{"repo": "chesser-y/noteclaw", "files": {"nanobot/nanobot/providers/registry.py": {"e93321748de7b073e943b94157f9226113957756": {"t": 1785595187.5422544, "n": 16786, "h": "2ebb538015870610", "c": "\"\"\"\nProvider Registry — single source of truth for LLM provider metadata.\n\nAdding a new provider:\n  1. Add a ProviderSpec to PROVIDERS below.\n  2. Add a field to ProvidersConfig in config/schema.py.\n  Done. Env vars, config matching, status display all derive from here.\n\nOrder matters — it controls match priority and fallback. Gateways first.\nEvery entry writes out all fields so you can copy-paste as a template.\n\"\"\"\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass\nfrom typing import Any\n\nfrom pydantic.alias_generators import to_snake\n\n\n@dataclass(frozen=True)\nclass ProviderSpec:\n    \"\"\"One LLM provider's metadata. See PROVIDERS below for real examples.\n\n    Placeholders in env_extras values:\n      {api_key}  — the user's API key\n      {api_base} — api_base from config, or this spec's default_api_base\n    \"\"\"\n\n    # identity\n    name: str  # config field name, e.g. \"dashscope\"\n    keywords: tuple[str, ...]  # model-name keywords for matching (lowercase)\n    env_key: str  # env var for API key, e.g. \"DASHSCOPE_API_KEY\"\n    display_name: str = \"\"  # shown in `nanobot status`\n\n    # which provider implementation to use\n    # \"openai_compat\" | \"anthropic\" | \"azure_openai\" | \"openai_codex\" | \"github_copilot\" | \"bedrock\"\n    backend: str = \"openai_compat\"\n\n    # extra env vars, e.g. ((\"ZHIPUAI_API_KEY\", \"{api_key}\"),)\n    env_extras: tuple[tuple[str, str], ...] = ()\n\n    # gateway / local detection\n    is_gateway: bool = False  # routes any model (OpenRouter, AiHubMix)\n    is_local: bool = False  # local deployment (vLLM, Ollama)\n    detect_by_key_prefix: str = \"\"  # match api_key prefix, e.g. \"sk-or-\"\n    detect_by_base_keyword: str = \"\"  # match substring in api_base URL\n    default_api_base: str = \"\"  # OpenAI-compatible base URL for this provider\n\n    # gateway behavior\n    strip_model_prefix: bool = False  # strip \"provider/\" before sending to gateway\n    supports_max_completion_tokens: bool = False\n\n    # per-model param overrides, e.g. ((\"kimi-k2.5\", {\"temperature\": 1.0}),)\n    model_overrides: tuple[tuple[str, dict[str, Any]], ...] = ()\n\n    # OAuth-based providers (e.g., OpenAI Codex) don't use API keys\n    is_oauth: bool = False\n\n    # Direct providers skip API-key validation (user supplies everything)\n    is_direct: bool = False\n\n    # Provider supports cache_control on content blocks (e.g. Anthropic prompt caching)\n    supports_prompt_caching: bool = False\n\n    # How to inject the thinking on/off toggle into extra_body.\n    # \"\"              — no extra_body needed (default)\n    # \"thinking_type\" — {\"thinking\": {\"type\": \"enabled\"/\"disabled\"}}\n    #                   (DeepSeek, VolcEngine, BytePlus)\n    # \"enable_thinking\" — {\"enable_thinking\": true/false}  (DashScope)\n    # \"reasoning_split\" — {\"reasoning_split\": true/false}  (MiniMax)\n    thinking_style: str = \"\"\n\n    # When True, treat the \"reasoning\" response field as formal content\n    # when \"content\" is empty.  Only set this for providers (e.g. StepFun)\n    # whose API returns the actual answer in \"reasoning\" instead of \"content\".\n    reasoning_as_content: bool = False\n\n    @property\n    def label(self) -> str:\n        return self.display_name or self.name.title()\n\n\n# ---------------------------------------------------------------------------\n# PROVIDERS — the registry. Order = priority. Copy any entry as template.\n# ---------------------------------------------------------------------------\n\nPROVIDERS: tuple[ProviderSpec, ...] = (\n    # === Custom (direct OpenAI-compatible endpoint) ========================\n    ProviderSpec(\n        name=\"custom\",\n        keywords=(),\n        env_key=\"\",\n        display_name=\"Custom\",\n        backend=\"openai_compat\",\n        is_direct=True,\n    ),\n\n    # === Azure OpenAI (direct API calls with API version 2024-10-21) =====\n    ProviderSpec(\n        name=\"azure_openai\",\n        keywords=(\"azure\", \"azure-openai\"),\n        env_key=\"\",\n        display_name=\"Azure OpenAI\",\n        backend=\"azure_openai\",\n        is_direct=True,\n    ),\n    # === AWS Bedrock (native Converse API via bedrock-runtime) =============\n    ProviderSpec(\n        name=\"bedrock\",\n        keywords=(\n            \"bedrock\",\n            \"anthropic.claude\",\n            \"amazon.nova\",\n            \"meta.\",\n            \"mistral.\",\n            \"cohere.\",\n            \"qwen.\",\n            \"deepseek.\",\n            \"openai.gpt-oss\",\n            \"ai21.\",\n            \"moonshot.\",\n            \"writer.\",\n            \"zai.\",\n        ),\n        env_key=\"AWS_BEARER_TOKEN_BEDROCK\",\n        display_name=\"AWS Bedrock\",\n        backend=\"bedrock\",\n        is_direct=True,\n    ),\n    # === Gateways (detected by api_key / api_base, not model name) =========\n    # Gateways can route any model, so they win in fallback.\n    # OpenRouter: global gateway, keys start with \"sk-or-\"\n    ProviderSpec(\n        name=\"openrouter\",\n        keywords=(\"openrouter\",),\n        env_key=\"OPENROUTER_API_KEY\",\n        display_name=\"OpenRouter\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_key_prefix=\"sk-or-\",\n        detect_by_base_keyword=\"openrouter\",\n        default_api_base=\"https://openrouter.ai/api/v1\",\n        supports_prompt_caching=True,\n    ),\n    # Hugging Face Inference Providers: OpenAI-compatible router for chat models.\n    ProviderSpec(\n        name=\"huggingface\",\n        keywords=(\"huggingface\", \"hugging-face\"),\n        env_key=\"HF_TOKEN\",\n        display_name=\"Hugging Face\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_key_prefix=\"hf_\",\n        detect_by_base_keyword=\"huggingface\",\n        default_api_base=\"https://router.huggingface.co/v1\",\n    ),\n    # AiHubMix: global gateway, OpenAI-compatible interface.\n    # strip_model_prefix=True: doesn't understand \"anthropic/claude-3\",\n    # strips to bare \"claude-3\".\n    ProviderSpec(\n        name=\"aihubmix\",\n        keywords=(\"aihubmix\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"AiHubMix\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"aihubmix\",\n        default_api_base=\"https://aihubmix.com/v1\",\n        strip_model_prefix=True,\n    ),\n    # SiliconFlow (硅基流动): OpenAI-compatible gateway, model names keep org prefix\n    ProviderSpec(\n        name=\"siliconflow\",\n        keywords=(\"siliconflow\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"SiliconFlow\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"siliconflow\",\n        default_api_base=\"https://api.siliconflow.cn/v1\",\n    ),\n\n    # VolcEngine (火山引擎): OpenAI-compatible gateway, pay-per-use models\n    ProviderSpec(\n        name=\"volcengine\",\n        keywords=(\"volcengine\", \"volces\", \"ark\"),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"VolcEngine\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"volces\",\n        default_api_base=\"https://ark.cn-beijing.volces.com/api/v3\",\n        thinking_style=\"thinking_type\",\n        supports_max_completion_tokens=True,\n    ),\n\n    # VolcEngine Coding Plan (火山引擎 Coding Plan): same key as volcengine\n    ProviderSpec(\n        name=\"volcengine_coding_plan\",\n        keywords=(\"volcengine-plan\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"VolcEngine Coding Plan\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        default_api_base=\"https://ark.cn-beijing.volces.com/api/coding/v3\",\n        strip_model_prefix=True,\n        thinking_style=\"thinking_type\",\n        supports_max_completion_tokens=True,\n    ),\n\n    # BytePlus: VolcEngine international, pay-per-use models\n    ProviderSpec(\n        name=\"byteplus\",\n        keywords=(\"byteplus\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"BytePlus\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"bytepluses\",\n        default_api_base=\"https://ark.ap-southeast.bytepluses.com/api/v3\",\n        strip_model_prefix=True,\n        thinking_style=\"thinking_type\",\n    ),\n\n    # BytePlus Coding Plan: same key as byteplus\n    ProviderSpec(\n        name=\"byteplus_coding_plan\",\n        keywords=(\"byteplus-plan\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"BytePlus Coding Plan\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        default_api_base=\"https://ark.ap-southeast.bytepluses.com/api/coding/v3\",\n        strip_model_prefix=True,\n        thinking_style=\"thinking_type\",\n    ),\n\n\n    # === Standard providers (matched by model-name keywords) ===============\n    # Anthropic: native Anthropic SDK\n    ProviderSpec(\n        name=\"anthropic\",\n        keywords=(\"anthropic\", \"claude\"),\n        env_key=\"ANTHROPIC_API_KEY\",\n        display_name=\"Anthropic\",\n        backend=\"anthropic\",\n        supports_prompt_caching=True,\n    ),\n    # OpenAI: SDK default base URL (no override needed)\n    ProviderSpec(\n        name=\"openai\",\n        keywords=(\"openai\", \"gpt\"),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"OpenAI\",\n        backend=\"openai_compat\",\n        supports_max_completion_tokens=True,\n    ),\n    # OpenAI Codex: OAuth-based, dedicated provider\n    ProviderSpec(\n        name=\"openai_codex\",\n        keywords=(\"openai-codex\",),\n        env_key=\"\",\n        display_name=\"OpenAI Codex\",\n        backend=\"openai_codex\",\n        detect_by_base_keyword=\"codex\",\n        default_api_base=\"https://chatgpt.com/backend-api\",\n        is_oauth=True,\n    ),\n    # GitHub Copilot: OAuth-based\n    ProviderSpec(\n        name=\"github_copilot\",\n        keywords=(\"github_copilot\", \"copilot\"),\n        env_key=\"\",\n        display_name=\"Github Copilot\",\n        backend=\"github_copilot\",\n        default_api_base=\"https://api.githubcopilot.com\",\n        strip_model_prefix=True,\n        is_oauth=True,\n        supports_max_completion_tokens=True,\n    ),\n    # DeepSeek: OpenAI-compatible at api.deepseek.com\n    ProviderSpec(\n        name=\"deepseek\",\n        keywords=(\"deepseek\",),\n        env_key=\"DEEPSEEK_API_KEY\",\n        display_name=\"DeepSeek\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.deepseek.com\",\n        thinking_style=\"thinking_type\",\n    ),\n    # Gemini: Google's OpenAI-compatible endpoint\n    ProviderSpec(\n        name=\"gemini\",\n        keywords=(\"gemini\", \"gemma\"),\n        env_key=\"GEMINI_API_KEY\",\n        display_name=\"Gemini\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://generativelanguage.googleapis.com/v1beta/openai/\",\n    ),\n    # Zhipu (智谱): OpenAI-compatible at open.bigmodel.cn\n    ProviderSpec(\n        name=\"zhipu\",\n        keywords=(\"zhipu\", \"glm\", \"zai\"),\n        env_key=\"ZAI_API_KEY\",\n        display_name=\"Zhipu AI\",\n        backend=\"openai_compat\",\n        env_extras=((\"ZHIPUAI_API_KEY\", \"{api_key}\"),),\n        default_api_base=\"https://open.bigmodel.cn/api/paas/v4\",\n    ),\n    # DashScope (通义): Qwen models, OpenAI-compatible endpoint\n    ProviderSpec(\n        name=\"dashscope\",\n        keywords=(\"qwen\", \"dashscope\"),\n        env_key=\"DASHSCOPE_API_KEY\",\n        display_name=\"DashScope\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://dashscope.aliyuncs.com/compatible-mode/v1\",\n        thinking_style=\"enable_thinking\",\n    ),\n    # Moonshot (月之暗面): Kimi K2.5 / K2.6 enforce temperature >= 1.0.\n    ProviderSpec(\n        name=\"moonshot\",\n        keywords=(\"moonshot\", \"kimi\"),\n        env_key=\"MOONSHOT_API_KEY\",\n        display_name=\"Moonshot\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.moonshot.ai/v1\",\n        model_overrides=(\n            (\"kimi-k2.5\", {\"temperature\": 1.0}),\n            (\"kimi-k2.6\", {\"temperature\": 1.0}),\n        ),\n    ),\n    # MiniMax: OpenAI-compatible API\n    ProviderSpec(\n        name=\"minimax\",\n        keywords=(\"minimax\",),\n        env_key=\"MINIMAX_API_KEY\",\n        display_name=\"MiniMax\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.minimax.io/v1\",\n        thinking_style=\"reasoning_split\",\n    ),\n    # MiniMax Anthropic-compatible endpoint: supports thinking mode\n    ProviderSpec(\n        name=\"minimax_anthropic\",\n        keywords=(\"minimax_anthropic\",),\n        env_key=\"MINIMAX_API_KEY\",\n        display_name=\"MiniMax (Anthropic)\",\n        backend=\"anthropic\",\n        default_api_base=\"https://api.minimax.io/anthropic\",\n    ),\n    # Mistral AI: OpenAI-compatible API\n    ProviderSpec(\n        name=\"mistral\",\n        keywords=(\"mistral\",),\n        env_key=\"MISTRAL_API_KEY\",\n        display_name=\"Mistral\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.mistral.ai/v1\",\n    ),\n    # Step Fun (阶跃星辰): OpenAI-compatible API\n    ProviderSpec(\n        name=\"stepfun\",\n        keywords=(\"stepfun\", \"step\"),\n        env_key=\"STEPFUN_API_KEY\",\n        display_name=\"Step Fun\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.stepfun.com/v1\",\n        reasoning_as_content=True,\n    ),\n    # Xiaomi MIMO (小米): OpenAI-compatible API\n    # Hosted API (api.xiaomimimo.com) accepts {\"thinking\": {\"type\": \"enabled\"|\"disabled\"}}\n    # to toggle reasoning, matching the existing thinking_type style.\n    ProviderSpec(\n        name=\"xiaomi_mimo\",\n        keywords=(\"xiaomi_mimo\", \"mimo\"),\n        env_key=\"XIAOMIMIMO_API_KEY\",\n        display_name=\"Xiaomi MIMO\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.xiaomimimo.com/v1\",\n        thinking_style=\"thinking_type\",\n    ),\n    # LongCat: OpenAI-compatible API\n    ProviderSpec(\n        name=\"longcat\",\n        keywords=(\"longcat\",),\n        env_key=\"LONGCAT_API_KEY\",\n        display_name=\"LongCat\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.longcat.chat/openai/v1\",\n    ),\n    # === Local deployment (matched by config key, NOT by api_base) =========\n    # vLLM / any OpenAI-compatible local server\n    ProviderSpec(\n        name=\"vllm\",\n        keywords=(\"vllm\",),\n        env_key=\"HOSTED_VLLM_API_KEY\",\n        display_name=\"vLLM/Local\",\n        backend=\"openai_compat\",\n        is_local=True,\n    ),\n    # Ollama (local, OpenAI-compatible)\n    ProviderSpec(\n        name=\"ollama\",\n        keywords=(\"ollama\", \"nemotron\"),\n        env_key=\"OLLAMA_API_KEY\",\n        display_name=\"Ollama\",\n        backend=\"openai_compat\",\n        is_local=True,\n        detect_by_base_keyword=\"11434\",\n        default_api_base=\"http://localhost:11434/v1\",\n    ),\n    # LM Studio (local, OpenAI-compatible)\n    ProviderSpec(\n        name=\"lm_studio\",\n        keywords=(\"lm-studio\", \"lmstudio\", \"lm_studio\"),\n        env_key=\"LM_STUDIO_API_KEY\",\n        display_name=\"LM Studio\",\n        backend=\"openai_compat\",\n        is_local=True,\n        detect_by_base_keyword=\"1234\",\n        default_api_base=\"http://localhost:1234/v1\",\n    ),\n    # === OpenVINO Model Server (direct, local, OpenAI-compatible at /v3) ===\n    ProviderSpec(\n        name=\"ovms\",\n        keywords=(\"openvino\", \"ovms\"),\n        env_key=\"\",\n        display_name=\"OpenVINO Model Server\",\n        backend=\"openai_compat\",\n        is_direct=True,\n        is_local=True,\n        default_api_base=\"http://localhost:8000/v3\",\n    ),\n    # === NVIDIA NIM (NVIDIA Inference Microservices) =======================\n    # Keys start with \"nvapi-\", base URL at integrate.api.nvidia.com\n    ProviderSpec(\n        name=\"nvidia\",\n        keywords=(\"nvidia\", \"nemotron\", \"nvapi\"),\n        env_key=\"NVIDIA_NIM_API_KEY\",\n        display_name=\"NVIDIA NIM\",\n        backend=\"openai_compat\",\n        is_gateway=False,\n        detect_by_key_prefix=\"nvapi-\",\n        detect_by_base_keyword=\"nvidia.com\",\n        default_api_base=\"https://integrate.api.nvidia.com/v1\",\n    ),\n    # === Auxiliary (not a primary LLM provider) ============================\n    # Groq: mainly used for Whisper voice transcription, also usable for LLM\n    ProviderSpec(\n        name=\"groq\",\n        keywords=(\"groq\",),\n        env_key=\"GROQ_API_KEY\",\n        display_name=\"Groq\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.groq.com/openai/v1\",\n    ),\n    # Qianfan (百度千帆): OpenAI-compatible API\n    ProviderSpec(\n        name=\"qianfan\",\n        keywords=(\"qianfan\", \"ernie\"),\n        env_key=\"QIANFAN_API_KEY\",\n        display_name=\"Qianfan\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://qianfan.baidubce.com/v2\"\n    ),\n)\n\n\n# ---------------------------------------------------------------------------\n# Lookup helpers\n# ---------------------------------------------------------------------------\n\n\ndef find_by_name(name: str) -> ProviderSpec | None:\n    \"\"\"Find a provider spec by config field name, e.g. \"dashscope\".\"\"\"\n    normalized = to_snake(name.replace(\"-\", \"_\"))\n    for spec in PROVIDERS:\n        if spec.name == normalized:\n            return spec\n    return None\n"}, "06489b8595e2c2a06bb17047c8ddd726384dff53": {"t": 1785763739.137999, "n": 16786, "h": "2ebb538015870610", "c": "\"\"\"\nProvider Registry — single source of truth for LLM provider metadata.\n\nAdding a new provider:\n  1. Add a ProviderSpec to PROVIDERS below.\n  2. Add a field to ProvidersConfig in config/schema.py.\n  Done. Env vars, config matching, status display all derive from here.\n\nOrder matters — it controls match priority and fallback. Gateways first.\nEvery entry writes out all fields so you can copy-paste as a template.\n\"\"\"\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass\nfrom typing import Any\n\nfrom pydantic.alias_generators import to_snake\n\n\n@dataclass(frozen=True)\nclass ProviderSpec:\n    \"\"\"One LLM provider's metadata. See PROVIDERS below for real examples.\n\n    Placeholders in env_extras values:\n      {api_key}  — the user's API key\n      {api_base} — api_base from config, or this spec's default_api_base\n    \"\"\"\n\n    # identity\n    name: str  # config field name, e.g. \"dashscope\"\n    keywords: tuple[str, ...]  # model-name keywords for matching (lowercase)\n    env_key: str  # env var for API key, e.g. \"DASHSCOPE_API_KEY\"\n    display_name: str = \"\"  # shown in `nanobot status`\n\n    # which provider implementation to use\n    # \"openai_compat\" | \"anthropic\" | \"azure_openai\" | \"openai_codex\" | \"github_copilot\" | \"bedrock\"\n    backend: str = \"openai_compat\"\n\n    # extra env vars, e.g. ((\"ZHIPUAI_API_KEY\", \"{api_key}\"),)\n    env_extras: tuple[tuple[str, str], ...] = ()\n\n    # gateway / local detection\n    is_gateway: bool = False  # routes any model (OpenRouter, AiHubMix)\n    is_local: bool = False  # local deployment (vLLM, Ollama)\n    detect_by_key_prefix: str = \"\"  # match api_key prefix, e.g. \"sk-or-\"\n    detect_by_base_keyword: str = \"\"  # match substring in api_base URL\n    default_api_base: str = \"\"  # OpenAI-compatible base URL for this provider\n\n    # gateway behavior\n    strip_model_prefix: bool = False  # strip \"provider/\" before sending to gateway\n    supports_max_completion_tokens: bool = False\n\n    # per-model param overrides, e.g. ((\"kimi-k2.5\", {\"temperature\": 1.0}),)\n    model_overrides: tuple[tuple[str, dict[str, Any]], ...] = ()\n\n    # OAuth-based providers (e.g., OpenAI Codex) don't use API keys\n    is_oauth: bool = False\n\n    # Direct providers skip API-key validation (user supplies everything)\n    is_direct: bool = False\n\n    # Provider supports cache_control on content blocks (e.g. Anthropic prompt caching)\n    supports_prompt_caching: bool = False\n\n    # How to inject the thinking on/off toggle into extra_body.\n    # \"\"              — no extra_body needed (default)\n    # \"thinking_type\" — {\"thinking\": {\"type\": \"enabled\"/\"disabled\"}}\n    #                   (DeepSeek, VolcEngine, BytePlus)\n    # \"enable_thinking\" — {\"enable_thinking\": true/false}  (DashScope)\n    # \"reasoning_split\" — {\"reasoning_split\": true/false}  (MiniMax)\n    thinking_style: str = \"\"\n\n    # When True, treat the \"reasoning\" response field as formal content\n    # when \"content\" is empty.  Only set this for providers (e.g. StepFun)\n    # whose API returns the actual answer in \"reasoning\" instead of \"content\".\n    reasoning_as_content: bool = False\n\n    @property\n    def label(self) -> str:\n        return self.display_name or self.name.title()\n\n\n# ---------------------------------------------------------------------------\n# PROVIDERS — the registry. Order = priority. Copy any entry as template.\n# ---------------------------------------------------------------------------\n\nPROVIDERS: tuple[ProviderSpec, ...] = (\n    # === Custom (direct OpenAI-compatible endpoint) ========================\n    ProviderSpec(\n        name=\"custom\",\n        keywords=(),\n        env_key=\"\",\n        display_name=\"Custom\",\n        backend=\"openai_compat\",\n        is_direct=True,\n    ),\n\n    # === Azure OpenAI (direct API calls with API version 2024-10-21) =====\n    ProviderSpec(\n        name=\"azure_openai\",\n        keywords=(\"azure\", \"azure-openai\"),\n        env_key=\"\",\n        display_name=\"Azure OpenAI\",\n        backend=\"azure_openai\",\n        is_direct=True,\n    ),\n    # === AWS Bedrock (native Converse API via bedrock-runtime) =============\n    ProviderSpec(\n        name=\"bedrock\",\n        keywords=(\n            \"bedrock\",\n            \"anthropic.claude\",\n            \"amazon.nova\",\n            \"meta.\",\n            \"mistral.\",\n            \"cohere.\",\n            \"qwen.\",\n            \"deepseek.\",\n            \"openai.gpt-oss\",\n            \"ai21.\",\n            \"moonshot.\",\n            \"writer.\",\n            \"zai.\",\n        ),\n        env_key=\"AWS_BEARER_TOKEN_BEDROCK\",\n        display_name=\"AWS Bedrock\",\n        backend=\"bedrock\",\n        is_direct=True,\n    ),\n    # === Gateways (detected by api_key / api_base, not model name) =========\n    # Gateways can route any model, so they win in fallback.\n    # OpenRouter: global gateway, keys start with \"sk-or-\"\n    ProviderSpec(\n        name=\"openrouter\",\n        keywords=(\"openrouter\",),\n        env_key=\"OPENROUTER_API_KEY\",\n        display_name=\"OpenRouter\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_key_prefix=\"sk-or-\",\n        detect_by_base_keyword=\"openrouter\",\n        default_api_base=\"https://openrouter.ai/api/v1\",\n        supports_prompt_caching=True,\n    ),\n    # Hugging Face Inference Providers: OpenAI-compatible router for chat models.\n    ProviderSpec(\n        name=\"huggingface\",\n        keywords=(\"huggingface\", \"hugging-face\"),\n        env_key=\"HF_TOKEN\",\n        display_name=\"Hugging Face\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_key_prefix=\"hf_\",\n        detect_by_base_keyword=\"huggingface\",\n        default_api_base=\"https://router.huggingface.co/v1\",\n    ),\n    # AiHubMix: global gateway, OpenAI-compatible interface.\n    # strip_model_prefix=True: doesn't understand \"anthropic/claude-3\",\n    # strips to bare \"claude-3\".\n    ProviderSpec(\n        name=\"aihubmix\",\n        keywords=(\"aihubmix\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"AiHubMix\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"aihubmix\",\n        default_api_base=\"https://aihubmix.com/v1\",\n        strip_model_prefix=True,\n    ),\n    # SiliconFlow (硅基流动): OpenAI-compatible gateway, model names keep org prefix\n    ProviderSpec(\n        name=\"siliconflow\",\n        keywords=(\"siliconflow\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"SiliconFlow\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"siliconflow\",\n        default_api_base=\"https://api.siliconflow.cn/v1\",\n    ),\n\n    # VolcEngine (火山引擎): OpenAI-compatible gateway, pay-per-use models\n    ProviderSpec(\n        name=\"volcengine\",\n        keywords=(\"volcengine\", \"volces\", \"ark\"),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"VolcEngine\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"volces\",\n        default_api_base=\"https://ark.cn-beijing.volces.com/api/v3\",\n        thinking_style=\"thinking_type\",\n        supports_max_completion_tokens=True,\n    ),\n\n    # VolcEngine Coding Plan (火山引擎 Coding Plan): same key as volcengine\n    ProviderSpec(\n        name=\"volcengine_coding_plan\",\n        keywords=(\"volcengine-plan\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"VolcEngine Coding Plan\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        default_api_base=\"https://ark.cn-beijing.volces.com/api/coding/v3\",\n        strip_model_prefix=True,\n        thinking_style=\"thinking_type\",\n        supports_max_completion_tokens=True,\n    ),\n\n    # BytePlus: VolcEngine international, pay-per-use models\n    ProviderSpec(\n        name=\"byteplus\",\n        keywords=(\"byteplus\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"BytePlus\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        detect_by_base_keyword=\"bytepluses\",\n        default_api_base=\"https://ark.ap-southeast.bytepluses.com/api/v3\",\n        strip_model_prefix=True,\n        thinking_style=\"thinking_type\",\n    ),\n\n    # BytePlus Coding Plan: same key as byteplus\n    ProviderSpec(\n        name=\"byteplus_coding_plan\",\n        keywords=(\"byteplus-plan\",),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"BytePlus Coding Plan\",\n        backend=\"openai_compat\",\n        is_gateway=True,\n        default_api_base=\"https://ark.ap-southeast.bytepluses.com/api/coding/v3\",\n        strip_model_prefix=True,\n        thinking_style=\"thinking_type\",\n    ),\n\n\n    # === Standard providers (matched by model-name keywords) ===============\n    # Anthropic: native Anthropic SDK\n    ProviderSpec(\n        name=\"anthropic\",\n        keywords=(\"anthropic\", \"claude\"),\n        env_key=\"ANTHROPIC_API_KEY\",\n        display_name=\"Anthropic\",\n        backend=\"anthropic\",\n        supports_prompt_caching=True,\n    ),\n    # OpenAI: SDK default base URL (no override needed)\n    ProviderSpec(\n        name=\"openai\",\n        keywords=(\"openai\", \"gpt\"),\n        env_key=\"OPENAI_API_KEY\",\n        display_name=\"OpenAI\",\n        backend=\"openai_compat\",\n        supports_max_completion_tokens=True,\n    ),\n    # OpenAI Codex: OAuth-based, dedicated provider\n    ProviderSpec(\n        name=\"openai_codex\",\n        keywords=(\"openai-codex\",),\n        env_key=\"\",\n        display_name=\"OpenAI Codex\",\n        backend=\"openai_codex\",\n        detect_by_base_keyword=\"codex\",\n        default_api_base=\"https://chatgpt.com/backend-api\",\n        is_oauth=True,\n    ),\n    # GitHub Copilot: OAuth-based\n    ProviderSpec(\n        name=\"github_copilot\",\n        keywords=(\"github_copilot\", \"copilot\"),\n        env_key=\"\",\n        display_name=\"Github Copilot\",\n        backend=\"github_copilot\",\n        default_api_base=\"https://api.githubcopilot.com\",\n        strip_model_prefix=True,\n        is_oauth=True,\n        supports_max_completion_tokens=True,\n    ),\n    # DeepSeek: OpenAI-compatible at api.deepseek.com\n    ProviderSpec(\n        name=\"deepseek\",\n        keywords=(\"deepseek\",),\n        env_key=\"DEEPSEEK_API_KEY\",\n        display_name=\"DeepSeek\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.deepseek.com\",\n        thinking_style=\"thinking_type\",\n    ),\n    # Gemini: Google's OpenAI-compatible endpoint\n    ProviderSpec(\n        name=\"gemini\",\n        keywords=(\"gemini\", \"gemma\"),\n        env_key=\"GEMINI_API_KEY\",\n        display_name=\"Gemini\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://generativelanguage.googleapis.com/v1beta/openai/\",\n    ),\n    # Zhipu (智谱): OpenAI-compatible at open.bigmodel.cn\n    ProviderSpec(\n        name=\"zhipu\",\n        keywords=(\"zhipu\", \"glm\", \"zai\"),\n        env_key=\"ZAI_API_KEY\",\n        display_name=\"Zhipu AI\",\n        backend=\"openai_compat\",\n        env_extras=((\"ZHIPUAI_API_KEY\", \"{api_key}\"),),\n        default_api_base=\"https://open.bigmodel.cn/api/paas/v4\",\n    ),\n    # DashScope (通义): Qwen models, OpenAI-compatible endpoint\n    ProviderSpec(\n        name=\"dashscope\",\n        keywords=(\"qwen\", \"dashscope\"),\n        env_key=\"DASHSCOPE_API_KEY\",\n        display_name=\"DashScope\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://dashscope.aliyuncs.com/compatible-mode/v1\",\n        thinking_style=\"enable_thinking\",\n    ),\n    # Moonshot (月之暗面): Kimi K2.5 / K2.6 enforce temperature >= 1.0.\n    ProviderSpec(\n        name=\"moonshot\",\n        keywords=(\"moonshot\", \"kimi\"),\n        env_key=\"MOONSHOT_API_KEY\",\n        display_name=\"Moonshot\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.moonshot.ai/v1\",\n        model_overrides=(\n            (\"kimi-k2.5\", {\"temperature\": 1.0}),\n            (\"kimi-k2.6\", {\"temperature\": 1.0}),\n        ),\n    ),\n    # MiniMax: OpenAI-compatible API\n    ProviderSpec(\n        name=\"minimax\",\n        keywords=(\"minimax\",),\n        env_key=\"MINIMAX_API_KEY\",\n        display_name=\"MiniMax\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.minimax.io/v1\",\n        thinking_style=\"reasoning_split\",\n    ),\n    # MiniMax Anthropic-compatible endpoint: supports thinking mode\n    ProviderSpec(\n        name=\"minimax_anthropic\",\n        keywords=(\"minimax_anthropic\",),\n        env_key=\"MINIMAX_API_KEY\",\n        display_name=\"MiniMax (Anthropic)\",\n        backend=\"anthropic\",\n        default_api_base=\"https://api.minimax.io/anthropic\",\n    ),\n    # Mistral AI: OpenAI-compatible API\n    ProviderSpec(\n        name=\"mistral\",\n        keywords=(\"mistral\",),\n        env_key=\"MISTRAL_API_KEY\",\n        display_name=\"Mistral\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.mistral.ai/v1\",\n    ),\n    # Step Fun (阶跃星辰): OpenAI-compatible API\n    ProviderSpec(\n        name=\"stepfun\",\n        keywords=(\"stepfun\", \"step\"),\n        env_key=\"STEPFUN_API_KEY\",\n        display_name=\"Step Fun\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.stepfun.com/v1\",\n        reasoning_as_content=True,\n    ),\n    # Xiaomi MIMO (小米): OpenAI-compatible API\n    # Hosted API (api.xiaomimimo.com) accepts {\"thinking\": {\"type\": \"enabled\"|\"disabled\"}}\n    # to toggle reasoning, matching the existing thinking_type style.\n    ProviderSpec(\n        name=\"xiaomi_mimo\",\n        keywords=(\"xiaomi_mimo\", \"mimo\"),\n        env_key=\"XIAOMIMIMO_API_KEY\",\n        display_name=\"Xiaomi MIMO\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.xiaomimimo.com/v1\",\n        thinking_style=\"thinking_type\",\n    ),\n    # LongCat: OpenAI-compatible API\n    ProviderSpec(\n        name=\"longcat\",\n        keywords=(\"longcat\",),\n        env_key=\"LONGCAT_API_KEY\",\n        display_name=\"LongCat\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.longcat.chat/openai/v1\",\n    ),\n    # === Local deployment (matched by config key, NOT by api_base) =========\n    # vLLM / any OpenAI-compatible local server\n    ProviderSpec(\n        name=\"vllm\",\n        keywords=(\"vllm\",),\n        env_key=\"HOSTED_VLLM_API_KEY\",\n        display_name=\"vLLM/Local\",\n        backend=\"openai_compat\",\n        is_local=True,\n    ),\n    # Ollama (local, OpenAI-compatible)\n    ProviderSpec(\n        name=\"ollama\",\n        keywords=(\"ollama\", \"nemotron\"),\n        env_key=\"OLLAMA_API_KEY\",\n        display_name=\"Ollama\",\n        backend=\"openai_compat\",\n        is_local=True,\n        detect_by_base_keyword=\"11434\",\n        default_api_base=\"http://localhost:11434/v1\",\n    ),\n    # LM Studio (local, OpenAI-compatible)\n    ProviderSpec(\n        name=\"lm_studio\",\n        keywords=(\"lm-studio\", \"lmstudio\", \"lm_studio\"),\n        env_key=\"LM_STUDIO_API_KEY\",\n        display_name=\"LM Studio\",\n        backend=\"openai_compat\",\n        is_local=True,\n        detect_by_base_keyword=\"1234\",\n        default_api_base=\"http://localhost:1234/v1\",\n    ),\n    # === OpenVINO Model Server (direct, local, OpenAI-compatible at /v3) ===\n    ProviderSpec(\n        name=\"ovms\",\n        keywords=(\"openvino\", \"ovms\"),\n        env_key=\"\",\n        display_name=\"OpenVINO Model Server\",\n        backend=\"openai_compat\",\n        is_direct=True,\n        is_local=True,\n        default_api_base=\"http://localhost:8000/v3\",\n    ),\n    # === NVIDIA NIM (NVIDIA Inference Microservices) =======================\n    # Keys start with \"nvapi-\", base URL at integrate.api.nvidia.com\n    ProviderSpec(\n        name=\"nvidia\",\n        keywords=(\"nvidia\", \"nemotron\", \"nvapi\"),\n        env_key=\"NVIDIA_NIM_API_KEY\",\n        display_name=\"NVIDIA NIM\",\n        backend=\"openai_compat\",\n        is_gateway=False,\n        detect_by_key_prefix=\"nvapi-\",\n        detect_by_base_keyword=\"nvidia.com\",\n        default_api_base=\"https://integrate.api.nvidia.com/v1\",\n    ),\n    # === Auxiliary (not a primary LLM provider) ============================\n    # Groq: mainly used for Whisper voice transcription, also usable for LLM\n    ProviderSpec(\n        name=\"groq\",\n        keywords=(\"groq\",),\n        env_key=\"GROQ_API_KEY\",\n        display_name=\"Groq\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://api.groq.com/openai/v1\",\n    ),\n    # Qianfan (百度千帆): OpenAI-compatible API\n    ProviderSpec(\n        name=\"qianfan\",\n        keywords=(\"qianfan\", \"ernie\"),\n        env_key=\"QIANFAN_API_KEY\",\n        display_name=\"Qianfan\",\n        backend=\"openai_compat\",\n        default_api_base=\"https://qianfan.baidubce.com/v2\"\n    ),\n)\n\n\n# ---------------------------------------------------------------------------\n# Lookup helpers\n# ---------------------------------------------------------------------------\n\n\ndef find_by_name(name: str) -> ProviderSpec | None:\n    \"\"\"Find a provider spec by config field name, e.g. \"dashscope\".\"\"\"\n    normalized = to_snake(name.replace(\"-\", \"_\"))\n    for spec in PROVIDERS:\n        if spec.name == normalized:\n            return spec\n    return None\n"}}}}