refactor: LLM 直接输出 Markdown，移除 JSON 解析层

服务端: - prompt 改为要求直接输出 Markdown，不再输出 JSON - GeneratePR 返回 string 而非 PRDescription 结构体 - 移除 JSON 解析、extractJSON、fixInlineCode 等逻辑 - ChatStream 的 content 事件直接携带 Markdown 片段流式传输前端: - generate() 仅处理 content + done + error 三个事件 - content 事件累积 Markdown 文本，实时渲染到两栏 - 移除 parsePartialJSON、buildMarkdown 等中间层 - 代码量从 ~100 行减至 ~50 行
2026-06-21 16:02:05 +08:00
parent 779bd9e332
commit 496e3ba71e
3 changed files with 39 additions and 152 deletions
@@ -75,19 +75,15 @@ func (h *GenerateHandler) Generate(c *gin.Context) {
 	// Update last_used
 	h.db.Exec(`UPDATE repositories SET last_used = NOW() WHERE id = ?`, id)

-	// Generate PR description (pass user ID for per-user LLM config)
-	pr, err := services.GeneratePR(h.db, localPath, req.Base, req.Head, user.ID, sendEvent)
+	// Generate PR description (streams markdown via content events)
+	markdown, err := services.GeneratePR(h.db, localPath, req.Base, req.Head, user.ID, sendEvent)
 	if err != nil {
 		sendEvent("error", map[string]interface{}{"message": err.Error()})
 		return
 	}

-	// Save analysis to DB with user_id
-	resultJSON, err := json.Marshal(pr)
-	if err != nil {
-		sendEvent("error", map[string]interface{}{"message": "序列化结果失败: " + err.Error()})
-		return
-	}
+	// Save analysis to DB
+	resultJSON, _ := json.Marshal(map[string]string{"markdown": markdown})
 	if _, err := h.db.Exec(`INSERT INTO analyses (user_id, repo_id, type, base_ref, head_ref, result) VALUES (?, ?, 'pr_description', ?, ?, ?)`,
 		user.ID, id, req.Base, req.Head, string(resultJSON)); err != nil {
 		sendEvent("error", map[string]interface{}{"message": "保存分析结果失败: " + err.Error()})
@@ -2,9 +2,7 @@ package services

 import (
 	"database/sql"
-	"encoding/json"
 	"fmt"
-	"regexp"
 	"strings"

 	"github.com/go-git/go-git/v5/plumbing"
@@ -12,34 +10,25 @@ import (
 	goopenai "github.com/sashabaranov/go-openai"
 )

-// PRDescription is the structured output from PR description generation.
-type PRDescription struct {
-	Title   string `json:"title"`
-	Type    string `json:"type"`
-	Summary string `json:"summary"`
-	Details string `json:"details"`
-	Impact  string `json:"impact"`
-}
-
-// GeneratePR generates a structured PR description from commit history and diff.
-// It streams progress via the callback and returns the parsed PR description.
-func GeneratePR(db *sql.DB, repoPath, base, head string, userID int64, callback StreamCallback) (*PRDescription, error) {
+// GeneratePR generates a PR description as Markdown from commit history and diff.
+// It streams the LLM output via callback ("content" events) and returns the full markdown.
+func GeneratePR(db *sql.DB, repoPath, base, head string, userID int64, callback StreamCallback) (string, error) {
 	// Read LLM config (per-user)
 	config, err := GetLLMConfig(db, userID)
 	if err != nil {
-		return nil, err
+		return "", err
 	}

 	// Open repo
 	repo, err := OpenRepo(repoPath)
 	if err != nil {
-		return nil, fmt.Errorf("open repo: %w", err)
+		return "", fmt.Errorf("open repo: %w", err)
 	}

 	// Get commits between base and head
 	commits, err := GetCommitLog(repo, head, 100)
 	if err != nil {
-		return nil, fmt.Errorf("get commits: %w", err)
+		return "", fmt.Errorf("get commits: %w", err)
 	}

 	// Filter commits to only those reachable from head but not from base
@@ -78,7 +67,7 @@ func GeneratePR(db *sql.DB, repoPath, base, head string, userID int64, callback
 	// Get diff
 	diff, err := GetDiff(repo, base, head)
 	if err != nil {
-		return nil, fmt.Errorf("get diff: %w", err)
+		return "", fmt.Errorf("get diff: %w", err)
 	}

 	// Build prompt
@@ -97,7 +86,7 @@ func GeneratePR(db *sql.DB, repoPath, base, head string, userID int64, callback
 		diff = diff[:60000] + "\n\n... [diff truncated due to size]"
 	}

-	prompt := fmt.Sprintf(`你是一个专业的技术文档撰写助手。根据以下 Git 变更信息，生成一份结构化的 PR 描述。
+	prompt := fmt.Sprintf(`你是一个专业的技术文档撰写助手。根据以下 Git 变更信息，生成一份 PR 描述。

 ## Commit 记录
 %s
@@ -105,65 +94,40 @@ func GeneratePR(db *sql.DB, repoPath, base, head string, userID int64, callback
 ## 代码变更 (Diff)
 %s

-请按以下 JSON 格式输出（直接输出 JSON，不要包含 markdown 代码块标记）：
-{
-  "title": "简洁的 PR 标题",
-  "type": "变更类型: feat|fix|refactor|docs|chore|style|test|perf",
-  "summary": "一段话概述变更内容",
-  "details": "详细的变更说明，按模块分组，使用 Markdown 格式",
-  "impact": "影响范围说明"
-}
+请直接输出 Markdown 格式的 PR 描述，包含以下部分：

-重要格式要求：
- details 字段中引用文件名、函数名、变量名等代码标识时，必须用反引号包裹，例如：`+"`services/generate.go`"+`, `+"`GeneratePR()`"+`
- 直接写出实际的代码名称，不要用任何占位符替代`, commitStr, diff)
+# 标题
+
+**类型**: feat|fix|refactor|docs|chore|style|test|perf
+
+## 概述
+一段话概述变更内容
+
+## 详细说明
+按模块分组的详细变更说明
+
+## 影响范围
+影响范围说明
+
+格式要求：
+- 引用文件名、函数名、变量名等代码标识时，必须用反引号包裹
+- 直接写出实际的代码名称，不要用任何占位符替代
+- 不要输出 JSON，直接输出 Markdown`, commitStr, diff)

 	messages := []goopenai.ChatCompletionMessage{
 		{Role: goopenai.ChatMessageRoleUser, Content: prompt},
 	}

-	// Call LLM with streaming
+	// Call LLM with streaming — content events carry markdown chunks in real-time
 	fullResponse, err := ChatStream(config, messages, callback)
 	if err != nil {
-		return nil, fmt.Errorf("LLM call: %w", err)
+		return "", fmt.Errorf("LLM call: %w", err)
 	}

-	// Parse JSON response
-	jsonStr := extractJSON(fullResponse)
-	var pr PRDescription
-	if err := json.Unmarshal([]byte(jsonStr), &pr); err != nil {
-		return nil, fmt.Errorf("parse PR description: %w (raw: %s)", err, truncateString(fullResponse, 200))
-	}
-
-	// Post-process: convert INLINECODE placeholders back to backtick-enclosed code
-	pr.Details = fixInlineCode(pr.Details)
-
-	// Send structured events
+	// Signal completion
 	if callback != nil {
-		callback("title", map[string]interface{}{"content": pr.Title})
-		callback("type", map[string]interface{}{"content": pr.Type})
-		callback("summary", map[string]interface{}{"content": pr.Summary})
-		callback("detail", map[string]interface{}{"content": pr.Details})
-		callback("impact", map[string]interface{}{"content": pr.Impact})
 		callback("done", map[string]interface{}{"content": ""})
 	}

-	return &pr, nil
-}
-
-// fixInlineCode converts LLM-generated INLINECODE placeholders back to backtick-enclosed inline code.
-// Some LLMs output INLINECODE0, INLINECODE1, etc. instead of `code` in JSON string values.
-// This wraps them in backticks so the frontend markdown renderer displays them as inline code.
-func fixInlineCode(text string) string {
-	re := regexp.MustCompile(`(?i)INLINECODE[_]?(\d+)`)
-	return re.ReplaceAllStringFunc(text, func(match string) string {
-		return "`" + match + "`"
-	})
-}
-
-func truncateString(s string, maxLen int) string {
-	if len(s) <= maxLen {
-		return s
-	}
-	return s[:maxLen] + "..."
+	return fullResponse, nil
 }
@@ -184,85 +184,27 @@
            indicator.classList.remove('hidden');

            // Clear previous results
-            document.getElementById('pr-rendered').innerHTML = '<span class="animate-pulse text-gray-400">正在生成...</span>';
+            document.getElementById('pr-rendered').innerHTML = '';
            document.getElementById('pr-markdown').textContent = '';
            markdownContent = '';

-            const fields = { title: '', type: '', summary: '', details: '', impact: '' };
-            let rawText = '';
-            let done = false;
-
-            function updateColumns() {
-                markdownContent = buildMarkdown(fields);
-                document.getElementById('pr-rendered').innerHTML = renderMarkdown(markdownContent);
-                document.getElementById('pr-markdown').textContent = markdownContent;
-            }
-
-            // Try to extract completed fields from partial JSON
-            function parsePartialJSON(text) {
-                const result = {};
-                for (const key of ['title', 'type', 'summary', 'details', 'impact']) {
-                    // Match "key": "value" where value's closing quote exists
-                    const re = new RegExp('"' + key + '"\\s*:\\s*"((?:[^"\\\\]|\\\\.)*)"');
-                    const m = text.match(re);
-                    if (m) result[key] = m[1].replace(/\\"/g, '"').replace(/\\n/g, '\n').replace(/\\\\/g, '\\');
-                }
-                return result;
-            }
-
+            // LLM outputs markdown directly — just accumulate and render
            SSE.post(`/api/repos/${repoId}/generate`, {
                base: baseRef,
                head: headRef,
            }, {
                content(data) {
-                    if (done) return;
-                    rawText += data.content || '';
-                    // Parse partial JSON to extract fields progressively
-                    const partial = parsePartialJSON(rawText);
-                    let changed = false;
-                    for (const key of ['title', 'type', 'summary', 'details', 'impact']) {
-                        if (partial[key] && partial[key] !== fields[key]) {
-                            fields[key] = partial[key];
-                            changed = true;
-                        }
-                    }
-                    if (changed) updateColumns();
-                },
-                title(data) {
-                    fields.title = data.content || fields.title;
-                    updateColumns();
-                    document.getElementById('streaming-status').textContent = '解析标题...';
-                },
-                type(data) {
-                    fields.type = data.content || fields.type;
-                    updateColumns();
-                    document.getElementById('streaming-status').textContent = '解析类型...';
-                },
-                summary(data) {
-                    fields.summary = data.content || fields.summary;
-                    updateColumns();
-                    document.getElementById('streaming-status').textContent = '解析摘要...';
-                },
-                detail(data) {
-                    fields.details = data.content || fields.details;
-                    updateColumns();
-                    document.getElementById('streaming-status').textContent = '解析详细说明...';
-                },
-                impact(data) {
-                    fields.impact = data.content || fields.impact;
-                    updateColumns();
-                    document.getElementById('streaming-status').textContent = '解析影响范围...';
+                    markdownContent += data.content || '';
+                    document.getElementById('pr-rendered').innerHTML = Markdown.render(markdownContent);
+                    document.getElementById('pr-markdown').textContent = markdownContent;
                },
                done() {
-                    done = true;
-                    updateColumns();
                    indicator.classList.add('hidden');
                    btn.disabled = false;
                    btn.textContent = '生成 PR 描述';
                    btn.classList.remove('opacity-70', 'cursor-not-allowed');
                },
                error(data) {
-                    done = true;
                    showToast('生成失败: ' + (data.message || '未知错误'), 'error');
                    indicator.classList.add('hidden');
                    btn.disabled = false;
@@ -272,21 +214,6 @@
            });
        }

-        // Markdown rendering (delegates to shared Markdown.render)
-        function renderMarkdown(text) {
-            return Markdown.render(text);
-        }
-
-        function buildMarkdown(f) {
-            const parts = [];
-            if (f.title) parts.push(`# ${f.title}`);
-            if (f.type) parts.push(`**类型**: ${f.type}`);
-            if (f.summary) parts.push(`## 摘要\n${f.summary}`);
-            if (f.details) parts.push(`## 详细说明\n${f.details}`);
-            if (f.impact) parts.push(`## 影响\n${f.impact}`);
-            return parts.join('\n\n');
-        }
-
        function copyMarkdown() {
            const markdown = markdownContent || document.getElementById('pr-markdown').textContent;
            if (!markdown.trim()) {