First build

nanobot/nanobot/providers/__init__.py

@@ -0,0 +1,8 @@
+"""LLM provider abstraction module."""
+
+from nanobot.providers.base import LLMProvider, LLMResponse
+from nanobot.providers.litellm_provider import LiteLLMProvider
+from nanobot.providers.openai_codex_provider import OpenAICodexProvider
+from nanobot.providers.azure_openai_provider import AzureOpenAIProvider
+
+__all__ = ["LLMProvider", "LLMResponse", "LiteLLMProvider", "OpenAICodexProvider", "AzureOpenAIProvider"]

nanobot/nanobot/providers/azure_openai_provider.py

@@ -0,0 +1,210 @@
+"""Azure OpenAI provider implementation with API version 2024-10-21."""
+
+from __future__ import annotations
+
+import uuid
+from typing import Any
+from urllib.parse import urljoin
+
+import httpx
+import json_repair
+
+from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+
+_AZURE_MSG_KEYS = frozenset({"role", "content", "tool_calls", "tool_call_id", "name"})
+
+
+class AzureOpenAIProvider(LLMProvider):
+    """
+    Azure OpenAI provider with API version 2024-10-21 compliance.
+    
+    Features:
+    - Hardcoded API version 2024-10-21
+    - Uses model field as Azure deployment name in URL path
+    - Uses api-key header instead of Authorization Bearer
+    - Uses max_completion_tokens instead of max_tokens
+    - Direct HTTP calls, bypasses LiteLLM
+    """
+
+    def __init__(
+        self,
+        api_key: str = "",
+        api_base: str = "",
+        default_model: str = "gpt-5.2-chat",
+    ):
+        super().__init__(api_key, api_base)
+        self.default_model = default_model
+        self.api_version = "2024-10-21"
+        
+        # Validate required parameters
+        if not api_key:
+            raise ValueError("Azure OpenAI api_key is required")
+        if not api_base:
+            raise ValueError("Azure OpenAI api_base is required")
+        
+        # Ensure api_base ends with /
+        if not api_base.endswith('/'):
+            api_base += '/'
+        self.api_base = api_base
+
+    def _build_chat_url(self, deployment_name: str) -> str:
+        """Build the Azure OpenAI chat completions URL."""
+        # Azure OpenAI URL format:
+        # https://{resource}.openai.azure.com/openai/deployments/{deployment}/chat/completions?api-version={version}
+        base_url = self.api_base
+        if not base_url.endswith('/'):
+            base_url += '/'
+        
+        url = urljoin(
+            base_url, 
+            f"openai/deployments/{deployment_name}/chat/completions"
+        )
+        return f"{url}?api-version={self.api_version}"
+
+    def _build_headers(self) -> dict[str, str]:
+        """Build headers for Azure OpenAI API with api-key header."""
+        return {
+            "Content-Type": "application/json",
+            "api-key": self.api_key,  # Azure OpenAI uses api-key header, not Authorization
+            "x-session-affinity": uuid.uuid4().hex,  # For cache locality
+        }
+
+    @staticmethod
+    def _supports_temperature(
+        deployment_name: str,
+        reasoning_effort: str | None = None,
+    ) -> bool:
+        """Return True when temperature is likely supported for this deployment."""
+        if reasoning_effort:
+            return False
+        name = deployment_name.lower()
+        return not any(token in name for token in ("gpt-5", "o1", "o3", "o4"))
+
+    def _prepare_request_payload(
+        self,
+        deployment_name: str,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
+        max_tokens: int = 4096,
+        temperature: float = 0.7,
+        reasoning_effort: str | None = None,
+    ) -> dict[str, Any]:
+        """Prepare the request payload with Azure OpenAI 2024-10-21 compliance."""
+        payload: dict[str, Any] = {
+            "messages": self._sanitize_request_messages(
+                self._sanitize_empty_content(messages),
+                _AZURE_MSG_KEYS,
+            ),
+            "max_completion_tokens": max(1, max_tokens),  # Azure API 2024-10-21 uses max_completion_tokens
+        }
+
+        if self._supports_temperature(deployment_name, reasoning_effort):
+            payload["temperature"] = temperature
+
+        if reasoning_effort:
+            payload["reasoning_effort"] = reasoning_effort
+
+        if tools:
+            payload["tools"] = tools
+            payload["tool_choice"] = "auto"
+
+        return payload
+
+    async def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
+        model: str | None = None,
+        max_tokens: int = 4096,
+        temperature: float = 0.7,
+        reasoning_effort: str | None = None,
+    ) -> LLMResponse:
+        """
+        Send a chat completion request to Azure OpenAI.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            tools: Optional list of tool definitions in OpenAI format.
+            model: Model identifier (used as deployment name).
+            max_tokens: Maximum tokens in response (mapped to max_completion_tokens).
+            temperature: Sampling temperature.
+            reasoning_effort: Optional reasoning effort parameter.
+
+        Returns:
+            LLMResponse with content and/or tool calls.
+        """
+        deployment_name = model or self.default_model
+        url = self._build_chat_url(deployment_name)
+        headers = self._build_headers()
+        payload = self._prepare_request_payload(
+            deployment_name, messages, tools, max_tokens, temperature, reasoning_effort
+        )
+
+        try:
+            async with httpx.AsyncClient(timeout=60.0, verify=True) as client:
+                response = await client.post(url, headers=headers, json=payload)
+                if response.status_code != 200:
+                    return LLMResponse(
+                        content=f"Azure OpenAI API Error {response.status_code}: {response.text}",
+                        finish_reason="error",
+                    )
+                
+                response_data = response.json()
+                return self._parse_response(response_data)
+
+        except Exception as e:
+            return LLMResponse(
+                content=f"Error calling Azure OpenAI: {repr(e)}",
+                finish_reason="error",
+            )
+
+    def _parse_response(self, response: dict[str, Any]) -> LLMResponse:
+        """Parse Azure OpenAI response into our standard format."""
+        try:
+            choice = response["choices"][0]
+            message = choice["message"]
+
+            tool_calls = []
+            if message.get("tool_calls"):
+                for tc in message["tool_calls"]:
+                    # Parse arguments from JSON string if needed
+                    args = tc["function"]["arguments"]
+                    if isinstance(args, str):
+                        args = json_repair.loads(args)
+
+                    tool_calls.append(
+                        ToolCallRequest(
+                            id=tc["id"],
+                            name=tc["function"]["name"],
+                            arguments=args,
+                        )
+                    )
+
+            usage = {}
+            if response.get("usage"):
+                usage_data = response["usage"]
+                usage = {
+                    "prompt_tokens": usage_data.get("prompt_tokens", 0),
+                    "completion_tokens": usage_data.get("completion_tokens", 0),
+                    "total_tokens": usage_data.get("total_tokens", 0),
+                }
+
+            reasoning_content = message.get("reasoning_content") or None
+
+            return LLMResponse(
+                content=message.get("content"),
+                tool_calls=tool_calls,
+                finish_reason=choice.get("finish_reason", "stop"),
+                usage=usage,
+                reasoning_content=reasoning_content,
+            )
+
+        except (KeyError, IndexError) as e:
+            return LLMResponse(
+                content=f"Error parsing Azure OpenAI response: {str(e)}",
+                finish_reason="error",
+            )
+
+    def get_default_model(self) -> str:
+        """Get the default model (also used as default deployment name)."""
+        return self.default_model

nanobot/nanobot/providers/base.py

@@ -0,0 +1,132 @@
+"""Base LLM provider interface."""
+
+from abc import ABC, abstractmethod
+from dataclasses import dataclass, field
+from typing import Any
+
+
+@dataclass
+class ToolCallRequest:
+    """A tool call request from the LLM."""
+    id: str
+    name: str
+    arguments: dict[str, Any]
+
+
+@dataclass
+class LLMResponse:
+    """Response from an LLM provider."""
+    content: str | None
+    tool_calls: list[ToolCallRequest] = field(default_factory=list)
+    finish_reason: str = "stop"
+    usage: dict[str, int] = field(default_factory=dict)
+    reasoning_content: str | None = None  # Kimi, DeepSeek-R1 etc.
+    thinking_blocks: list[dict] | None = None  # Anthropic extended thinking
+    
+    @property
+    def has_tool_calls(self) -> bool:
+        """Check if response contains tool calls."""
+        return len(self.tool_calls) > 0
+
+
+class LLMProvider(ABC):
+    """
+    Abstract base class for LLM providers.
+    
+    Implementations should handle the specifics of each provider's API
+    while maintaining a consistent interface.
+    """
+
+    def __init__(self, api_key: str | None = None, api_base: str | None = None):
+        self.api_key = api_key
+        self.api_base = api_base
+
+    @staticmethod
+    def _sanitize_empty_content(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
+        """Replace empty text content that causes provider 400 errors.
+
+        Empty content can appear when MCP tools return nothing. Most providers
+        reject empty-string content or empty text blocks in list content.
+        """
+        result: list[dict[str, Any]] = []
+        for msg in messages:
+            content = msg.get("content")
+
+            if isinstance(content, str) and not content:
+                clean = dict(msg)
+                clean["content"] = None if (msg.get("role") == "assistant" and msg.get("tool_calls")) else "(empty)"
+                result.append(clean)
+                continue
+
+            if isinstance(content, list):
+                filtered = [
+                    item for item in content
+                    if not (
+                        isinstance(item, dict)
+                        and item.get("type") in ("text", "input_text", "output_text")
+                        and not item.get("text")
+                    )
+                ]
+                if len(filtered) != len(content):
+                    clean = dict(msg)
+                    if filtered:
+                        clean["content"] = filtered
+                    elif msg.get("role") == "assistant" and msg.get("tool_calls"):
+                        clean["content"] = None
+                    else:
+                        clean["content"] = "(empty)"
+                    result.append(clean)
+                    continue
+
+            if isinstance(content, dict):
+                clean = dict(msg)
+                clean["content"] = [content]
+                result.append(clean)
+                continue
+
+            result.append(msg)
+        return result
+
+    @staticmethod
+    def _sanitize_request_messages(
+        messages: list[dict[str, Any]],
+        allowed_keys: frozenset[str],
+    ) -> list[dict[str, Any]]:
+        """Keep only provider-safe message keys and normalize assistant content."""
+        sanitized = []
+        for msg in messages:
+            clean = {k: v for k, v in msg.items() if k in allowed_keys}
+            if clean.get("role") == "assistant" and "content" not in clean:
+                clean["content"] = None
+            sanitized.append(clean)
+        return sanitized
+
+    @abstractmethod
+    async def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
+        model: str | None = None,
+        max_tokens: int = 4096,
+        temperature: float = 0.7,
+        reasoning_effort: str | None = None,
+    ) -> LLMResponse:
+        """
+        Send a chat completion request.
+        
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            tools: Optional list of tool definitions.
+            model: Model identifier (provider-specific).
+            max_tokens: Maximum tokens in response.
+            temperature: Sampling temperature.
+        
+        Returns:
+            LLMResponse with content and/or tool calls.
+        """
+        pass
+
+    @abstractmethod
+    def get_default_model(self) -> str:
+        """Get the default model for this provider."""
+        pass

nanobot/nanobot/providers/custom_provider.py

@@ -0,0 +1,61 @@
+"""Direct OpenAI-compatible provider — bypasses LiteLLM."""
+
+from __future__ import annotations
+
+import uuid
+from typing import Any
+
+import json_repair
+from openai import AsyncOpenAI
+
+from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+
+
+class CustomProvider(LLMProvider):
+
+    def __init__(self, api_key: str = "no-key", api_base: str = "http://localhost:8000/v1", default_model: str = "default"):
+        super().__init__(api_key, api_base)
+        self.default_model = default_model
+        # Keep affinity stable for this provider instance to improve backend cache locality.
+        self._client = AsyncOpenAI(
+            api_key=api_key,
+            base_url=api_base,
+            default_headers={"x-session-affinity": uuid.uuid4().hex},
+        )
+
+    async def chat(self, messages: list[dict[str, Any]], tools: list[dict[str, Any]] | None = None,
+                   model: str | None = None, max_tokens: int = 4096, temperature: float = 0.7,
+                   reasoning_effort: str | None = None) -> LLMResponse:
+        kwargs: dict[str, Any] = {
+            "model": model or self.default_model,
+            "messages": self._sanitize_empty_content(messages),
+            "max_tokens": max(1, max_tokens),
+            "temperature": temperature,
+        }
+        if reasoning_effort:
+            kwargs["reasoning_effort"] = reasoning_effort
+        if tools:
+            kwargs.update(tools=tools, tool_choice="auto")
+        try:
+            return self._parse(await self._client.chat.completions.create(**kwargs))
+        except Exception as e:
+            return LLMResponse(content=f"Error: {e}", finish_reason="error")
+
+    def _parse(self, response: Any) -> LLMResponse:
+        choice = response.choices[0]
+        msg = choice.message
+        tool_calls = [
+            ToolCallRequest(id=tc.id, name=tc.function.name,
+                            arguments=json_repair.loads(tc.function.arguments) if isinstance(tc.function.arguments, str) else tc.function.arguments)
+            for tc in (msg.tool_calls or [])
+        ]
+        u = response.usage
+        return LLMResponse(
+            content=msg.content, tool_calls=tool_calls, finish_reason=choice.finish_reason or "stop",
+            usage={"prompt_tokens": u.prompt_tokens, "completion_tokens": u.completion_tokens, "total_tokens": u.total_tokens} if u else {},
+            reasoning_content=getattr(msg, "reasoning_content", None) or None,
+        )
+
+    def get_default_model(self) -> str:
+        return self.default_model
+

nanobot/nanobot/providers/litellm_provider.py

@@ -0,0 +1,340 @@
+"""LiteLLM provider implementation for multi-provider support."""
+
+import hashlib
+import os
+import secrets
+import string
+from typing import Any
+
+import json_repair
+import litellm
+from litellm import acompletion
+from loguru import logger
+
+from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+from nanobot.providers.registry import find_by_model, find_gateway
+
+# Standard chat-completion message keys.
+_ALLOWED_MSG_KEYS = frozenset({"role", "content", "tool_calls", "tool_call_id", "name", "reasoning_content"})
+_ANTHROPIC_EXTRA_KEYS = frozenset({"thinking_blocks"})
+_ALNUM = string.ascii_letters + string.digits
+
+def _short_tool_id() -> str:
+    """Generate a 9-char alphanumeric ID compatible with all providers (incl. Mistral)."""
+    return "".join(secrets.choice(_ALNUM) for _ in range(9))
+
+
+class LiteLLMProvider(LLMProvider):
+    """
+    LLM provider using LiteLLM for multi-provider support.
+    
+    Supports OpenRouter, Anthropic, OpenAI, Gemini, MiniMax, and many other providers through
+    a unified interface.  Provider-specific logic is driven by the registry
+    (see providers/registry.py) — no if-elif chains needed here.
+    """
+
+    def __init__(
+        self,
+        api_key: str | None = None,
+        api_base: str | None = None,
+        default_model: str = "anthropic/claude-opus-4-5",
+        extra_headers: dict[str, str] | None = None,
+        provider_name: str | None = None,
+    ):
+        super().__init__(api_key, api_base)
+        self.default_model = default_model
+        self.extra_headers = extra_headers or {}
+
+        # Detect gateway / local deployment.
+        # provider_name (from config key) is the primary signal;
+        # api_key / api_base are fallback for auto-detection.
+        self._gateway = find_gateway(provider_name, api_key, api_base)
+
+        # Configure environment variables
+        if api_key:
+            self._setup_env(api_key, api_base, default_model)
+
+        if api_base:
+            litellm.api_base = api_base
+
+        # Disable LiteLLM logging noise
+        litellm.suppress_debug_info = True
+        # Drop unsupported parameters for providers (e.g., gpt-5 rejects some params)
+        litellm.drop_params = True
+
+    def _setup_env(self, api_key: str, api_base: str | None, model: str) -> None:
+        """Set environment variables based on detected provider."""
+        spec = self._gateway or find_by_model(model)
+        if not spec:
+            return
+        if not spec.env_key:
+            # OAuth/provider-only specs (for example: openai_codex)
+            return
+
+        # Gateway/local overrides existing env; standard provider doesn't
+        if self._gateway:
+            os.environ[spec.env_key] = api_key
+        else:
+            os.environ.setdefault(spec.env_key, api_key)
+
+        # Resolve env_extras placeholders:
+        #   {api_key}  → user's API key
+        #   {api_base} → user's api_base, falling back to spec.default_api_base
+        effective_base = api_base or spec.default_api_base
+        for env_name, env_val in spec.env_extras:
+            resolved = env_val.replace("{api_key}", api_key)
+            resolved = resolved.replace("{api_base}", effective_base)
+            os.environ.setdefault(env_name, resolved)
+
+    def _resolve_model(self, model: str) -> str:
+        """Resolve model name by applying provider/gateway prefixes."""
+        if self._gateway:
+            # Gateway mode: apply gateway prefix, skip provider-specific prefixes
+            prefix = self._gateway.litellm_prefix
+            if self._gateway.strip_model_prefix:
+                model = model.split("/")[-1]
+            if prefix and not model.startswith(f"{prefix}/"):
+                model = f"{prefix}/{model}"
+            return model
+
+        # Standard mode: auto-prefix for known providers
+        spec = find_by_model(model)
+        if spec and spec.litellm_prefix:
+            model = self._canonicalize_explicit_prefix(model, spec.name, spec.litellm_prefix)
+            if not any(model.startswith(s) for s in spec.skip_prefixes):
+                model = f"{spec.litellm_prefix}/{model}"
+
+        return model
+
+    @staticmethod
+    def _canonicalize_explicit_prefix(model: str, spec_name: str, canonical_prefix: str) -> str:
+        """Normalize explicit provider prefixes like `github-copilot/...`."""
+        if "/" not in model:
+            return model
+        prefix, remainder = model.split("/", 1)
+        if prefix.lower().replace("-", "_") != spec_name:
+            return model
+        return f"{canonical_prefix}/{remainder}"
+
+    def _supports_cache_control(self, model: str) -> bool:
+        """Return True when the provider supports cache_control on content blocks."""
+        if self._gateway is not None:
+            return self._gateway.supports_prompt_caching
+        spec = find_by_model(model)
+        return spec is not None and spec.supports_prompt_caching
+
+    def _apply_cache_control(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None,
+    ) -> tuple[list[dict[str, Any]], list[dict[str, Any]] | None]:
+        """Return copies of messages and tools with cache_control injected."""
+        new_messages = []
+        for msg in messages:
+            if msg.get("role") == "system":
+                content = msg["content"]
+                if isinstance(content, str):
+                    new_content = [{"type": "text", "text": content, "cache_control": {"type": "ephemeral"}}]
+                else:
+                    new_content = list(content)
+                    new_content[-1] = {**new_content[-1], "cache_control": {"type": "ephemeral"}}
+                new_messages.append({**msg, "content": new_content})
+            else:
+                new_messages.append(msg)
+
+        new_tools = tools
+        if tools:
+            new_tools = list(tools)
+            new_tools[-1] = {**new_tools[-1], "cache_control": {"type": "ephemeral"}}
+
+        return new_messages, new_tools
+
+    def _apply_model_overrides(self, model: str, kwargs: dict[str, Any]) -> None:
+        """Apply model-specific parameter overrides from the registry."""
+        model_lower = model.lower()
+        spec = find_by_model(model)
+        if spec:
+            for pattern, overrides in spec.model_overrides:
+                if pattern in model_lower:
+                    kwargs.update(overrides)
+                    return
+
+    @staticmethod
+    def _extra_msg_keys(original_model: str, resolved_model: str) -> frozenset[str]:
+        """Return provider-specific extra keys to preserve in request messages."""
+        spec = find_by_model(original_model) or find_by_model(resolved_model)
+        if (spec and spec.name == "anthropic") or "claude" in original_model.lower() or resolved_model.startswith("anthropic/"):
+            return _ANTHROPIC_EXTRA_KEYS
+        return frozenset()
+
+    @staticmethod
+    def _normalize_tool_call_id(tool_call_id: Any) -> Any:
+        """Normalize tool_call_id to a provider-safe 9-char alphanumeric form."""
+        if not isinstance(tool_call_id, str):
+            return tool_call_id
+        if len(tool_call_id) == 9 and tool_call_id.isalnum():
+            return tool_call_id
+        return hashlib.sha1(tool_call_id.encode()).hexdigest()[:9]
+
+    @staticmethod
+    def _sanitize_messages(messages: list[dict[str, Any]], extra_keys: frozenset[str] = frozenset()) -> list[dict[str, Any]]:
+        """Strip non-standard keys and ensure assistant messages have a content key."""
+        allowed = _ALLOWED_MSG_KEYS | extra_keys
+        sanitized = LLMProvider._sanitize_request_messages(messages, allowed)
+        id_map: dict[str, str] = {}
+
+        def map_id(value: Any) -> Any:
+            if not isinstance(value, str):
+                return value
+            return id_map.setdefault(value, LiteLLMProvider._normalize_tool_call_id(value))
+
+        for clean in sanitized:
+            # Keep assistant tool_calls[].id and tool tool_call_id in sync after
+            # shortening, otherwise strict providers reject the broken linkage.
+            if isinstance(clean.get("tool_calls"), list):
+                normalized_tool_calls = []
+                for tc in clean["tool_calls"]:
+                    if not isinstance(tc, dict):
+                        normalized_tool_calls.append(tc)
+                        continue
+                    tc_clean = dict(tc)
+                    tc_clean["id"] = map_id(tc_clean.get("id"))
+                    normalized_tool_calls.append(tc_clean)
+                clean["tool_calls"] = normalized_tool_calls
+
+            if "tool_call_id" in clean and clean["tool_call_id"]:
+                clean["tool_call_id"] = map_id(clean["tool_call_id"])
+        return sanitized
+
+    async def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
+        model: str | None = None,
+        max_tokens: int = 4096,
+        temperature: float = 0.7,
+        reasoning_effort: str | None = None,
+    ) -> LLMResponse:
+        """
+        Send a chat completion request via LiteLLM.
+
+        Args:
+            messages: List of message dicts with 'role' and 'content'.
+            tools: Optional list of tool definitions in OpenAI format.
+            model: Model identifier (e.g., 'anthropic/claude-sonnet-4-5').
+            max_tokens: Maximum tokens in response.
+            temperature: Sampling temperature.
+
+        Returns:
+            LLMResponse with content and/or tool calls.
+        """
+        original_model = model or self.default_model
+        model = self._resolve_model(original_model)
+        extra_msg_keys = self._extra_msg_keys(original_model, model)
+
+        if self._supports_cache_control(original_model):
+            messages, tools = self._apply_cache_control(messages, tools)
+
+        # Clamp max_tokens to at least 1 — negative or zero values cause
+        # LiteLLM to reject the request with "max_tokens must be at least 1".
+        max_tokens = max(1, max_tokens)
+
+        kwargs: dict[str, Any] = {
+            "model": model,
+            "messages": self._sanitize_messages(self._sanitize_empty_content(messages), extra_keys=extra_msg_keys),
+            "max_tokens": max_tokens,
+            "temperature": temperature,
+        }
+
+        # Apply model-specific overrides (e.g. kimi-k2.5 temperature)
+        self._apply_model_overrides(model, kwargs)
+
+        # Pass api_key directly — more reliable than env vars alone
+        if self.api_key:
+            kwargs["api_key"] = self.api_key
+
+        # Pass api_base for custom endpoints
+        if self.api_base:
+            kwargs["api_base"] = self.api_base
+
+        # Pass extra headers (e.g. APP-Code for AiHubMix)
+        if self.extra_headers:
+            kwargs["extra_headers"] = self.extra_headers
+        
+        if reasoning_effort:
+            kwargs["reasoning_effort"] = reasoning_effort
+            kwargs["drop_params"] = True
+        
+        if tools:
+            kwargs["tools"] = tools
+            kwargs["tool_choice"] = "auto"
+
+        try:
+            response = await acompletion(**kwargs)
+            return self._parse_response(response)
+        except Exception as e:
+            # Return error as content for graceful handling
+            return LLMResponse(
+                content=f"Error calling LLM: {str(e)}",
+                finish_reason="error",
+            )
+
+    def _parse_response(self, response: Any) -> LLMResponse:
+        """Parse LiteLLM response into our standard format."""
+        choice = response.choices[0]
+        message = choice.message
+        content = message.content
+        finish_reason = choice.finish_reason
+
+        # Some providers (e.g. GitHub Copilot) split content and tool_calls
+        # across multiple choices. Merge them so tool_calls are not lost.
+        raw_tool_calls = []
+        for ch in response.choices:
+            msg = ch.message
+            if hasattr(msg, "tool_calls") and msg.tool_calls:
+                raw_tool_calls.extend(msg.tool_calls)
+                if ch.finish_reason in ("tool_calls", "stop"):
+                    finish_reason = ch.finish_reason
+            if not content and msg.content:
+                content = msg.content
+
+        if len(response.choices) > 1:
+            logger.debug("LiteLLM response has {} choices, merged {} tool_calls",
+                         len(response.choices), len(raw_tool_calls))
+
+        tool_calls = []
+        for tc in raw_tool_calls:
+            # Parse arguments from JSON string if needed
+            args = tc.function.arguments
+            if isinstance(args, str):
+                args = json_repair.loads(args)
+
+            tool_calls.append(ToolCallRequest(
+                id=_short_tool_id(),
+                name=tc.function.name,
+                arguments=args,
+            ))
+
+        usage = {}
+        if hasattr(response, "usage") and response.usage:
+            usage = {
+                "prompt_tokens": response.usage.prompt_tokens,
+                "completion_tokens": response.usage.completion_tokens,
+                "total_tokens": response.usage.total_tokens,
+            }
+
+        reasoning_content = getattr(message, "reasoning_content", None) or None
+        thinking_blocks = getattr(message, "thinking_blocks", None) or None
+
+        return LLMResponse(
+            content=content,
+            tool_calls=tool_calls,
+            finish_reason=finish_reason or "stop",
+            usage=usage,
+            reasoning_content=reasoning_content,
+            thinking_blocks=thinking_blocks,
+        )
+
+    def get_default_model(self) -> str:
+        """Get the default model."""
+        return self.default_model

nanobot/nanobot/providers/openai_codex_provider.py

@@ -0,0 +1,316 @@
+"""OpenAI Codex Responses Provider."""
+
+from __future__ import annotations
+
+import asyncio
+import hashlib
+import json
+from typing import Any, AsyncGenerator
+
+import httpx
+from loguru import logger
+from oauth_cli_kit import get_token as get_codex_token
+
+from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
+
+DEFAULT_CODEX_URL = "https://chatgpt.com/backend-api/codex/responses"
+DEFAULT_ORIGINATOR = "nanobot"
+
+
+class OpenAICodexProvider(LLMProvider):
+    """Use Codex OAuth to call the Responses API."""
+
+    def __init__(self, default_model: str = "openai-codex/gpt-5.1-codex"):
+        super().__init__(api_key=None, api_base=None)
+        self.default_model = default_model
+
+    async def chat(
+        self,
+        messages: list[dict[str, Any]],
+        tools: list[dict[str, Any]] | None = None,
+        model: str | None = None,
+        max_tokens: int = 4096,
+        temperature: float = 0.7,
+        reasoning_effort: str | None = None,
+    ) -> LLMResponse:
+        model = model or self.default_model
+        system_prompt, input_items = _convert_messages(messages)
+
+        token = await asyncio.to_thread(get_codex_token)
+        headers = _build_headers(token.account_id, token.access)
+
+        body: dict[str, Any] = {
+            "model": _strip_model_prefix(model),
+            "store": False,
+            "stream": True,
+            "instructions": system_prompt,
+            "input": input_items,
+            "text": {"verbosity": "medium"},
+            "include": ["reasoning.encrypted_content"],
+            "prompt_cache_key": _prompt_cache_key(messages),
+            "tool_choice": "auto",
+            "parallel_tool_calls": True,
+        }
+
+        if reasoning_effort:
+            body["reasoning"] = {"effort": reasoning_effort}
+
+        if tools:
+            body["tools"] = _convert_tools(tools)
+
+        url = DEFAULT_CODEX_URL
+
+        try:
+            try:
+                content, tool_calls, finish_reason = await _request_codex(url, headers, body, verify=True)
+            except Exception as e:
+                if "CERTIFICATE_VERIFY_FAILED" not in str(e):
+                    raise
+                logger.warning("SSL certificate verification failed for Codex API; retrying with verify=False")
+                content, tool_calls, finish_reason = await _request_codex(url, headers, body, verify=False)
+            return LLMResponse(
+                content=content,
+                tool_calls=tool_calls,
+                finish_reason=finish_reason,
+            )
+        except Exception as e:
+            return LLMResponse(
+                content=f"Error calling Codex: {str(e)}",
+                finish_reason="error",
+            )
+
+    def get_default_model(self) -> str:
+        return self.default_model
+
+
+def _strip_model_prefix(model: str) -> str:
+    if model.startswith("openai-codex/") or model.startswith("openai_codex/"):
+        return model.split("/", 1)[1]
+    return model
+
+
+def _build_headers(account_id: str, token: str) -> dict[str, str]:
+    return {
+        "Authorization": f"Bearer {token}",
+        "chatgpt-account-id": account_id,
+        "OpenAI-Beta": "responses=experimental",
+        "originator": DEFAULT_ORIGINATOR,
+        "User-Agent": "nanobot (python)",
+        "accept": "text/event-stream",
+        "content-type": "application/json",
+    }
+
+
+async def _request_codex(
+    url: str,
+    headers: dict[str, str],
+    body: dict[str, Any],
+    verify: bool,
+) -> tuple[str, list[ToolCallRequest], str]:
+    async with httpx.AsyncClient(timeout=60.0, verify=verify) as client:
+        async with client.stream("POST", url, headers=headers, json=body) as response:
+            if response.status_code != 200:
+                text = await response.aread()
+                raise RuntimeError(_friendly_error(response.status_code, text.decode("utf-8", "ignore")))
+            return await _consume_sse(response)
+
+
+def _convert_tools(tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
+    """Convert OpenAI function-calling schema to Codex flat format."""
+    converted: list[dict[str, Any]] = []
+    for tool in tools:
+        fn = (tool.get("function") or {}) if tool.get("type") == "function" else tool
+        name = fn.get("name")
+        if not name:
+            continue
+        params = fn.get("parameters") or {}
+        converted.append({
+            "type": "function",
+            "name": name,
+            "description": fn.get("description") or "",
+            "parameters": params if isinstance(params, dict) else {},
+        })
+    return converted
+
+
+def _convert_messages(messages: list[dict[str, Any]]) -> tuple[str, list[dict[str, Any]]]:
+    system_prompt = ""
+    input_items: list[dict[str, Any]] = []
+
+    for idx, msg in enumerate(messages):
+        role = msg.get("role")
+        content = msg.get("content")
+
+        if role == "system":
+            system_prompt = content if isinstance(content, str) else ""
+            continue
+
+        if role == "user":
+            input_items.append(_convert_user_message(content))
+            continue
+
+        if role == "assistant":
+            # Handle text first.
+            if isinstance(content, str) and content:
+                input_items.append(
+                    {
+                        "type": "message",
+                        "role": "assistant",
+                        "content": [{"type": "output_text", "text": content}],
+                        "status": "completed",
+                        "id": f"msg_{idx}",
+                    }
+                )
+            # Then handle tool calls.
+            for tool_call in msg.get("tool_calls", []) or []:
+                fn = tool_call.get("function") or {}
+                call_id, item_id = _split_tool_call_id(tool_call.get("id"))
+                call_id = call_id or f"call_{idx}"
+                item_id = item_id or f"fc_{idx}"
+                input_items.append(
+                    {
+                        "type": "function_call",
+                        "id": item_id,
+                        "call_id": call_id,
+                        "name": fn.get("name"),
+                        "arguments": fn.get("arguments") or "{}",
+                    }
+                )
+            continue
+
+        if role == "tool":
+            call_id, _ = _split_tool_call_id(msg.get("tool_call_id"))
+            output_text = content if isinstance(content, str) else json.dumps(content, ensure_ascii=False)
+            input_items.append(
+                {
+                    "type": "function_call_output",
+                    "call_id": call_id,
+                    "output": output_text,
+                }
+            )
+            continue
+
+    return system_prompt, input_items
+
+
+def _convert_user_message(content: Any) -> dict[str, Any]:
+    if isinstance(content, str):
+        return {"role": "user", "content": [{"type": "input_text", "text": content}]}
+    if isinstance(content, list):
+        converted: list[dict[str, Any]] = []
+        for item in content:
+            if not isinstance(item, dict):
+                continue
+            if item.get("type") == "text":
+                converted.append({"type": "input_text", "text": item.get("text", "")})
+            elif item.get("type") == "image_url":
+                url = (item.get("image_url") or {}).get("url")
+                if url:
+                    converted.append({"type": "input_image", "image_url": url, "detail": "auto"})
+        if converted:
+            return {"role": "user", "content": converted}
+    return {"role": "user", "content": [{"type": "input_text", "text": ""}]}
+
+
+def _split_tool_call_id(tool_call_id: Any) -> tuple[str, str | None]:
+    if isinstance(tool_call_id, str) and tool_call_id:
+        if "|" in tool_call_id:
+            call_id, item_id = tool_call_id.split("|", 1)
+            return call_id, item_id or None
+        return tool_call_id, None
+    return "call_0", None
+
+
+def _prompt_cache_key(messages: list[dict[str, Any]]) -> str:
+    raw = json.dumps(messages, ensure_ascii=True, sort_keys=True)
+    return hashlib.sha256(raw.encode("utf-8")).hexdigest()
+
+
+async def _iter_sse(response: httpx.Response) -> AsyncGenerator[dict[str, Any], None]:
+    buffer: list[str] = []
+    async for line in response.aiter_lines():
+        if line == "":
+            if buffer:
+                data_lines = [l[5:].strip() for l in buffer if l.startswith("data:")]
+                buffer = []
+                if not data_lines:
+                    continue
+                data = "\n".join(data_lines).strip()
+                if not data or data == "[DONE]":
+                    continue
+                try:
+                    yield json.loads(data)
+                except Exception:
+                    continue
+            continue
+        buffer.append(line)
+
+
+async def _consume_sse(response: httpx.Response) -> tuple[str, list[ToolCallRequest], str]:
+    content = ""
+    tool_calls: list[ToolCallRequest] = []
+    tool_call_buffers: dict[str, dict[str, Any]] = {}
+    finish_reason = "stop"
+
+    async for event in _iter_sse(response):
+        event_type = event.get("type")
+        if event_type == "response.output_item.added":
+            item = event.get("item") or {}
+            if item.get("type") == "function_call":
+                call_id = item.get("call_id")
+                if not call_id:
+                    continue
+                tool_call_buffers[call_id] = {
+                    "id": item.get("id") or "fc_0",
+                    "name": item.get("name"),
+                    "arguments": item.get("arguments") or "",
+                }
+        elif event_type == "response.output_text.delta":
+            content += event.get("delta") or ""
+        elif event_type == "response.function_call_arguments.delta":
+            call_id = event.get("call_id")
+            if call_id and call_id in tool_call_buffers:
+                tool_call_buffers[call_id]["arguments"] += event.get("delta") or ""
+        elif event_type == "response.function_call_arguments.done":
+            call_id = event.get("call_id")
+            if call_id and call_id in tool_call_buffers:
+                tool_call_buffers[call_id]["arguments"] = event.get("arguments") or ""
+        elif event_type == "response.output_item.done":
+            item = event.get("item") or {}
+            if item.get("type") == "function_call":
+                call_id = item.get("call_id")
+                if not call_id:
+                    continue
+                buf = tool_call_buffers.get(call_id) or {}
+                args_raw = buf.get("arguments") or item.get("arguments") or "{}"
+                try:
+                    args = json.loads(args_raw)
+                except Exception:
+                    args = {"raw": args_raw}
+                tool_calls.append(
+                    ToolCallRequest(
+                        id=f"{call_id}|{buf.get('id') or item.get('id') or 'fc_0'}",
+                        name=buf.get("name") or item.get("name"),
+                        arguments=args,
+                    )
+                )
+        elif event_type == "response.completed":
+            status = (event.get("response") or {}).get("status")
+            finish_reason = _map_finish_reason(status)
+        elif event_type in {"error", "response.failed"}:
+            raise RuntimeError("Codex response failed")
+
+    return content, tool_calls, finish_reason
+
+
+_FINISH_REASON_MAP = {"completed": "stop", "incomplete": "length", "failed": "error", "cancelled": "error"}
+
+
+def _map_finish_reason(status: str | None) -> str:
+    return _FINISH_REASON_MAP.get(status or "completed", "stop")
+
+
+def _friendly_error(status_code: int, raw: str) -> str:
+    if status_code == 429:
+        return "ChatGPT usage quota exceeded or rate limit triggered. Please try again later."
+    return f"HTTP {status_code}: {raw}"

nanobot/nanobot/providers/registry.py

@@ -0,0 +1,448 @@
+"""
+Provider Registry — single source of truth for LLM provider metadata.
+
+Adding a new provider:
+  1. Add a ProviderSpec to PROVIDERS below.
+  2. Add a field to ProvidersConfig in config/schema.py.
+  Done. Env vars, prefixing, config matching, status display all derive from here.
+
+Order matters — it controls match priority and fallback. Gateways first.
+Every entry writes out all fields so you can copy-paste as a template.
+"""
+
+from __future__ import annotations
+
+from dataclasses import dataclass
+from typing import Any
+
+
+@dataclass(frozen=True)
+class ProviderSpec:
+    """One LLM provider's metadata. See PROVIDERS below for real examples.
+
+    Placeholders in env_extras values:
+      {api_key}  — the user's API key
+      {api_base} — api_base from config, or this spec's default_api_base
+    """
+
+    # identity
+    name: str  # config field name, e.g. "dashscope"
+    keywords: tuple[str, ...]  # model-name keywords for matching (lowercase)
+    env_key: str  # LiteLLM env var, e.g. "DASHSCOPE_API_KEY"
+    display_name: str = ""  # shown in `nanobot status`
+
+    # model prefixing
+    litellm_prefix: str = ""  # "dashscope" → model becomes "dashscope/{model}"
+    skip_prefixes: tuple[str, ...] = ()  # don't prefix if model already starts with these
+
+    # extra env vars, e.g. (("ZHIPUAI_API_KEY", "{api_key}"),)
+    env_extras: tuple[tuple[str, str], ...] = ()
+
+    # gateway / local detection
+    is_gateway: bool = False  # routes any model (OpenRouter, AiHubMix)
+    is_local: bool = False  # local deployment (vLLM, Ollama)
+    detect_by_key_prefix: str = ""  # match api_key prefix, e.g. "sk-or-"
+    detect_by_base_keyword: str = ""  # match substring in api_base URL
+    default_api_base: str = ""  # fallback base URL
+
+    # gateway behavior
+    strip_model_prefix: bool = False  # strip "provider/" before re-prefixing
+
+    # per-model param overrides, e.g. (("kimi-k2.5", {"temperature": 1.0}),)
+    model_overrides: tuple[tuple[str, dict[str, Any]], ...] = ()
+
+    # OAuth-based providers (e.g., OpenAI Codex) don't use API keys
+    is_oauth: bool = False  # if True, uses OAuth flow instead of API key
+
+    # Direct providers bypass LiteLLM entirely (e.g., CustomProvider)
+    is_direct: bool = False
+
+    # Provider supports cache_control on content blocks (e.g. Anthropic prompt caching)
+    supports_prompt_caching: bool = False
+
+    @property
+    def label(self) -> str:
+        return self.display_name or self.name.title()
+
+
+# ---------------------------------------------------------------------------
+# PROVIDERS — the registry. Order = priority. Copy any entry as template.
+# ---------------------------------------------------------------------------
+
+PROVIDERS: tuple[ProviderSpec, ...] = (
+    # === Custom (direct OpenAI-compatible endpoint, bypasses LiteLLM) ======
+    ProviderSpec(
+        name="custom",
+        keywords=(),
+        env_key="",
+        display_name="Custom",
+        litellm_prefix="",
+        is_direct=True,
+    ),
+
+    # === Azure OpenAI (direct API calls with API version 2024-10-21) =====
+    ProviderSpec(
+        name="azure_openai",
+        keywords=("azure", "azure-openai"),
+        env_key="",
+        display_name="Azure OpenAI",
+        litellm_prefix="",
+        is_direct=True,
+    ),
+    # === Gateways (detected by api_key / api_base, not model name) =========
+    # Gateways can route any model, so they win in fallback.
+    # OpenRouter: global gateway, keys start with "sk-or-"
+    ProviderSpec(
+        name="openrouter",
+        keywords=("openrouter",),
+        env_key="OPENROUTER_API_KEY",
+        display_name="OpenRouter",
+        litellm_prefix="openrouter",  # claude-3 → openrouter/claude-3
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=True,
+        is_local=False,
+        detect_by_key_prefix="sk-or-",
+        detect_by_base_keyword="openrouter",
+        default_api_base="https://openrouter.ai/api/v1",
+        strip_model_prefix=False,
+        model_overrides=(),
+        supports_prompt_caching=True,
+    ),
+    # AiHubMix: global gateway, OpenAI-compatible interface.
+    # strip_model_prefix=True: it doesn't understand "anthropic/claude-3",
+    # so we strip to bare "claude-3" then re-prefix as "openai/claude-3".
+    ProviderSpec(
+        name="aihubmix",
+        keywords=("aihubmix",),
+        env_key="OPENAI_API_KEY",  # OpenAI-compatible
+        display_name="AiHubMix",
+        litellm_prefix="openai",  # → openai/{model}
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=True,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="aihubmix",
+        default_api_base="https://aihubmix.com/v1",
+        strip_model_prefix=True,  # anthropic/claude-3 → claude-3 → openai/claude-3
+        model_overrides=(),
+    ),
+    # SiliconFlow (硅基流动): OpenAI-compatible gateway, model names keep org prefix
+    ProviderSpec(
+        name="siliconflow",
+        keywords=("siliconflow",),
+        env_key="OPENAI_API_KEY",
+        display_name="SiliconFlow",
+        litellm_prefix="openai",
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=True,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="siliconflow",
+        default_api_base="https://api.siliconflow.cn/v1",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # VolcEngine (火山引擎): OpenAI-compatible gateway
+    ProviderSpec(
+        name="volcengine",
+        keywords=("volcengine", "volces", "ark"),
+        env_key="OPENAI_API_KEY",
+        display_name="VolcEngine",
+        litellm_prefix="volcengine",
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=True,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="volces",
+        default_api_base="https://ark.cn-beijing.volces.com/api/v3",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # === Standard providers (matched by model-name keywords) ===============
+    # Anthropic: LiteLLM recognizes "claude-*" natively, no prefix needed.
+    ProviderSpec(
+        name="anthropic",
+        keywords=("anthropic", "claude"),
+        env_key="ANTHROPIC_API_KEY",
+        display_name="Anthropic",
+        litellm_prefix="",
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+        supports_prompt_caching=True,
+    ),
+    # OpenAI: LiteLLM recognizes "gpt-*" natively, no prefix needed.
+    ProviderSpec(
+        name="openai",
+        keywords=("openai", "gpt"),
+        env_key="OPENAI_API_KEY",
+        display_name="OpenAI",
+        litellm_prefix="",
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # OpenAI Codex: uses OAuth, not API key.
+    ProviderSpec(
+        name="openai_codex",
+        keywords=("openai-codex",),
+        env_key="",  # OAuth-based, no API key
+        display_name="OpenAI Codex",
+        litellm_prefix="",  # Not routed through LiteLLM
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="codex",
+        default_api_base="https://chatgpt.com/backend-api",
+        strip_model_prefix=False,
+        model_overrides=(),
+        is_oauth=True,  # OAuth-based authentication
+    ),
+    # Github Copilot: uses OAuth, not API key.
+    ProviderSpec(
+        name="github_copilot",
+        keywords=("github_copilot", "copilot"),
+        env_key="",  # OAuth-based, no API key
+        display_name="Github Copilot",
+        litellm_prefix="github_copilot",  # github_copilot/model → github_copilot/model
+        skip_prefixes=("github_copilot/",),
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+        is_oauth=True,  # OAuth-based authentication
+    ),
+    # DeepSeek: needs "deepseek/" prefix for LiteLLM routing.
+    ProviderSpec(
+        name="deepseek",
+        keywords=("deepseek",),
+        env_key="DEEPSEEK_API_KEY",
+        display_name="DeepSeek",
+        litellm_prefix="deepseek",  # deepseek-chat → deepseek/deepseek-chat
+        skip_prefixes=("deepseek/",),  # avoid double-prefix
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # Gemini: needs "gemini/" prefix for LiteLLM.
+    ProviderSpec(
+        name="gemini",
+        keywords=("gemini",),
+        env_key="GEMINI_API_KEY",
+        display_name="Gemini",
+        litellm_prefix="gemini",  # gemini-pro → gemini/gemini-pro
+        skip_prefixes=("gemini/",),  # avoid double-prefix
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # Zhipu: LiteLLM uses "zai/" prefix.
+    # Also mirrors key to ZHIPUAI_API_KEY (some LiteLLM paths check that).
+    # skip_prefixes: don't add "zai/" when already routed via gateway.
+    ProviderSpec(
+        name="zhipu",
+        keywords=("zhipu", "glm", "zai"),
+        env_key="ZAI_API_KEY",
+        display_name="Zhipu AI",
+        litellm_prefix="zai",  # glm-4 → zai/glm-4
+        skip_prefixes=("zhipu/", "zai/", "openrouter/", "hosted_vllm/"),
+        env_extras=(("ZHIPUAI_API_KEY", "{api_key}"),),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # DashScope: Qwen models, needs "dashscope/" prefix.
+    ProviderSpec(
+        name="dashscope",
+        keywords=("qwen", "dashscope"),
+        env_key="DASHSCOPE_API_KEY",
+        display_name="DashScope",
+        litellm_prefix="dashscope",  # qwen-max → dashscope/qwen-max
+        skip_prefixes=("dashscope/", "openrouter/"),
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # Moonshot: Kimi models, needs "moonshot/" prefix.
+    # LiteLLM requires MOONSHOT_API_BASE env var to find the endpoint.
+    # Kimi K2.5 API enforces temperature >= 1.0.
+    ProviderSpec(
+        name="moonshot",
+        keywords=("moonshot", "kimi"),
+        env_key="MOONSHOT_API_KEY",
+        display_name="Moonshot",
+        litellm_prefix="moonshot",  # kimi-k2.5 → moonshot/kimi-k2.5
+        skip_prefixes=("moonshot/", "openrouter/"),
+        env_extras=(("MOONSHOT_API_BASE", "{api_base}"),),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="https://api.moonshot.ai/v1",  # intl; use api.moonshot.cn for China
+        strip_model_prefix=False,
+        model_overrides=(("kimi-k2.5", {"temperature": 1.0}),),
+    ),
+    # MiniMax: needs "minimax/" prefix for LiteLLM routing.
+    # Uses OpenAI-compatible API at api.minimax.io/v1.
+    ProviderSpec(
+        name="minimax",
+        keywords=("minimax",),
+        env_key="MINIMAX_API_KEY",
+        display_name="MiniMax",
+        litellm_prefix="minimax",  # MiniMax-M2.1 → minimax/MiniMax-M2.1
+        skip_prefixes=("minimax/", "openrouter/"),
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="https://api.minimax.io/v1",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # === Local deployment (matched by config key, NOT by api_base) =========
+    # vLLM / any OpenAI-compatible local server.
+    # Detected when config key is "vllm" (provider_name="vllm").
+    ProviderSpec(
+        name="vllm",
+        keywords=("vllm",),
+        env_key="HOSTED_VLLM_API_KEY",
+        display_name="vLLM/Local",
+        litellm_prefix="hosted_vllm",  # Llama-3-8B → hosted_vllm/Llama-3-8B
+        skip_prefixes=(),
+        env_extras=(),
+        is_gateway=False,
+        is_local=True,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",  # user must provide in config
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+    # === Auxiliary (not a primary LLM provider) ============================
+    # Groq: mainly used for Whisper voice transcription, also usable for LLM.
+    # Needs "groq/" prefix for LiteLLM routing. Placed last — it rarely wins fallback.
+    ProviderSpec(
+        name="groq",
+        keywords=("groq",),
+        env_key="GROQ_API_KEY",
+        display_name="Groq",
+        litellm_prefix="groq",  # llama3-8b-8192 → groq/llama3-8b-8192
+        skip_prefixes=("groq/",),  # avoid double-prefix
+        env_extras=(),
+        is_gateway=False,
+        is_local=False,
+        detect_by_key_prefix="",
+        detect_by_base_keyword="",
+        default_api_base="",
+        strip_model_prefix=False,
+        model_overrides=(),
+    ),
+)
+
+
+# ---------------------------------------------------------------------------
+# Lookup helpers
+# ---------------------------------------------------------------------------
+
+
+def find_by_model(model: str) -> ProviderSpec | None:
+    """Match a standard provider by model-name keyword (case-insensitive).
+    Skips gateways/local — those are matched by api_key/api_base instead."""
+    model_lower = model.lower()
+    model_normalized = model_lower.replace("-", "_")
+    model_prefix = model_lower.split("/", 1)[0] if "/" in model_lower else ""
+    normalized_prefix = model_prefix.replace("-", "_")
+    std_specs = [s for s in PROVIDERS if not s.is_gateway and not s.is_local]
+
+    # Prefer explicit provider prefix — prevents `github-copilot/...codex` matching openai_codex.
+    for spec in std_specs:
+        if model_prefix and normalized_prefix == spec.name:
+            return spec
+
+    for spec in std_specs:
+        if any(
+            kw in model_lower or kw.replace("-", "_") in model_normalized for kw in spec.keywords
+        ):
+            return spec
+    return None
+
+
+def find_gateway(
+    provider_name: str | None = None,
+    api_key: str | None = None,
+    api_base: str | None = None,
+) -> ProviderSpec | None:
+    """Detect gateway/local provider.
+
+    Priority:
+      1. provider_name — if it maps to a gateway/local spec, use it directly.
+      2. api_key prefix — e.g. "sk-or-" → OpenRouter.
+      3. api_base keyword — e.g. "aihubmix" in URL → AiHubMix.
+
+    A standard provider with a custom api_base (e.g. DeepSeek behind a proxy)
+    will NOT be mistaken for vLLM — the old fallback is gone.
+    """
+    # 1. Direct match by config key
+    if provider_name:
+        spec = find_by_name(provider_name)
+        if spec and (spec.is_gateway or spec.is_local):
+            return spec
+
+    # 2. Auto-detect by api_key prefix / api_base keyword
+    for spec in PROVIDERS:
+        if spec.detect_by_key_prefix and api_key and api_key.startswith(spec.detect_by_key_prefix):
+            return spec
+        if spec.detect_by_base_keyword and api_base and spec.detect_by_base_keyword in api_base:
+            return spec
+
+    return None
+
+
+def find_by_name(name: str) -> ProviderSpec | None:
+    """Find a provider spec by config field name, e.g. "dashscope"."""
+    for spec in PROVIDERS:
+        if spec.name == name:
+            return spec
+    return None

nanobot/nanobot/providers/transcription.py

@@ -0,0 +1,64 @@
+"""Voice transcription provider using Groq."""
+
+import os
+from pathlib import Path
+
+import httpx
+from loguru import logger
+
+
+class GroqTranscriptionProvider:
+    """
+    Voice transcription provider using Groq's Whisper API.
+
+    Groq offers extremely fast transcription with a generous free tier.
+    """
+
+    def __init__(self, api_key: str | None = None):
+        self.api_key = api_key or os.environ.get("GROQ_API_KEY")
+        self.api_url = "https://api.groq.com/openai/v1/audio/transcriptions"
+
+    async def transcribe(self, file_path: str | Path) -> str:
+        """
+        Transcribe an audio file using Groq.
+
+        Args:
+            file_path: Path to the audio file.
+
+        Returns:
+            Transcribed text.
+        """
+        if not self.api_key:
+            logger.warning("Groq API key not configured for transcription")
+            return ""
+
+        path = Path(file_path)
+        if not path.exists():
+            logger.error("Audio file not found: {}", file_path)
+            return ""
+
+        try:
+            async with httpx.AsyncClient() as client:
+                with open(path, "rb") as f:
+                    files = {
+                        "file": (path.name, f),
+                        "model": (None, "whisper-large-v3"),
+                    }
+                    headers = {
+                        "Authorization": f"Bearer {self.api_key}",
+                    }
+
+                    response = await client.post(
+                        self.api_url,
+                        headers=headers,
+                        files=files,
+                        timeout=60.0
+                    )
+
+                    response.raise_for_status()
+                    data = response.json()
+                    return data.get("text", "")
+
+        except Exception as e:
+            logger.error("Groq transcription error: {}", e)
+            return ""