hack/tools/scripts/llm-key-hunter/results/_cache/content/agentic-in_elephant-agent.json

{"repo": "agentic-in/elephant-agent", "files": {"packages/models/model_metadata.py": {"b342fa5f7cceb2efc4945b50b54b6e8233452272": {"t": 1785595148.7759643, "n": 23169, "h": "8080afce24003935", "c": "\"\"\"Provider-aware model metadata fallback resolution.\"\"\"\n\nfrom __future__ import annotations\n\nfrom collections.abc import Mapping\nfrom dataclasses import dataclass\nimport ipaddress\nimport json\nimport os\nfrom pathlib import Path\nimport tempfile\nimport time\nfrom typing import Any\nfrom urllib import error, request\nfrom urllib.parse import urlparse\n\nfrom packages.runtime_layout import default_cli_state_dir\n\n\nMODELS_DEV_URL = \"https://models.dev/api.json\"\nOPENROUTER_MODELS_URL = \"https://openrouter.ai/api/v1/models\"\n_MODELS_DEV_CACHE_TTL_SECONDS = 3600\n_OPENROUTER_CACHE_TTL_SECONDS = 3600\n_ENDPOINT_CACHE_TTL_SECONDS = 300\n\n_LOCAL_HOSTS = frozenset({\"localhost\", \"127.0.0.1\", \"::1\", \"0.0.0.0\"})\n_CONTAINER_LOCAL_SUFFIXES = (\".docker.internal\", \".containers.internal\", \".lima.internal\")\n\n_CONTEXT_KEYS = (\n    \"context_length\",\n    \"context_window\",\n    \"max_context_length\",\n    \"max_position_embeddings\",\n    \"max_model_len\",\n    \"max_input_tokens\",\n    \"max_sequence_length\",\n    \"max_seq_len\",\n    \"n_ctx\",\n    \"n_ctx_train\",\n)\n_OUTPUT_KEYS = (\n    \"max_completion_tokens\",\n    \"max_output_tokens\",\n    \"max_tokens\",\n)\n\n_PROVIDER_TO_MODELS_DEV: dict[str, str] = {\n    \"alibaba\": \"alibaba\",\n    \"anthropic\": \"anthropic\",\n    \"baidu-qianfan\": \"baidu\",\n    \"cerebras\": \"cerebras\",\n    \"claude-code\": \"anthropic\",\n    \"cohere\": \"cohere\",\n    \"copilot\": \"github-copilot\",\n    \"deepseek\": \"deepseek\",\n    \"fireworks\": \"fireworks-ai\",\n    \"google\": \"google\",\n    \"groq\": \"groq\",\n    \"huggingface\": \"huggingface\",\n    \"kilocode\": \"kilo\",\n    \"minimax\": \"minimax\",\n    \"minimax-cn\": \"minimax-cn\",\n    \"modelscope\": \"modelscope\",\n    \"mistral\": \"mistral\",\n    \"moonshot\": \"kimi-for-coding\",\n    \"moonshot-cn\": \"kimi-for-coding\",\n    \"nvidia\": \"nvidia\",\n    \"openai\": \"openai\",\n    \"openai-codex\": \"openai\",\n    \"opencode-go\": \"opencode-go\",\n    \"opencode-zen\": \"opencode\",\n    \"openrouter\": \"openrouter\",\n    \"perplexity\": \"perplexity\",\n    \"qwen-oauth\": \"alibaba\",\n    \"siliconflow\": \"siliconflow\",\n    \"stepfun\": \"stepfun\",\n    \"tencent-hunyuan\": \"tencent\",\n    \"tencent-tokenhub\": \"tencent\",\n    \"together\": \"togetherai\",\n    \"volcengine\": \"volcengine\",\n    \"xai\": \"xai\",\n    \"xiaomi\": \"xiaomi\",\n    \"zai\": \"zai\",\n}\n\n_URL_TO_PROVIDER: dict[str, str] = {\n    \"api-inference.modelscope.cn\": \"modelscope\",\n    \"api.anthropic.com\": \"anthropic\",\n    \"api.cerebras.ai\": \"cerebras\",\n    \"api.cohere.ai\": \"cohere\",\n    \"api.deepseek.com\": \"deepseek\",\n    \"api.githubcopilot.com\": \"copilot\",\n    \"api.groq.com\": \"groq\",\n    \"api.hunyuan.cloud.tencent.com\": \"tencent-hunyuan\",\n    \"api.minimax\": \"minimax\",\n    \"api.mistral.ai\": \"mistral\",\n    \"api.moonshot.ai\": \"moonshot\",\n    \"api.moonshot.cn\": \"moonshot-cn\",\n    \"api.openai.com\": \"openai\",\n    \"api.perplexity.ai\": \"perplexity\",\n    \"api.siliconflow.cn\": \"siliconflow\",\n    \"api.siliconflow.com\": \"siliconflow\",\n    \"api.stepfun.ai\": \"stepfun\",\n    \"api.stepfun.com\": \"stepfun\",\n    \"api.together.ai\": \"together\",\n    \"api.x.ai\": \"xai\",\n    \"api.xiaomimimo.com\": \"xiaomi\",\n    \"api.z.ai\": \"zai\",\n    \"ark.cn-beijing.volces.com\": \"volcengine\",\n    \"chatgpt.com\": \"openai-codex\",\n    \"dashscope-intl.aliyuncs.com\": \"alibaba\",\n    \"dashscope.aliyuncs.com\": \"alibaba\",\n    \"generativelanguage.googleapis.com\": \"google\",\n    \"integrate.api.nvidia.com\": \"nvidia\",\n    \"models.github.ai\": \"copilot\",\n    \"openrouter.ai\": \"openrouter\",\n    \"opencode.ai\": \"opencode-go\",\n    \"portal.qwen.ai\": \"qwen-oauth\",\n    \"qianfan.baidubce.com\": \"baidu-qianfan\",\n    \"tokenhub-intl.tencentcloudmaas.com\": \"tencent-tokenhub\",\n    \"tokenhub-intl.tencentmaas.cn\": \"tencent-tokenhub\",\n    \"tokenhub-intl.tencentmaas.com\": \"tencent-tokenhub\",\n    \"tokenhub.tencentmaas.cn\": \"tencent-tokenhub\",\n    \"tokenhub.tencentmaas.com\": \"tencent-tokenhub\",\n    \"xiaomimimo.com\": \"xiaomi\",\n}\n\n_models_dev_cache: dict[str, Any] = {}\n_models_dev_cache_time = 0.0\n_openrouter_cache: dict[str, \"ResolvedModelMetadata\"] = {}\n_openrouter_cache_time = 0.0\n_endpoint_metadata_cache: dict[str, dict[str, \"ResolvedModelMetadata\"]] = {}\n_endpoint_metadata_cache_time: dict[str, float] = {}\n\n\n@dataclass(frozen=True, slots=True)\nclass ResolvedModelMetadata:\n    context_window_tokens: int | None = None\n    max_output_tokens: int | None = None\n    source: str = \"unknown\"\n\n\ndef resolve_provider_model_metadata(\n    *,\n    provider_id: str,\n    model_id: str,\n    base_url: str | None = None,\n    api_key: str | None = None,\n) -> ResolvedModelMetadata | None:\n    \"\"\"Resolve metadata for one selected model from local, registry, or routed catalogs.\"\"\"\n\n    normalized_model = str(model_id or \"\").strip()\n    if not normalized_model:\n        return None\n\n    normalized_base_url = _normalize_base_url(base_url)\n    cached_context = get_cached_context_length(normalized_model, normalized_base_url)\n    if cached_context is not None:\n        return ResolvedModelMetadata(\n            context_window_tokens=cached_context,\n            source=\"context-length-cache\",\n        )\n\n    if normalized_base_url and is_local_endpoint(normalized_base_url):\n        local = query_local_endpoint_metadata(\n            model_id=normalized_model,\n            base_url=normalized_base_url,\n            api_key=api_key,\n        )\n        if local is not None and local.context_window_tokens is not None:\n            save_context_length(normalized_model, normalized_base_url, local.context_window_tokens)\n            return local\n\n    effective_provider = _effective_provider_id(provider_id=provider_id, base_url=normalized_base_url)\n    models_dev = lookup_models_dev_metadata(effective_provider, normalized_model)\n    if models_dev is not None and models_dev.context_window_tokens is not None:\n        return models_dev\n\n    openrouter = lookup_openrouter_metadata(normalized_model)\n    if openrouter is not None and openrouter.context_window_tokens is not None:\n        return openrouter\n\n    return None\n\n\ndef lookup_models_dev_metadata(provider_id: str, model_id: str) -> ResolvedModelMetadata | None:\n    models_dev_provider = _PROVIDER_TO_MODELS_DEV.get(provider_id.strip().lower())\n    if not models_dev_provider:\n        return None\n    data = fetch_models_dev_registry()\n    provider_data = data.get(models_dev_provider)\n    if not isinstance(provider_data, Mapping):\n        return None\n    models = provider_data.get(\"models\")\n    if not isinstance(models, Mapping):\n        return None\n    entry = _find_model_entry(models, model_id)\n    if entry is None and \"/\" in model_id:\n        entry = _find_model_entry(models, model_id.rsplit(\"/\", 1)[1])\n    if entry is None:\n        return None\n    limit = entry.get(\"limit\")\n    context = _coerce_reasonable_int(limit.get(\"context\") if isinstance(limit, Mapping) else None)\n    output = _coerce_reasonable_int(limit.get(\"output\") if isinstance(limit, Mapping) else None)\n    if context is None:\n        context = _extract_nested_int(entry, _CONTEXT_KEYS)\n    if output is None:\n        output = _extract_nested_int(entry, _OUTPUT_KEYS)\n    if context is None and output is None:\n        return None\n    return ResolvedModelMetadata(\n        context_window_tokens=context,\n        max_output_tokens=output,\n        source=f\"models.dev:{models_dev_provider}\",\n    )\n\n\ndef lookup_openrouter_metadata(model_id: str) -> ResolvedModelMetadata | None:\n    metadata = fetch_openrouter_model_metadata()\n    return _find_metadata_entry(metadata, model_id)\n\n\ndef get_cached_context_length(model_id: str, base_url: str | None) -> int | None:\n    normalized_model = str(model_id or \"\").strip()\n    normalized_base_url = _normalize_base_url(base_url)\n    if not normalized_model or not normalized_base_url:\n        return None\n    return _load_context_length_cache().get(_context_length_cache_key(normalized_model, normalized_base_url))\n\n\ndef save_context_length(model_id: str, base_url: str | None, context_window_tokens: int) -> None:\n    normalized_model = str(model_id or \"\").strip()\n    normalized_base_url = _normalize_base_url(base_url)\n    parsed = _coerce_reasonable_int(context_window_tokens)\n    if not normalized_model or not normalized_base_url or parsed is None:\n        return\n    cache = _load_context_length_cache()\n    key = _context_length_cache_key(normalized_model, normalized_base_url)\n    if cache.get(key) == parsed:\n        return\n    cache[key] = parsed\n    _save_context_length_cache(cache)\n\n\ndef query_local_endpoint_metadata(\n    *,\n    model_id: str,\n    base_url: str,\n    api_key: str | None = None,\n) -> ResolvedModelMetadata | None:\n    normalized = _normalize_base_url(base_url)\n    server_root = normalized[:-3] if normalized.endswith(\"/v1\") else normalized\n\n    lm_studio = _request_json_object(f\"{server_root}/api/v1/models\", timeout_seconds=2.0)\n    if lm_studio:\n        for model in _model_items(lm_studio):\n            candidate_id = str(model.get(\"key\") or model.get(\"id\") or \"\").strip()\n            if not _model_id_matches(candidate_id, model_id):\n                continue\n            loaded_instances = model.get(\"loaded_instances\")\n            if isinstance(loaded_instances, list):\n                for instance in loaded_instances:\n                    if not isinstance(instance, Mapping):\n                        continue\n                    config = instance.get(\"config\")\n                    if not isinstance(config, Mapping):\n                        continue\n                    context = _coerce_reasonable_int(config.get(\"context_length\"))\n                    if context is not None:\n                        return ResolvedModelMetadata(context_window_tokens=context, source=\"local:lm-studio\")\n            context = _extract_nested_int(model, _CONTEXT_KEYS)\n            output = _extract_nested_int(model, _OUTPUT_KEYS)\n            if context is not None or output is not None:\n                return ResolvedModelMetadata(context, output, \"local:lm-studio\")\n\n    detail = _request_json_object(\n        f\"{server_root}/v1/models/{model_id}\",\n        headers=_bearer_headers(api_key),\n        timeout_seconds=2.0,\n    )\n    if detail:\n        context = _extract_nested_int(detail, _CONTEXT_KEYS)\n        output = _extract_nested_int(detail, _OUTPUT_KEYS)\n        if context is not None or output is not None:\n            return ResolvedModelMetadata(context, output, \"local:model-detail\")\n\n    listed = fetch_endpoint_model_metadata(base_url=normalized, api_key=api_key)\n    entry = _find_metadata_entry(listed, model_id)\n    if entry is not None:\n        return entry\n\n    props = _request_json_object(f\"{server_root}/v1/props\", headers=_bearer_headers(api_key), timeout_seconds=2.0)\n    if not props:\n        props = _request_json_object(f\"{server_root}/props\", headers=_bearer_headers(api_key), timeout_seconds=2.0)\n    if props:\n        context = _extract_nested_int(props, _CONTEXT_KEYS)\n        if context is not None:\n            return ResolvedModelMetadata(context_window_tokens=context, source=\"local:llamacpp\")\n\n    return None\n\n\ndef fetch_models_dev_registry(*, force_refresh: bool = False) -> Mapping[str, Any]:\n    global _models_dev_cache, _models_dev_cache_time\n    if (\n        not force_refresh\n        and _models_dev_cache\n        and (time.time() - _models_dev_cache_time) < _MODELS_DEV_CACHE_TTL_SECONDS\n    ):\n        return _models_dev_cache\n\n    try:\n        payload = _request_json_object(MODELS_DEV_URL, timeout_seconds=10.0)\n        if payload:\n            _models_dev_cache = dict(payload)\n            _models_dev_cache_time = time.time()\n            _save_models_dev_disk_cache(_models_dev_cache)\n            return _models_dev_cache\n    except RuntimeError:\n        pass\n\n    if not _models_dev_cache:\n        _models_dev_cache = _load_models_dev_disk_cache()\n        if _models_dev_cache:\n            _models_dev_cache_time = time.time() - _MODELS_DEV_CACHE_TTL_SECONDS + 300\n    return _models_dev_cache\n\n\ndef fetch_openrouter_model_metadata(*, force_refresh: bool = False) -> Mapping[str, ResolvedModelMetadata]:\n    global _openrouter_cache, _openrouter_cache_time\n    if (\n        not force_refresh\n        and _openrouter_cache\n        and (time.time() - _openrouter_cache_time) < _OPENROUTER_CACHE_TTL_SECONDS\n    ):\n        return _openrouter_cache\n\n    try:\n        payload = _request_json_object(OPENROUTER_MODELS_URL, timeout_seconds=10.0)\n    except RuntimeError:\n        return _openrouter_cache\n    cache: dict[str, ResolvedModelMetadata] = {}\n    for model in _model_items(payload):\n        model_id = str(model.get(\"id\") or \"\").strip()\n        if not model_id:\n            continue\n        top_provider = model.get(\"top_provider\")\n        output = None\n        if isinstance(top_provider, Mapping):\n            output = _coerce_reasonable_int(top_provider.get(\"max_completion_tokens\"))\n        entry = ResolvedModelMetadata(\n            context_window_tokens=_extract_nested_int(model, _CONTEXT_KEYS),\n            max_output_tokens=output or _extract_nested_int(model, _OUTPUT_KEYS),\n            source=\"openrouter\",\n        )\n        _add_metadata_aliases(cache, model_id, entry)\n        canonical_slug = str(model.get(\"canonical_slug\") or \"\").strip()\n        if canonical_slug:\n            _add_metadata_aliases(cache, canonical_slug, entry)\n    _openrouter_cache = cache\n    _openrouter_cache_time = time.time()\n    return _openrouter_cache\n\n\ndef fetch_endpoint_model_metadata(\n    *,\n    base_url: str,\n    api_key: str | None = None,\n    force_refresh: bool = False,\n) -> Mapping[str, ResolvedModelMetadata]:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return {}\n    if not force_refresh:\n        cached = _endpoint_metadata_cache.get(normalized)\n        cached_at = _endpoint_metadata_cache_time.get(normalized, 0)\n        if cached is not None and (time.time() - cached_at) < _ENDPOINT_CACHE_TTL_SECONDS:\n            return cached\n\n    candidates = [normalized]\n    alternate = normalized[:-3].rstrip(\"/\") if normalized.endswith(\"/v1\") else f\"{normalized}/v1\"\n    if alternate and alternate not in candidates:\n        candidates.append(alternate)\n\n    for candidate in candidates:\n        payload = _request_json_object(\n            f\"{candidate.rstrip('/')}/models\",\n            headers=_bearer_headers(api_key),\n            timeout_seconds=5.0,\n        )\n        if not payload:\n            continue\n        cache: dict[str, ResolvedModelMetadata] = {}\n        for model in _model_items(payload):\n            model_id = str(model.get(\"id\") or model.get(\"key\") or \"\").strip()\n            if not model_id:\n                continue\n            entry = ResolvedModelMetadata(\n                context_window_tokens=_extract_nested_int(model, _CONTEXT_KEYS),\n                max_output_tokens=_extract_nested_int(model, _OUTPUT_KEYS),\n                source=\"endpoint:/models\",\n            )\n            _add_metadata_aliases(cache, model_id, entry)\n        if cache:\n            _endpoint_metadata_cache[normalized] = cache\n            _endpoint_metadata_cache_time[normalized] = time.time()\n            return cache\n\n    _endpoint_metadata_cache[normalized] = {}\n    _endpoint_metadata_cache_time[normalized] = time.time()\n    return {}\n\n\ndef is_local_endpoint(base_url: str | None) -> bool:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return False\n    try:\n        parsed = urlparse(normalized if \"://\" in normalized else f\"http://{normalized}\")\n    except ValueError:\n        return False\n    host = parsed.hostname or \"\"\n    if host in _LOCAL_HOSTS or any(host.endswith(suffix) for suffix in _CONTAINER_LOCAL_SUFFIXES):\n        return True\n    try:\n        address = ipaddress.ip_address(host)\n    except ValueError:\n        return False\n    return address.is_private or address.is_loopback or address.is_link_local\n\n\ndef infer_provider_from_url(base_url: str | None) -> str | None:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return None\n    parsed = urlparse(normalized if \"://\" in normalized else f\"https://{normalized}\")\n    host = parsed.netloc.lower() or parsed.path.lower()\n    for url_part, provider_id in _URL_TO_PROVIDER.items():\n        if url_part in host:\n            return provider_id\n    return None\n\n\ndef _effective_provider_id(*, provider_id: str, base_url: str | None) -> str:\n    normalized_provider = provider_id.strip().lower()\n    if normalized_provider in {\"\", \"openai-compatible\"}:\n        return infer_provider_from_url(base_url) or normalized_provider\n    return normalized_provider\n\n\ndef _request_json_object(\n    url: str,\n    *,\n    headers: Mapping[str, str] | None = None,\n    timeout_seconds: float = 10.0,\n) -> Mapping[str, Any] | None:\n    http_request = request.Request(url, headers=dict(headers or {}), method=\"GET\")\n    try:\n        with request.urlopen(http_request, timeout=timeout_seconds) as response:\n            raw_body = response.read().decode(\"utf-8\")\n    except (error.HTTPError, error.URLError, TimeoutError) as exc:\n        return None\n    try:\n        payload = json.loads(raw_body) if raw_body else {}\n    except json.JSONDecodeError:\n        return None\n    if isinstance(payload, list):\n        return {\"data\": payload}\n    if not isinstance(payload, Mapping):\n        return None\n    return payload\n\n\ndef _model_items(payload: Mapping[str, Any]) -> tuple[Mapping[str, Any], ...]:\n    for key in (\"data\", \"models\"):\n        value = payload.get(key)\n        if isinstance(value, list):\n            return tuple(item for item in value if isinstance(item, Mapping))\n    return ()\n\n\ndef _iter_nested_mappings(value: Any):\n    if isinstance(value, Mapping):\n        yield value\n        for nested in value.values():\n            yield from _iter_nested_mappings(nested)\n    elif isinstance(value, list):\n        for item in value:\n            yield from _iter_nested_mappings(item)\n\n\ndef _extract_nested_int(payload: Mapping[str, Any], keys: tuple[str, ...]) -> int | None:\n    keyset = {key.casefold() for key in keys}\n    for mapping in _iter_nested_mappings(payload):\n        for key, value in mapping.items():\n            if str(key).casefold() not in keyset:\n                continue\n            parsed = _coerce_reasonable_int(value)\n            if parsed is not None:\n                return parsed\n    return None\n\n\ndef _coerce_reasonable_int(value: Any, *, minimum: int = 1024, maximum: int = 10_000_000) -> int | None:\n    try:\n        if isinstance(value, bool):\n            return None\n        if isinstance(value, str):\n            value = value.strip().replace(\",\", \"\")\n        parsed = int(value)\n    except (TypeError, ValueError):\n        return None\n    if minimum <= parsed <= maximum:\n        return parsed\n    return None\n\n\ndef _find_model_entry(models: Mapping[str, Any], model_id: str) -> Mapping[str, Any] | None:\n    entry = models.get(model_id)\n    if isinstance(entry, Mapping):\n        return entry\n    model_lower = model_id.casefold()\n    for candidate_id, candidate in models.items():\n        if str(candidate_id).casefold() == model_lower and isinstance(candidate, Mapping):\n            return candidate\n    return None\n\n\ndef _find_metadata_entry(\n    metadata: Mapping[str, ResolvedModelMetadata],\n    model_id: str,\n) -> ResolvedModelMetadata | None:\n    entry = metadata.get(model_id)\n    if entry is not None:\n        return entry\n    if \"/\" not in model_id:\n        for candidate_id, candidate in metadata.items():\n            if candidate_id.rsplit(\"/\", 1)[-1] == model_id:\n                return candidate\n    model_lower = model_id.casefold()\n    for candidate_id, candidate in metadata.items():\n        if candidate_id.casefold() == model_lower:\n            return candidate\n    return None\n\n\ndef _add_metadata_aliases(\n    cache: dict[str, ResolvedModelMetadata],\n    model_id: str,\n    entry: ResolvedModelMetadata,\n) -> None:\n    cache[model_id] = entry\n    if \"/\" in model_id:\n        cache.setdefault(model_id.rsplit(\"/\", 1)[1], entry)\n\n\ndef _model_id_matches(candidate_id: str, lookup_model: str) -> bool:\n    if candidate_id == lookup_model:\n        return True\n    return \"/\" in candidate_id and candidate_id.rsplit(\"/\", 1)[1] == lookup_model\n\n\ndef _bearer_headers(api_key: str | None) -> dict[str, str]:\n    resolved = str(api_key or \"\").strip()\n    return {\"Authorization\": f\"Bearer {resolved}\"} if resolved else {}\n\n\ndef _normalize_base_url(base_url: str | None) -> str:\n    return str(base_url or \"\").strip().rstrip(\"/\")\n\n\ndef _models_dev_disk_cache_path() -> Path:\n    override = os.environ.get(\"ELEPHANT_MODELS_DEV_CACHE_PATH\", \"\").strip()\n    if override:\n        return Path(override).expanduser()\n    return default_cli_state_dir() / \"models-dev-cache.json\"\n\n\ndef _context_length_cache_path() -> Path:\n    override = os.environ.get(\"ELEPHANT_CONTEXT_LENGTH_CACHE_PATH\", \"\").strip()\n    if override:\n        return Path(override).expanduser()\n    return default_cli_state_dir() / \"context-length-cache.json\"\n\n\ndef _context_length_cache_key(model_id: str, base_url: str) -> str:\n    return f\"{model_id}@{base_url}\"\n\n\ndef _load_context_length_cache() -> dict[str, int]:\n    path = _context_length_cache_path()\n    try:\n        payload = json.loads(path.read_text(encoding=\"utf-8\"))\n    except (OSError, json.JSONDecodeError):\n        return {}\n    raw_cache = payload.get(\"context_lengths\") if isinstance(payload, Mapping) else None\n    if not isinstance(raw_cache, Mapping):\n        return {}\n    cache: dict[str, int] = {}\n    for key, value in raw_cache.items():\n        parsed = _coerce_reasonable_int(value)\n        if parsed is not None:\n            cache[str(key)] = parsed\n    return cache\n\n\ndef _save_context_length_cache(cache: Mapping[str, int]) -> None:\n    path = _context_length_cache_path()\n    try:\n        path.parent.mkdir(parents=True, exist_ok=True)\n        with tempfile.NamedTemporaryFile(\"w\", encoding=\"utf-8\", dir=path.parent, delete=False) as handle:\n            json.dump({\"context_lengths\": dict(cache)}, handle, separators=(\",\", \":\"))\n            temp_name = handle.name\n        Path(temp_name).replace(path)\n    except OSError:\n        return\n\n\ndef _load_models_dev_disk_cache() -> dict[str, Any]:\n    path = _models_dev_disk_cache_path()\n    try:\n        payload = json.loads(path.read_text(encoding=\"utf-8\"))\n    except (OSError, json.JSONDecodeError):\n        return {}\n    return dict(payload) if isinstance(payload, Mapping) else {}\n\n\ndef _save_models_dev_disk_cache(payload: Mapping[str, Any]) -> None:\n    path = _models_dev_disk_cache_path()\n    try:\n        path.parent.mkdir(parents=True, exist_ok=True)\n        with tempfile.NamedTemporaryFile(\"w\", encoding=\"utf-8\", dir=path.parent, delete=False) as handle:\n            json.dump(dict(payload), handle, separators=(\",\", \":\"))\n            temp_name = handle.name\n        Path(temp_name).replace(path)\n    except OSError:\n        return\n\n\n__all__ = [\n    \"ResolvedModelMetadata\",\n    \"fetch_endpoint_model_metadata\",\n    \"fetch_models_dev_registry\",\n    \"fetch_openrouter_model_metadata\",\n    \"get_cached_context_length\",\n    \"infer_provider_from_url\",\n    \"is_local_endpoint\",\n    \"lookup_models_dev_metadata\",\n    \"lookup_openrouter_metadata\",\n    \"query_local_endpoint_metadata\",\n    \"resolve_provider_model_metadata\",\n    \"save_context_length\",\n]\n"}, "aa68047a34022a7a1e0ba91218bd90cb46204257": {"t": 1785758536.773551, "n": 23169, "h": "8080afce24003935", "c": "\"\"\"Provider-aware model metadata fallback resolution.\"\"\"\n\nfrom __future__ import annotations\n\nfrom collections.abc import Mapping\nfrom dataclasses import dataclass\nimport ipaddress\nimport json\nimport os\nfrom pathlib import Path\nimport tempfile\nimport time\nfrom typing import Any\nfrom urllib import error, request\nfrom urllib.parse import urlparse\n\nfrom packages.runtime_layout import default_cli_state_dir\n\n\nMODELS_DEV_URL = \"https://models.dev/api.json\"\nOPENROUTER_MODELS_URL = \"https://openrouter.ai/api/v1/models\"\n_MODELS_DEV_CACHE_TTL_SECONDS = 3600\n_OPENROUTER_CACHE_TTL_SECONDS = 3600\n_ENDPOINT_CACHE_TTL_SECONDS = 300\n\n_LOCAL_HOSTS = frozenset({\"localhost\", \"127.0.0.1\", \"::1\", \"0.0.0.0\"})\n_CONTAINER_LOCAL_SUFFIXES = (\".docker.internal\", \".containers.internal\", \".lima.internal\")\n\n_CONTEXT_KEYS = (\n    \"context_length\",\n    \"context_window\",\n    \"max_context_length\",\n    \"max_position_embeddings\",\n    \"max_model_len\",\n    \"max_input_tokens\",\n    \"max_sequence_length\",\n    \"max_seq_len\",\n    \"n_ctx\",\n    \"n_ctx_train\",\n)\n_OUTPUT_KEYS = (\n    \"max_completion_tokens\",\n    \"max_output_tokens\",\n    \"max_tokens\",\n)\n\n_PROVIDER_TO_MODELS_DEV: dict[str, str] = {\n    \"alibaba\": \"alibaba\",\n    \"anthropic\": \"anthropic\",\n    \"baidu-qianfan\": \"baidu\",\n    \"cerebras\": \"cerebras\",\n    \"claude-code\": \"anthropic\",\n    \"cohere\": \"cohere\",\n    \"copilot\": \"github-copilot\",\n    \"deepseek\": \"deepseek\",\n    \"fireworks\": \"fireworks-ai\",\n    \"google\": \"google\",\n    \"groq\": \"groq\",\n    \"huggingface\": \"huggingface\",\n    \"kilocode\": \"kilo\",\n    \"minimax\": \"minimax\",\n    \"minimax-cn\": \"minimax-cn\",\n    \"modelscope\": \"modelscope\",\n    \"mistral\": \"mistral\",\n    \"moonshot\": \"kimi-for-coding\",\n    \"moonshot-cn\": \"kimi-for-coding\",\n    \"nvidia\": \"nvidia\",\n    \"openai\": \"openai\",\n    \"openai-codex\": \"openai\",\n    \"opencode-go\": \"opencode-go\",\n    \"opencode-zen\": \"opencode\",\n    \"openrouter\": \"openrouter\",\n    \"perplexity\": \"perplexity\",\n    \"qwen-oauth\": \"alibaba\",\n    \"siliconflow\": \"siliconflow\",\n    \"stepfun\": \"stepfun\",\n    \"tencent-hunyuan\": \"tencent\",\n    \"tencent-tokenhub\": \"tencent\",\n    \"together\": \"togetherai\",\n    \"volcengine\": \"volcengine\",\n    \"xai\": \"xai\",\n    \"xiaomi\": \"xiaomi\",\n    \"zai\": \"zai\",\n}\n\n_URL_TO_PROVIDER: dict[str, str] = {\n    \"api-inference.modelscope.cn\": \"modelscope\",\n    \"api.anthropic.com\": \"anthropic\",\n    \"api.cerebras.ai\": \"cerebras\",\n    \"api.cohere.ai\": \"cohere\",\n    \"api.deepseek.com\": \"deepseek\",\n    \"api.githubcopilot.com\": \"copilot\",\n    \"api.groq.com\": \"groq\",\n    \"api.hunyuan.cloud.tencent.com\": \"tencent-hunyuan\",\n    \"api.minimax\": \"minimax\",\n    \"api.mistral.ai\": \"mistral\",\n    \"api.moonshot.ai\": \"moonshot\",\n    \"api.moonshot.cn\": \"moonshot-cn\",\n    \"api.openai.com\": \"openai\",\n    \"api.perplexity.ai\": \"perplexity\",\n    \"api.siliconflow.cn\": \"siliconflow\",\n    \"api.siliconflow.com\": \"siliconflow\",\n    \"api.stepfun.ai\": \"stepfun\",\n    \"api.stepfun.com\": \"stepfun\",\n    \"api.together.ai\": \"together\",\n    \"api.x.ai\": \"xai\",\n    \"api.xiaomimimo.com\": \"xiaomi\",\n    \"api.z.ai\": \"zai\",\n    \"ark.cn-beijing.volces.com\": \"volcengine\",\n    \"chatgpt.com\": \"openai-codex\",\n    \"dashscope-intl.aliyuncs.com\": \"alibaba\",\n    \"dashscope.aliyuncs.com\": \"alibaba\",\n    \"generativelanguage.googleapis.com\": \"google\",\n    \"integrate.api.nvidia.com\": \"nvidia\",\n    \"models.github.ai\": \"copilot\",\n    \"openrouter.ai\": \"openrouter\",\n    \"opencode.ai\": \"opencode-go\",\n    \"portal.qwen.ai\": \"qwen-oauth\",\n    \"qianfan.baidubce.com\": \"baidu-qianfan\",\n    \"tokenhub-intl.tencentcloudmaas.com\": \"tencent-tokenhub\",\n    \"tokenhub-intl.tencentmaas.cn\": \"tencent-tokenhub\",\n    \"tokenhub-intl.tencentmaas.com\": \"tencent-tokenhub\",\n    \"tokenhub.tencentmaas.cn\": \"tencent-tokenhub\",\n    \"tokenhub.tencentmaas.com\": \"tencent-tokenhub\",\n    \"xiaomimimo.com\": \"xiaomi\",\n}\n\n_models_dev_cache: dict[str, Any] = {}\n_models_dev_cache_time = 0.0\n_openrouter_cache: dict[str, \"ResolvedModelMetadata\"] = {}\n_openrouter_cache_time = 0.0\n_endpoint_metadata_cache: dict[str, dict[str, \"ResolvedModelMetadata\"]] = {}\n_endpoint_metadata_cache_time: dict[str, float] = {}\n\n\n@dataclass(frozen=True, slots=True)\nclass ResolvedModelMetadata:\n    context_window_tokens: int | None = None\n    max_output_tokens: int | None = None\n    source: str = \"unknown\"\n\n\ndef resolve_provider_model_metadata(\n    *,\n    provider_id: str,\n    model_id: str,\n    base_url: str | None = None,\n    api_key: str | None = None,\n) -> ResolvedModelMetadata | None:\n    \"\"\"Resolve metadata for one selected model from local, registry, or routed catalogs.\"\"\"\n\n    normalized_model = str(model_id or \"\").strip()\n    if not normalized_model:\n        return None\n\n    normalized_base_url = _normalize_base_url(base_url)\n    cached_context = get_cached_context_length(normalized_model, normalized_base_url)\n    if cached_context is not None:\n        return ResolvedModelMetadata(\n            context_window_tokens=cached_context,\n            source=\"context-length-cache\",\n        )\n\n    if normalized_base_url and is_local_endpoint(normalized_base_url):\n        local = query_local_endpoint_metadata(\n            model_id=normalized_model,\n            base_url=normalized_base_url,\n            api_key=api_key,\n        )\n        if local is not None and local.context_window_tokens is not None:\n            save_context_length(normalized_model, normalized_base_url, local.context_window_tokens)\n            return local\n\n    effective_provider = _effective_provider_id(provider_id=provider_id, base_url=normalized_base_url)\n    models_dev = lookup_models_dev_metadata(effective_provider, normalized_model)\n    if models_dev is not None and models_dev.context_window_tokens is not None:\n        return models_dev\n\n    openrouter = lookup_openrouter_metadata(normalized_model)\n    if openrouter is not None and openrouter.context_window_tokens is not None:\n        return openrouter\n\n    return None\n\n\ndef lookup_models_dev_metadata(provider_id: str, model_id: str) -> ResolvedModelMetadata | None:\n    models_dev_provider = _PROVIDER_TO_MODELS_DEV.get(provider_id.strip().lower())\n    if not models_dev_provider:\n        return None\n    data = fetch_models_dev_registry()\n    provider_data = data.get(models_dev_provider)\n    if not isinstance(provider_data, Mapping):\n        return None\n    models = provider_data.get(\"models\")\n    if not isinstance(models, Mapping):\n        return None\n    entry = _find_model_entry(models, model_id)\n    if entry is None and \"/\" in model_id:\n        entry = _find_model_entry(models, model_id.rsplit(\"/\", 1)[1])\n    if entry is None:\n        return None\n    limit = entry.get(\"limit\")\n    context = _coerce_reasonable_int(limit.get(\"context\") if isinstance(limit, Mapping) else None)\n    output = _coerce_reasonable_int(limit.get(\"output\") if isinstance(limit, Mapping) else None)\n    if context is None:\n        context = _extract_nested_int(entry, _CONTEXT_KEYS)\n    if output is None:\n        output = _extract_nested_int(entry, _OUTPUT_KEYS)\n    if context is None and output is None:\n        return None\n    return ResolvedModelMetadata(\n        context_window_tokens=context,\n        max_output_tokens=output,\n        source=f\"models.dev:{models_dev_provider}\",\n    )\n\n\ndef lookup_openrouter_metadata(model_id: str) -> ResolvedModelMetadata | None:\n    metadata = fetch_openrouter_model_metadata()\n    return _find_metadata_entry(metadata, model_id)\n\n\ndef get_cached_context_length(model_id: str, base_url: str | None) -> int | None:\n    normalized_model = str(model_id or \"\").strip()\n    normalized_base_url = _normalize_base_url(base_url)\n    if not normalized_model or not normalized_base_url:\n        return None\n    return _load_context_length_cache().get(_context_length_cache_key(normalized_model, normalized_base_url))\n\n\ndef save_context_length(model_id: str, base_url: str | None, context_window_tokens: int) -> None:\n    normalized_model = str(model_id or \"\").strip()\n    normalized_base_url = _normalize_base_url(base_url)\n    parsed = _coerce_reasonable_int(context_window_tokens)\n    if not normalized_model or not normalized_base_url or parsed is None:\n        return\n    cache = _load_context_length_cache()\n    key = _context_length_cache_key(normalized_model, normalized_base_url)\n    if cache.get(key) == parsed:\n        return\n    cache[key] = parsed\n    _save_context_length_cache(cache)\n\n\ndef query_local_endpoint_metadata(\n    *,\n    model_id: str,\n    base_url: str,\n    api_key: str | None = None,\n) -> ResolvedModelMetadata | None:\n    normalized = _normalize_base_url(base_url)\n    server_root = normalized[:-3] if normalized.endswith(\"/v1\") else normalized\n\n    lm_studio = _request_json_object(f\"{server_root}/api/v1/models\", timeout_seconds=2.0)\n    if lm_studio:\n        for model in _model_items(lm_studio):\n            candidate_id = str(model.get(\"key\") or model.get(\"id\") or \"\").strip()\n            if not _model_id_matches(candidate_id, model_id):\n                continue\n            loaded_instances = model.get(\"loaded_instances\")\n            if isinstance(loaded_instances, list):\n                for instance in loaded_instances:\n                    if not isinstance(instance, Mapping):\n                        continue\n                    config = instance.get(\"config\")\n                    if not isinstance(config, Mapping):\n                        continue\n                    context = _coerce_reasonable_int(config.get(\"context_length\"))\n                    if context is not None:\n                        return ResolvedModelMetadata(context_window_tokens=context, source=\"local:lm-studio\")\n            context = _extract_nested_int(model, _CONTEXT_KEYS)\n            output = _extract_nested_int(model, _OUTPUT_KEYS)\n            if context is not None or output is not None:\n                return ResolvedModelMetadata(context, output, \"local:lm-studio\")\n\n    detail = _request_json_object(\n        f\"{server_root}/v1/models/{model_id}\",\n        headers=_bearer_headers(api_key),\n        timeout_seconds=2.0,\n    )\n    if detail:\n        context = _extract_nested_int(detail, _CONTEXT_KEYS)\n        output = _extract_nested_int(detail, _OUTPUT_KEYS)\n        if context is not None or output is not None:\n            return ResolvedModelMetadata(context, output, \"local:model-detail\")\n\n    listed = fetch_endpoint_model_metadata(base_url=normalized, api_key=api_key)\n    entry = _find_metadata_entry(listed, model_id)\n    if entry is not None:\n        return entry\n\n    props = _request_json_object(f\"{server_root}/v1/props\", headers=_bearer_headers(api_key), timeout_seconds=2.0)\n    if not props:\n        props = _request_json_object(f\"{server_root}/props\", headers=_bearer_headers(api_key), timeout_seconds=2.0)\n    if props:\n        context = _extract_nested_int(props, _CONTEXT_KEYS)\n        if context is not None:\n            return ResolvedModelMetadata(context_window_tokens=context, source=\"local:llamacpp\")\n\n    return None\n\n\ndef fetch_models_dev_registry(*, force_refresh: bool = False) -> Mapping[str, Any]:\n    global _models_dev_cache, _models_dev_cache_time\n    if (\n        not force_refresh\n        and _models_dev_cache\n        and (time.time() - _models_dev_cache_time) < _MODELS_DEV_CACHE_TTL_SECONDS\n    ):\n        return _models_dev_cache\n\n    try:\n        payload = _request_json_object(MODELS_DEV_URL, timeout_seconds=10.0)\n        if payload:\n            _models_dev_cache = dict(payload)\n            _models_dev_cache_time = time.time()\n            _save_models_dev_disk_cache(_models_dev_cache)\n            return _models_dev_cache\n    except RuntimeError:\n        pass\n\n    if not _models_dev_cache:\n        _models_dev_cache = _load_models_dev_disk_cache()\n        if _models_dev_cache:\n            _models_dev_cache_time = time.time() - _MODELS_DEV_CACHE_TTL_SECONDS + 300\n    return _models_dev_cache\n\n\ndef fetch_openrouter_model_metadata(*, force_refresh: bool = False) -> Mapping[str, ResolvedModelMetadata]:\n    global _openrouter_cache, _openrouter_cache_time\n    if (\n        not force_refresh\n        and _openrouter_cache\n        and (time.time() - _openrouter_cache_time) < _OPENROUTER_CACHE_TTL_SECONDS\n    ):\n        return _openrouter_cache\n\n    try:\n        payload = _request_json_object(OPENROUTER_MODELS_URL, timeout_seconds=10.0)\n    except RuntimeError:\n        return _openrouter_cache\n    cache: dict[str, ResolvedModelMetadata] = {}\n    for model in _model_items(payload):\n        model_id = str(model.get(\"id\") or \"\").strip()\n        if not model_id:\n            continue\n        top_provider = model.get(\"top_provider\")\n        output = None\n        if isinstance(top_provider, Mapping):\n            output = _coerce_reasonable_int(top_provider.get(\"max_completion_tokens\"))\n        entry = ResolvedModelMetadata(\n            context_window_tokens=_extract_nested_int(model, _CONTEXT_KEYS),\n            max_output_tokens=output or _extract_nested_int(model, _OUTPUT_KEYS),\n            source=\"openrouter\",\n        )\n        _add_metadata_aliases(cache, model_id, entry)\n        canonical_slug = str(model.get(\"canonical_slug\") or \"\").strip()\n        if canonical_slug:\n            _add_metadata_aliases(cache, canonical_slug, entry)\n    _openrouter_cache = cache\n    _openrouter_cache_time = time.time()\n    return _openrouter_cache\n\n\ndef fetch_endpoint_model_metadata(\n    *,\n    base_url: str,\n    api_key: str | None = None,\n    force_refresh: bool = False,\n) -> Mapping[str, ResolvedModelMetadata]:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return {}\n    if not force_refresh:\n        cached = _endpoint_metadata_cache.get(normalized)\n        cached_at = _endpoint_metadata_cache_time.get(normalized, 0)\n        if cached is not None and (time.time() - cached_at) < _ENDPOINT_CACHE_TTL_SECONDS:\n            return cached\n\n    candidates = [normalized]\n    alternate = normalized[:-3].rstrip(\"/\") if normalized.endswith(\"/v1\") else f\"{normalized}/v1\"\n    if alternate and alternate not in candidates:\n        candidates.append(alternate)\n\n    for candidate in candidates:\n        payload = _request_json_object(\n            f\"{candidate.rstrip('/')}/models\",\n            headers=_bearer_headers(api_key),\n            timeout_seconds=5.0,\n        )\n        if not payload:\n            continue\n        cache: dict[str, ResolvedModelMetadata] = {}\n        for model in _model_items(payload):\n            model_id = str(model.get(\"id\") or model.get(\"key\") or \"\").strip()\n            if not model_id:\n                continue\n            entry = ResolvedModelMetadata(\n                context_window_tokens=_extract_nested_int(model, _CONTEXT_KEYS),\n                max_output_tokens=_extract_nested_int(model, _OUTPUT_KEYS),\n                source=\"endpoint:/models\",\n            )\n            _add_metadata_aliases(cache, model_id, entry)\n        if cache:\n            _endpoint_metadata_cache[normalized] = cache\n            _endpoint_metadata_cache_time[normalized] = time.time()\n            return cache\n\n    _endpoint_metadata_cache[normalized] = {}\n    _endpoint_metadata_cache_time[normalized] = time.time()\n    return {}\n\n\ndef is_local_endpoint(base_url: str | None) -> bool:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return False\n    try:\n        parsed = urlparse(normalized if \"://\" in normalized else f\"http://{normalized}\")\n    except ValueError:\n        return False\n    host = parsed.hostname or \"\"\n    if host in _LOCAL_HOSTS or any(host.endswith(suffix) for suffix in _CONTAINER_LOCAL_SUFFIXES):\n        return True\n    try:\n        address = ipaddress.ip_address(host)\n    except ValueError:\n        return False\n    return address.is_private or address.is_loopback or address.is_link_local\n\n\ndef infer_provider_from_url(base_url: str | None) -> str | None:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return None\n    parsed = urlparse(normalized if \"://\" in normalized else f\"https://{normalized}\")\n    host = parsed.netloc.lower() or parsed.path.lower()\n    for url_part, provider_id in _URL_TO_PROVIDER.items():\n        if url_part in host:\n            return provider_id\n    return None\n\n\ndef _effective_provider_id(*, provider_id: str, base_url: str | None) -> str:\n    normalized_provider = provider_id.strip().lower()\n    if normalized_provider in {\"\", \"openai-compatible\"}:\n        return infer_provider_from_url(base_url) or normalized_provider\n    return normalized_provider\n\n\ndef _request_json_object(\n    url: str,\n    *,\n    headers: Mapping[str, str] | None = None,\n    timeout_seconds: float = 10.0,\n) -> Mapping[str, Any] | None:\n    http_request = request.Request(url, headers=dict(headers or {}), method=\"GET\")\n    try:\n        with request.urlopen(http_request, timeout=timeout_seconds) as response:\n            raw_body = response.read().decode(\"utf-8\")\n    except (error.HTTPError, error.URLError, TimeoutError) as exc:\n        return None\n    try:\n        payload = json.loads(raw_body) if raw_body else {}\n    except json.JSONDecodeError:\n        return None\n    if isinstance(payload, list):\n        return {\"data\": payload}\n    if not isinstance(payload, Mapping):\n        return None\n    return payload\n\n\ndef _model_items(payload: Mapping[str, Any]) -> tuple[Mapping[str, Any], ...]:\n    for key in (\"data\", \"models\"):\n        value = payload.get(key)\n        if isinstance(value, list):\n            return tuple(item for item in value if isinstance(item, Mapping))\n    return ()\n\n\ndef _iter_nested_mappings(value: Any):\n    if isinstance(value, Mapping):\n        yield value\n        for nested in value.values():\n            yield from _iter_nested_mappings(nested)\n    elif isinstance(value, list):\n        for item in value:\n            yield from _iter_nested_mappings(item)\n\n\ndef _extract_nested_int(payload: Mapping[str, Any], keys: tuple[str, ...]) -> int | None:\n    keyset = {key.casefold() for key in keys}\n    for mapping in _iter_nested_mappings(payload):\n        for key, value in mapping.items():\n            if str(key).casefold() not in keyset:\n                continue\n            parsed = _coerce_reasonable_int(value)\n            if parsed is not None:\n                return parsed\n    return None\n\n\ndef _coerce_reasonable_int(value: Any, *, minimum: int = 1024, maximum: int = 10_000_000) -> int | None:\n    try:\n        if isinstance(value, bool):\n            return None\n        if isinstance(value, str):\n            value = value.strip().replace(\",\", \"\")\n        parsed = int(value)\n    except (TypeError, ValueError):\n        return None\n    if minimum <= parsed <= maximum:\n        return parsed\n    return None\n\n\ndef _find_model_entry(models: Mapping[str, Any], model_id: str) -> Mapping[str, Any] | None:\n    entry = models.get(model_id)\n    if isinstance(entry, Mapping):\n        return entry\n    model_lower = model_id.casefold()\n    for candidate_id, candidate in models.items():\n        if str(candidate_id).casefold() == model_lower and isinstance(candidate, Mapping):\n            return candidate\n    return None\n\n\ndef _find_metadata_entry(\n    metadata: Mapping[str, ResolvedModelMetadata],\n    model_id: str,\n) -> ResolvedModelMetadata | None:\n    entry = metadata.get(model_id)\n    if entry is not None:\n        return entry\n    if \"/\" not in model_id:\n        for candidate_id, candidate in metadata.items():\n            if candidate_id.rsplit(\"/\", 1)[-1] == model_id:\n                return candidate\n    model_lower = model_id.casefold()\n    for candidate_id, candidate in metadata.items():\n        if candidate_id.casefold() == model_lower:\n            return candidate\n    return None\n\n\ndef _add_metadata_aliases(\n    cache: dict[str, ResolvedModelMetadata],\n    model_id: str,\n    entry: ResolvedModelMetadata,\n) -> None:\n    cache[model_id] = entry\n    if \"/\" in model_id:\n        cache.setdefault(model_id.rsplit(\"/\", 1)[1], entry)\n\n\ndef _model_id_matches(candidate_id: str, lookup_model: str) -> bool:\n    if candidate_id == lookup_model:\n        return True\n    return \"/\" in candidate_id and candidate_id.rsplit(\"/\", 1)[1] == lookup_model\n\n\ndef _bearer_headers(api_key: str | None) -> dict[str, str]:\n    resolved = str(api_key or \"\").strip()\n    return {\"Authorization\": f\"Bearer {resolved}\"} if resolved else {}\n\n\ndef _normalize_base_url(base_url: str | None) -> str:\n    return str(base_url or \"\").strip().rstrip(\"/\")\n\n\ndef _models_dev_disk_cache_path() -> Path:\n    override = os.environ.get(\"ELEPHANT_MODELS_DEV_CACHE_PATH\", \"\").strip()\n    if override:\n        return Path(override).expanduser()\n    return default_cli_state_dir() / \"models-dev-cache.json\"\n\n\ndef _context_length_cache_path() -> Path:\n    override = os.environ.get(\"ELEPHANT_CONTEXT_LENGTH_CACHE_PATH\", \"\").strip()\n    if override:\n        return Path(override).expanduser()\n    return default_cli_state_dir() / \"context-length-cache.json\"\n\n\ndef _context_length_cache_key(model_id: str, base_url: str) -> str:\n    return f\"{model_id}@{base_url}\"\n\n\ndef _load_context_length_cache() -> dict[str, int]:\n    path = _context_length_cache_path()\n    try:\n        payload = json.loads(path.read_text(encoding=\"utf-8\"))\n    except (OSError, json.JSONDecodeError):\n        return {}\n    raw_cache = payload.get(\"context_lengths\") if isinstance(payload, Mapping) else None\n    if not isinstance(raw_cache, Mapping):\n        return {}\n    cache: dict[str, int] = {}\n    for key, value in raw_cache.items():\n        parsed = _coerce_reasonable_int(value)\n        if parsed is not None:\n            cache[str(key)] = parsed\n    return cache\n\n\ndef _save_context_length_cache(cache: Mapping[str, int]) -> None:\n    path = _context_length_cache_path()\n    try:\n        path.parent.mkdir(parents=True, exist_ok=True)\n        with tempfile.NamedTemporaryFile(\"w\", encoding=\"utf-8\", dir=path.parent, delete=False) as handle:\n            json.dump({\"context_lengths\": dict(cache)}, handle, separators=(\",\", \":\"))\n            temp_name = handle.name\n        Path(temp_name).replace(path)\n    except OSError:\n        return\n\n\ndef _load_models_dev_disk_cache() -> dict[str, Any]:\n    path = _models_dev_disk_cache_path()\n    try:\n        payload = json.loads(path.read_text(encoding=\"utf-8\"))\n    except (OSError, json.JSONDecodeError):\n        return {}\n    return dict(payload) if isinstance(payload, Mapping) else {}\n\n\ndef _save_models_dev_disk_cache(payload: Mapping[str, Any]) -> None:\n    path = _models_dev_disk_cache_path()\n    try:\n        path.parent.mkdir(parents=True, exist_ok=True)\n        with tempfile.NamedTemporaryFile(\"w\", encoding=\"utf-8\", dir=path.parent, delete=False) as handle:\n            json.dump(dict(payload), handle, separators=(\",\", \":\"))\n            temp_name = handle.name\n        Path(temp_name).replace(path)\n    except OSError:\n        return\n\n\n__all__ = [\n    \"ResolvedModelMetadata\",\n    \"fetch_endpoint_model_metadata\",\n    \"fetch_models_dev_registry\",\n    \"fetch_openrouter_model_metadata\",\n    \"get_cached_context_length\",\n    \"infer_provider_from_url\",\n    \"is_local_endpoint\",\n    \"lookup_models_dev_metadata\",\n    \"lookup_openrouter_metadata\",\n    \"query_local_endpoint_metadata\",\n    \"resolve_provider_model_metadata\",\n    \"save_context_length\",\n]\n"}, "80147155ed93065761d7d90b52b875505760d9b4": {"t": 1785758538.5653996, "n": 22066, "h": "eb2c81c8a4a135b4", "c": "\"\"\"Provider-aware model metadata fallback resolution.\"\"\"\n\nfrom __future__ import annotations\n\nfrom collections.abc import Mapping\nfrom dataclasses import dataclass\nimport ipaddress\nimport json\nimport os\nfrom pathlib import Path\nimport tempfile\nimport time\nfrom typing import Any\nfrom urllib import error, request\nfrom urllib.parse import urlparse\n\nfrom packages.runtime_layout import default_cli_state_dir\n\n\nMODELS_DEV_URL = \"https://models.dev/api.json\"\nOPENROUTER_MODELS_URL = \"https://openrouter.ai/api/v1/models\"\n_MODELS_DEV_CACHE_TTL_SECONDS = 3600\n_OPENROUTER_CACHE_TTL_SECONDS = 3600\n_ENDPOINT_CACHE_TTL_SECONDS = 300\n\n_LOCAL_HOSTS = frozenset({\"localhost\", \"127.0.0.1\", \"::1\", \"0.0.0.0\"})\n_CONTAINER_LOCAL_SUFFIXES = (\".docker.internal\", \".containers.internal\", \".lima.internal\")\n\n_CONTEXT_KEYS = (\n    \"context_length\",\n    \"context_window\",\n    \"max_context_length\",\n    \"max_position_embeddings\",\n    \"max_model_len\",\n    \"max_input_tokens\",\n    \"max_sequence_length\",\n    \"max_seq_len\",\n    \"n_ctx\",\n    \"n_ctx_train\",\n)\n_OUTPUT_KEYS = (\n    \"max_completion_tokens\",\n    \"max_output_tokens\",\n    \"max_tokens\",\n)\n\n_PROVIDER_TO_MODELS_DEV: dict[str, str] = {\n    \"alibaba\": \"alibaba\",\n    \"anthropic\": \"anthropic\",\n    \"claude-code\": \"anthropic\",\n    \"copilot\": \"github-copilot\",\n    \"deepseek\": \"deepseek\",\n    \"fireworks\": \"fireworks-ai\",\n    \"google\": \"google\",\n    \"groq\": \"groq\",\n    \"huggingface\": \"huggingface\",\n    \"kilocode\": \"kilo\",\n    \"minimax\": \"minimax\",\n    \"minimax-cn\": \"minimax-cn\",\n    \"mistral\": \"mistral\",\n    \"moonshot\": \"kimi-for-coding\",\n    \"moonshot-cn\": \"kimi-for-coding\",\n    \"openai\": \"openai\",\n    \"openai-codex\": \"openai\",\n    \"opencode-go\": \"opencode-go\",\n    \"opencode-zen\": \"opencode\",\n    \"openrouter\": \"openrouter\",\n    \"qwen-oauth\": \"alibaba\",\n    \"together\": \"togetherai\",\n    \"xai\": \"xai\",\n    \"xiaomi\": \"xiaomi\",\n    \"zai\": \"zai\",\n}\n\n_URL_TO_PROVIDER: dict[str, str] = {\n    \"api.anthropic.com\": \"anthropic\",\n    \"api.deepseek.com\": \"deepseek\",\n    \"api.githubcopilot.com\": \"copilot\",\n    \"api.groq.com\": \"groq\",\n    \"api.minimax\": \"minimax\",\n    \"api.mistral.ai\": \"mistral\",\n    \"api.moonshot.ai\": \"moonshot\",\n    \"api.moonshot.cn\": \"moonshot-cn\",\n    \"api.openai.com\": \"openai\",\n    \"api.together.ai\": \"together\",\n    \"api.x.ai\": \"xai\",\n    \"api.xiaomimimo.com\": \"xiaomi\",\n    \"api.z.ai\": \"zai\",\n    \"chatgpt.com\": \"openai-codex\",\n    \"dashscope-intl.aliyuncs.com\": \"alibaba\",\n    \"dashscope.aliyuncs.com\": \"alibaba\",\n    \"generativelanguage.googleapis.com\": \"google\",\n    \"models.github.ai\": \"copilot\",\n    \"openrouter.ai\": \"openrouter\",\n    \"opencode.ai\": \"opencode-go\",\n    \"portal.qwen.ai\": \"qwen-oauth\",\n    \"xiaomimimo.com\": \"xiaomi\",\n}\n\n_models_dev_cache: dict[str, Any] = {}\n_models_dev_cache_time = 0.0\n_openrouter_cache: dict[str, \"ResolvedModelMetadata\"] = {}\n_openrouter_cache_time = 0.0\n_endpoint_metadata_cache: dict[str, dict[str, \"ResolvedModelMetadata\"]] = {}\n_endpoint_metadata_cache_time: dict[str, float] = {}\n\n\n@dataclass(frozen=True, slots=True)\nclass ResolvedModelMetadata:\n    context_window_tokens: int | None = None\n    max_output_tokens: int | None = None\n    source: str = \"unknown\"\n\n\ndef resolve_provider_model_metadata(\n    *,\n    provider_id: str,\n    model_id: str,\n    base_url: str | None = None,\n    api_key: str | None = None,\n) -> ResolvedModelMetadata | None:\n    \"\"\"Resolve metadata for one selected model from local, registry, or routed catalogs.\"\"\"\n\n    normalized_model = str(model_id or \"\").strip()\n    if not normalized_model:\n        return None\n\n    normalized_base_url = _normalize_base_url(base_url)\n    cached_context = get_cached_context_length(normalized_model, normalized_base_url)\n    if cached_context is not None:\n        return ResolvedModelMetadata(\n            context_window_tokens=cached_context,\n            source=\"context-length-cache\",\n        )\n\n    if normalized_base_url and is_local_endpoint(normalized_base_url):\n        local = query_local_endpoint_metadata(\n            model_id=normalized_model,\n            base_url=normalized_base_url,\n            api_key=api_key,\n        )\n        if local is not None and local.context_window_tokens is not None:\n            save_context_length(normalized_model, normalized_base_url, local.context_window_tokens)\n            return local\n\n    effective_provider = _effective_provider_id(provider_id=provider_id, base_url=normalized_base_url)\n    models_dev = lookup_models_dev_metadata(effective_provider, normalized_model)\n    if models_dev is not None and models_dev.context_window_tokens is not None:\n        return models_dev\n\n    openrouter = lookup_openrouter_metadata(normalized_model)\n    if openrouter is not None and openrouter.context_window_tokens is not None:\n        return openrouter\n\n    return None\n\n\ndef lookup_models_dev_metadata(provider_id: str, model_id: str) -> ResolvedModelMetadata | None:\n    models_dev_provider = _PROVIDER_TO_MODELS_DEV.get(provider_id.strip().lower())\n    if not models_dev_provider:\n        return None\n    data = fetch_models_dev_registry()\n    provider_data = data.get(models_dev_provider)\n    if not isinstance(provider_data, Mapping):\n        return None\n    models = provider_data.get(\"models\")\n    if not isinstance(models, Mapping):\n        return None\n    entry = _find_model_entry(models, model_id)\n    if entry is None and \"/\" in model_id:\n        entry = _find_model_entry(models, model_id.rsplit(\"/\", 1)[1])\n    if entry is None:\n        return None\n    limit = entry.get(\"limit\")\n    context = _coerce_reasonable_int(limit.get(\"context\") if isinstance(limit, Mapping) else None)\n    output = _coerce_reasonable_int(limit.get(\"output\") if isinstance(limit, Mapping) else None)\n    if context is None:\n        context = _extract_nested_int(entry, _CONTEXT_KEYS)\n    if output is None:\n        output = _extract_nested_int(entry, _OUTPUT_KEYS)\n    if context is None and output is None:\n        return None\n    return ResolvedModelMetadata(\n        context_window_tokens=context,\n        max_output_tokens=output,\n        source=f\"models.dev:{models_dev_provider}\",\n    )\n\n\ndef lookup_openrouter_metadata(model_id: str) -> ResolvedModelMetadata | None:\n    metadata = fetch_openrouter_model_metadata()\n    return _find_metadata_entry(metadata, model_id)\n\n\ndef get_cached_context_length(model_id: str, base_url: str | None) -> int | None:\n    normalized_model = str(model_id or \"\").strip()\n    normalized_base_url = _normalize_base_url(base_url)\n    if not normalized_model or not normalized_base_url:\n        return None\n    return _load_context_length_cache().get(_context_length_cache_key(normalized_model, normalized_base_url))\n\n\ndef save_context_length(model_id: str, base_url: str | None, context_window_tokens: int) -> None:\n    normalized_model = str(model_id or \"\").strip()\n    normalized_base_url = _normalize_base_url(base_url)\n    parsed = _coerce_reasonable_int(context_window_tokens)\n    if not normalized_model or not normalized_base_url or parsed is None:\n        return\n    cache = _load_context_length_cache()\n    key = _context_length_cache_key(normalized_model, normalized_base_url)\n    if cache.get(key) == parsed:\n        return\n    cache[key] = parsed\n    _save_context_length_cache(cache)\n\n\ndef query_local_endpoint_metadata(\n    *,\n    model_id: str,\n    base_url: str,\n    api_key: str | None = None,\n) -> ResolvedModelMetadata | None:\n    normalized = _normalize_base_url(base_url)\n    server_root = normalized[:-3] if normalized.endswith(\"/v1\") else normalized\n\n    lm_studio = _request_json_object(f\"{server_root}/api/v1/models\", timeout_seconds=2.0)\n    if lm_studio:\n        for model in _model_items(lm_studio):\n            candidate_id = str(model.get(\"key\") or model.get(\"id\") or \"\").strip()\n            if not _model_id_matches(candidate_id, model_id):\n                continue\n            loaded_instances = model.get(\"loaded_instances\")\n            if isinstance(loaded_instances, list):\n                for instance in loaded_instances:\n                    if not isinstance(instance, Mapping):\n                        continue\n                    config = instance.get(\"config\")\n                    if not isinstance(config, Mapping):\n                        continue\n                    context = _coerce_reasonable_int(config.get(\"context_length\"))\n                    if context is not None:\n                        return ResolvedModelMetadata(context_window_tokens=context, source=\"local:lm-studio\")\n            context = _extract_nested_int(model, _CONTEXT_KEYS)\n            output = _extract_nested_int(model, _OUTPUT_KEYS)\n            if context is not None or output is not None:\n                return ResolvedModelMetadata(context, output, \"local:lm-studio\")\n\n    detail = _request_json_object(\n        f\"{server_root}/v1/models/{model_id}\",\n        headers=_bearer_headers(api_key),\n        timeout_seconds=2.0,\n    )\n    if detail:\n        context = _extract_nested_int(detail, _CONTEXT_KEYS)\n        output = _extract_nested_int(detail, _OUTPUT_KEYS)\n        if context is not None or output is not None:\n            return ResolvedModelMetadata(context, output, \"local:model-detail\")\n\n    listed = fetch_endpoint_model_metadata(base_url=normalized, api_key=api_key)\n    entry = _find_metadata_entry(listed, model_id)\n    if entry is not None:\n        return entry\n\n    props = _request_json_object(f\"{server_root}/v1/props\", headers=_bearer_headers(api_key), timeout_seconds=2.0)\n    if not props:\n        props = _request_json_object(f\"{server_root}/props\", headers=_bearer_headers(api_key), timeout_seconds=2.0)\n    if props:\n        context = _extract_nested_int(props, _CONTEXT_KEYS)\n        if context is not None:\n            return ResolvedModelMetadata(context_window_tokens=context, source=\"local:llamacpp\")\n\n    return None\n\n\ndef fetch_models_dev_registry(*, force_refresh: bool = False) -> Mapping[str, Any]:\n    global _models_dev_cache, _models_dev_cache_time\n    if (\n        not force_refresh\n        and _models_dev_cache\n        and (time.time() - _models_dev_cache_time) < _MODELS_DEV_CACHE_TTL_SECONDS\n    ):\n        return _models_dev_cache\n\n    try:\n        payload = _request_json_object(MODELS_DEV_URL, timeout_seconds=10.0)\n        if payload:\n            _models_dev_cache = dict(payload)\n            _models_dev_cache_time = time.time()\n            _save_models_dev_disk_cache(_models_dev_cache)\n            return _models_dev_cache\n    except RuntimeError:\n        pass\n\n    if not _models_dev_cache:\n        _models_dev_cache = _load_models_dev_disk_cache()\n        if _models_dev_cache:\n            _models_dev_cache_time = time.time() - _MODELS_DEV_CACHE_TTL_SECONDS + 300\n    return _models_dev_cache\n\n\ndef fetch_openrouter_model_metadata(*, force_refresh: bool = False) -> Mapping[str, ResolvedModelMetadata]:\n    global _openrouter_cache, _openrouter_cache_time\n    if (\n        not force_refresh\n        and _openrouter_cache\n        and (time.time() - _openrouter_cache_time) < _OPENROUTER_CACHE_TTL_SECONDS\n    ):\n        return _openrouter_cache\n\n    try:\n        payload = _request_json_object(OPENROUTER_MODELS_URL, timeout_seconds=10.0)\n    except RuntimeError:\n        return _openrouter_cache\n    cache: dict[str, ResolvedModelMetadata] = {}\n    for model in _model_items(payload):\n        model_id = str(model.get(\"id\") or \"\").strip()\n        if not model_id:\n            continue\n        top_provider = model.get(\"top_provider\")\n        output = None\n        if isinstance(top_provider, Mapping):\n            output = _coerce_reasonable_int(top_provider.get(\"max_completion_tokens\"))\n        entry = ResolvedModelMetadata(\n            context_window_tokens=_extract_nested_int(model, _CONTEXT_KEYS),\n            max_output_tokens=output or _extract_nested_int(model, _OUTPUT_KEYS),\n            source=\"openrouter\",\n        )\n        _add_metadata_aliases(cache, model_id, entry)\n        canonical_slug = str(model.get(\"canonical_slug\") or \"\").strip()\n        if canonical_slug:\n            _add_metadata_aliases(cache, canonical_slug, entry)\n    _openrouter_cache = cache\n    _openrouter_cache_time = time.time()\n    return _openrouter_cache\n\n\ndef fetch_endpoint_model_metadata(\n    *,\n    base_url: str,\n    api_key: str | None = None,\n    force_refresh: bool = False,\n) -> Mapping[str, ResolvedModelMetadata]:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return {}\n    if not force_refresh:\n        cached = _endpoint_metadata_cache.get(normalized)\n        cached_at = _endpoint_metadata_cache_time.get(normalized, 0)\n        if cached is not None and (time.time() - cached_at) < _ENDPOINT_CACHE_TTL_SECONDS:\n            return cached\n\n    candidates = [normalized]\n    alternate = normalized[:-3].rstrip(\"/\") if normalized.endswith(\"/v1\") else f\"{normalized}/v1\"\n    if alternate and alternate not in candidates:\n        candidates.append(alternate)\n\n    for candidate in candidates:\n        payload = _request_json_object(\n            f\"{candidate.rstrip('/')}/models\",\n            headers=_bearer_headers(api_key),\n            timeout_seconds=5.0,\n        )\n        if not payload:\n            continue\n        cache: dict[str, ResolvedModelMetadata] = {}\n        for model in _model_items(payload):\n            model_id = str(model.get(\"id\") or model.get(\"key\") or \"\").strip()\n            if not model_id:\n                continue\n            entry = ResolvedModelMetadata(\n                context_window_tokens=_extract_nested_int(model, _CONTEXT_KEYS),\n                max_output_tokens=_extract_nested_int(model, _OUTPUT_KEYS),\n                source=\"endpoint:/models\",\n            )\n            _add_metadata_aliases(cache, model_id, entry)\n        if cache:\n            _endpoint_metadata_cache[normalized] = cache\n            _endpoint_metadata_cache_time[normalized] = time.time()\n            return cache\n\n    _endpoint_metadata_cache[normalized] = {}\n    _endpoint_metadata_cache_time[normalized] = time.time()\n    return {}\n\n\ndef is_local_endpoint(base_url: str | None) -> bool:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return False\n    try:\n        parsed = urlparse(normalized if \"://\" in normalized else f\"http://{normalized}\")\n    except ValueError:\n        return False\n    host = parsed.hostname or \"\"\n    if host in _LOCAL_HOSTS or any(host.endswith(suffix) for suffix in _CONTAINER_LOCAL_SUFFIXES):\n        return True\n    try:\n        address = ipaddress.ip_address(host)\n    except ValueError:\n        return False\n    return address.is_private or address.is_loopback or address.is_link_local\n\n\ndef infer_provider_from_url(base_url: str | None) -> str | None:\n    normalized = _normalize_base_url(base_url)\n    if not normalized:\n        return None\n    parsed = urlparse(normalized if \"://\" in normalized else f\"https://{normalized}\")\n    host = parsed.netloc.lower() or parsed.path.lower()\n    for url_part, provider_id in _URL_TO_PROVIDER.items():\n        if url_part in host:\n            return provider_id\n    return None\n\n\ndef _effective_provider_id(*, provider_id: str, base_url: str | None) -> str:\n    normalized_provider = provider_id.strip().lower()\n    if normalized_provider in {\"\", \"openai-compatible\"}:\n        return infer_provider_from_url(base_url) or normalized_provider\n    return normalized_provider\n\n\ndef _request_json_object(\n    url: str,\n    *,\n    headers: Mapping[str, str] | None = None,\n    timeout_seconds: float = 10.0,\n) -> Mapping[str, Any] | None:\n    http_request = request.Request(url, headers=dict(headers or {}), method=\"GET\")\n    try:\n        with request.urlopen(http_request, timeout=timeout_seconds) as response:\n            raw_body = response.read().decode(\"utf-8\")\n    except (error.HTTPError, error.URLError, TimeoutError) as exc:\n        return None\n    try:\n        payload = json.loads(raw_body) if raw_body else {}\n    except json.JSONDecodeError:\n        return None\n    if isinstance(payload, list):\n        return {\"data\": payload}\n    if not isinstance(payload, Mapping):\n        return None\n    return payload\n\n\ndef _model_items(payload: Mapping[str, Any]) -> tuple[Mapping[str, Any], ...]:\n    for key in (\"data\", \"models\"):\n        value = payload.get(key)\n        if isinstance(value, list):\n            return tuple(item for item in value if isinstance(item, Mapping))\n    return ()\n\n\ndef _iter_nested_mappings(value: Any):\n    if isinstance(value, Mapping):\n        yield value\n        for nested in value.values():\n            yield from _iter_nested_mappings(nested)\n    elif isinstance(value, list):\n        for item in value:\n            yield from _iter_nested_mappings(item)\n\n\ndef _extract_nested_int(payload: Mapping[str, Any], keys: tuple[str, ...]) -> int | None:\n    keyset = {key.casefold() for key in keys}\n    for mapping in _iter_nested_mappings(payload):\n        for key, value in mapping.items():\n            if str(key).casefold() not in keyset:\n                continue\n            parsed = _coerce_reasonable_int(value)\n            if parsed is not None:\n                return parsed\n    return None\n\n\ndef _coerce_reasonable_int(value: Any, *, minimum: int = 1024, maximum: int = 10_000_000) -> int | None:\n    try:\n        if isinstance(value, bool):\n            return None\n        if isinstance(value, str):\n            value = value.strip().replace(\",\", \"\")\n        parsed = int(value)\n    except (TypeError, ValueError):\n        return None\n    if minimum <= parsed <= maximum:\n        return parsed\n    return None\n\n\ndef _find_model_entry(models: Mapping[str, Any], model_id: str) -> Mapping[str, Any] | None:\n    entry = models.get(model_id)\n    if isinstance(entry, Mapping):\n        return entry\n    model_lower = model_id.casefold()\n    for candidate_id, candidate in models.items():\n        if str(candidate_id).casefold() == model_lower and isinstance(candidate, Mapping):\n            return candidate\n    return None\n\n\ndef _find_metadata_entry(\n    metadata: Mapping[str, ResolvedModelMetadata],\n    model_id: str,\n) -> ResolvedModelMetadata | None:\n    entry = metadata.get(model_id)\n    if entry is not None:\n        return entry\n    if \"/\" not in model_id:\n        for candidate_id, candidate in metadata.items():\n            if candidate_id.rsplit(\"/\", 1)[-1] == model_id:\n                return candidate\n    model_lower = model_id.casefold()\n    for candidate_id, candidate in metadata.items():\n        if candidate_id.casefold() == model_lower:\n            return candidate\n    return None\n\n\ndef _add_metadata_aliases(\n    cache: dict[str, ResolvedModelMetadata],\n    model_id: str,\n    entry: ResolvedModelMetadata,\n) -> None:\n    cache[model_id] = entry\n    if \"/\" in model_id:\n        cache.setdefault(model_id.rsplit(\"/\", 1)[1], entry)\n\n\ndef _model_id_matches(candidate_id: str, lookup_model: str) -> bool:\n    if candidate_id == lookup_model:\n        return True\n    return \"/\" in candidate_id and candidate_id.rsplit(\"/\", 1)[1] == lookup_model\n\n\ndef _bearer_headers(api_key: str | None) -> dict[str, str]:\n    resolved = str(api_key or \"\").strip()\n    return {\"Authorization\": f\"Bearer {resolved}\"} if resolved else {}\n\n\ndef _normalize_base_url(base_url: str | None) -> str:\n    return str(base_url or \"\").strip().rstrip(\"/\")\n\n\ndef _models_dev_disk_cache_path() -> Path:\n    override = os.environ.get(\"ELEPHANT_MODELS_DEV_CACHE_PATH\", \"\").strip()\n    if override:\n        return Path(override).expanduser()\n    return default_cli_state_dir() / \"models-dev-cache.json\"\n\n\ndef _context_length_cache_path() -> Path:\n    override = os.environ.get(\"ELEPHANT_CONTEXT_LENGTH_CACHE_PATH\", \"\").strip()\n    if override:\n        return Path(override).expanduser()\n    return default_cli_state_dir() / \"context-length-cache.json\"\n\n\ndef _context_length_cache_key(model_id: str, base_url: str) -> str:\n    return f\"{model_id}@{base_url}\"\n\n\ndef _load_context_length_cache() -> dict[str, int]:\n    path = _context_length_cache_path()\n    try:\n        payload = json.loads(path.read_text(encoding=\"utf-8\"))\n    except (OSError, json.JSONDecodeError):\n        return {}\n    raw_cache = payload.get(\"context_lengths\") if isinstance(payload, Mapping) else None\n    if not isinstance(raw_cache, Mapping):\n        return {}\n    cache: dict[str, int] = {}\n    for key, value in raw_cache.items():\n        parsed = _coerce_reasonable_int(value)\n        if parsed is not None:\n            cache[str(key)] = parsed\n    return cache\n\n\ndef _save_context_length_cache(cache: Mapping[str, int]) -> None:\n    path = _context_length_cache_path()\n    try:\n        path.parent.mkdir(parents=True, exist_ok=True)\n        with tempfile.NamedTemporaryFile(\"w\", encoding=\"utf-8\", dir=path.parent, delete=False) as handle:\n            json.dump({\"context_lengths\": dict(cache)}, handle, separators=(\",\", \":\"))\n            temp_name = handle.name\n        Path(temp_name).replace(path)\n    except OSError:\n        return\n\n\ndef _load_models_dev_disk_cache() -> dict[str, Any]:\n    path = _models_dev_disk_cache_path()\n    try:\n        payload = json.loads(path.read_text(encoding=\"utf-8\"))\n    except (OSError, json.JSONDecodeError):\n        return {}\n    return dict(payload) if isinstance(payload, Mapping) else {}\n\n\ndef _save_models_dev_disk_cache(payload: Mapping[str, Any]) -> None:\n    path = _models_dev_disk_cache_path()\n    try:\n        path.parent.mkdir(parents=True, exist_ok=True)\n        with tempfile.NamedTemporaryFile(\"w\", encoding=\"utf-8\", dir=path.parent, delete=False) as handle:\n            json.dump(dict(payload), handle, separators=(\",\", \":\"))\n            temp_name = handle.name\n        Path(temp_name).replace(path)\n    except OSError:\n        return\n\n\n__all__ = [\n    \"ResolvedModelMetadata\",\n    \"fetch_endpoint_model_metadata\",\n    \"fetch_models_dev_registry\",\n    \"fetch_openrouter_model_metadata\",\n    \"get_cached_context_length\",\n    \"infer_provider_from_url\",\n    \"is_local_endpoint\",\n    \"lookup_models_dev_metadata\",\n    \"lookup_openrouter_metadata\",\n    \"query_local_endpoint_metadata\",\n    \"resolve_provider_model_metadata\",\n    \"save_context_length\",\n]\n"}}, "packages/models/auth_headers.py": {"b342fa5f7cceb2efc4945b50b54b6e8233452272": {"t": 1785674430.8893015, "n": 6810, "h": "c7a32da98d3dc950", "c": "\"\"\"Auth-header strategies for provider transports.\"\"\"\n\nfrom __future__ import annotations\n\nfrom dataclasses import dataclass, field\nfrom typing import Mapping, Protocol, runtime_checkable\n\nANTHROPIC_OAUTH_BETA_HEADER = (\n    \"interleaved-thinking-2025-05-14,\"\n    \"fine-grained-tool-streaming-2025-05-14,\"\n    \"claude-code-20250219,\"\n    \"oauth-2025-04-20\"\n)\n\n\n@dataclass(frozen=True, slots=True)\nclass AuthHeaderContext:\n    provider_id: str\n    request_family: str\n    api_key: str | None\n    anthropic_version: str = \"2023-06-01\"\n    metadata: Mapping[str, str] = field(default_factory=dict)\n\n\n@runtime_checkable\nclass AuthHeaderStrategy(Protocol):\n    strategy_id: str\n\n    def supports(self, context: AuthHeaderContext) -> bool:\n        \"\"\"Return whether this strategy can build headers for the context.\"\"\"\n\n    def build_headers(self, context: AuthHeaderContext) -> Mapping[str, str]:\n        \"\"\"Build auth headers for the context.\"\"\"\n\n\nclass InMemoryAuthHeaderStrategyRegistry:\n    def __init__(self, strategies: tuple[AuthHeaderStrategy, ...] = ()) -> None:\n        self._strategies: dict[str, AuthHeaderStrategy] = {}\n        for strategy in strategies:\n            self.register(strategy)\n\n    def register(self, strategy: AuthHeaderStrategy) -> None:\n        self._strategies[strategy.strategy_id] = strategy\n\n    def get(self, strategy_id: str) -> AuthHeaderStrategy | None:\n        return self._strategies.get(strategy_id)\n\n    def list(self) -> tuple[AuthHeaderStrategy, ...]:\n        return tuple(self._strategies.values())\n\n    def select(self, context: AuthHeaderContext) -> AuthHeaderStrategy:\n        configured = str(context.metadata.get(\"auth_header_strategy\") or \"\").strip()\n        if configured:\n            strategy = self.get(configured)\n            if strategy is None:\n                raise LookupError(f\"no auth-header strategy registered for id: {configured}\")\n            return strategy\n        for strategy in self._strategies.values():\n            if strategy.supports(context):\n                return strategy\n        raise LookupError(\n            \"no auth-header strategy registered for \"\n            f\"provider={context.provider_id} request_family={context.request_family}\"\n        )\n\n    @classmethod\n    def default(cls) -> \"InMemoryAuthHeaderStrategyRegistry\":\n        return cls(\n            (\n                _CopilotBearerStrategy(),\n                _AnthropicOauthBearerStrategy(),\n                _AnthropicApiKeyStrategy(),\n                _BearerAuthHeaderStrategy(),\n            )\n        )\n\n\ndef _is_anthropic_oauth_token(api_key: str) -> bool:\n    return (\n        (api_key.startswith(\"sk-ant-\") and not api_key.startswith(\"sk-ant-api\"))\n        or api_key.startswith(\"eyJ\")\n    )\n\n\nclass _CopilotBearerStrategy:\n    strategy_id = \"copilot-bearer\"\n\n    def supports(self, context: AuthHeaderContext) -> bool:\n        return bool(context.api_key) and context.provider_id.strip().lower() == \"copilot\"\n\n    def build_headers(self, context: AuthHeaderContext) -> Mapping[str, str]:\n        api_key = str(context.api_key or \"\").strip()\n        if not api_key:\n            return {}\n        headers = {\"Authorization\": f\"Bearer {api_key}\"}\n        if context.request_family.strip().lower() == \"messages\":\n            headers[\"anthropic-version\"] = context.anthropic_version\n        return headers\n\n\nclass _AnthropicOauthBearerStrategy:\n    strategy_id = \"anthropic-oauth-bearer\"\n\n    def supports(self, context: AuthHeaderContext) -> bool:\n        api_key = str(context.api_key or \"\").strip()\n        provider_id = context.provider_id.strip().lower()\n        return bool(api_key) and (\n            provider_id == \"claude-code\"\n            or (provider_id == \"anthropic\" and _is_anthropic_oauth_token(api_key))\n        )\n\n    def build_headers(self, context: AuthHeaderContext) -> Mapping[str, str]:\n        api_key = str(context.api_key or \"\").strip()\n        if not api_key:\n            return {}\n        headers = {\n            \"Authorization\": f\"Bearer {api_key}\",\n            \"anthropic-beta\": ANTHROPIC_OAUTH_BETA_HEADER,\n            \"user-agent\": \"claude-cli/2.1.74 (external, cli)\",\n            \"x-app\": \"cli\",\n        }\n        if context.provider_id.strip().lower() == \"anthropic\":\n            headers[\"anthropic-version\"] = context.anthropic_version\n        return headers\n\n\nclass _AnthropicApiKeyStrategy:\n    strategy_id = \"anthropic-x-api-key\"\n\n    def supports(self, context: AuthHeaderContext) -> bool:\n        api_key = str(context.api_key or \"\").strip()\n        provider_id = context.provider_id.strip().lower()\n        request_family = context.request_family.strip().lower()\n        return bool(api_key) and (\n            request_family == \"messages\" or provider_id == \"anthropic\"\n        )\n\n    def build_headers(self, context: AuthHeaderContext) -> Mapping[str, str]:\n        api_key = str(context.api_key or \"\").strip()\n        if not api_key:\n            return {}\n        return {\n            \"anthropic-version\": context.anthropic_version,\n            \"x-api-key\": api_key,\n        }\n\n\nclass _BearerAuthHeaderStrategy:\n    strategy_id = \"bearer\"\n\n    def supports(self, context: AuthHeaderContext) -> bool:\n        return bool(str(context.api_key or \"\").strip())\n\n    def build_headers(self, context: AuthHeaderContext) -> Mapping[str, str]:\n        api_key = str(context.api_key or \"\").strip()\n        if not api_key:\n            return {}\n        return {\"Authorization\": f\"Bearer {api_key}\"}\n\n\n_DEFAULT_AUTH_HEADER_STRATEGY_REGISTRY: InMemoryAuthHeaderStrategyRegistry | None = None\n\n\ndef default_auth_header_strategy_registry() -> InMemoryAuthHeaderStrategyRegistry:\n    global _DEFAULT_AUTH_HEADER_STRATEGY_REGISTRY\n    if _DEFAULT_AUTH_HEADER_STRATEGY_REGISTRY is None:\n        _DEFAULT_AUTH_HEADER_STRATEGY_REGISTRY = InMemoryAuthHeaderStrategyRegistry.default()\n    return _DEFAULT_AUTH_HEADER_STRATEGY_REGISTRY\n\n\ndef build_provider_auth_headers(\n    *,\n    provider_id: str,\n    request_family: str,\n    api_key: str | None,\n    anthropic_version: str = \"2023-06-01\",\n    metadata: Mapping[str, str] | None = None,\n    registry: InMemoryAuthHeaderStrategyRegistry | None = None,\n) -> dict[str, str]:\n    context = AuthHeaderContext(\n        provider_id=provider_id,\n        request_family=request_family,\n        api_key=api_key,\n        anthropic_version=anthropic_version,\n        metadata=dict(metadata or {}),\n    )\n    resolved_registry = registry or default_auth_header_strategy_registry()\n    if not api_key:\n        return {}\n    return dict(resolved_registry.select(context).build_headers(context))\n\n\n__all__ = [\n    \"ANTHROPIC_OAUTH_BETA_HEADER\",\n    \"AuthHeaderContext\",\n    \"AuthHeaderStrategy\",\n    \"InMemoryAuthHeaderStrategyRegistry\",\n    \"build_provider_auth_headers\",\n    \"default_auth_header_strategy_registry\",\n]\n"}}, "packages/models/provider_catalog_extra.py": {"b342fa5f7cceb2efc4945b50b54b6e8233452272": {"t": 1785809847.1282167, "n": 14376, "h": "061296d1ab0fca7a", "c": "\"\"\"Additional popular provider catalog entries.\"\"\"\n\nfrom __future__ import annotations\n\nfrom .provider_catalog import ProviderDefinition\n\n\ndef extra_provider_definitions() -> tuple[ProviderDefinition, ...]:\n    return (\n        ProviderDefinition(\n            provider_id=\"cohere\",\n            display_name=\"Cohere\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use Cohere models through Cohere's OpenAI-compatible Compatibility API.\",\n            onboarding_hint=\"Choose a Command model, attach a Cohere API key, and validate the Compatibility API route.\",\n            default_base_url=\"https://api.cohere.ai/compatibility/v1\",\n            default_model_id=\"command-a-03-2025\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\", \"embeddings\"),\n            model_hints=(\"command-a-03-2025\", \"command-a-plus-05-2026\", \"command-r-plus-08-2024\"),\n            supports_custom_base_url=False,\n            listing_priority=56,\n            docs_url=\"https://docs.cohere.com/docs/compatibility-api\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"COHERE_API_KEY\",),\n            base_url_env_var=\"COHERE_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"reasoning\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"perplexity\",\n            display_name=\"Perplexity Agent API\",\n            transport_id=\"openai_responses\",\n            catalog_summary=\"Use Perplexity's OpenAI-compatible Agent API for grounded model routing.\",\n            onboarding_hint=\"Attach a Perplexity API key and use the Responses-compatible Agent API endpoint.\",\n            default_base_url=\"https://api.perplexity.ai/v1\",\n            default_model_id=\"openai/gpt-5-mini\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\",),\n            model_hints=(\"openai/gpt-5.5\", \"openai/gpt-5-mini\", \"anthropic/claude-sonnet-4.6\"),\n            supports_custom_base_url=False,\n            listing_priority=57,\n            docs_url=\"https://docs.perplexity.ai/docs/agent-api/openai-compatibility\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"PERPLEXITY_API_KEY\",),\n            base_url_env_var=\"PERPLEXITY_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"embeddings\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"cerebras\",\n            display_name=\"Cerebras\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use Cerebras Inference through its OpenAI-compatible high-speed endpoint.\",\n            onboarding_hint=\"Attach a Cerebras API key and choose a production or preview Cerebras-served model.\",\n            default_base_url=\"https://api.cerebras.ai/v1\",\n            default_model_id=\"gpt-oss-120b\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\",),\n            model_hints=(\"gpt-oss-120b\", \"zai-glm-4.7\", \"llama3.1-8b\"),\n            supports_custom_base_url=False,\n            listing_priority=58,\n            docs_url=\"https://inference-docs.cerebras.ai/resources/openai\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"CEREBRAS_API_KEY\",),\n            base_url_env_var=\"CEREBRAS_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"embeddings\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"nvidia\",\n            display_name=\"NVIDIA NIM\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use NVIDIA API Catalog / NIM hosted models through the OpenAI-compatible endpoint.\",\n            onboarding_hint=\"Attach an NVIDIA API key and choose a NIM-hosted model from the API catalog.\",\n            default_base_url=\"https://integrate.api.nvidia.com/v1\",\n            default_model_id=\"nvidia/llama-3.3-nemotron-super-49b-v1\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\",),\n            model_hints=(\n                \"nvidia/llama-3.3-nemotron-super-49b-v1\",\n                \"nvidia/nemotron-3-super-120b-a12b\",\n                \"nvidia/llama-3.1-nemotron-nano-vl-8b-v1\",\n            ),\n            supports_custom_base_url=False,\n            listing_priority=59,\n            docs_url=\"https://docs.api.nvidia.com/nim/reference/llm-apis\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"NVIDIA_API_KEY\", \"NVIDIA_NIM_API_KEY\"),\n            base_url_env_var=\"NVIDIA_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"embeddings\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"siliconflow\",\n            display_name=\"SiliconFlow\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use SiliconFlow's OpenAI-compatible model cloud for Chinese and open-source model families.\",\n            onboarding_hint=\"Attach a SiliconFlow key and choose a Qwen, DeepSeek, Kimi, GLM, or other hosted model.\",\n            default_base_url=\"https://api.siliconflow.cn/v1\",\n            default_model_id=\"Qwen/Qwen3-32B\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\", \"embeddings\"),\n            model_hints=(\n                \"Qwen/Qwen3-32B\",\n                \"Qwen/Qwen3-235B-A22B\",\n                \"deepseek-ai/DeepSeek-V3.2\",\n                \"moonshotai/Kimi-K2-Instruct\",\n            ),\n            supports_custom_base_url=False,\n            listing_priority=63,\n            docs_url=\"https://docs.siliconflow.com/cn/api-reference/chat-completions/chat-completions\",\n            provider_kind=\"aggregator\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"SILICONFLOW_API_KEY\",),\n            base_url_env_var=\"SILICONFLOW_BASE_URL\",\n            metadata={\"surface\": \"aggregator\", \"unsupported_capabilities\": \"reasoning\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"volcengine\",\n            display_name=\"Volcengine Ark / Doubao\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use Volcengine Ark and Doubao models through Ark's OpenAI-compatible endpoint.\",\n            onboarding_hint=\"Attach an Ark API key, choose a Doubao model or endpoint id, and validate the China-region route.\",\n            default_base_url=\"https://ark.cn-beijing.volces.com/api/v3\",\n            default_model_id=\"doubao-1-5-pro-32k-250115\",\n            endpoint_path_override=\"/chat/completions\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\",),\n            model_hints=(\n                \"doubao-1-5-pro-32k-250115\",\n                \"doubao-1-5-pro-256k-250115\",\n                \"doubao-seed-1-6-250615\",\n                \"doubao-seed-1-6-thinking-250615\",\n            ),\n            supports_custom_base_url=False,\n            listing_priority=64,\n            docs_url=\"https://www.volcengine.com/docs/82379/1330626\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"ARK_API_KEY\", \"VOLCENGINE_API_KEY\"),\n            base_url_env_var=\"VOLCENGINE_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"reasoning,embeddings\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"baidu-qianfan\",\n            display_name=\"Baidu Qianfan / Wenxin\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use Baidu Qianfan ModelBuilder and Wenxin-compatible models through the v2 OpenAI-compatible API.\",\n            onboarding_hint=\"Attach a Qianfan API key, select an ERNIE or DeepSeek model, and validate the v2 chat route.\",\n            default_base_url=\"https://qianfan.baidubce.com/v2\",\n            default_model_id=\"ernie-4.0-turbo-8k\",\n            endpoint_path_override=\"/chat/completions\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\",),\n            model_hints=(\"ernie-4.0-turbo-8k\", \"ernie-3.5-8k\", \"deepseek-v3.2\", \"deepseek-r1\"),\n            supports_custom_base_url=False,\n            listing_priority=67,\n            docs_url=\"https://cloud.baidu.com/doc/qianfan-docs/s/Fm9l6ocai\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"QIANFAN_API_KEY\", \"BAIDU_QIANFAN_API_KEY\"),\n            base_url_env_var=\"QIANFAN_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"reasoning,embeddings\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"tencent-hunyuan\",\n            display_name=\"Tencent Hunyuan\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use Tencent Hunyuan through Tencent Cloud's OpenAI-compatible API.\",\n            onboarding_hint=\"Attach a Hunyuan API key and choose a Tencent-hosted model such as Hunyuan Turbos.\",\n            default_base_url=\"https://api.hunyuan.cloud.tencent.com/v1\",\n            default_model_id=\"hunyuan-turbos-latest\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\", \"embeddings\"),\n            model_hints=(\"hunyuan-turbos-latest\", \"hunyuan-t1-latest\", \"hunyuan-a13b-instruct\"),\n            supports_custom_base_url=False,\n            listing_priority=68,\n            docs_url=\"https://cloud.tencent.com/document/product/1729/111007\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"HUNYUAN_API_KEY\", \"TENCENT_HUNYUAN_API_KEY\"),\n            base_url_env_var=\"HUNYUAN_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"reasoning\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"tencent-tokenhub\",\n            display_name=\"Tencent Cloud TokenHub\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use Tencent Cloud TokenHub as an OpenAI-compatible gateway for Hunyuan and third-party models.\",\n            onboarding_hint=\"Attach a TokenHub API key and choose a TokenHub service ID such as hy3-preview or deepseek-v3.2.\",\n            default_base_url=\"https://tokenhub.tencentmaas.com/v1\",\n            default_model_id=\"hy3-preview\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\",),\n            model_hints=(\n                \"hy3-preview\",\n                \"deepseek-v3.2\",\n                \"deepseek-r1\",\n                \"glm-5\",\n                \"kimi-k2.5\",\n                \"minimax-m2.5\",\n            ),\n            supports_custom_base_url=False,\n            listing_priority=68,\n            docs_url=\"https://cloud.tencent.com/document/product/1823/130079\",\n            provider_kind=\"aggregator\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"TOKENHUB_API_KEY\", \"TENCENT_TOKENHUB_API_KEY\"),\n            base_url_env_var=\"TOKENHUB_BASE_URL\",\n            metadata={\"surface\": \"aggregator\", \"unsupported_capabilities\": \"reasoning,embeddings\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"stepfun\",\n            display_name=\"StepFun\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use StepFun's Step models through the OpenAI-compatible chat endpoint.\",\n            onboarding_hint=\"Attach a StepFun API key, select a Step model, and validate the compatible chat route.\",\n            default_base_url=\"https://api.stepfun.com/v1\",\n            default_model_id=\"step-3.5-flash\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\",),\n            model_hints=(\"step-3.5-flash\", \"step-3.5\", \"step-2-16k\", \"step-1v-8k\"),\n            supports_custom_base_url=False,\n            listing_priority=69,\n            docs_url=\"https://platform.stepfun.com/docs/guide/openai\",\n            provider_kind=\"first_party\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"STEPFUN_API_KEY\", \"STEP_API_KEY\"),\n            base_url_env_var=\"STEPFUN_BASE_URL\",\n            metadata={\"surface\": \"first_party\", \"unsupported_capabilities\": \"reasoning,embeddings\"},\n        ),\n        ProviderDefinition(\n            provider_id=\"modelscope\",\n            display_name=\"ModelScope\",\n            transport_id=\"openai_chat_compatible\",\n            catalog_summary=\"Use ModelScope API Inference as an OpenAI-compatible route to popular open-source models.\",\n            onboarding_hint=\"Attach a ModelScope access token and choose a hosted Qwen, GLM, DeepSeek, or Kimi model.\",\n            default_base_url=\"https://api-inference.modelscope.cn/v1\",\n            default_model_id=\"Qwen/Qwen3-235B-A22B-Instruct-2507\",\n            required_secret_keys=(\"api_key\",),\n            required_config_keys=(\"model_id\",),\n            capability_flags=(\"chat\", \"embeddings\"),\n            model_hints=(\n                \"Qwen/Qwen3-235B-A22B-Instruct-2507\",\n                \"Qwen/Qwen3-Coder-30B-A3B-Instruct\",\n                \"ZhipuAI/GLM-4.6\",\n            ),\n            supports_custom_base_url=False,\n            listing_priority=77,\n            docs_url=\"https://www.modelscope.cn/docs/model-service/API-Inference/intro\",\n            provider_kind=\"aggregator\",\n            auth_method=\"api_key\",\n            auth_type=\"api_key\",\n            env_var_names=(\"MODELSCOPE_API_KEY\",),\n            base_url_env_var=\"MODELSCOPE_BASE_URL\",\n            metadata={\"surface\": \"aggregator\", \"unsupported_capabilities\": \"reasoning\"},\n        ),\n    )\n"}}}}