.0 h1:apOUWs51W5PlhuyGyz9FCeeBIOUDA/6nW8Oi/yOhh5U=
 github.com/opencontainers/go-digest v1.0.0/go.mod h1:0JzlMkj0TRzQZfJkVvzbP0HBR3IKzErnv2BNG4W4MAM=
 github.com/opencontainers/image-spec v1.1.1 h1:y0fUlFfIZhPF1W537XOLg0/fcx6zcHCJwooC2xJA040=
@@ -131,6 +135,8 @@ github.com/pelletier/go-toml/v2 v2.2.4 h1:mye9XuhQ6gvn5h28+VilKrrPoQVanw5PMw/TB0
 github.com/pelletier/go-toml/v2 v2.2.4/go.mod h1:2gIqNv+qfxSVS7cM2xJQKtLSTLUE9V8t9Stt+h56mCY=
 github.com/pkg/errors v0.9.1 h1:FEBLx1zS214owpjy7qsBeixbURkuhQAwrK5UwLGTwt4=
 github.com/pkg/errors v0.9.1/go.mod h1:bwawxfHBFNV+L2hUp1rHADufV3IMtnDRdf1r5NINEl0=
+github.com/pkoukk/tiktoken-go v0.1.8 h1:85ENo+3FpWgAACBaEUVp+lctuTcYUO7BtmfhlN/QTRo=
+github.com/pkoukk/tiktoken-go v0.1.8/go.mod h1:9NiV+i9mJKGj1rYOT+njbv+ZwA/zJxYdewGl6qVatpg=
 github.com/pmezard/go-difflib v1.0.0/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
 github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2 h1:Jamvg5psRIccs7FGNTlIRMkT8wgtp5eCXdBlqhYGL6U=
 github.com/pmezard/go-difflib v1.0.1-0.20181226105442-5d4384ee4fb2/go.mod h1:iKH77koFhYxTK1pcRnkKkqfTogsbg7gZNVY4sRDYZ/4=
diff --git a/internal/infra/llm/openai.go b/internal/infra/llm/openai.go
new file mode 100644
index 0000000..820c6e9
--- /dev/null
+++ b/internal/infra/llm/openai.go
@@ -0,0 +1,201 @@
+package llm
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+
+	"github.com/openai/openai-go"
+	"github.com/openai/openai-go/option"
+	"github.com/openai/openai-go/packages/param"
+	tiktoken "github.com/pkoukk/tiktoken-go"
+)
+
+// Compile-time assertion: OpenAIClient must satisfy LLMClient.
+var _ LLMClient = (*OpenAIClient)(nil)
+
+// OpenAIClient 通过官方 SDK 调用 Chat Completions API。
+// baseURL 改写后兼容 DeepSeek / Qwen / Ollama / vLLM / Moonshot 等 OpenAI-compatible 端点。
+// 流式响应中识别 reasoning_content 扩展字段以支持 DeepSeek / Qwen 的 thinking 输出。
+type OpenAIClient struct {
+	sdk openai.Client // SDK v1.12: NewClient returns value, not pointer
+}
+
+// NewOpenAI 构造客户端。baseURL 可空（用 SDK 默认）。
+func NewOpenAI(apiKey, baseURL string) (*OpenAIClient, error) {
+	opts := []option.RequestOption{option.WithAPIKey(apiKey)}
+	if baseURL != "" {
+		opts = append(opts, option.WithBaseURL(baseURL))
+	}
+	c := openai.NewClient(opts...)
+	return &OpenAIClient{sdk: c}, nil
+}
+
+func (c *OpenAIClient) Provider() string { return "openai" }
+
+// toOpenAIMessages 把中立 Message 切片转换为 SDK 类型。
+// document/PDF 块由上层能力检查拦截，此处不翻译（OpenAI Chat Completions 不原生支持 PDF）。
+// image 块编码为 base64 data URI 传入。
+func toOpenAIMessages(msgs []Message) []openai.ChatCompletionMessageParamUnion {
+	out := make([]openai.ChatCompletionMessageParamUnion, 0, len(msgs))
+	for _, m := range msgs {
+		switch m.Role {
+		case RoleUser:
+			parts := make([]openai.ChatCompletionContentPartUnionParam, 0, len(m.Blocks))
+			for _, b := range m.Blocks {
+				switch b.Type {
+				case "text":
+					parts = append(parts, openai.TextContentPart(b.Text))
+				case "image":
+					// SDK v1.12: ImageContentPart 接受 ChatCompletionContentPartImageImageURLParam
+					encoded := base64.StdEncoding.EncodeToString(b.Data)
+					dataURI := fmt.Sprintf("data:%s;base64,%s", b.MimeType, encoded)
+					parts = append(parts, openai.ImageContentPart(openai.ChatCompletionContentPartImageImageURLParam{
+						URL: dataURI,
+					}))
+				// document/PDF 不翻译，由上层能力检查拦截
+				}
+			}
+			out = append(out, openai.UserMessage(parts))
+		case RoleAssistant:
+			// 助手消息只取第一个 text block（assistant 输出 PDF 场景不存在）
+			text := ""
+			for _, b := range m.Blocks {
+				if b.Type == "text" {
+					text = b.Text
+					break
+				}
+			}
+			out = append(out, openai.AssistantMessage(text))
+		}
+	}
+	return out
+}
+
+// Stream 启动流式请求，将 SDK chunk 转换为中立 StreamEvent。
+// 使用 ctx-aware send 辅助函数防止下游 reader 退出后 goroutine 永久阻塞（同 anthropic.go 模式）。
+func (c *OpenAIClient) Stream(ctx context.Context, modelID string, req Request) (<-chan StreamEvent, error) {
+	msgs := make([]openai.ChatCompletionMessageParamUnion, 0)
+	if req.SystemPrompt != "" {
+		msgs = append(msgs, openai.SystemMessage(req.SystemPrompt))
+	}
+	msgs = append(msgs, toOpenAIMessages(req.Messages)...)
+
+	params := openai.ChatCompletionNewParams{
+		Model:    openai.ChatModel(modelID),
+		Messages: msgs,
+		// SDK v1.12: MaxCompletionTokens 替代已弃用的 MaxTokens
+		StreamOptions: openai.ChatCompletionStreamOptionsParam{
+			IncludeUsage: param.NewOpt(true),
+		},
+	}
+	if req.MaxOutputTokens > 0 {
+		params.MaxCompletionTokens = param.NewOpt(int64(req.MaxOutputTokens))
+	}
+
+	stream := c.sdk.Chat.Completions.NewStreaming(ctx, params)
+	out := make(chan StreamEvent, 16)
+
+	// send 在 ctx 取消时立即放弃发送，防止下游 reader 已退出导致 goroutine 永久阻塞。
+	send := func(ev StreamEvent) bool {
+		select {
+		case <-ctx.Done():
+			return false
+		case out <- ev:
+			return true
+		}
+	}
+
+	go func() {
+		defer close(out)
+		var usageSent bool
+		for stream.Next() {
+			chunk := stream.Current()
+
+			// 处理 usage（最后一个含 usage 的 chunk，choices 为空）
+			if chunk.JSON.Usage.Valid() && chunk.Usage.PromptTokens > 0 {
+				if !send(StreamEvent{Type: EventUsage, Usage: &Usage{
+					PromptTokens:     int(chunk.Usage.PromptTokens),
+					CompletionTokens: int(chunk.Usage.CompletionTokens),
+				}}) {
+					return
+				}
+				usageSent = true
+			}
+
+			for _, choice := range chunk.Choices {
+				delta := choice.Delta
+
+				// reasoning_content 扩展字段（DeepSeek-Reasoner / Qwen-Thinking）
+				// SDK v1.12: 非标准字段通过 delta.JSON.ExtraFields["reasoning_content"] 访问。
+				// ExtraFields 中的条目 status=invalid（不可 Valid()），但 Raw() 非空时表示字段存在。
+				// Raw() 返回带引号的原始 JSON 字符串，需 json.Unmarshal 解析为 Go string。
+				if rcField, ok := delta.JSON.ExtraFields["reasoning_content"]; ok && rcField.Raw() != "" {
+					var rcText string
+					if err := json.Unmarshal([]byte(rcField.Raw()), &rcText); err == nil && rcText != "" {
+						if !send(StreamEvent{Type: EventThinkingDelta, Text: rcText}) {
+							return
+						}
+					}
+				}
+
+				// 普通文本 delta
+				if delta.Content != "" {
+					if !send(StreamEvent{Type: EventTextDelta, Text: delta.Content}) {
+						return
+					}
+				}
+
+				// 流结束
+				if choice.FinishReason != "" {
+					// usage 可能在同一 chunk 的 choices 之后，也可能在独立 chunk；
+					// 此处仅在尚未发送 usage 时检查本 chunk 的 usage
+					if !usageSent && chunk.JSON.Usage.Valid() && chunk.Usage.PromptTokens > 0 {
+						if !send(StreamEvent{Type: EventUsage, Usage: &Usage{
+							PromptTokens:     int(chunk.Usage.PromptTokens),
+							CompletionTokens: int(chunk.Usage.CompletionTokens),
+						}}) {
+							return
+						}
+						usageSent = true
+					}
+					if !send(StreamEvent{Type: EventDone, Reason: choice.FinishReason}) {
+						return
+					}
+				}
+			}
+		}
+		if err := stream.Err(); err != nil {
+			send(StreamEvent{Type: EventError, Err: fmt.Errorf("openai stream: %w", err)})
+		}
+	}()
+
+	return out, nil
+}
+
+// CountTokens 使用本地 tiktoken-go 统计 token 数，不发起远程 API 调用。
+// 当模型不可识别时，自动回退到 cl100k_base 编码。
+func (c *OpenAIClient) CountTokens(_ context.Context, modelID string, req Request) (int, error) {
+	enc, err := tiktoken.EncodingForModel(modelID)
+	if err != nil {
+		// SDK v1.12 / tiktoken-go v0.1.x: 未知模型时回退到 cl100k_base
+		enc, err = tiktoken.GetEncoding("cl100k_base")
+		if err != nil {
+			return 0, fmt.Errorf("tiktoken fallback cl100k_base: %w", err)
+		}
+	}
+
+	total := 0
+	if req.SystemPrompt != "" {
+		total += len(enc.Encode(req.SystemPrompt, nil, nil))
+	}
+	for _, m := range req.Messages {
+		for _, b := range m.Blocks {
+			if b.Type == "text" {
+				total += len(enc.Encode(b.Text, nil, nil))
+			}
+		}
+	}
+	return total, nil
+}
diff --git a/internal/infra/llm/openai_test.go b/internal/infra/llm/openai_test.go
new file mode 100644
index 0000000..daa9e93
--- /dev/null
+++ b/internal/infra/llm/openai_test.go
@@ -0,0 +1,96 @@
+package llm
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/stretchr/testify/require"
+)
+
+const openaiStreamFixture = `data: {"id":"c1","object":"chat.completion.chunk","model":"gpt-4o","choices":[{"index":0,"delta":{"role":"assistant","content":"Hello"},"finish_reason":null}]}
+
+data: {"id":"c1","object":"chat.completion.chunk","model":"gpt-4o","choices":[{"index":0,"delta":{"content":" world"},"finish_reason":null}]}
+
+data: {"id":"c1","object":"chat.completion.chunk","model":"gpt-4o","choices":[{"index":0,"delta":{},"finish_reason":"stop"}],"usage":{"prompt_tokens":10,"completion_tokens":2,"total_tokens":12}}
+
+data: [DONE]
+
+`
+
+const openaiReasoningFixture = `data: {"id":"c1","object":"chat.completion.chunk","model":"deepseek-reasoner","choices":[{"index":0,"delta":{"reasoning_content":"Let me think"},"finish_reason":null}]}
+
+data: {"id":"c1","object":"chat.completion.chunk","model":"deepseek-reasoner","choices":[{"index":0,"delta":{"content":"Done"},"finish_reason":"stop"}],"usage":{"prompt_tokens":5,"completion_tokens":1}}
+
+data: [DONE]
+
+`
+
+func TestOpenAI_Stream_TextDelta(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte(openaiStreamFixture))
+	}))
+	defer srv.Close()
+
+	c, _ := NewOpenAI("k", srv.URL)
+	ch, err := c.Stream(context.Background(), "gpt-4o", Request{
+		Messages:        []Message{{Role: RoleUser, Blocks: []ContentBlock{{Type: "text", Text: "hi"}}}},
+		MaxOutputTokens: 100,
+	})
+	require.NoError(t, err)
+
+	var deltas []string
+	var sawDone, sawUsage bool
+	for ev := range ch {
+		switch ev.Type {
+		case EventTextDelta:
+			deltas = append(deltas, ev.Text)
+		case EventUsage:
+			sawUsage = true
+			require.Equal(t, 10, ev.Usage.PromptTokens)
+			require.Equal(t, 2, ev.Usage.CompletionTokens)
+		case EventDone:
+			sawDone = true
+		}
+	}
+	require.Equal(t, "Hello world", strings.Join(deltas, ""))
+	require.True(t, sawUsage)
+	require.True(t, sawDone)
+}
+
+func TestOpenAI_Stream_ReasoningContent_Compatible(t *testing.T) {
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "text/event-stream")
+		_, _ = w.Write([]byte(openaiReasoningFixture))
+	}))
+	defer srv.Close()
+
+	c, _ := NewOpenAI("k", srv.URL)
+	ch, _ := c.Stream(context.Background(), "deepseek-reasoner", Request{
+		Messages: []Message{{Role: RoleUser, Blocks: []ContentBlock{{Type: "text", Text: "hi"}}}},
+	})
+	var thinking, text string
+	for ev := range ch {
+		switch ev.Type {
+		case EventThinkingDelta:
+			thinking += ev.Text
+		case EventTextDelta:
+			text += ev.Text
+		}
+	}
+	require.Equal(t, "Let me think", thinking)
+	require.Equal(t, "Done", text)
+}
+
+func TestOpenAI_CountTokens_LocalTiktoken(t *testing.T) {
+	c, _ := NewOpenAI("k", "")
+	got, err := c.CountTokens(context.Background(), "gpt-4o", Request{
+		SystemPrompt: "You are helpful.",
+		Messages:     []Message{{Role: RoleUser, Blocks: []ContentBlock{{Type: "text", Text: "hello world"}}}},
+	})
+	require.NoError(t, err)
+	require.Greater(t, got, 0)
+}