hack/tools/scripts/llm-key-hunter/results/_cache/content/alextangson_autocrew.json

{"repo": "alextangson/autocrew", "files": {"src/modules/learnings/style-distiller.test.ts": {"49ae2f8ffc27af44d6f91d69e477ec55f5c146c6": {"t": 1786068277.7035098, "n": 16703, "h": "eed110a01033be64", "c": "/**\n * style-distiller.test.ts — 全 mock，零网络\n *\n * 假 runLoopImpl 实际调用 submit_rules 工具的 execute 处理器，\n * 这是捕获机制的关键：execute 闭包把 rules 写进外部变量。\n */\nimport { describe, it, expect, beforeEach, afterEach } from \"vitest\";\nimport fs from \"node:fs/promises\";\nimport os from \"node:os\";\nimport path from \"node:path\";\n\nimport {\n  distillStyleRules,\n  analyzeStyleSamples,\n  shouldDistillStyle,\n  verifyVerbatimExcerpts,\n  fallbackExcerpts,\n} from \"./style-distiller.js\";\nimport { recordDiff, listDiffs } from \"./diff-tracker.js\";\nimport { loadProfile, addWritingRule } from \"../profile/creator-profile.js\";\nimport type { LoopResult, LoopTool, LoopOptions } from \"../../engine/loop.js\";\nimport type { EngineConfig } from \"../../engine/config.js\";\n\n// ─── Helpers ─────────────────────────────────────────────────────────────────\n\nlet testDir: string;\n\nconst ENV_KEYS = [\"DEEPSEEK_API_KEY\", \"DEEPSEEK_BASE_URL\"] as const;\nconst saved: Record<string, string | undefined> = {};\n\nbeforeEach(async () => {\n  testDir = await fs.mkdtemp(path.join(os.tmpdir(), \"autocrew-style-test-\"));\n  await fs.writeFile(\n    path.join(testDir, \"engine.json\"),\n    JSON.stringify({ apiKey: \"sk-test\", strongModel: \"m-strong\", fastModel: \"m-fast\" }),\n  );\n  for (const k of ENV_KEYS) {\n    saved[k] = process.env[k];\n    delete process.env[k];\n  }\n});\n\nafterEach(async () => {\n  await fs.rm(testDir, { recursive: true, force: true, maxRetries: 3, retryDelay: 50 });\n  for (const k of ENV_KEYS) {\n    if (saved[k] === undefined) delete process.env[k];\n    else process.env[k] = saved[k];\n  }\n});\n\n/**\n * Write a distill-state.json with the given ISO timestamp.\n * Pass \"\" to write an empty object (= no previous distillation).\n */\nasync function writeDistillState(lastDistilledAt: string): Promise<void> {\n  const dir = path.join(testDir, \"learnings\");\n  await fs.mkdir(dir, { recursive: true });\n  await fs.writeFile(\n    path.join(dir, \"distill-state.json\"),\n    JSON.stringify(lastDistilledAt ? { lastDistilledAt } : {}),\n  );\n}\n\nasync function readDistillState(): Promise<{ lastDistilledAt?: string }> {\n  try {\n    const raw = await fs.readFile(path.join(testDir, \"learnings\", \"distill-state.json\"), \"utf-8\");\n    return JSON.parse(raw);\n  } catch {\n    return {};\n  }\n}\n\n/**\n * Build a fake runLoopImpl that calls submit_rules with the given rule payloads.\n * Each execute return value is pushed into execResults.\n */\nfunction makeRunLoop(\n  ruleBatches: Array<Array<{ rule: string; evidence: string; confidence: number; scope?: string }>>,\n  execResults: string[] = [],\n): (_cfg: EngineConfig, opts: LoopOptions) => Promise<LoopResult> {\n  return async (_cfg, opts) => {\n    const submitTool = (opts.tools ?? []).find((t: LoopTool) => t.name === \"submit_rules\");\n    if (!submitTool) throw new Error(\"submit_rules tool not found in opts\");\n\n    for (const rules of ruleBatches) {\n      execResults.push(await submitTool.execute({ rules }));\n    }\n\n    return {\n      finalMessage: \"规则已提交\",\n      turns: ruleBatches.length + 1,\n      totalTokens: 100,\n      toolCallCount: ruleBatches.length,\n      stopReason: \"no_tool_calls\",\n    } satisfies LoopResult;\n  };\n}\n\n/**\n * Variant that also exposes callCount.\n */\nfunction makeRunLoopWithCount(\n  ruleBatches: Array<Array<{ rule: string; evidence: string; confidence: number }>>,\n): {\n  impl: (_cfg: EngineConfig, opts: LoopOptions) => Promise<LoopResult>;\n  getCallCount: () => number;\n} {\n  let callCount = 0;\n  const impl = async (_cfg: EngineConfig, opts: LoopOptions): Promise<LoopResult> => {\n    callCount++;\n    const submitTool = (opts.tools ?? []).find((t: LoopTool) => t.name === \"submit_rules\");\n    if (!submitTool) throw new Error(\"submit_rules tool not found in opts\");\n\n    for (const rules of ruleBatches) {\n      await submitTool.execute({ rules });\n    }\n\n    return {\n      finalMessage: \"规则已提交\",\n      turns: ruleBatches.length + 1,\n      totalTokens: 100,\n      toolCallCount: ruleBatches.length,\n      stopReason: \"no_tool_calls\",\n    } satisfies LoopResult;\n  };\n  return { impl, getCallCount: () => callCount };\n}\n\nconst GOOD_RULES = [\n  { rule: \"用口语化短句，避免书面长句\", evidence: \"before: 鉴于...因此 after: 说白了...\", confidence: 0.9 },\n  { rule: \"段落不超过三行\", evidence: \"before: 5行段落 after: 2行段落\", confidence: 0.8 },\n];\n\n// ─── Tests ────────────────────────────────────────────────────────────────────\n\ndescribe(\"distillStyleRules\", () => {\n  // 1. Happy path\n  it(\"happy path: 2 new diffs + 2 rules → profile updated, state written, summary correct\", async () => {\n    // Record 2 diffs\n    await recordDiff(\"c1\", \"body\", \"鉴于某某原因，因此可以考虑\", \"说白了，直接做就行\", testDir);\n    await recordDiff(\"c1\", \"body\", \"首先其次最后罗列\", \"三点建议：一 二 三\", testDir);\n\n    const runLoopImpl = makeRunLoop([GOOD_RULES]);\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(result.newRules).toHaveLength(2);\n    expect(result.diffsAnalyzed).toBe(2);\n    expect(result.skippedDuplicates).toBe(0);\n    expect(result.summary).toContain(\"2\");\n\n    // Rules written to profile\n    const profile = await loadProfile(testDir);\n    expect(profile!.writingRules).toHaveLength(2);\n    expect(profile!.writingRules.every((r) => r.source === \"auto_distilled\")).toBe(true);\n    expect(profile!.writingRules[0].confidence).toBe(0.9);\n\n    // State written\n    const state = await readDistillState();\n    expect(state.lastDistilledAt).toBeTruthy();\n  });\n\n  // 2. Self-correction: confidence out of range → error, then valid → success\n  it(\"self-correction: confidence 1.5 → execute returns error, valid retry succeeds\", async () => {\n    await recordDiff(\"c1\", \"body\", \"before text\", \"after text\", testDir);\n\n    const execResults: string[] = [];\n    const badRules = [{ rule: \"测试规则\", evidence: \"证据\", confidence: 1.5 }];\n    const runLoopImpl: (_cfg: EngineConfig, opts: LoopOptions) => Promise<LoopResult> = async (\n      _cfg,\n      opts,\n    ) => {\n      const submitTool = (opts.tools ?? []).find((t: LoopTool) => t.name === \"submit_rules\");\n      if (!submitTool) throw new Error(\"submit_rules tool not found\");\n\n      // First call rejected (out-of-range confidence), second call valid — self-correction channel\n      execResults.push(await submitTool.execute({ rules: badRules }));\n      execResults.push(await submitTool.execute({ rules: GOOD_RULES }));\n\n      return {\n        finalMessage: \"done\",\n        turns: 3,\n        totalTokens: 120,\n        toolCallCount: 2,\n        stopReason: \"no_tool_calls\",\n      } satisfies LoopResult;\n    };\n\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(execResults[0]).toContain(\"请修正\");\n    expect(result.newRules).toHaveLength(2);\n  });\n\n  // 3. Exact-duplicate guard: rule matches existing → skippedDuplicates=1\n  it(\"exact duplicate: 1 rule same as existing → skippedDuplicates=1, not re-added\", async () => {\n    await recordDiff(\"c1\", \"body\", \"before\", \"after\", testDir);\n    // Pre-seed profile with the same rule text\n    const existingRule = GOOD_RULES[0].rule;\n    await addWritingRule({ rule: existingRule, source: \"user_explicit\", confidence: 1 }, testDir);\n\n    const runLoopImpl = makeRunLoop([GOOD_RULES]); // submits 2 rules, 1 is duplicate\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(result.skippedDuplicates).toBe(1);\n    expect(result.newRules).toHaveLength(1);\n\n    const profile = await loadProfile(testDir);\n    // 1 pre-existing + 1 new = 2 total\n    expect(profile!.writingRules).toHaveLength(2);\n  });\n\n  // 3b. Evidence attribution: duplicate dropped mid-array must not shift evidences (I1)\n  it(\"evidence attribution: [dup, 新A, 新B] → summary pairs each new rule with its own evidence\", async () => {\n    await recordDiff(\"c1\", \"body\", \"before\", \"after\", testDir);\n    const dupRule = \"已有规则：保持简短\";\n    await addWritingRule({ rule: dupRule, source: \"user_explicit\", confidence: 1 }, testDir);\n\n    const submitted = [\n      { rule: dupRule, evidence: \"重复证据\", confidence: 0.9 },\n      { rule: \"新规则A：开头用提问\", evidence: \"证据A：before陈述 after提问\", confidence: 0.8 },\n      { rule: \"新规则B：结尾留钩子\", evidence: \"证据B：after新增悬念句\", confidence: 0.7 },\n    ];\n    const runLoopImpl = makeRunLoop([submitted]);\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(result.skippedDuplicates).toBe(1);\n    expect(result.newRules).toHaveLength(2);\n    expect(result.summary).toContain(\"新规则A：开头用提问（依据：证据A：before陈述 after提问）\");\n    expect(result.summary).toContain(\"新规则B：结尾留钩子（依据：证据B：after新增悬念句）\");\n    expect(result.summary).not.toContain(\"重复证据\");\n  });\n\n  // 2b. NaN confidence must be rejected at the trust boundary (I2)\n  it(\"confidence NaN → execute returns 自纠 error\", async () => {\n    await recordDiff(\"c1\", \"body\", \"before\", \"after\", testDir);\n\n    const execResults: string[] = [];\n    const nanRules = [{ rule: \"测试规则\", evidence: \"证据\", confidence: NaN }];\n    const runLoopImpl = makeRunLoop([nanRules, GOOD_RULES], execResults);\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(execResults[0]).toContain(\"请修正\");\n    expect(result.newRules).toHaveLength(2);\n  });\n\n  // 4. No new diffs → model not called, state unchanged\n  it(\"no new diffs → model not called (callCount=0), state unchanged\", async () => {\n    // Record a diff, then set lastDistilledAt in the future so it counts as already distilled\n    await recordDiff(\"c1\", \"body\", \"before\", \"after\", testDir);\n    await writeDistillState(new Date(Date.now() + 60000).toISOString());\n\n    const { impl: runLoopImpl, getCallCount } = makeRunLoopWithCount([]);\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(getCallCount()).toBe(0);\n    expect(result.newRules).toHaveLength(0);\n    expect(result.diffsAnalyzed).toBe(0);\n    // State should remain as-is (not updated)\n    const state = await readDistillState();\n    expect(state.lastDistilledAt).toBeTruthy();\n  });\n\n  // 5. Truncation: 5 rules submitted → only 3 stored\n  it(\"truncation: fake submits 5 rules → only 3 stored\", async () => {\n    await recordDiff(\"c1\", \"body\", \"before\", \"after\", testDir);\n\n    const fiveRules = Array.from({ length: 5 }, (_, i) => ({\n      rule: `规则${i + 1}：具体要求${i + 1}`,\n      evidence: `证据${i + 1}`,\n      confidence: 0.8,\n    }));\n    const runLoopImpl = makeRunLoop([fiveRules]);\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(result.newRules).toHaveLength(3);\n    const profile = await loadProfile(testDir);\n    expect(profile!.writingRules).toHaveLength(3);\n  });\n\n  // 6. Never submits → throws, state not updated\n  it(\"never submits → throws, state not updated\", async () => {\n    await recordDiff(\"c1\", \"body\", \"before\", \"after\", testDir);\n\n    const runLoopImpl: (_cfg: EngineConfig, opts: LoopOptions) => Promise<LoopResult> = async () => {\n      return {\n        finalMessage: \"不想调工具\",\n        turns: 3,\n        totalTokens: 50,\n        toolCallCount: 0,\n        stopReason: \"max_turns\",\n      } satisfies LoopResult;\n    };\n\n    await expect(distillStyleRules(testDir, { runLoopImpl })).rejects.toThrow(\n      /submit_rules|未提交|规则/,\n    );\n\n    // State must not be written\n    const state = await readDistillState();\n    expect(state.lastDistilledAt).toBeUndefined();\n  });\n\n  // 7. shouldDistillStyle: 2 new diffs → false; 3 new diffs → true\n  it(\"shouldDistillStyle: 2 diffs → false; 3 diffs → true\", async () => {\n    await recordDiff(\"c1\", \"body\", \"before\", \"after\", testDir);\n    await recordDiff(\"c2\", \"body\", \"before\", \"after\", testDir);\n    expect(await shouldDistillStyle(testDir)).toBe(false);\n\n    await recordDiff(\"c3\", \"body\", \"before\", \"after\", testDir);\n    expect(await shouldDistillStyle(testDir)).toBe(true);\n  });\n});\n\ndescribe(\"analyzeStyleSamples\", () => {\n  // 8. Happy path: samples analyzed, no state change\n  it(\"happy path: 2 samples → rules returned, distill-state not written\", async () => {\n    const runLoopImpl = makeRunLoop([GOOD_RULES]);\n    const result = await analyzeStyleSamples(\n      [\"爆款样本1：今天分享三个绝绝子的用法\", \"爆款样本2：你们有没有发现这个细节\"],\n      testDir,\n      { runLoopImpl },\n    );\n\n    expect(result.newRules).toHaveLength(2);\n    expect(result.diffsAnalyzed).toBe(2); // samplesAnalyzed reuses diffsAnalyzed field\n    expect(result.summary).toBeTruthy();\n\n    // distill-state must NOT be touched\n    const state = await readDistillState();\n    expect(state.lastDistilledAt).toBeUndefined();\n  });\n\n  // 9. Empty samples → throws parameter error\n  it(\"empty samples array → throws parameter error\", async () => {\n    const runLoopImpl = makeRunLoop([]);\n    await expect(analyzeStyleSamples([], testDir, { runLoopImpl })).rejects.toThrow(\n      /samples|样本/i,\n    );\n  });\n\n  // 10. Calibration samples always land as voice_core, even if the model tags a platform\n  it(\"forces voice_core scope on sample-derived rules (§4.3: 校准样本 = 声音内核种子)\", async () => {\n    const scoped = [\n      { rule: \"标题带数字\", evidence: \"样本1标题\", confidence: 0.9, scope: \"platform:wechat_mp\" },\n    ];\n    const runLoopImpl = makeRunLoop([scoped]);\n    await analyzeStyleSamples([\"样本正文\"], testDir, { runLoopImpl });\n\n    const profile = JSON.parse(await fs.readFile(path.join(testDir, \"creator-profile.json\"), \"utf-8\"));\n    expect(profile.writingRules[0].scope).toBe(\"voice_core\");\n  });\n});\n\ndescribe(\"voice excerpts (V5.7 声音样本)\", () => {\n  const SAMPLE =\n    \"创业者找我聊的时候,十有八九第一句话是问该买哪个课。我的答案从来只有一个:先别买,把你上周真实卡住的那个问题写下来。\\n\\n第二段讲方法论,这里的内容也足够长,超过二十个字没有问题。\";\n\n  function makeRunLoopWithExcerpts(\n    rules: Array<{ rule: string; evidence: string; confidence: number }>,\n    excerpts: string[],\n  ): (_cfg: EngineConfig, opts: LoopOptions) => Promise<LoopResult> {\n    return async (_cfg, opts) => {\n      const submitTool = (opts.tools ?? []).find((t: LoopTool) => t.name === \"submit_rules\");\n      if (!submitTool) throw new Error(\"submit_rules tool not found\");\n      await submitTool.execute({ rules, excerpts });\n      return { finalMessage: \"done\", turns: 2, totalTokens: 100, toolCallCount: 1, stopReason: \"no_tool_calls\" } satisfies LoopResult;\n    };\n  }\n\n  it(\"persists verbatim excerpts, drops fabricated ones\", async () => {\n    const verbatim = \"先别买,把你上周真实卡住的那个问题写下来。\";\n    const fabricated = \"这一段是模型自己编的,样本里根本没有这句话,超过二十个字。\";\n    const runLoopImpl = makeRunLoopWithExcerpts(GOOD_RULES, [verbatim, fabricated]);\n\n    const result = await analyzeStyleSamples([SAMPLE], testDir, { runLoopImpl });\n\n    const profile = await loadProfile(testDir);\n    expect(profile!.voiceSamples).toContain(verbatim);\n    expect(profile!.voiceSamples).not.toContain(fabricated);\n    expect(result.voiceSamplesSaved).toBeGreaterThan(0);\n    expect(result.summary).toContain(\"声音样本\");\n  });\n\n  it(\"falls back to first paragraphs when model submits no usable excerpts\", async () => {\n    const runLoopImpl = makeRunLoop([GOOD_RULES]);\n    await analyzeStyleSamples([SAMPLE], testDir, { runLoopImpl });\n\n    const profile = await loadProfile(testDir);\n    expect(profile!.voiceSamples!.length).toBeGreaterThan(0);\n    expect(profile!.voiceSamples![0]).toContain(\"创业者找我聊的时候\");\n  });\n\n  it(\"verifyVerbatimExcerpts: tolerates whitespace differences, rejects short/foreign text\", () => {\n    const samples = [\"第一句话。第二句话继续继续继续继续继续继续。\"];\n    const withNewline = \"第一句话。\\n第二句话继续继续继续继续继续继续。\";\n    expect(verifyVerbatimExcerpts([withNewline], samples)).toEqual([withNewline]);\n    expect(verifyVerbatimExcerpts([\"太短\"], samples)).toEqual([]);\n    expect(verifyVerbatimExcerpts([\"完全不在样本里的一段文字并且长度超过二十个字符啊\"], samples)).toEqual([]);\n  });\n\n  it(\"fallbackExcerpts: takes first paragraph ≥20 chars, slices to 200\", () => {\n    const long = \"长\".repeat(300);\n    const out = fallbackExcerpts([`短段。\\n\\n${long}`], 2);\n    expect(out).toHaveLength(1);\n    expect(out[0]).toHaveLength(200);\n  });\n});\n\ndescribe(\"scope routing through distillStyleRules\", () => {\n  it(\"persists the scope submitted by the model (纠正记入发生平台)\", async () => {\n    await recordDiff(\"c1\", \"body\", \"旧正文\", \"新正文\", testDir, undefined, \"wechat_mp\");\n\n    const scoped = [\n      { rule: \"公众号正文空行分段\", evidence: \"diff1\", confidence: 0.8, scope: \"platform:wechat_mp\" },\n    ];\n    const runLoopImpl = makeRunLoop([scoped]);\n    const result = await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(result.newRules).toHaveLength(1);\n    expect(result.newRules[0].scope).toBe(\"platform:wechat_mp\");\n  });\n\n  it(\"rejects malformed scope via submit_rules self-correction\", async () => {\n    await recordDiff(\"c1\", \"body\", \"旧\", \"新\", testDir);\n\n    const bad = [{ rule: \"规则\", evidence: \"e\", confidence: 0.5, scope: \"wechat_mp\" }];\n    const execResults: string[] = [];\n    const runLoopImpl = makeRunLoop([bad, GOOD_RULES], execResults);\n    await distillStyleRules(testDir, { runLoopImpl });\n\n    expect(execResults[0]).toContain(\"scope\");\n  });\n});\n"}}}}