{"repo": "caiyilian/anime-accurate-sub", "files": {"scripts/season_monitor.py": {"4d507921cd926e659e5a5a22d58eec34686302b8": {"t": 1785768246.6065085, "n": 21253, "h": "c8d3c05e16668cd8", "c": "\"\"\"Read-only progress collector and dashboard for a full-season pipeline run.\"\"\"\n\n# The embedded, self-contained dashboard intentionally keeps CSS and JS compact.\n# ruff: noqa: E501\n\nfrom __future__ import annotations\n\nimport json\nimport re\nimport time\nfrom collections import Counter\nfrom datetime import datetime\nfrom pathlib import Path\nfrom typing import Any\n\nSEASON_STAGES = (\n (\"japanese_subtitle\", \"日文字幕\", 2),\n (\"translate\", \"Sakura 翻译\", 10),\n (\"multi_agent_review\", \"多 Agent 审查\", 38),\n (\"mqm_quality_review\", \"双裁判 MQM\", 42),\n (\"subtitle\", \"生成字幕\", 1),\n (\"embed_subtitle\", \"压制视频\", 6),\n (\"quality_check\", \"规则质检\", 1),\n)\nSTAGE_LABELS = dict((stage, label) for stage, label, _ in SEASON_STAGES)\nSTAGE_WEIGHTS = dict((stage, weight) for stage, _, weight in SEASON_STAGES)\nREVIEW_FILES = {\n \"multi_agent_review\": (\"multi_agent_review.progress.jsonl\", \"translated.json\"),\n \"mqm_quality_review\": (\"mqm_quality_review.progress.jsonl\", \"reviewed.json\"),\n}\nEPISODE_PATTERN = re.compile(r\"第\\s*(\\d{1,3})\\s*集\")\nANSI_PATTERN = re.compile(r\"\\x1b\\[[0-?]*[ -/]*[@-~]\")\nSECRET_PATTERNS = (\n re.compile(r\"(?i)(authorization\\s*[:=]\\s*bearer\\s+)[^\\s,;]+\"),\n re.compile(r\"(?i)((?:api[_-]?key|token|secret)\\s*[:=]\\s*)[^\\s,;]+\"),\n re.compile(r\"\\bsk-[A-Za-z0-9_-]{8,}\\b\"),\n)\n\n\ndef _iso_timestamp(timestamp: float | None) -> str | None:\n if timestamp is None:\n return None\n return datetime.fromtimestamp(timestamp).astimezone().isoformat(timespec=\"seconds\")\n\n\ndef _read_json(path: Path) -> Any:\n try:\n return json.loads(path.read_text(encoding=\"utf-8\"))\n except (OSError, UnicodeError, json.JSONDecodeError):\n return None\n\n\ndef _redact_log(text: str) -> str:\n redacted = ANSI_PATTERN.sub(\"\", text)\n for pattern in SECRET_PATTERNS:\n if pattern.groups:\n redacted = pattern.sub(r\"\\1[已隐藏]\", redacted)\n else:\n redacted = pattern.sub(\"[已隐藏]\", redacted)\n return redacted\n\n\nclass SeasonProgressMonitor:\n \"\"\"Summarize durable pipeline artifacts without controlling the process.\"\"\"\n\n def __init__(\n self,\n root: str | Path,\n episode_count: int = 14,\n stale_after_seconds: int = 300,\n log_line_count: int = 36,\n ) -> None:\n self.root = Path(root).resolve()\n self.episode_count = max(1, int(episode_count))\n self.stale_after_seconds = max(30, int(stale_after_seconds))\n self.log_line_count = max(5, int(log_line_count))\n self._jsonl_cache: dict[Path, tuple[tuple[int, int], dict[str, Any]]] = {}\n self._count_cache: dict[Path, tuple[tuple[int, int], int]] = {}\n\n @staticmethod\n def _signature(path: Path) -> tuple[int, int] | None:\n try:\n stat = path.stat()\n return stat.st_size, stat.st_mtime_ns\n except OSError:\n return None\n\n def _segment_count(self, path: Path) -> int:\n signature = self._signature(path)\n if signature is None:\n return 0\n cached = self._count_cache.get(path)\n if cached and cached[0] == signature:\n return cached[1]\n payload = _read_json(path)\n count = len(payload) if isinstance(payload, list) else 0\n self._count_cache[path] = (signature, count)\n return count\n\n def _review_progress(self, path: Path, total: int) -> dict[str, Any]:\n signature = self._signature(path)\n empty = {\"processed\": 0, \"total\": total, \"percent\": 0, \"statuses\": {}}\n if signature is None:\n return empty\n cached = self._jsonl_cache.get(path)\n if cached and cached[0] == signature:\n summary = dict(cached[1])\n summary[\"total\"] = total\n summary[\"percent\"] = round(100 * summary[\"processed\"] / total, 1) if total else 0\n return summary\n\n records: dict[int, dict[str, Any]] = {}\n try:\n with path.open(\"r\", encoding=\"utf-8\", errors=\"replace\") as source:\n for line in source:\n try:\n row = json.loads(line)\n except json.JSONDecodeError:\n continue\n position = row.get(\"position\") if isinstance(row, dict) else None\n if isinstance(position, int) and position >= 0:\n records[position] = row\n except OSError:\n return empty\n\n statuses = Counter(str(row.get(\"status\") or \"unknown\") for row in records.values())\n processed = len(records)\n summary = {\n \"processed\": processed,\n \"total\": total,\n \"percent\": round(100 * processed / total, 1) if total else 0,\n \"statuses\": dict(sorted(statuses.items())),\n }\n self._jsonl_cache[path] = (signature, dict(summary))\n return summary\n\n def _episode_directories(self) -> dict[int, Path]:\n result: dict[int, Path] = {}\n if not self.root.is_dir():\n return result\n try:\n children = self.root.iterdir()\n except OSError:\n return result\n for path in children:\n if not path.is_dir():\n continue\n match = EPISODE_PATTERN.search(path.name)\n if match:\n result[int(match.group(1))] = path\n return result\n\n @staticmethod\n def _newest_mtime(path: Path | None) -> float | None:\n if path is None or not path.is_dir():\n return None\n newest: float | None = None\n try:\n for child in path.iterdir():\n if child.is_file():\n modified = child.stat().st_mtime\n newest = modified if newest is None else max(newest, modified)\n except OSError:\n return newest\n return newest\n\n def _quality_summary(self, episode_dir: Path | None) -> dict[str, Any]:\n if episode_dir is None:\n return {}\n result: dict[str, Any] = {}\n mqm = _read_json(episode_dir / \"mqm_quality_report.json\")\n if isinstance(mqm, dict) and isinstance(mqm.get(\"summary\"), dict):\n result[\"mqm\"] = {\n key: mqm[\"summary\"].get(key, 0)\n for key in (\"approved\", \"corrected\", \"needs_review\", \"errors\", \"applied\")\n }\n final = _read_json(episode_dir / \"final_adjudication_summary.json\")\n if isinstance(final, dict) and isinstance(final.get(\"summary\"), dict):\n final_summary = final[\"summary\"]\n resolved = min(\n int(result[\"mqm\"][\"needs_review\"]),\n int(final_summary.get(\"resolved_needs_review\", 0)),\n )\n kept = min(resolved, int(final_summary.get(\"needs_review_kept\", 0)))\n revised = min(\n resolved - kept,\n int(final_summary.get(\"needs_review_revised\", 0)),\n )\n result[\"mqm\"][\"approved\"] += kept\n result[\"mqm\"][\"corrected\"] += revised\n result[\"mqm\"][\"needs_review\"] -= resolved\n result[\"mqm\"][\"final_reviewed\"] = int(\n final_summary.get(\"reviewed\", resolved)\n )\n quality = _read_json(episode_dir / \"quality_report.json\")\n if isinstance(quality, dict) and isinstance(quality.get(\"stats\"), dict):\n result[\"rules\"] = {\n key: quality[\"stats\"].get(key, 0)\n for key in (\"total_issues\", \"errors\", \"warnings\", \"info\")\n }\n return result\n\n def _episode_snapshot(self, number: int, episode_dir: Path | None) -> dict[str, Any]:\n checkpoint = _read_json(episode_dir / \"checkpoint.json\") if episode_dir else None\n checkpoint = checkpoint if isinstance(checkpoint, dict) else {}\n completed = {\n stage for stage, _, _ in SEASON_STAGES\n if isinstance(checkpoint.get(stage), dict)\n and checkpoint[stage].get(\"status\") == \"completed\"\n }\n failed = {\n stage for stage, _, _ in SEASON_STAGES\n if isinstance(checkpoint.get(stage), dict)\n and checkpoint[stage].get(\"status\") == \"failed\"\n }\n active = next((stage for stage, _, _ in SEASON_STAGES if stage not in completed), None)\n review = {\"processed\": 0, \"total\": 0, \"percent\": 0, \"statuses\": {}}\n if active in REVIEW_FILES and episode_dir:\n progress_name, source_name = REVIEW_FILES[active]\n total = self._segment_count(episode_dir / source_name)\n review = self._review_progress(episode_dir / progress_name, total)\n\n stages = []\n weighted_progress = 0.0\n for stage, label, weight in SEASON_STAGES:\n if stage in completed:\n state, fraction = \"completed\", 1.0\n elif stage in failed:\n state, fraction = \"failed\", 0.0\n elif stage == active:\n state = \"active\"\n fraction = review[\"percent\"] / 100 if stage in REVIEW_FILES else 0.0\n else:\n state, fraction = \"pending\", 0.0\n weighted_progress += weight * fraction\n stages.append({\"id\": stage, \"label\": label, \"state\": state})\n\n if len(completed) == len(SEASON_STAGES):\n state = \"completed\"\n elif failed:\n state = \"failed\"\n elif checkpoint or review[\"processed\"]:\n state = \"running\"\n else:\n state = \"pending\"\n updated = self._newest_mtime(episode_dir)\n return {\n \"number\": number,\n \"name\": episode_dir.name if episode_dir else f\"第{number:02d}集\",\n \"state\": state,\n \"percent\": round(weighted_progress, 1),\n \"active_stage\": active,\n \"active_stage_label\": STAGE_LABELS.get(active, \"已完成\"),\n \"completed_stages\": len(completed),\n \"total_stages\": len(SEASON_STAGES),\n \"stages\": stages,\n \"review\": review,\n \"quality\": self._quality_summary(episode_dir),\n \"updated_at\": _iso_timestamp(updated),\n \"_updated_epoch\": updated,\n }\n\n def _log_tail(self) -> tuple[str, float | None]:\n paths = [self.root / \"full_run_v3.stdout.log\", self.root / \"full_run.stdout.log\"]\n path = next((candidate for candidate in paths if candidate.is_file()), None)\n if path is None:\n return \"\", None\n try:\n size = path.stat().st_size\n with path.open(\"rb\") as source:\n source.seek(max(0, size - 65_536))\n lines = source.read().decode(\"utf-8\", errors=\"replace\").splitlines()\n return _redact_log(\"\\n\".join(lines[-self.log_line_count :])), path.stat().st_mtime\n except OSError:\n return \"\", None\n\n def snapshot(self) -> dict[str, Any]:\n directories = self._episode_directories()\n episodes = [\n self._episode_snapshot(number, directories.get(number))\n for number in range(1, self.episode_count + 1)\n ]\n log_tail, log_mtime = self._log_tail()\n activity_times = [\n timestamp for timestamp in [log_mtime, *(item[\"_updated_epoch\"] for item in episodes)]\n if timestamp is not None\n ]\n last_activity = max(activity_times) if activity_times else None\n now = time.time()\n completed_episodes = sum(item[\"state\"] == \"completed\" for item in episodes)\n failed = any(item[\"state\"] == \"failed\" for item in episodes)\n if completed_episodes == self.episode_count:\n status = \"completed\"\n elif failed:\n status = \"failed\"\n elif last_activity is not None and now - last_activity <= self.stale_after_seconds:\n status = \"running\"\n elif last_activity is not None:\n status = \"stale\"\n else:\n status = \"waiting\"\n\n active_episode = next((item for item in episodes if item[\"state\"] == \"running\"), None)\n if active_episode is None:\n active_episode = next((item for item in episodes if item[\"state\"] != \"completed\"), None)\n overall = sum(item[\"percent\"] for item in episodes) / self.episode_count\n for episode in episodes:\n episode.pop(\"_updated_epoch\", None)\n return {\n \"status\": status,\n \"status_label\": {\n \"running\": \"运行正常\",\n \"completed\": \"全季完成\",\n \"failed\": \"发现失败阶段\",\n \"stale\": \"等待新进度\",\n \"waiting\": \"尚未开始\",\n }[status],\n \"overall_percent\": round(overall, 1),\n \"completed_episodes\": completed_episodes,\n \"episode_count\": self.episode_count,\n \"active_episode\": active_episode[\"number\"] if active_episode else None,\n \"last_activity_at\": _iso_timestamp(last_activity),\n \"captured_at\": _iso_timestamp(now),\n \"root_name\": self.root.name,\n \"episodes\": episodes,\n \"log_tail\": log_tail,\n }\n\n\nMONITOR_HTML = \"\"\"\n\n轻音少女全季进度 · Anime Accurate Sub\n
\n
Anime Accurate Sub · Quality Run

轻音少女 S1
全流程监控

Sakura 翻译、SenseNova 多 Agent 审查与双裁判 MQM 的实时执行状态。页面每 4 秒自动读取持久化进度。

正在连接
\n
正在读取全季进度……
\n
\n
\"\"\"\n"}}, "config/quality_review.sensenova.json": {"4d507921cd926e659e5a5a22d58eec34686302b8": {"t": 1785768255.8765476, "n": 856, "h": "4e5a517e9ae23875", "c": "{\n \"provider\": \"openai\",\n \"base_url\": \"https://token.sensenova.cn/v1\",\n \"api_key_file\": \"config/sensenova_apikeys\",\n \"review_model\": \"sensenova-6.7-flash-lite\",\n \"review_fallback_model\": \"deepseek-v4-flash\",\n \"agent_models\": {\n \"consistency\": \"sensenova-6.7-flash-lite\",\n \"naturalness\": \"sensenova-6.7-flash-lite\",\n \"accuracy\": \"deepseek-v4-flash\",\n \"asr_check\": \"sensenova-6.7-flash-lite\",\n \"style\": \"sensenova-6.7-flash-lite\"\n },\n \"editor_provider\": \"openai\",\n \"editor_base_url\": \"https://token.sensenova.cn/v1\",\n \"editor_api_key_file\": \"config/sensenova_apikeys\",\n \"editor_model\": \"sensenova-6.7-flash-lite\",\n \"editor_fallback_model\": \"deepseek-v4-flash\",\n \"max_workers\": 2,\n \"context_window\": 5,\n \"min_fix_votes\": 2,\n \"min_reviewer_confidence\": 0.75,\n \"min_editor_confidence\": 0.85,\n \"retries\": 6,\n \"timeout_s\": 600\n}\n"}}, "config/translator.sakura-remote.example.json": {"4d507921cd926e659e5a5a22d58eec34686302b8": {"t": 1785768255.9559946, "n": 997, "h": "7f4907ead49e90fa", "c": "{\n \"backend\": \"sakura\",\n \"sakura\": {\n \"model\": \"crosery/sakura-14b-qwen2.5-v1.0-q6k:latest\",\n \"host\": \"172.31.102.189\",\n \"timeout_s\": 300,\n \"max_retries\": 4,\n \"validation_retries\": 2,\n \"validation_fallback_backend\": \"galtransl\",\n \"validation_rescue_models\": [\n \"sensenova-6.7-flash-lite\",\n \"deepseek-v4-flash\"\n ],\n \"validation_rescue_base_url\": \"https://token.sensenova.cn/v1\",\n \"validation_rescue_api_key_file\": \"config/sensenova_apikeys\",\n \"validation_rescue_attempts\": 2,\n \"validation_quarantine_on_failure\": true,\n \"batch_size\": 16,\n \"num_ctx\": 4096,\n \"temperature\": 0.1,\n \"top_p\": 0.3,\n \"repeat_penalty\": 1.0,\n \"frequency_penalty\": 0.1\n },\n \"galtransl\": {\n \"model\": \"crosery/GalTransl-7B-v2.6:Q6_k\",\n \"host\": \"172.31.102.189\",\n \"timeout_s\": 300,\n \"max_retries\": 4,\n \"batch_size\": 10,\n \"num_ctx\": 4096,\n \"temperature\": 0.2,\n \"top_p\": 0.8,\n \"repeat_penalty\": 1.1,\n \"frequency_penalty\": 0.1\n }\n}\n"}}, "config/quality_final_adjudication.sensenova.json": {"4d507921cd926e659e5a5a22d58eec34686302b8": {"t": 1785768257.9338837, "n": 529, "h": "d6107fc4d7253030", "c": "{\n \"provider\": \"openai\",\n \"base_url\": \"https://token.sensenova.cn/v1\",\n \"api_key_file\": \"config/sensenova_apikeys\",\n \"review_models\": [\n \"sensenova-6.7-flash-lite\",\n \"deepseek-v4-flash\"\n ],\n \"review_fallback_models\": {\n \"sensenova-6.7-flash-lite\": \"glm-5.2\",\n \"deepseek-v4-flash\": \"glm-5.2\"\n },\n \"adjudicator_model\": \"sensenova-6.7-flash-lite\",\n \"adjudicator_fallback_model\": \"deepseek-v4-flash\",\n \"max_workers\": 6,\n \"context_window\": 5,\n \"retries\": 3,\n \"timeout_s\": 600,\n \"min_apply_confidence\": 0.9\n}\n"}}, "scripts/review_agents.py": {"4d507921cd926e659e5a5a22d58eec34686302b8": {"t": 1785768271.7009218, "n": 35793, "h": "23793332ccfffab1", "c": "\"\"\"S11.2: auditable multi-agent subtitle review.\n\nFive reviewer prompts inspect one translated segment in parallel. A separate\neditor may replace the translation only when multiple reviewers agree and the\neditor returns a high-confidence, strictly structured decision. The module is\nsafe to resume on a season-sized job and never mutates the source file.\n\"\"\"\n\nfrom __future__ import annotations\n\nimport argparse\nimport hashlib\nimport itertools\nimport json\nimport os\nimport re\nimport threading\nimport time\nimport urllib.error\nimport urllib.request\nfrom concurrent.futures import ThreadPoolExecutor, as_completed\nfrom dataclasses import asdict, dataclass, field\nfrom pathlib import Path\nfrom typing import Callable, Optional\n\n\nDEFAULT_MODEL = \"EasonONLINE/Sakura-qwen2.5-v1.0:7b\"\nDEFAULT_SENSENOVA_URL = \"https://token.sensenova.cn/v1\"\nVALID_VERDICTS = {\"ok\", \"fix\", \"suspicious\"}\nVALID_EDITOR_DECISIONS = {\"keep\", \"replace\"}\nMAX_INCOMPLETE_PASSES = 3\nAGENT_SCHEMA = {\n \"type\": \"object\",\n \"properties\": {\n \"verdict\": {\"type\": \"string\", \"enum\": sorted(VALID_VERDICTS)},\n \"suggested_zh\": {\"type\": \"string\"},\n \"reason\": {\"type\": \"string\"},\n \"confidence\": {\"type\": \"number\", \"minimum\": 0, \"maximum\": 1},\n },\n \"required\": [\"verdict\", \"suggested_zh\", \"reason\", \"confidence\"],\n \"additionalProperties\": False,\n}\nEDITOR_SCHEMA = {\n \"type\": \"object\",\n \"properties\": {\n \"decision\": {\"type\": \"string\", \"enum\": sorted(VALID_EDITOR_DECISIONS)},\n \"corrected_zh\": {\"type\": \"string\"},\n \"reason\": {\"type\": \"string\"},\n \"confidence\": {\"type\": \"number\", \"minimum\": 0, \"maximum\": 1},\n },\n \"required\": [\"decision\", \"corrected_zh\", \"reason\", \"confidence\"],\n \"additionalProperties\": False,\n}\n\n\nAGENTS = {\n \"consistency\": {\n \"name\": \"Consistency Checker\",\n \"focus\": \"角色名、称呼、术语和前后台词是否一致\",\n },\n \"naturalness\": {\n \"name\": \"Naturalness Checker\",\n \"focus\": \"中文是否自然、简洁、符合口语和字幕阅读习惯\",\n },\n \"accuracy\": {\n \"name\": \"Accuracy Checker\",\n \"focus\": \"中文是否忠实表达日文含义、语气、主客体，是否漏译或增译\",\n },\n \"asr_check\": {\n \"name\": \"ASR Quality Checker\",\n \"focus\": \"日文 ASR 是否存在误听、断裂、重复或与上下文矛盾\",\n },\n \"style\": {\n \"name\": \"Anime Style Checker\",\n \"focus\": \"译文是否符合轻松动漫对白的语气、人物关系和场景情绪\",\n },\n}\n\n\n@dataclass\nclass ReviewConfig:\n provider: str = field(\n default_factory=lambda: os.environ.get(\"REVIEW_PROVIDER\", \"ollama\")\n )\n host: str = field(\n default_factory=lambda: os.environ.get(\n \"REVIEW_HOST\", os.environ.get(\"OLLAMA_HOST\", \"localhost\")\n )\n )\n review_model: str = field(\n default_factory=lambda: os.environ.get(\"REVIEW_MODEL\", DEFAULT_MODEL)\n )\n review_fallback_model: str = field(\n default_factory=lambda: os.environ.get(\"REVIEW_FALLBACK_MODEL\", \"\")\n )\n base_url: str = field(\n default_factory=lambda: os.environ.get(\"REVIEW_BASE_URL\", \"\")\n )\n api_key_file: str = field(\n default_factory=lambda: os.environ.get(\"REVIEW_API_KEY_FILE\", \"\")\n )\n editor_provider: str = field(\n default_factory=lambda: os.environ.get(\"EDITOR_PROVIDER\", \"\")\n )\n editor_host: str = field(\n default_factory=lambda: os.environ.get(\n \"EDITOR_HOST\",\n os.environ.get(\"REVIEW_HOST\", os.environ.get(\"OLLAMA_HOST\", \"localhost\")),\n )\n )\n editor_model: str = field(\n default_factory=lambda: os.environ.get(\n \"EDITOR_MODEL\", os.environ.get(\"REVIEW_MODEL\", DEFAULT_MODEL)\n )\n )\n editor_fallback_model: str = field(\n default_factory=lambda: os.environ.get(\"EDITOR_FALLBACK_MODEL\", \"\")\n )\n editor_base_url: str = field(\n default_factory=lambda: os.environ.get(\"EDITOR_BASE_URL\", \"\")\n )\n editor_api_key_file: str = field(\n default_factory=lambda: os.environ.get(\"EDITOR_API_KEY_FILE\", \"\")\n )\n agent_models: dict[str, str] = field(default_factory=dict)\n max_workers: int = 2\n context_window: int = 3\n min_fix_votes: int = 2\n min_reviewer_confidence: float = 0.75\n min_editor_confidence: float = 0.80\n retries: int = 3\n timeout_s: int = 300\n\n @classmethod\n def from_dict(cls, data: Optional[dict]) -> \"ReviewConfig\":\n if not data:\n return cls()\n known = {item.name for item in cls.__dataclass_fields__.values()}\n return cls(**{key: value for key, value in data.items() if key in known})\n\n def validate(self) -> \"ReviewConfig\":\n self.provider = self.provider.strip().lower()\n self.review_fallback_model = self.review_fallback_model.strip()\n self.editor_fallback_model = self.editor_fallback_model.strip()\n if self.provider not in {\"ollama\", \"openai\"}:\n raise ValueError(\"provider must be ollama or openai\")\n self.editor_provider = (self.editor_provider or self.provider).strip().lower()\n if self.editor_provider not in {\"ollama\", \"openai\"}:\n raise ValueError(\"editor_provider must be ollama or openai\")\n if self.provider == \"openai\" and not (self.api_key_file or os.environ.get(\"REVIEW_API_KEY\")):\n raise ValueError(\"openai review provider requires api_key_file or REVIEW_API_KEY\")\n if self.editor_provider == \"openai\":\n self.editor_base_url = self.editor_base_url or self.base_url\n self.editor_api_key_file = self.editor_api_key_file or self.api_key_file\n if self.max_workers < 1:\n raise ValueError(\"max_workers must be >= 1\")\n if self.context_window < 0:\n raise ValueError(\"context_window must be >= 0\")\n if not 1 <= self.min_fix_votes <= len(AGENTS):\n raise ValueError(f\"min_fix_votes must be between 1 and {len(AGENTS)}\")\n if not 0 <= self.min_reviewer_confidence <= 1:\n raise ValueError(\"min_reviewer_confidence must be between 0 and 1\")\n if not 0 <= self.min_editor_confidence <= 1:\n raise ValueError(\"min_editor_confidence must be between 0 and 1\")\n if self.retries < 1:\n raise ValueError(\"retries must be >= 1\")\n return self\n\n\ndef _api_url(host: str) -> str:\n host = host.strip().rstrip(\"/\")\n if not host.startswith((\"http://\", \"https://\")):\n host = f\"http://{host}\"\n authority = host.split(\"//\", 1)[1].split(\"/\", 1)[0]\n if \":\" not in authority:\n host = f\"{host}:11434\"\n if host.endswith(\"/api\"):\n return host\n return f\"{host}/api\"\n\n\ndef ollama_chat(\n messages: list[dict],\n *,\n temperature: float = 0.1,\n model: str = DEFAULT_MODEL,\n host: str = \"localhost\",\n json_mode: bool = True,\n json_schema: Optional[dict] = None,\n timeout_s: int = 300,\n) -> str:\n payload = {\n \"model\": model,\n \"messages\": messages,\n \"stream\": False,\n \"options\": {\"temperature\": temperature, \"num_predict\": 768},\n }\n if json_mode:\n payload[\"format\"] = json_schema or \"json\"\n request = urllib.request.Request(\n f\"{_api_url(host)}/chat\",\n data=json.dumps(payload, ensure_ascii=False).encode(\"utf-8\"),\n headers={\"Content-Type\": \"application/json; charset=utf-8\"},\n )\n with urllib.request.urlopen(request, timeout=timeout_s) as response:\n result = json.loads(response.read().decode(\"utf-8\"))\n return result.get(\"message\", {}).get(\"content\", \"\").strip()\n\n\n_KEY_SEQUENCE = itertools.count()\n_KEY_LOCK = threading.Lock()\n\n\ndef _api_keys(path: str, environment_name: str = \"REVIEW_API_KEY\") -> list[str]:\n keys = []\n environment_value = os.environ.get(environment_name, \"\").strip()\n if environment_value:\n keys.append(environment_value)\n if path:\n keys.extend(\n line.strip()\n for line in Path(path).read_text(encoding=\"utf-8\").splitlines()\n if line.strip()\n )\n # Preserve order while removing accidental duplicates.\n return list(dict.fromkeys(keys))\n\n\ndef _next_api_key(path: str) -> str:\n keys = _api_keys(path)\n if not keys:\n raise ValueError(\"no API key is available\")\n with _KEY_LOCK:\n position = next(_KEY_SEQUENCE)\n return keys[position % len(keys)]\n\n\ndef openai_chat(\n messages: list[dict],\n *,\n temperature: float = 0.1,\n model: str,\n base_url: str,\n api_key_file: str,\n json_mode: bool = True,\n timeout_s: int = 300,\n **_kwargs,\n) -> str:\n payload = {\n \"model\": model,\n \"messages\": messages,\n \"stream\": False,\n \"temperature\": temperature,\n \"max_tokens\": 2048,\n }\n if json_mode:\n # SenseNova's OpenAI-compatible gateway supports json_object. The\n # semantic schema remains in the prompt and is validated locally.\n payload[\"response_format\"] = {\"type\": \"json_object\"}\n endpoint = (base_url or DEFAULT_SENSENOVA_URL).rstrip(\"/\") + \"/chat/completions\"\n request = urllib.request.Request(\n endpoint,\n data=json.dumps(payload, ensure_ascii=False).encode(\"utf-8\"),\n headers={\n \"Authorization\": f\"Bearer {_next_api_key(api_key_file)}\",\n \"Content-Type\": \"application/json; charset=utf-8\",\n \"Accept\": \"application/json\",\n },\n )\n try:\n with urllib.request.urlopen(request, timeout=timeout_s) as response:\n result = json.loads(response.read().decode(\"utf-8\"))\n except urllib.error.HTTPError as error:\n # Do not include response bodies: providers sometimes echo request\n # metadata, and review logs must never risk persisting credentials.\n raise RuntimeError(f\"OpenAI-compatible API returned HTTP {error.code}\") from error\n choices = result.get(\"choices\") or []\n if not choices:\n raise ValueError(\"OpenAI-compatible API response has no choices\")\n return str(choices[0].get(\"message\", {}).get(\"content\", \"\")).strip()\n\n\ndef model_chat(\n messages: list[dict],\n *,\n provider: str = \"ollama\",\n base_url: str = \"\",\n api_key_file: str = \"\",\n **kwargs,\n) -> str:\n if provider == \"openai\":\n return openai_chat(\n messages,\n base_url=base_url,\n api_key_file=api_key_file,\n **kwargs,\n )\n return ollama_chat(messages, **kwargs)\n\n\ndef _json_object(response: str) -> tuple[dict, bool]:\n \"\"\"Parse one JSON object with only deterministic syntax repairs.\n\n Sakura occasionally stops after a complete final value but before the\n closing brace. Appending that single brace (or removing a trailing comma)\n cannot invent a verdict; all semantic fields are still schema-validated by\n the caller. Natural-language keyword guessing remains deliberately banned.\n \"\"\"\n text = (response or \"\").strip()\n fenced = re.fullmatch(r\"```(?:json)?\\s*(\\{.*\\})\\s*```\", text, re.DOTALL)\n if fenced:\n text = fenced.group(1)\n repaired = False\n try:\n value = json.loads(text)\n except json.JSONDecodeError as original_error:\n candidates = []\n without_trailing_comma = re.sub(r\",\\s*}$\", \"}\", text)\n if without_trailing_comma != text:\n candidates.append(without_trailing_comma)\n if (\n text.startswith(\"{\")\n and text.count(\"{\") == text.count(\"}\") + 1\n and len(re.findall(r'(? None:\n keys = set(value)\n missing = required - keys\n extra = keys - required\n if missing:\n raise ValueError(f\"{label} response is missing fields: {sorted(missing)}\")\n if extra:\n raise ValueError(f\"{label} response has unexpected fields: {sorted(extra)}\")\n\n\ndef _strict_confidence(value, label: str) -> float:\n if isinstance(value, bool) or not isinstance(value, (int, float)):\n raise ValueError(f\"{label} confidence must be a JSON number\")\n confidence = float(value)\n if not 0 <= confidence <= 1:\n raise ValueError(f\"{label} confidence must be between 0 and 1\")\n return confidence\n\n\ndef parse_agent_response(response: str) -> dict:\n value, repaired = _json_object(response)\n _strict_object(\n value, {\"verdict\", \"suggested_zh\", \"reason\", \"confidence\"}, \"reviewer\"\n )\n if not all(isinstance(value[key], str) for key in (\"verdict\", \"suggested_zh\", \"reason\")):\n raise ValueError(\"reviewer verdict, suggested_zh and reason must be strings\")\n verdict = str(value.get(\"verdict\", \"\")).strip().lower()\n if verdict not in VALID_VERDICTS:\n raise ValueError(f\"invalid reviewer verdict: {verdict!r}\")\n suggested = str(value.get(\"suggested_zh\") or \"\").strip()\n if verdict == \"fix\" and not suggested:\n raise ValueError(\"fix verdict requires suggested_zh\")\n return {\n \"verdict\": verdict,\n \"suggested_zh\": suggested or None,\n \"reason\": str(value.get(\"reason\") or \"\").strip(),\n \"confidence\": _strict_confidence(value[\"confidence\"], \"reviewer\"),\n \"response_repaired\": repaired,\n }\n\n\ndef parse_editor_response(response: str) -> dict:\n value, repaired = _json_object(response)\n _strict_object(\n value, {\"decision\", \"corrected_zh\", \"reason\", \"confidence\"}, \"editor\"\n )\n if not all(isinstance(value[key], str) for key in (\"decision\", \"corrected_zh\", \"reason\")):\n raise ValueError(\"editor decision, corrected_zh and reason must be strings\")\n decision = str(value.get(\"decision\", \"\")).strip().lower()\n if decision not in VALID_EDITOR_DECISIONS:\n raise ValueError(f\"invalid editor decision: {decision!r}\")\n corrected = str(value.get(\"corrected_zh\") or \"\").strip()\n if decision == \"replace\" and not corrected:\n raise ValueError(\"replace decision requires corrected_zh\")\n return {\n \"decision\": decision,\n \"corrected_zh\": corrected or None,\n \"reason\": str(value.get(\"reason\") or \"\").strip(),\n \"confidence\": _strict_confidence(value[\"confidence\"], \"editor\"),\n \"response_repaired\": repaired,\n }\n\n\ndef _segment_text(segment: dict) -> str:\n return str(segment.get(\"text\", segment.get(\"zh\", \"\"))).strip()\n\n\ndef _segment_key(position: int, segment: dict) -> str:\n payload = {\n \"position\": position,\n \"id\": segment.get(\"id\", segment.get(\"index\")),\n \"ja\": str(segment.get(\"ja\", \"\")),\n \"text\": _segment_text(segment),\n }\n digest = hashlib.sha256(\n json.dumps(payload, ensure_ascii=False, sort_keys=True).encode(\"utf-8\")\n ).hexdigest()[:20]\n return f\"{position}:{digest}\"\n\n\ndef _context_for(segments: list[dict], position: int, window: int) -> str:\n start = max(0, position - window)\n end = min(len(segments), position + window + 1)\n lines = []\n for index in range(start, end):\n marker = \"目标\" if index == position else \"上下文\"\n segment = segments[index]\n lines.append(\n f\"[{marker} {index}] 日文：{segment.get('ja', '')}\\n\"\n f\"[{marker} {index}] 中文：{_segment_text(segment)}\"\n )\n return \"\\n\".join(lines)\n\n\ndef load_glossary_prompt(path: str = \"\") -> str:\n if not path:\n return \"（未提供术语表）\"\n value = json.loads(Path(path).read_text(encoding=\"utf-8\"))\n terms = value.get(\"terms\", value) if isinstance(value, dict) else value\n rows = []\n if isinstance(terms, list):\n for item in terms:\n if isinstance(item, dict) and item.get(\"ja\") and item.get(\"zh\"):\n rows.append(f\"{item['ja']} => {item['zh']}\")\n elif isinstance(terms, dict):\n rows.extend(f\"{ja} => {zh}\" for ja, zh in terms.items())\n return \"\\n\".join(rows) if rows else \"（术语表为空）\"\n\n\ndef _call_json_with_retries(\n messages: list[dict],\n *,\n model: str,\n host: str,\n provider: str,\n base_url: str,\n api_key_file: str,\n config: ReviewConfig,\n parser: Callable[[str], dict],\n json_schema: dict,\n chat_fn: Callable = model_chat,\n temperature: float = 0.1,\n) -> tuple[Optional[dict], str, str, int, float]:\n last_response = \"\"\n last_error = \"\"\n started = time.time()\n for attempt in range(1, config.retries + 1):\n try:\n last_response = chat_fn(\n messages,\n temperature=temperature,\n model=model,\n host=host,\n provider=provider,\n base_url=base_url,\n api_key_file=api_key_file,\n json_mode=True,\n json_schema=json_schema,\n timeout_s=config.timeout_s,\n )\n return parser(last_response), last_response, \"\", attempt, time.time() - started\n except Exception as error: # network and model-format failures are retryable\n last_error = str(error)\n if attempt < config.retries:\n time.sleep(min(2 ** (attempt - 1), 4))\n return None, last_response, last_error, config.retries, time.time() - started\n\n\ndef _call_json_with_fallback(\n messages: list[dict],\n *,\n model: str,\n fallback_model: str,\n host: str,\n provider: str,\n base_url: str,\n api_key_file: str,\n config: ReviewConfig,\n parser: Callable[[str], dict],\n json_schema: dict,\n chat_fn: Callable = model_chat,\n temperature: float = 0.1,\n) -> tuple[Optional[dict], str, str, int, float, str, bool, str]:\n primary = _call_json_with_retries(\n messages,\n model=model,\n host=host,\n provider=provider,\n base_url=base_url,\n api_key_file=api_key_file,\n config=config,\n parser=parser,\n json_schema=json_schema,\n chat_fn=chat_fn,\n temperature=temperature,\n )\n parsed, raw, error, attempts, elapsed = primary\n fallback_model = fallback_model.strip()\n if parsed is not None or not fallback_model or fallback_model == model:\n return parsed, raw, error, attempts, elapsed, model, False, \"\"\n\n primary_error = error\n fallback = _call_json_with_retries(\n messages,\n model=fallback_model,\n host=host,\n provider=provider,\n base_url=base_url,\n api_key_file=api_key_file,\n config=config,\n parser=parser,\n json_schema=json_schema,\n chat_fn=chat_fn,\n temperature=temperature,\n )\n parsed, raw, fallback_error, fallback_attempts, fallback_elapsed = fallback\n if parsed is None:\n error = (\n f\"primary {model}: {primary_error}; \"\n f\"fallback {fallback_model}: {fallback_error}\"\n )\n else:\n error = \"\"\n return (\n parsed,\n raw,\n error,\n attempts + fallback_attempts,\n elapsed + fallback_elapsed,\n fallback_model,\n True,\n primary_error,\n )\n\n\ndef run_agent(\n agent_id: str,\n segment: dict,\n context: str,\n glossary: str,\n config: ReviewConfig,\n chat_fn: Callable = model_chat,\n) -> dict:\n agent = AGENTS[agent_id]\n ja = str(segment.get(\"ja\", \"\")).strip()\n zh = _segment_text(segment)\n system = (\n f\"你是{agent['name']}，只负责检查：{agent['focus']}。\"\n \"不要为了追求不同而改写正确译文。只能输出一个 JSON 对象，不要复述提示词。\"\n )\n prompt = f\"\"\"请审查目标字幕。\n\n上下文：\n{context}\n\n目标日文：{ja}\n目标中文：{zh}\n\n必须遵守的术语：\n{glossary}\n\n只输出：\n{{\"verdict\":\"ok|fix|suspicious\",\"suggested_zh\":\"完整建议译文，ok 时为空字符串\",\"reason\":\"简短依据\",\"confidence\":0.0}}\nconfidence 范围为 0 到 1。ASR 可疑用 suspicious；翻译确有错误才用 fix。\"\"\"\n model = config.agent_models.get(agent_id, config.review_model)\n (\n parsed,\n raw,\n error,\n attempts,\n elapsed,\n used_model,\n fallback_used,\n primary_error,\n ) = _call_json_with_fallback(\n [{\"role\": \"system\", \"content\": system}, {\"role\": \"user\", \"content\": prompt}],\n model=model,\n fallback_model=config.review_fallback_model,\n host=config.host,\n provider=config.provider,\n base_url=config.base_url,\n api_key_file=config.api_key_file,\n config=config,\n parser=parse_agent_response,\n json_schema=AGENT_SCHEMA,\n chat_fn=chat_fn,\n )\n result = {\n \"agent\": agent_id,\n \"agent_name\": agent[\"name\"],\n \"model\": used_model,\n \"primary_model\": model,\n \"fallback_used\": fallback_used,\n \"primary_error\": primary_error or None,\n \"attempts\": attempts,\n \"time_s\": round(elapsed, 2),\n \"raw_response\": raw,\n }\n if parsed is None:\n result.update(\n verdict=\"error\", suggested_zh=None, reason=error,\n confidence=0.0, error=error,\n )\n else:\n result.update(parsed)\n return result\n\n\ndef run_editor(\n segment: dict,\n context: str,\n glossary: str,\n agent_results: dict[str, dict],\n config: ReviewConfig,\n chat_fn: Callable = model_chat,\n) -> dict:\n feedback = [\n {\n \"agent\": agent_id,\n \"verdict\": result.get(\"verdict\"),\n \"suggested_zh\": result.get(\"suggested_zh\"),\n \"reason\": result.get(\"reason\"),\n \"confidence\": result.get(\"confidence\"),\n }\n for agent_id in AGENTS\n for result in [agent_results[agent_id]]\n ]\n system = (\n \"你是保守的字幕总编。只有审查意见揭示真实语义、术语、ASR 或中文表达问题时才替换；\"\n \"意见冲突、证据不足或只是同义改写时必须保留原译。只能输出一个 JSON 对象。\"\n )\n prompt = f\"\"\"上下文：\n{context}\n\n目标日文：{segment.get('ja', '')}\n原中文：{_segment_text(segment)}\n\n术语：\n{glossary}\n\n审查意见：\n{json.dumps(feedback, ensure_ascii=False)}\n\n只输出：\n{{\"decision\":\"keep|replace\",\"corrected_zh\":\"完整最终译文，keep 时可为空\",\"reason\":\"依据\",\"confidence\":0.0}}\"\"\"\n (\n parsed,\n raw,\n error,\n attempts,\n elapsed,\n used_model,\n fallback_used,\n primary_error,\n ) = _call_json_with_fallback(\n [{\"role\": \"system\", \"content\": system}, {\"role\": \"user\", \"content\": prompt}],\n model=config.editor_model,\n fallback_model=config.editor_fallback_model,\n host=config.editor_host,\n provider=config.editor_provider,\n base_url=config.editor_base_url,\n api_key_file=config.editor_api_key_file,\n config=config,\n parser=parse_editor_response,\n json_schema=EDITOR_SCHEMA,\n chat_fn=chat_fn,\n temperature=0.05,\n )\n result = {\n \"model\": used_model,\n \"primary_model\": config.editor_model,\n \"fallback_used\": fallback_used,\n \"primary_error\": primary_error or None,\n \"attempts\": attempts,\n \"time_s\": round(elapsed, 2),\n \"raw_response\": raw,\n }\n if parsed is None:\n result.update(\n decision=\"error\", corrected_zh=None, reason=error,\n confidence=0.0, error=error,\n )\n else:\n result.update(parsed)\n return result\n\n\ndef review_segment(\n position: int,\n segment: dict,\n context: str,\n glossary: str,\n config: ReviewConfig,\n chat_fn: Callable = model_chat,\n) -> dict:\n results = {}\n with ThreadPoolExecutor(max_workers=len(AGENTS)) as executor:\n futures = {\n executor.submit(\n run_agent, agent_id, segment, context, glossary, config, chat_fn\n ): agent_id\n for agent_id in AGENTS\n }\n for future in as_completed(futures):\n agent_id = futures[future]\n try:\n results[agent_id] = future.result()\n except Exception as error:\n results[agent_id] = {\n \"agent\": agent_id,\n \"agent_name\": AGENTS[agent_id][\"name\"],\n \"verdict\": \"error\",\n \"reason\": str(error),\n \"confidence\": 0.0,\n \"suggested_zh\": None,\n }\n\n results = {agent_id: results[agent_id] for agent_id in AGENTS}\n fix_votes = sum(\n result.get(\"verdict\") in {\"fix\", \"suspicious\"}\n and result.get(\"confidence\", 0.0) >= config.min_reviewer_confidence\n for result in results.values()\n )\n editor = None\n if fix_votes >= config.min_fix_votes:\n editor = run_editor(segment, context, glossary, results, config, chat_fn)\n\n original = _segment_text(segment)\n errors = sum(result.get(\"verdict\") == \"error\" for result in results.values())\n editor_error = bool(editor and editor.get(\"decision\") == \"error\")\n applied = bool(\n editor\n and errors == 0\n and not editor_error\n and editor.get(\"decision\") == \"replace\"\n and editor.get(\"corrected_zh\")\n and editor.get(\"corrected_zh\") != original\n and editor.get(\"confidence\", 0.0) >= config.min_editor_confidence\n )\n status = \"corrected\" if applied else \"approved\"\n if errors or editor_error:\n status = \"error\"\n elif fix_votes and not applied:\n status = \"needs_review\"\n\n return {\n \"position\": position,\n \"segment_key\": _segment_key(position, segment),\n \"ja\": str(segment.get(\"ja\", \"\")),\n \"original_zh\": original,\n \"final_zh\": editor[\"corrected_zh\"] if applied else original,\n \"status\": status,\n \"fix_votes\": fix_votes,\n \"applied\": applied,\n \"agent_results\": results,\n \"editor_result\": editor,\n }\n\n\ndef _load_progress(path: Optional[Path]) -> dict[str, dict]:\n completed = {}\n if not path or not path.exists():\n return completed\n with path.open(encoding=\"utf-8\") as file:\n for line in file:\n try:\n item = json.loads(line)\n except json.JSONDecodeError:\n continue\n key = item.get(\"segment_key\")\n if not key:\n continue\n editor_result = item.get(\"editor_result\") or {}\n if (\n item.get(\"status\") == \"error\"\n or editor_result.get(\"decision\") == \"error\"\n ):\n completed.pop(key, None)\n else:\n completed[key] = item\n return completed\n\n\ndef _append_progress(path: Optional[Path], item: dict) -> None:\n if not path:\n return\n path.parent.mkdir(parents=True, exist_ok=True)\n with path.open(\"a\", encoding=\"utf-8\") as file:\n file.write(json.dumps(item, ensure_ascii=False) + \"\\n\")\n file.flush()\n\n\ndef review_batch(\n segments: list[dict],\n max_workers: Optional[int] = None,\n *,\n config: Optional[ReviewConfig] = None,\n glossary: str = \"（未提供术语表）\",\n progress_path: str = \"\",\n completed_results: Optional[dict[str, dict]] = None,\n chat_fn: Callable = model_chat,\n) -> list[dict]:\n config = config or ReviewConfig()\n if max_workers is not None:\n config.max_workers = max_workers\n config.validate()\n progress = Path(progress_path) if progress_path else None\n cached = _load_progress(progress)\n cached.update(completed_results or {})\n results: list[Optional[dict]] = [None] * len(segments)\n pending = []\n for position, segment in enumerate(segments):\n key = _segment_key(position, segment)\n if key in cached:\n results[position] = cached[key]\n else:\n pending.append(position)\n\n print(\n f\"Reviewing {len(segments)} segments with {len(AGENTS)} agents; \"\n f\"resume={len(segments) - len(pending)}, pending={len(pending)}\"\n )\n print(f\"Reviewer model: {config.review_model}; editor: {config.editor_model}\")\n\n with ThreadPoolExecutor(max_workers=config.max_workers) as executor:\n futures = {\n executor.submit(\n review_segment,\n position,\n segments[position],\n _context_for(segments, position, config.context_window),\n glossary,\n config,\n chat_fn,\n ): position\n for position in pending\n }\n for future in as_completed(futures):\n position = futures[future]\n item = future.result()\n results[position] = item\n _append_progress(progress, item)\n print(\n f\" [{position + 1}/{len(segments)}] {item['status']} \"\n f\"votes={item['fix_votes']}\"\n )\n\n return [item for item in results if item is not None]\n\n\ndef _atomic_json(path: Path, value) -> None:\n path.parent.mkdir(parents=True, exist_ok=True)\n temporary = path.with_suffix(path.suffix + \".tmp\")\n temporary.write_text(\n json.dumps(value, ensure_ascii=False, indent=2), encoding=\"utf-8\"\n )\n temporary.replace(path)\n\n\ndef review_translation_file(\n input_path: str,\n reviewed_path: str,\n report_path: str,\n *,\n progress_path: str = \"\",\n glossary_path: str = \"\",\n config: Optional[ReviewConfig] = None,\n apply_fixes: bool = True,\n chat_fn: Callable = model_chat,\n) -> dict:\n source = Path(input_path)\n segments = json.loads(source.read_text(encoding=\"utf-8\"))\n if not isinstance(segments, list):\n raise ValueError(\"translated input must be a JSON list\")\n config = (config or ReviewConfig()).validate()\n glossary = load_glossary_prompt(glossary_path)\n reviews = []\n completed_results = {}\n for pass_index in range(1, MAX_INCOMPLETE_PASSES + 1):\n reviews = review_batch(\n segments,\n config=config,\n glossary=glossary,\n progress_path=progress_path,\n completed_results=completed_results,\n chat_fn=chat_fn,\n )\n error_count = sum(item[\"status\"] == \"error\" for item in reviews)\n completed_results.update(\n (item[\"segment_key\"], item)\n for item in reviews\n if item[\"status\"] != \"error\"\n )\n if error_count == 0:\n break\n if pass_index < MAX_INCOMPLETE_PASSES:\n print(\n f\"Retrying {error_count} incomplete multi-agent review(s); \"\n f\"pass {pass_index + 1}/{MAX_INCOMPLETE_PASSES}\"\n )\n\n reviewed_segments = []\n for segment, review in zip(segments, reviews):\n output = dict(segment)\n if apply_fixes and review[\"applied\"]:\n output[\"text\"] = review[\"final_zh\"]\n output[\"multi_agent_review\"] = {\n \"segment_key\": review[\"segment_key\"],\n \"status\": review[\"status\"],\n \"fix_votes\": review[\"fix_votes\"],\n \"applied\": bool(apply_fixes and review[\"applied\"]),\n \"original_zh\": review[\"original_zh\"],\n \"final_zh\": output.get(\"text\", _segment_text(output)),\n \"editor\": review.get(\"editor_result\"),\n }\n reviewed_segments.append(output)\n\n summary = {\n \"total\": len(reviews),\n \"approved\": sum(item[\"status\"] == \"approved\" for item in reviews),\n \"corrected\": sum(item[\"status\"] == \"corrected\" for item in reviews),\n \"needs_review\": sum(item[\"status\"] == \"needs_review\" for item in reviews),\n \"errors\": sum(item[\"status\"] == \"error\" for item in reviews),\n \"applied\": sum(bool(apply_fixes and item[\"applied\"]) for item in reviews),\n }\n report = {\n \"schema\": \"anime-accurate-sub/multi-agent-review-v2\",\n \"source\": str(source.resolve()),\n \"config\": asdict(config),\n \"apply_fixes\": apply_fixes,\n \"summary\": summary,\n \"reviews\": reviews,\n }\n _atomic_json(Path(reviewed_path), reviewed_segments)\n _atomic_json(Path(report_path), report)\n if summary[\"errors\"]:\n raise RuntimeError(\n f\"multi-agent review has {summary['errors']} incomplete segment(s); \"\n \"rerun with the same progress path\"\n )\n return report\n\n\ndef create_test_segments() -> list[dict]:\n return [\n {\"ja\": \"おはよう、唯\", \"text\": \"早安，唯\"},\n {\"ja\": \"今日も元気だね\", \"text\": \"今天也很有精神呢\"},\n {\"ja\": \"澪ちゃん、一緒に練習しない？\", \"text\": \"澪，要不要一起练习？\"},\n {\"ja\": \"ありがとう、唯。じゃあ放課後ね\", \"text\": \"谢谢你，唯。那放学后见\"},\n ]\n\n\ndef evaluate() -> None:\n output_root = Path(__file__).resolve().parent.parent / \"docs\" / \"evaluation\"\n source = output_root / \"S11.2_test_segments.json\"\n reviewed = output_root / \"S11.2_reviewed_segments.json\"\n report = output_root / \"S11.2_review_results.json\"\n _atomic_json(source, create_test_segments())\n result = review_translation_file(\n str(source), str(reviewed), str(report), apply_fixes=True\n )\n print(json.dumps(result[\"summary\"], ensure_ascii=False, indent=2))\n\n\ndef _load_config(path: str, args) -> ReviewConfig:\n data = {}\n if path:\n data = json.loads(Path(path).read_text(encoding=\"utf-8\"))\n overrides = {\n \"provider\": args.provider,\n \"host\": args.host,\n \"review_model\": args.model,\n \"base_url\": args.base_url,\n \"api_key_file\": args.api_key_file,\n \"editor_provider\": args.editor_provider,\n \"editor_host\": args.editor_host,\n \"editor_model\": args.editor_model,\n \"editor_base_url\": args.editor_base_url,\n \"editor_api_key_file\": args.editor_api_key_file,\n \"max_workers\": args.workers,\n \"min_fix_votes\": args.min_fix_votes,\n \"min_reviewer_confidence\": args.min_reviewer_confidence,\n \"min_editor_confidence\": args.min_editor_confidence,\n \"context_window\": args.context_window,\n }\n data.update({key: value for key, value in overrides.items() if value is not None})\n return ReviewConfig.from_dict(data).validate()\n\n\ndef main() -> None:\n parser = argparse.ArgumentParser(description=\"S11.2 auditable multi-agent review\")\n parser.add_argument(\"--input\", help=\"translated JSON list\")\n parser.add_argument(\"--output\", default=\"reviewed.json\", help=\"reviewed JSON\")\n parser.add_argument(\"--report\", default=\"multi_agent_review.json\")\n parser.add_argument(\"--progress\", default=\"\", help=\"resumable JSONL progress\")\n parser.add_argument(\"--glossary\", default=\"\")\n parser.add_argument(\"--config\", default=\"\", help=\"review configuration JSON\")\n parser.add_argument(\"--host\")\n parser.add_argument(\"--model\")\n parser.add_argument(\"--provider\", choices=(\"ollama\", \"openai\"))\n parser.add_argument(\"--base-url\")\n parser.add_argument(\"--api-key-file\")\n parser.add_argument(\"--editor-host\")\n parser.add_argument(\"--editor-model\")\n parser.add_argument(\"--editor-provider\", choices=(\"ollama\", \"openai\"))\n parser.add_argument(\"--editor-base-url\")\n parser.add_argument(\"--editor-api-key-file\")\n parser.add_argument(\"--workers\", type=int)\n parser.add_argument(\"--min-fix-votes\", type=int)\n parser.add_argument(\"--min-reviewer-confidence\", type=float)\n parser.add_argument(\"--min-editor-confidence\", type=float)\n parser.add_argument(\"--context-window\", type=int)\n parser.add_argument(\n \"--dry-run\", action=\"store_true\", help=\"review but do not apply corrections\"\n )\n parser.add_argument(\"--evaluate\", action=\"store_true\")\n args = parser.parse_args()\n\n if args.evaluate:\n evaluate()\n return\n if not args.input:\n parser.error(\"--input is required unless --evaluate is used\")\n result = review_translation_file(\n args.input,\n args.output,\n args.report,\n progress_path=args.progress,\n glossary_path=args.glossary,\n config=_load_config(args.config, args),\n apply_fixes=not args.dry_run,\n )\n print(json.dumps(result[\"summary\"], ensure_ascii=False, indent=2))\n\n\nif __name__ == \"__main__\":\n main()\n"}}, "scripts/translator_adapter.py": {"4d507921cd926e659e5a5a22d58eec34686302b8": {"t": 1785768275.1200373, "n": 35279, "h": "bf9c45edb7a58a03", "c": "# S14.1: Translator Adapter - abstract base + multi-backend implementations\n#\n# Architecture:\n# TranslatorAdapter (ABC)\n# +-- SakuraAdapter (Ollama Sakura-7B/14B)\n# +-- QwenAdapter (Ollama Qwen2.5)\n# +-- GalTranslAdapter (Ollama GalTransl-7B)\n# +-- ExternalAdapter (OpenAI-compatible API)\n#\n# Config file format:\n# {\n# \"backend\": \"sakura\",\n# \"sakura\": {\"model\": \"crosery/sakura-14b-qwen2.5-v1.0-q6k:latest\", \"host\": \"172.31.102.189\"},\n# \"qwen\": {\"model\": \"qwen2.5:7b\", \"host\": \"localhost\"},\n# \"galtransl\": {\"model\": \"crosery/GalTransl-7B-v2.6:Q6_k\", \"host\": \"172.31.102.189\"},\n# \"external\": {\"api_url\": \"...\", \"api_key\": \"...\", \"model\": \"...\"}\n# }\n#\n# Usage:\n# python scripts/translator_adapter.py --text \"おはよう\" --backend sakura\n# python scripts/translator_adapter.py --text \"おはよう\" --config translator_config.json\n# python scripts/translator_adapter.py --list-backends\n# python scripts/translator_adapter.py --evaluate\n\nimport abc\nimport argparse\nimport inspect\nimport json\nimport re\nimport socket\nimport sys\nimport time\nimport unicodedata\nimport urllib.error\nimport urllib.request\nfrom pathlib import Path\nfrom typing import Sequence\n\nproject_root = Path(__file__).resolve().parent.parent\nsys.path.insert(0, str(project_root))\n\nfrom scripts.plugin_system import load_plugins, plugin_registry # noqa: E402\n\nDEFAULT_CONFIG = {\n \"backend\": \"sakura\",\n \"sakura\": {\n \"model\": \"crosery/sakura-14b-qwen2.5-v1.0-q6k:latest\",\n \"host\": \"172.31.102.189\",\n \"timeout_s\": 300,\n \"max_retries\": 4,\n \"validation_retries\": 2,\n \"validation_fallback_backend\": \"galtransl\",\n \"validation_rescue_models\": [\n \"sensenova-6.7-flash-lite\",\n \"deepseek-v4-flash\",\n ],\n \"validation_rescue_base_url\": \"https://token.sensenova.cn/v1\",\n \"validation_rescue_api_key_file\": \"config/sensenova_apikeys\",\n \"validation_rescue_attempts\": 2,\n \"validation_quarantine_on_failure\": True,\n \"batch_size\": 16,\n \"num_ctx\": 4096,\n \"temperature\": 0.1,\n \"top_p\": 0.3,\n \"repeat_penalty\": 1.0,\n \"frequency_penalty\": 0.1,\n },\n \"qwen\": {\n \"model\": \"qwen2.5:7b\",\n \"host\": \"localhost\",\n },\n \"galtransl\": {\n \"model\": \"crosery/GalTransl-7B-v2.6:Q6_k\",\n \"host\": \"172.31.102.189\",\n \"timeout_s\": 300,\n \"max_retries\": 4,\n \"batch_size\": 10,\n \"num_ctx\": 4096,\n \"temperature\": 0.2,\n \"top_p\": 0.8,\n \"repeat_penalty\": 1.1,\n \"frequency_penalty\": 0.1,\n },\n \"external\": {\n \"api_url\": \"\",\n \"api_key\": \"\",\n \"model\": \"\",\n },\n}\n\n\ndef _call_batch_with_optional_context(\n adapter,\n texts,\n glossary_terms,\n context_before=None,\n context_after=None,\n):\n \"\"\"Call old and new translator plugins without masking adapter errors.\"\"\"\n method = adapter.translate_batch\n try:\n parameters = inspect.signature(method).parameters.values()\n supports_keywords = any(\n parameter.kind == inspect.Parameter.VAR_KEYWORD\n for parameter in parameters\n )\n parameter_names = {parameter.name for parameter in parameters}\n except (TypeError, ValueError):\n supports_keywords = False\n parameter_names = set()\n if supports_keywords or {\"context_before\", \"context_after\"} <= parameter_names:\n return method(\n texts,\n glossary_terms,\n context_before=context_before,\n context_after=context_after,\n )\n return method(texts, glossary_terms)\n\n\n# ============ Abstract Base ============\n\nclass TranslatorAdapter(abc.ABC):\n \"\"\"Abstract base class for all translation backends.\"\"\"\n\n def __init__(self, config: dict):\n self.config = config\n self.series_info = \"\"\n self._translation_models: dict[str, str] = {}\n self._translation_fallbacks: dict[str, bool] = {}\n self._translation_errors: dict[str, str] = {}\n\n @abc.abstractmethod\n def translate(self, text: str, context_before=None, context_after=None) -> str:\n \"\"\"Translate Japanese text to Chinese.\"\"\"\n pass\n\n def translate_batch(\n self,\n texts: Sequence[str],\n glossary_terms=None,\n context_before=None,\n context_after=None,\n ) -> list[str]:\n \"\"\"Translate ordered lines, falling back to the single-line interface.\"\"\"\n texts = list(texts)\n outer_before = list(context_before or [])\n outer_after = list(context_after or [])\n return [\n self.translate(\n text,\n context_before=outer_before + texts[:index],\n context_after=texts[index + 1:] + outer_after,\n )\n for index, text in enumerate(texts)\n ]\n\n def result_model(self, source: str) -> str:\n \"\"\"Return the model that produced a source line's latest translation.\"\"\"\n return self._translation_models.get(\n source,\n getattr(self, \"model\", self.name()),\n )\n\n def result_is_fallback(self, source: str) -> bool:\n return self._translation_fallbacks.get(source, False)\n\n def result_error(self, source: str) -> str | None:\n \"\"\"Return a terminal recovery error when a line was quarantined.\"\"\"\n return self._translation_errors.get(source)\n\n @abc.abstractmethod\n def name(self) -> str:\n \"\"\"Return backend name for display.\"\"\"\n pass\n\n @classmethod\n def from_config(cls, config: dict) -> \"TranslatorAdapter\":\n \"\"\"Factory: create adapter from config.\"\"\"\n _ensure_builtin_translator_plugins()\n backend = str(config.get(\"backend\", \"sakura\")).lower()\n return plugin_registry.create(\"translator\", backend, config)\n\n\n# ============ Ollama Base ============\n\nclass OllamaAdapter(TranslatorAdapter):\n \"\"\"Base for Ollama-based adapters.\"\"\"\n\n def __init__(self, config: dict, backend_key: str):\n super().__init__(config)\n backend_cfg = config.get(backend_key, {})\n self.model = backend_cfg.get(\"model\", \"\")\n self.host = str(backend_cfg.get(\"host\", \"localhost\")).rstrip(\"/\")\n if \"://\" in self.host:\n base_url = self.host\n elif re.search(r\":\\d+$\", self.host):\n base_url = f\"http://{self.host}\"\n else:\n base_url = f\"http://{self.host}:11434\"\n self.api_url = f\"{base_url}/api/chat\"\n self.timeout_s = int(backend_cfg.get(\"timeout_s\", 300))\n self.max_retries = int(backend_cfg.get(\"max_retries\", 4))\n self.retry_backoff_s = float(backend_cfg.get(\"retry_backoff_s\", 2.0))\n self.batch_size = int(backend_cfg.get(\"batch_size\", 16))\n self.num_ctx = int(backend_cfg.get(\"num_ctx\", 4096))\n self.temperature = float(backend_cfg.get(\"temperature\", 0.1))\n self.top_p = float(backend_cfg.get(\"top_p\", 0.3))\n self.repeat_penalty = float(backend_cfg.get(\"repeat_penalty\", 1.0))\n self.frequency_penalty = float(backend_cfg.get(\"frequency_penalty\", 0.1))\n self.validation_retries = int(backend_cfg.get(\"validation_retries\", 2))\n\n def _call(self, messages, num_predict=1024):\n payload = {\n \"model\": self.model,\n \"messages\": messages,\n \"stream\": False,\n \"options\": {\n \"temperature\": self.temperature,\n \"top_p\": self.top_p,\n \"repeat_penalty\": self.repeat_penalty,\n \"num_predict\": num_predict,\n \"frequency_penalty\": self.frequency_penalty,\n \"num_ctx\": self.num_ctx,\n },\n \"keep_alive\": \"30m\",\n }\n data = json.dumps(payload, ensure_ascii=False).encode(\"utf-8\")\n last_error = None\n for attempt in range(1, self.max_retries + 1):\n req = urllib.request.Request(\n self.api_url,\n data=data,\n headers={\"Content-Type\": \"application/json; charset=utf-8\"},\n )\n try:\n with urllib.request.urlopen(req, timeout=self.timeout_s) as response:\n result = json.loads(response.read().decode(\"utf-8\"))\n content = result.get(\"message\", {}).get(\"content\", \"\").strip()\n if not content:\n raise RuntimeError(\"Ollama returned an empty translation\")\n return content\n except (\n urllib.error.HTTPError,\n urllib.error.URLError,\n TimeoutError,\n socket.timeout,\n json.JSONDecodeError,\n RuntimeError,\n ) as error:\n last_error = error\n if attempt >= self.max_retries:\n break\n delay = min(self.retry_backoff_s * (2 ** (attempt - 1)), 20.0)\n print(\n f\" Translation request failed ({attempt}/{self.max_retries}): \"\n f\"{error}; retrying in {delay:.1f}s\"\n )\n time.sleep(delay)\n raise RuntimeError(\n f\"Ollama translation failed after {self.max_retries} attempts: {last_error}\"\n )\n\n @staticmethod\n def _system_prompt(series_info: str = \"\") -> str:\n prompt = (\n \"你是一个轻小说翻译模型，可以流畅通顺地以日本轻小说的风格将日文翻译成\"\n \"简体中文，并联系上下文正确使用人称代词，不擅自添加原文中没有的代词。\"\n )\n if series_info:\n prompt += f\"\\n\\n系列设定与角色信息：\\n{series_info}\"\n return prompt\n\n @staticmethod\n def _user_prompt(\n texts: Sequence[str],\n glossary_terms=None,\n ) -> str:\n raw_text = \"\\n\".join(texts)\n terms = list(glossary_terms or [])\n if not terms:\n return f\"将下面的待翻译日文逐行翻译成中文：\\n{raw_text}\"\n glossary = \"\\n\".join(f\"{source}->{target}\" for source, target in terms)\n return (\n \"根据以下术语表（可以为空）：\\n\"\n f\"{glossary}\\n\"\n \"将下面的待翻译日文根据对应关系和备注逐行翻译成中文：\\n\"\n f\"{raw_text}\"\n )\n\n @staticmethod\n def _context_block(context_before=None, context_after=None) -> str:\n parts = []\n if context_before:\n parts.append(\"已翻译前文（简体中文）：\\n\" + \"\\n\".join(context_before))\n if context_after:\n parts.append(\"参考下文：\\n\" + \"\\n\".join(context_after))\n if not parts:\n return \"\"\n return (\n \"\\n\\n参考对话（只供理解待翻译句子的指代、语气和省略信息，\"\n \"禁止翻译、复述或输出）：\\n\"\n + \"\\n\".join(parts)\n + \"\\n只输出 user 消息中待翻译日文的简体中文译文。\"\n )\n\n @staticmethod\n def _parse_lines(raw: str, expected: int) -> list[str] | None:\n raw = raw.strip()\n raw = re.sub(r\"^```(?:text|json)?\\s*\", \"\", raw, flags=re.IGNORECASE)\n raw = re.sub(r\"\\s*```$\", \"\", raw)\n if expected == 1:\n return [re.sub(r\"\\s+\", \" \", raw).strip()] if raw else None\n lines = [line.strip() for line in raw.splitlines() if line.strip()]\n if len(lines) == expected + 1 and lines[0].rstrip(\"：:\") in {\n \"翻译\",\n \"译文\",\n \"翻译结果\",\n }:\n lines = lines[1:]\n if len(lines) != expected:\n return None\n if all(re.match(r\"^\\d+[.、]\\s*\", line) for line in lines):\n lines = [re.sub(r\"^\\d+[.、]\\s*\", \"\", line) for line in lines]\n return lines\n\n @staticmethod\n def _tagged_texts(texts: Sequence[str]) -> list[str]:\n \"\"\"Attach stable IDs so a model cannot silently swap batch lines.\"\"\"\n return [f\"[[L{index:03d}]] {text}\" for index, text in enumerate(texts)]\n\n @staticmethod\n def _parse_tagged_lines(raw: str, expected: int) -> list[str] | None:\n \"\"\"Parse tagged translations, accepting brackets commonly normalized by LLMs.\"\"\"\n raw = raw.strip()\n raw = re.sub(r\"^```(?:text|json)?\\s*\", \"\", raw, flags=re.IGNORECASE)\n raw = re.sub(r\"\\s*```$\", \"\", raw)\n tagged = {}\n pattern = re.compile(\n r\"^\\s*(?:\\[\\[|\\[|【)\\s*L(\\d{1,4})\\s*(?:\\]\\]|\\]|】)\\s*(.*?)\\s*$\"\n )\n for line in raw.splitlines():\n if not line.strip():\n continue\n match = pattern.match(line)\n if not match:\n return None\n index = int(match.group(1))\n target = match.group(2).strip()\n if index >= expected or index in tagged or not target:\n return None\n tagged[index] = target\n if set(tagged) != set(range(expected)):\n return None\n return [tagged[index] for index in range(expected)]\n\n @staticmethod\n def _is_nonverbal_source(source: str) -> bool:\n \"\"\"Return whether a cue contains only vocal markers and punctuation.\"\"\"\n vocal_markers = {\"っ\", \"ッ\", \"ー\"}\n return bool(source) and all(\n character in vocal_markers\n or unicodedata.category(character)[0] in {\"P\", \"S\", \"Z\"}\n for character in source\n )\n\n @staticmethod\n def _valid_translation(source: str, target: str) -> bool:\n if not target or target.startswith(\"[API Error\"):\n return False\n if any(\n marker in target\n for marker in (\n \"将下面\",\n \"术语表\",\n \"翻译结果如下\",\n \"上文：\",\n \"上文:\",\n \"下文：\",\n \"下文:\",\n \"参考对话\",\n \"只供理解\",\n \"只输出 user 消息\",\n \"待翻译日文的简体中文译文\",\n )\n ):\n return False\n nonverbal_source = OllamaAdapter._is_nonverbal_source(source)\n if (\n not re.search(r\"[\\u3400-\\u9fffA-Za-z0-9]\", target)\n and not nonverbal_source\n ):\n return False\n if len(target) > max(40, len(source) * 5):\n return False\n if re.search(r\"(.)\\1{15,}\", target):\n return False\n if OllamaAdapter._has_suspicious_repeated_phrase(source, target):\n return False\n has_kana = bool(re.search(r\"[\\u3040-\\u30ff]\", source))\n if has_kana and target.strip() == source.strip():\n return False\n return True\n\n @staticmethod\n def _has_suspicious_repeated_phrase(source: str, target: str) -> bool:\n \"\"\"Detect three merged neighboring lines with a repeated Chinese lead-in.\"\"\"\n compact_target = re.sub(r\"[^\\u3400-\\u9fffA-Za-z0-9]\", \"\", target)\n compact_source = re.sub(r\"\\s+\", \"\", source)\n source_has_triplet = bool(re.search(r\"(.{1,4})\\1\\1\", compact_source))\n if source_has_triplet:\n return False\n for size in range(8, 2, -1):\n for start in range(0, max(0, len(compact_target) - size + 1)):\n phrase = compact_target[start:start + size]\n if compact_target.count(phrase) >= 3:\n return True\n return False\n\n def translate_batch(\n self,\n texts: Sequence[str],\n glossary_terms=None,\n context_before=None,\n context_after=None,\n ) -> list[str]:\n texts = [str(text).strip() for text in texts]\n if not texts:\n return []\n glossary_terms = self._matching_terms(texts, glossary_terms)\n context_enabled = context_before is not None or context_after is not None\n if context_enabled and len(texts) > 1:\n translated = []\n outer_before = list(context_before or [])\n outer_after = list(context_after or [])\n for index, text in enumerate(texts):\n translated.extend(\n self.translate_batch(\n [text],\n self._matching_terms([text], glossary_terms),\n context_before=outer_before + texts[:index],\n context_after=texts[index + 1:] + outer_after,\n )\n )\n return translated\n tagged_batch = len(texts) > 1\n prompt_texts = self._tagged_texts(texts) if tagged_batch else texts\n system_prompt = self._system_prompt(self.series_info) + self._context_block(\n context_before,\n context_after,\n )\n if tagged_batch:\n system_prompt += (\n \"\\n逐行翻译。每行开头的 [[Lnnn]] 是不可翻译的行号；译文必须保留\"\n \"相同行号，不得交换、合并或遗漏。\"\n )\n messages = [\n {\"role\": \"system\", \"content\": system_prompt},\n {\n \"role\": \"user\",\n \"content\": self._user_prompt(prompt_texts, glossary_terms),\n },\n ]\n try:\n raw = self._call(\n messages,\n num_predict=min(4096, max(512, len(texts) * 128)),\n )\n except Exception as error:\n if len(texts) == 1:\n return [\n self._fallback_invalid_single(\n texts[0],\n glossary_terms,\n f\"primary translation request failed: {error}\",\n context_before=context_before,\n context_after=context_after,\n )\n ]\n # A batch-level transport failure may be request-size-specific. Split\n # it before allowing each individual line to use its fallback chain.\n middle = len(texts) // 2\n return self.translate_batch(\n texts[:middle],\n glossary_terms,\n context_before=context_before,\n context_after=texts[middle:] + list(context_after or []),\n ) + self.translate_batch(\n texts[middle:],\n glossary_terms,\n context_before=list(context_before or []) + texts[:middle],\n context_after=context_after,\n )\n translated = (\n self._parse_tagged_lines(raw, len(texts))\n if tagged_batch\n else self._parse_lines(raw, len(texts))\n )\n if translated and all(\n self._valid_translation(source, target)\n for source, target in zip(texts, translated)\n ):\n for source in texts:\n self._translation_models[source] = self.model\n self._translation_fallbacks[source] = False\n return translated\n\n if len(texts) == 1:\n return [\n self._recover_invalid_single(\n texts[0],\n glossary_terms,\n raw,\n context_before=context_before,\n context_after=context_after,\n )\n ]\n # A model may occasionally merge or drop a line. Recursively reduce the\n # batch instead of guessing which result belongs to which timestamp.\n middle = len(texts) // 2\n return self.translate_batch(\n texts[:middle],\n glossary_terms,\n context_before=context_before,\n context_after=texts[middle:] + list(context_after or []),\n ) + self.translate_batch(\n texts[middle:],\n glossary_terms,\n context_before=list(context_before or []) + texts[:middle],\n context_after=context_after,\n )\n\n def _recover_invalid_single(\n self,\n source: str,\n glossary_terms,\n raw: str,\n context_before=None,\n context_after=None,\n ) -> str:\n \"\"\"Retry a rejected line with a constrained prompt before using a fallback.\"\"\"\n last_raw = raw\n for attempt in range(1, self.validation_retries + 1):\n messages = [\n {\n \"role\": \"system\",\n \"content\": (\n self._system_prompt(self.series_info)\n + self._context_block(context_before, context_after)\n + \"\\n只输出一行简洁的简体中文译文；禁止解释、重复、引号包裹和遗漏。\"\n ),\n },\n {\n \"role\": \"user\",\n \"content\": (\n \"上一次译文未通过字幕校验。请重新翻译且只输出一行：\\n\"\n + self._user_prompt([source], glossary_terms)\n ),\n },\n ]\n try:\n last_raw = self._call(messages, num_predict=256)\n except Exception as error:\n last_raw = (\n f\"validation retry {attempt}/{self.validation_retries} \"\n f\"request failed: {error}\"\n )\n print(f\" Translation {last_raw}\")\n break\n parsed = self._parse_lines(last_raw, 1)\n if parsed and self._valid_translation(source, parsed[0]):\n self._translation_models[source] = self.model\n self._translation_fallbacks[source] = False\n print(\n f\" Translation validation recovered on retry \"\n f\"{attempt}/{self.validation_retries}\"\n )\n return parsed[0]\n return self._fallback_invalid_single(\n source,\n glossary_terms,\n last_raw,\n context_before=context_before,\n context_after=context_after,\n )\n\n def _fallback_invalid_single(\n self,\n source: str,\n glossary_terms,\n raw: str,\n context_before=None,\n context_after=None,\n ) -> str:\n raise RuntimeError(\n f\"{self.name()} returned an invalid single-line translation after \"\n f\"{self.validation_retries} validation retries: {raw[:200]}\"\n )\n\n @staticmethod\n def _matching_terms(texts: Sequence[str], glossary_terms=None) -> list[tuple[str, str]]:\n joined = \"\\n\".join(texts)\n matches = [\n (source, target)\n for source, target in (glossary_terms or [])\n if source in joined\n ]\n return sorted(matches, key=lambda item: len(item[0]), reverse=True)\n\n def translate(self, text, context_before=None, context_after=None):\n return self.translate_batch(\n [text],\n context_before=context_before,\n context_after=context_after,\n )[0]\n\n\n# ============ Sakura ============\n\nclass SakuraAdapter(OllamaAdapter):\n def __init__(self, config: dict):\n super().__init__(config, \"sakura\")\n sakura_config = config.get(\"sakura\", {})\n fallback_name = sakura_config.get(\n \"validation_fallback_backend\", \"galtransl\"\n )\n fallback_config = config.get(fallback_name, {})\n self.fallback_adapter = None\n if fallback_name == \"galtransl\" and fallback_config.get(\"model\"):\n self.fallback_adapter = GalTranslAdapter(config)\n self.rescue_models = tuple(sakura_config.get(\"validation_rescue_models\", []))\n self.rescue_base_url = str(\n sakura_config.get(\"validation_rescue_base_url\", \"\")\n ).rstrip(\"/\")\n self.rescue_api_key_file = str(\n sakura_config.get(\"validation_rescue_api_key_file\", \"\")\n )\n self.rescue_attempts = max(\n 1, int(sakura_config.get(\"validation_rescue_attempts\", 2))\n )\n self.quarantine_on_failure = bool(\n sakura_config.get(\"validation_quarantine_on_failure\", True)\n )\n\n def _rescue_invalid_single(\n self,\n source: str,\n glossary_terms,\n context_before=None,\n context_after=None,\n ) -> str | None:\n \"\"\"Use the rotating SenseNova accounts only after both local models fail.\"\"\"\n if not self.rescue_models or not self.rescue_base_url:\n return None\n # Import lazily so local-only translation has no review-provider dependency.\n from scripts.review_agents import model_chat\n\n messages = [\n {\n \"role\": \"system\",\n \"content\": (\n self._system_prompt(self.series_info)\n + self._context_block(context_before, context_after)\n + \"\\n本地翻译模型未通过校验。只输出一行简洁准确的简体中文译文；\"\n \"禁止解释、引号包裹、重复或输出日文。\"\n ),\n },\n {\n \"role\": \"user\",\n \"content\": self._user_prompt([source], glossary_terms),\n },\n ]\n for model in self.rescue_models:\n for attempt in range(1, self.rescue_attempts + 1):\n try:\n raw = model_chat(\n messages,\n provider=\"openai\",\n base_url=self.rescue_base_url,\n api_key_file=self.rescue_api_key_file,\n model=model,\n json_mode=False,\n temperature=0.1,\n timeout_s=self.timeout_s,\n )\n except Exception as error:\n print(\n f\" Translation rescue {model} failed \"\n f\"({attempt}/{self.rescue_attempts}): {error}\"\n )\n continue\n parsed = self._parse_lines(raw, 1)\n if parsed and self._valid_translation(source, parsed[0]):\n self._translation_models[source] = model\n self._translation_fallbacks[source] = True\n self._translation_errors.pop(source, None)\n print(f\" Translation validation rescue: {self.model} -> {model}\")\n return parsed[0]\n print(\n f\" Translation rescue {model} returned invalid output \"\n f\"({attempt}/{self.rescue_attempts})\"\n )\n return None\n\n def _quarantine_invalid_single(self, source: str, reason: str) -> str:\n \"\"\"Preserve a failed line for downstream AI review without poisoning TM.\"\"\"\n if not self.quarantine_on_failure:\n raise RuntimeError(reason)\n self._translation_models[source] = \"untranslated-source-quarantine\"\n self._translation_fallbacks[source] = True\n self._translation_errors[source] = reason[:1000]\n print(\n \" WARNING: translation quarantined for multi-agent/MQM review: \"\n f\"{source[:80]}\"\n )\n return source\n\n def _fallback_invalid_single(\n self,\n source: str,\n glossary_terms,\n raw: str,\n context_before=None,\n context_after=None,\n ) -> str:\n fallback_error = \"configured GalTransl fallback is unavailable\"\n if self.fallback_adapter is not None:\n self.fallback_adapter.series_info = self.series_info\n try:\n target = _call_batch_with_optional_context(\n self.fallback_adapter,\n [source],\n glossary_terms,\n context_before=context_before,\n context_after=context_after,\n )[0]\n except Exception as error:\n fallback_error = str(error)\n else:\n self._translation_models[source] = self.fallback_adapter.result_model(source)\n self._translation_fallbacks[source] = True\n self._translation_errors.pop(source, None)\n print(\n f\" Translation validation fallback: {self.model} -> \"\n f\"{self._translation_models[source]}\"\n )\n return target\n\n rescued = self._rescue_invalid_single(\n source,\n glossary_terms,\n context_before=context_before,\n context_after=context_after,\n )\n if rescued is not None:\n return rescued\n reason = (\n \"Sakura validation retries, configured GalTransl fallback, and \"\n f\"SenseNova rescue all failed; GalTransl error: {fallback_error}; \"\n f\"last Sakura output: {raw[:200]}\"\n )\n return self._quarantine_invalid_single(source, reason)\n\n def name(self):\n return f\"Sakura ({self.model})\"\n\n\n# ============ Qwen ============\n\nclass QwenAdapter(OllamaAdapter):\n def __init__(self, config: dict):\n super().__init__(config, \"qwen\")\n\n def name(self):\n return f\"Qwen ({self.model})\"\n\n\n# ============ GalTransl ============\n\nclass GalTranslAdapter(OllamaAdapter):\n def __init__(self, config: dict):\n super().__init__(config, \"galtransl\")\n\n def name(self):\n return f\"GalTransl ({self.model})\"\n\n\n# ============ External API ============\n\nclass ExternalAdapter(TranslatorAdapter):\n \"\"\"OpenAI-compatible API adapter.\"\"\"\n\n def __init__(self, config: dict):\n super().__init__(config)\n ext = config.get(\"external\", {})\n self.api_url = ext.get(\"api_url\", \"\")\n self.api_key = ext.get(\"api_key\", \"\")\n self.model = ext.get(\"model\", \"\")\n\n def name(self):\n return f\"External ({self.model})\"\n\n def translate(self, text, context_before=None, context_after=None):\n if not self.api_url:\n return \"[External API not configured]\"\n\n headers = {\n \"Content-Type\": \"application/json\",\n \"Authorization\": f\"Bearer {self.api_key}\",\n }\n payload = {\n \"model\": self.model,\n \"messages\": [\n {\n \"role\": \"system\",\n \"content\": self._external_system_prompt()\n + OllamaAdapter._context_block(context_before, context_after),\n },\n {\n \"role\": \"user\",\n \"content\": OllamaAdapter._user_prompt(\n [text],\n ),\n },\n ],\n \"temperature\": 0.1,\n }\n data = json.dumps(payload).encode(\"utf-8\")\n req = urllib.request.Request(\n self.api_url, data=data, headers=headers,\n )\n try:\n resp = urllib.request.urlopen(req, timeout=120)\n result = json.loads(resp.read().decode(\"utf-8\"))\n return result.get(\"choices\", [{}])[0].get(\"message\", {}).get(\"content\", \"\").strip()\n except Exception as e:\n return f\"[API Error: {e}]\"\n\n def _external_system_prompt(self) -> str:\n prompt = \"你是一个轻小说翻译模型，可以将日语翻译成简体中文。\"\n if self.series_info:\n prompt += f\"\\n\\n系列设定与角色信息：\\n{self.series_info}\"\n return prompt\n\n\ndef _ensure_builtin_translator_plugins() -> None:\n for name, adapter_class in {\n \"sakura\": SakuraAdapter,\n \"qwen\": QwenAdapter,\n \"galtransl\": GalTranslAdapter,\n \"external\": ExternalAdapter,\n }.items():\n plugin_registry.register_if_missing(\n \"translator\",\n name,\n adapter_class,\n source=\"builtin:translator_adapter\",\n description=f\"Built-in {adapter_class.__name__}\",\n )\n\n\n# ============ Config ============\n\ndef load_config(path: str = \"\") -> dict:\n \"\"\"Load translator config from file, or return default.\"\"\"\n if path:\n config_path = Path(path)\n if config_path.exists():\n with open(config_path, encoding=\"utf-8\") as f:\n return json.load(f)\n return dict(DEFAULT_CONFIG)\n\n\ndef save_default_config(path: str = \"translator_config.json\"):\n \"\"\"Save default config to file.\"\"\"\n config_path = Path(path)\n with open(config_path, \"w\", encoding=\"utf-8\") as f:\n json.dump(DEFAULT_CONFIG, f, ensure_ascii=False, indent=2)\n print(f\"Default config saved: {path}\")\n\n\n# ============ Evaluate ============\n\ndef evaluate():\n print(\"\\n============================================================\")\n print(\"S14.1 TRANSLATOR ADAPTER EVALUATION\")\n print(\"============================================================\")\n\n config = load_config()\n\n test_texts = [\n \"おはよう、唯\",\n \"今日も元気だね\",\n \"澪ちゃん、一緒に練習しない？\",\n ]\n\n backends_to_test = [\"sakura\", \"qwen\", \"galtransl\"]\n results = {}\n\n for backend_name in backends_to_test:\n cfg = dict(config)\n cfg[\"backend\"] = backend_name\n try:\n adapter = TranslatorAdapter.from_config(cfg)\n print(f\"\\n--- Backend: {adapter.name()} ---\")\n\n backend_results = []\n for text in test_texts:\n t0 = time.time()\n try:\n result = adapter.translate(text)\n elapsed = time.time() - t0\n except Exception as e:\n result = f\"[ERROR: {e}]\"\n elapsed = time.time() - t0\n backend_results.append({\n \"ja\": text,\n \"zh\": result,\n \"time_s\": round(elapsed, 2),\n })\n print(f\" {text[:20]} -> {result[:30]} ({elapsed:.1f}s)\")\n\n results[backend_name] = {\n \"name\": adapter.name(),\n \"samples\": backend_results,\n }\n except Exception as e:\n print(f\" SKIP: {e}\")\n results[backend_name] = {\"name\": backend_name, \"error\": str(e)}\n\n # Save\n out_path = project_root / \"docs\" / \"evaluation\" / \"S14.1_adapter_results.json\"\n with open(out_path, \"w\", encoding=\"utf-8\") as f:\n json.dump({\n \"config\": {k: v for k, v in config.items() if k != \"external\"},\n \"results\": results,\n }, f, ensure_ascii=False, indent=2)\n print(f\"\\nResults saved: {out_path}\")\n\n print(\"\\n============================================================\")\n print(\"EVALUATION COMPLETE\")\n print(\"============================================================\")\n\n\n# ============ CLI ============\n\ndef main():\n parser = argparse.ArgumentParser(description=\"S14.1 Translator Adapter\")\n parser.add_argument(\"--text\", type=str, help=\"Japanese text to translate\")\n parser.add_argument(\"--backend\", type=str, help=\"Translation backend or plugin name\")\n parser.add_argument(\"--config\", type=str, default=\"\", help=\"Config file path\")\n parser.add_argument(\"--plugin\", action=\"append\", default=[],\n help=\"Trusted local plugin .py file (repeatable)\")\n parser.add_argument(\"--list-backends\", action=\"store_true\", help=\"List available backends\")\n parser.add_argument(\"--save-config\", type=str, help=\"Save default config to file\")\n parser.add_argument(\"--evaluate\", action=\"store_true\")\n args = parser.parse_args()\n\n load_plugins(args.plugin)\n _ensure_builtin_translator_plugins()\n\n if args.list_backends:\n print(\"Available backends:\")\n for spec in plugin_registry.specs(\"translator\"):\n print(f\" {spec['name']:<12} {spec['description']} [{spec['source']}]\")\n return\n\n if args.save_config:\n save_default_config(args.save_config)\n return\n\n if args.evaluate:\n evaluate()\n return\n\n if args.text:\n config = load_config(args.config)\n if args.backend:\n config[\"backend\"] = args.backend\n adapter = TranslatorAdapter.from_config(config)\n result = adapter.translate(args.text)\n print(result)\n return\n\n parser.print_help()\n\n\nif __name__ == \"__main__\":\n main()\n"}}, "config/quality_mqm.sensenova.json": {"4d507921cd926e659e5a5a22d58eec34686302b8": {"t": 1785768284.0659661, "n": 580, "h": "4e97e01b0de54431", "c": "{\n \"provider\": \"openai\",\n \"base_url\": \"https://token.sensenova.cn/v1\",\n \"api_key_file\": \"config/sensenova_apikeys\",\n \"judge_models\": [\n \"sensenova-6.7-flash-lite\",\n \"deepseek-v4-flash\"\n ],\n \"judge_fallback_models\": {\n \"sensenova-6.7-flash-lite\": \"glm-5.2\"\n },\n \"editor_model\": \"sensenova-6.7-flash-lite\",\n \"editor_fallback_model\": \"glm-5.2\",\n \"max_workers\": 2,\n \"context_window\": 5,\n \"review_threshold\": 80,\n \"min_judge_confidence\": 0.75,\n \"min_editor_confidence\": 0.85,\n \"min_final_score\": 80,\n \"min_improvement\": 5,\n \"retries\": 6,\n \"timeout_s\": 600\n}\n"}}}}