| 1 | import fs from 'fs' |
| 2 | import path from 'path' |
| 3 | import log from 'electron-log/main.js' |
| 4 | import { nanoid } from 'nanoid' |
| 5 | import { FilesystemBackend, createDeepAgent } from 'deepagents' |
| 6 | import { resolveModelTimeoutMs } from '@shared/model-timeout' |
| 7 | import { extractJsonBlock, extractModelText, resolveModel } from '../../agent-runtime/model' |
| 8 | import type { ModelRuntimeConfig } from '../../agent-runtime/model' |
| 9 | import { importPptxToEditableHtml } from '../../io/pptx-import' |
| 10 | import { buildStylePptxImportPrompt } from '../../agent-runtime/prompt' |
| 11 | import { logAgentToolEvents } from '../../utils/agent-tool-logger' |
| 12 | import type { StyleParseResult } from './file' |
| 13 | |
| 14 | const MAX_PPTX_SIZE_MB = 500 |
| 15 | const MAX_PPTX_SIZE = MAX_PPTX_SIZE_MB * 1024 * 1024 |
| 16 | const MAX_IMPORT_PAGES = 40 |
| 17 | |
| 18 | export async function parseStylePptx(args: { |
| 19 | filePath: string |
| 20 | provider: string |
| 21 | apiKey: string |
| 22 | model: string |
| 23 | baseUrl: string |
| 24 | maxTokens?: number |
| 25 | modelRuntime?: ModelRuntimeConfig |
| 26 | modelTimeoutMs: number |
| 27 | tmpRootDir: string |
| 28 | }): Promise<StyleParseResult> { |
| 29 | const sourcePath = path.resolve(args.filePath) |
| 30 | const ext = path.extname(sourcePath).toLowerCase() |
| 31 | if (ext !== '.pptx') throw new Error('不支持的文件格式，仅支持 .pptx') |
| 32 | const stat = await fs.promises.stat(sourcePath) |
| 33 | if (!stat.isFile()) throw new Error(`路径不是文件：${sourcePath}`) |
| 34 | if (stat.size > MAX_PPTX_SIZE) { |
| 35 | throw new Error( |
| 36 | `文件过大（${(stat.size / 1024 / 1024).toFixed(1)}MB），PPTX 上限 ${MAX_PPTX_SIZE_MB}MB` |
| 37 | ) |
| 38 | } |
| 39 | |
| 40 | await fs.promises.mkdir(args.tmpRootDir, { recursive: true }) |
| 41 | const taskDir = path.join(args.tmpRootDir, `${Date.now()}-${nanoid(8)}`) |
| 42 | await fs.promises.mkdir(taskDir, { recursive: true }) |
| 43 | |
| 44 | try { |
| 45 | const imported = await importPptxToEditableHtml({ |
| 46 | filePath: sourcePath, |
| 47 | projectDir: taskDir, |
| 48 | title: path.basename(sourcePath, path.extname(sourcePath)), |
| 49 | maxPages: MAX_IMPORT_PAGES |
| 50 | }) |
| 51 | |
| 52 | const samplePages = selectSamplePagePaths( |
| 53 | imported.pages.map((page) => `/${path.basename(page.htmlPath)}`) |
| 54 | ) |
| 55 | const response = await runStylePptxImportAgent({ |
| 56 | provider: args.provider, |
| 57 | apiKey: args.apiKey, |
| 58 | model: args.model, |
| 59 | baseUrl: args.baseUrl, |
| 60 | maxTokens: args.maxTokens, |
| 61 | modelRuntime: args.modelRuntime, |
| 62 | modelTimeoutMs: args.modelTimeoutMs, |
| 63 | workspaceDir: taskDir, |
| 64 | prompt: buildStylePptxImportPrompt({ |
| 65 | deckRootPath: '/', |
| 66 | indexPath: '/index.html', |
| 67 | samplePagePaths: samplePages |
| 68 | }) |
| 69 | }) |
| 70 | |
| 71 | try { |
| 72 | return parseStyleImportResponse(response) |
| 73 | } catch (parseError) { |
| 74 | const reason = parseError instanceof Error ? parseError.message : String(parseError) |
| 75 | log.info('[styles:parsePptx] first parse failed, retrying with fix prompt', { reason }) |
| 76 | |
| 77 | const fixedResponse = await retryFixJson({ |
| 78 | provider: args.provider, |
| 79 | apiKey: args.apiKey, |
| 80 | model: args.model, |
| 81 | baseUrl: args.baseUrl, |
| 82 | maxTokens: args.maxTokens, |
| 83 | modelRuntime: args.modelRuntime, |
| 84 | modelTimeoutMs: args.modelTimeoutMs, |
| 85 | brokenResponse: response, |
| 86 | parseError: reason |
| 87 | }) |
| 88 | return parseStyleImportResponse(fixedResponse) |
| 89 | } |
| 90 | } finally { |
| 91 | await fs.promises.rm(taskDir, { recursive: true, force: true }).catch((error) => { |
| 92 | log.warn('[styles:parsePptx] cleanup failed', { |
| 93 | taskDir, |
| 94 | message: error instanceof Error ? error.message : String(error) |
| 95 | }) |
| 96 | }) |
| 97 | } |
| 98 | } |
| 99 | |
| 100 | export function selectSamplePagePaths(pagePaths: string[]): string[] { |
| 101 | if (pagePaths.length <= 4) return pagePaths |
| 102 | const sorted = [...pagePaths].sort() |
| 103 | const first = sorted[0] |
| 104 | const last = sorted[sorted.length - 1] |
| 105 | const middle = sorted.slice(1, -1) |
| 106 | if (middle.length === 0) return [first, last] |
| 107 | |
| 108 | // 等距抽样中间页：将中间页均分为 N 个桶，每桶取一页，保证从头到尾均匀覆盖，最多 20 页 |
| 109 | const maxSamples = Math.min( |
| 110 | middle.length, |
| 111 | Math.max(2, Math.min(20, Math.ceil(Math.sqrt(pagePaths.length)))) |
| 112 | ) |
| 113 | const sampled: string[] = [] |
| 114 | for (let i = 0; i < maxSamples; i++) { |
| 115 | const idx = Math.floor((i + 0.5) * middle.length / maxSamples) |
| 116 | sampled.push(middle[idx]) |
| 117 | } |
| 118 | |
| 119 | return [first, ...sampled, last] |
| 120 | } |
| 121 | |
| 122 | async function runStylePptxImportAgent(args: { |
| 123 | provider: string |
| 124 | apiKey: string |
| 125 | model: string |
| 126 | baseUrl: string |
| 127 | maxTokens?: number |
| 128 | modelRuntime?: ModelRuntimeConfig |
| 129 | modelTimeoutMs: number |
| 130 | workspaceDir: string |
| 131 | prompt: string |
| 132 | }): Promise<string> { |
| 133 | const model = resolveModel( |
| 134 | args.provider, |
| 135 | args.apiKey, |
| 136 | args.model, |
| 137 | args.baseUrl, |
| 138 | 0.2, |
| 139 | args.maxTokens, |
| 140 | args.modelRuntime |
| 141 | ) |
| 142 | const agent = createDeepAgent({ |
| 143 | model, |
| 144 | backend: new FilesystemBackend({ |
| 145 | rootDir: args.workspaceDir, |
| 146 | virtualMode: true |
| 147 | }), |
| 148 | systemPrompt: |
| 149 | 'You are a style-import parsing agent for PPTX-derived HTML files. You must use grep and read_file tools before generating the result. Return strict JSON only: label, description, category, aliases, styleCase, styleSkill.' |
| 150 | }) |
| 151 | |
| 152 | const stream = await agent.stream( |
| 153 | { |
| 154 | messages: [ |
| 155 | { |
| 156 | role: 'user', |
| 157 | content: args.prompt |
| 158 | } |
| 159 | ] |
| 160 | }, |
| 161 | { |
| 162 | streamMode: ['updates', 'messages'], |
| 163 | subgraphs: true, |
| 164 | signal: AbortSignal.timeout(resolveModelTimeoutMs(args.modelTimeoutMs, 'document')) |
| 165 | } |
| 166 | ) |
| 167 | |
| 168 | let messageBuffer = '' |
| 169 | let latestAssistantStateText = '' |
| 170 | const seenToolEvents = new Set<string>() |
| 171 | for await (const chunk of stream as AsyncIterable<unknown>) { |
| 172 | if (!Array.isArray(chunk) || chunk.length < 3) continue |
| 173 | const mode = chunk[1] |
| 174 | const data = chunk[2] |
| 175 | |
| 176 | if (mode === 'updates') { |
| 177 | logAgentToolEvents(data, seenToolEvents, { tag: 'styles:parsePptx', source: 'updates' }) |
| 178 | const assistantTexts = extractAssistantTextsFromState(data) |
| 179 | const longestText = assistantTexts.sort((a, b) => b.length - a.length)[0] || '' |
| 180 | if (longestText.length >= latestAssistantStateText.length) { |
| 181 | latestAssistantStateText = longestText |
| 182 | } |
| 183 | continue |
| 184 | } |
| 185 | |
| 186 | if (mode !== 'messages' || !Array.isArray(data)) continue |
| 187 | logAgentToolEvents(data, seenToolEvents, { tag: 'styles:parsePptx', source: 'messages' }) |
| 188 | for (const message of data as Array<Record<string, unknown>>) { |
| 189 | const content = extractModelText(message).trim() |
| 190 | if (content) messageBuffer += content |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | return latestAssistantStateText.length > messageBuffer.length ? latestAssistantStateText : messageBuffer |
| 195 | } |
| 196 | |
| 197 | export function parseStyleImportResponse(response: unknown): StyleParseResult { |
| 198 | const text = extractModelText(response) || (typeof response === 'string' ? response : JSON.stringify(response)) |
| 199 | const jsonText = extractJsonBlock(text).trim() |
| 200 | if (!jsonText) throw new Error('LLM 返回格式异常：未找到 JSON') |
| 201 | |
| 202 | let parsed: Record<string, unknown> |
| 203 | try { |
| 204 | parsed = JSON.parse(jsonText) |
| 205 | } catch (parseError) { |
| 206 | const hint = jsonText.length > 200 ? `${jsonText.slice(0, 200)}...` : jsonText |
| 207 | const reason = parseError instanceof Error ? parseError.message : String(parseError) |
| 208 | log.warn('[styles:parsePptx] JSON parse failed', { reason, jsonPreview: hint }) |
| 209 | throw new Error(`LLM 返回的 JSON 格式异常：${reason}`) |
| 210 | } |
| 211 | |
| 212 | const label = String(parsed.label || '').trim() |
| 213 | const styleSkill = String(parsed.styleSkill || '').trim() |
| 214 | if (!label || !styleSkill) { |
| 215 | throw new Error('LLM 返回缺少必填字段（label / styleSkill）') |
| 216 | } |
| 217 | |
| 218 | return { |
| 219 | label, |
| 220 | description: String(parsed.description || '').trim(), |
| 221 | category: String(parsed.category || '自定义').trim(), |
| 222 | aliases: Array.isArray(parsed.aliases) |
| 223 | ? parsed.aliases.map((item) => String(item || '').trim()).filter((item) => item.length > 0) |
| 224 | : [], |
| 225 | styleSkill, |
| 226 | styleCase: String(parsed.styleCase || '').trim() |
| 227 | } |
| 228 | } |
| 229 | |
| 230 | export async function retryFixJson(args: { |
| 231 | provider: string |
| 232 | apiKey: string |
| 233 | model: string |
| 234 | baseUrl: string |
| 235 | maxTokens?: number |
| 236 | modelRuntime?: ModelRuntimeConfig |
| 237 | modelTimeoutMs: number |
| 238 | brokenResponse: string |
| 239 | parseError: string |
| 240 | }): Promise<string> { |
| 241 | const model = resolveModel( |
| 242 | args.provider, |
| 243 | args.apiKey, |
| 244 | args.model, |
| 245 | args.baseUrl, |
| 246 | 0.2, |
| 247 | args.maxTokens, |
| 248 | args.modelRuntime |
| 249 | ) |
| 250 | const result = await model.invoke([ |
| 251 | { |
| 252 | role: 'user', |
| 253 | content: `你上次输出的 JSON 格式有误，解析报错：${args.parseError} |
| 254 | |
| 255 | 请修复 JSON 格式并重新输出完整的 JSON（用 \`\`\`json ... \`\`\` 包裹）。不要修改内容，只修格式。 |
| 256 | |
| 257 | 原始输出： |
| 258 | ${args.brokenResponse}` |
| 259 | } |
| 260 | ], { |
| 261 | signal: AbortSignal.timeout(resolveModelTimeoutMs(args.modelTimeoutMs, 'document')) |
| 262 | }) |
| 263 | return extractModelText(result) |
| 264 | } |
| 265 | |
| 266 | export async function extractStyleFromExistingHtml(args: { |
| 267 | projectDir: string |
| 268 | pageHtmlPaths: string[] |
| 269 | sourceFilePath: string |
| 270 | provider: string |
| 271 | apiKey: string |
| 272 | model: string |
| 273 | baseUrl: string |
| 274 | maxTokens?: number |
| 275 | modelTimeoutMs: number |
| 276 | }): Promise<StyleParseResult> { |
| 277 | const samplePages = selectSamplePagePaths( |
| 278 | args.pageHtmlPaths.map((p) => `/${path.basename(p)}`) |
| 279 | ) |
| 280 | |
| 281 | const response = await runStylePptxImportAgent({ |
| 282 | provider: args.provider, |
| 283 | apiKey: args.apiKey, |
| 284 | model: args.model, |
| 285 | baseUrl: args.baseUrl, |
| 286 | maxTokens: args.maxTokens, |
| 287 | modelTimeoutMs: args.modelTimeoutMs, |
| 288 | workspaceDir: args.projectDir, |
| 289 | prompt: buildStylePptxImportPrompt({ |
| 290 | deckRootPath: '/', |
| 291 | indexPath: '/index.html', |
| 292 | samplePagePaths: samplePages |
| 293 | }) |
| 294 | }) |
| 295 | |
| 296 | try { |
| 297 | return parseStyleImportResponse(response) |
| 298 | } catch (parseError) { |
| 299 | const reason = parseError instanceof Error ? parseError.message : String(parseError) |
| 300 | log.info('[styles:extractFromHtml] first parse failed, retrying with fix prompt', { reason }) |
| 301 | |
| 302 | const fixedResponse = await retryFixJson({ |
| 303 | provider: args.provider, |
| 304 | apiKey: args.apiKey, |
| 305 | model: args.model, |
| 306 | baseUrl: args.baseUrl, |
| 307 | maxTokens: args.maxTokens, |
| 308 | modelTimeoutMs: args.modelTimeoutMs, |
| 309 | brokenResponse: response, |
| 310 | parseError: reason |
| 311 | }) |
| 312 | return parseStyleImportResponse(fixedResponse) |
| 313 | } |
| 314 | } |
| 315 | |
| 316 | function extractAssistantTextsFromState(data: unknown): string[] { |
| 317 | const texts: string[] = [] |
| 318 | const seen = new Set<object>() |
| 319 | const getObject = (value: unknown): Record<string, unknown> | null => |
| 320 | value && typeof value === 'object' && !Array.isArray(value) ? (value as Record<string, unknown>) : null |
| 321 | |
| 322 | const visit = (value: unknown): void => { |
| 323 | if (!value || typeof value !== 'object') return |
| 324 | if (seen.has(value as object)) return |
| 325 | seen.add(value as object) |
| 326 | |
| 327 | if (Array.isArray(value)) { |
| 328 | value.forEach(visit) |
| 329 | return |
| 330 | } |
| 331 | |
| 332 | const record = value as Record<string, unknown> |
| 333 | const role = String(record.role || '').toLowerCase() |
| 334 | const type = String(record.type || '').toLowerCase() |
| 335 | const constructorName = String( |
| 336 | getObject(record.lc_kwargs)?.type ?? getObject(record.kwargs)?.type ?? '' |
| 337 | ).toLowerCase() |
| 338 | const isAssistant = |
| 339 | role === 'assistant' || type === 'ai' || type === 'assistant' || constructorName === 'ai' |
| 340 | const isToolOrHuman = |
| 341 | role === 'tool' || |
| 342 | role === 'user' || |
| 343 | role === 'system' || |
| 344 | type === 'tool' || |
| 345 | type === 'human' || |
| 346 | type === 'system' |
| 347 | if (!isAssistant || isToolOrHuman) { |
| 348 | for (const nested of Object.values(record)) { |
| 349 | if (nested && typeof nested === 'object') visit(nested) |
| 350 | } |
| 351 | return |
| 352 | } |
| 353 | const text = extractModelText(record).trim() |
| 354 | if (text) texts.push(text) |
| 355 | |
| 356 | for (const nested of Object.values(record)) { |
| 357 | if (nested && typeof nested === 'object') visit(nested) |
| 358 | } |
| 359 | } |
| 360 | |
| 361 | visit(data) |
| 362 | return texts |
| 363 | } |
| 364 |