Grayscale/nanobot
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge origin/main into fix/sanitize-messages-non-claude

Browse Source 
Resolved conflict in azure_openai_provider.py by keeping main's
Responses API implementation (role alternation not needed for the
Responses API input format).

Made-with: Cursor
This commit is contained in:
Xubin Ren
2026-04-09 04:45:45 +00:00
parent 26ae906116 1e3057d0d6
commit dadf453097
172 changed files with 23012 additions and 3414 deletions
Download Patch File Download Diff File Expand all files Collapse all files
nanobot/providers/__init__.py
+3
View File
@@ -13,6 +13,7 @@ __all__ = [
"AnthropicProvider",
"OpenAICompatProvider",
"OpenAICodexProvider",
"GitHubCopilotProvider",
"AzureOpenAIProvider",
]
@@ -20,12 +21,14 @@ _LAZY_IMPORTS = {
"AnthropicProvider": ".anthropic_provider",
"OpenAICompatProvider": ".openai_compat_provider",
"OpenAICodexProvider": ".openai_codex_provider",
"GitHubCopilotProvider": ".github_copilot_provider",
"AzureOpenAIProvider": ".azure_openai_provider",
}
if TYPE_CHECKING:
from nanobot.providers.anthropic_provider import AnthropicProvider
from nanobot.providers.azure_openai_provider import AzureOpenAIProvider
from nanobot.providers.github_copilot_provider import GitHubCopilotProvider
from nanobot.providers.openai_compat_provider import OpenAICompatProvider
from nanobot.providers.openai_codex_provider import OpenAICodexProvider
nanobot/providers/anthropic_provider.py
+106 -11
View File
@@ -2,6 +2,8 @@
from __future__ import annotations
import asyncio
import os
import re
import secrets
import string
@@ -9,7 +11,6 @@ from collections.abc import Awaitable, Callable
from typing import Any
import json_repair
from loguru import logger
from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
@@ -47,8 +48,66 @@ class AnthropicProvider(LLMProvider):
client_kw["base_url"] = api_base
if extra_headers:
client_kw["default_headers"] = extra_headers
# Keep retries centralized in LLMProvider._run_with_retry to avoid retry amplification.
client_kw["max_retries"] = 0
self._client = AsyncAnthropic(**client_kw)
@classmethod
def _handle_error(cls, e: Exception) -> LLMResponse:
response = getattr(e, "response", None)
headers = getattr(response, "headers", None)
payload = (
getattr(e, "body", None)
or getattr(e, "doc", None)
or getattr(response, "text", None)
)
if payload is None and response is not None:
response_json = getattr(response, "json", None)
if callable(response_json):
try:
payload = response_json()
except Exception:
payload = None
payload_text = payload if isinstance(payload, str) else str(payload) if payload is not None else ""
msg = f"Error: {payload_text.strip()[:500]}" if payload_text.strip() else f"Error calling LLM: {e}"
retry_after = cls._extract_retry_after_from_headers(headers)
if retry_after is None:
retry_after = LLMProvider._extract_retry_after(msg)
status_code = getattr(e, "status_code", None)
if status_code is None and response is not None:
status_code = getattr(response, "status_code", None)
should_retry: bool | None = None
if headers is not None:
raw = headers.get("x-should-retry")
if isinstance(raw, str):
lowered = raw.strip().lower()
if lowered == "true":
should_retry = True
elif lowered == "false":
should_retry = False
error_kind: str | None = None
error_name = e.__class__.__name__.lower()
if "timeout" in error_name:
error_kind = "timeout"
elif "connection" in error_name:
error_kind = "connection"
error_type, error_code = LLMProvider._extract_error_type_code(payload)
return LLMResponse(
content=msg,
finish_reason="error",
retry_after=retry_after,
error_status_code=int(status_code) if status_code is not None else None,
error_kind=error_kind,
error_type=error_type,
error_code=error_code,
error_retry_after_s=retry_after,
error_should_retry=should_retry,
)
@staticmethod
def _strip_prefix(model: str) -> str:
if model.startswith("anthropic/"):
@@ -251,8 +310,9 @@ class AnthropicProvider(LLMProvider):
# Prompt caching
# ------------------------------------------------------------------
@staticmethod
@classmethod
def _apply_cache_control(
cls,
system: str | list[dict[str, Any]],
messages: list[dict[str, Any]],
tools: list[dict[str, Any]] | None,
@@ -279,7 +339,8 @@ class AnthropicProvider(LLMProvider):
new_tools = tools
if tools:
new_tools = list(tools)
new_tools[-1] = {**new_tools[-1], "cache_control": marker}
for idx in cls._tool_cache_marker_indices(new_tools):
new_tools[idx] = {**new_tools[idx], "cache_control": marker}
return system, new_msgs, new_tools
@@ -319,9 +380,15 @@ class AnthropicProvider(LLMProvider):
if system:
kwargs["system"] = system
if thinking_enabled:
if reasoning_effort == "adaptive":
# Adaptive thinking: model decides when and how much to think
# Supported on claude-sonnet-4-6 and claude-opus-4-6.
# Also auto-enables interleaved thinking between tool calls.
kwargs["thinking"] = {"type": "adaptive"}
kwargs["temperature"] = 1.0
elif thinking_enabled:
budget_map = {"low": 1024, "medium": 4096, "high": max(8192, max_tokens)}
budget = budget_map.get(reasoning_effort.lower(), 4096) # type: ignore[union-attr]
budget = budget_map.get(reasoning_effort.lower(), 4096)
kwargs["thinking"] = {"type": "enabled", "budget_tokens": budget}
kwargs["max_tokens"] = max(max_tokens, budget + 4096)
kwargs["temperature"] = 1.0
@@ -370,15 +437,22 @@ class AnthropicProvider(LLMProvider):
usage: dict[str, int] = {}
if response.usage:
input_tokens = response.usage.input_tokens
cache_creation = getattr(response.usage, "cache_creation_input_tokens", 0) or 0
cache_read = getattr(response.usage, "cache_read_input_tokens", 0) or 0
total_prompt_tokens = input_tokens + cache_creation + cache_read
usage = {
"prompt_tokens": response.usage.input_tokens,
"prompt_tokens": total_prompt_tokens,
"completion_tokens": response.usage.output_tokens,
"total_tokens": response.usage.input_tokens + response.usage.output_tokens,
"total_tokens": total_prompt_tokens + response.usage.output_tokens,
}
for attr in ("cache_creation_input_tokens", "cache_read_input_tokens"):
val = getattr(response.usage, attr, 0)
if val:
usage[attr] = val
# Normalize to cached_tokens for downstream consistency.
if cache_read:
usage["cached_tokens"] = cache_read
return LLMResponse(
content="".join(content_parts) or None,
@@ -410,7 +484,7 @@ class AnthropicProvider(LLMProvider):
response = await self._client.messages.create(**kwargs)
return self._parse_response(response)
except Exception as e:
return LLMResponse(content=f"Error calling LLM: {e}", finish_reason="error")
return self._handle_error(e)
async def chat_stream(
self,
@@ -427,15 +501,36 @@ class AnthropicProvider(LLMProvider):
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
idle_timeout_s = int(os.environ.get("NANOBOT_STREAM_IDLE_TIMEOUT_S", "90"))
try:
async with self._client.messages.stream(**kwargs) as stream:
if on_content_delta:
async for text in stream.text_stream:
stream_iter = stream.text_stream.__aiter__()
while True:
try:
text = await asyncio.wait_for(
stream_iter.__anext__(),
timeout=idle_timeout_s,
)
except StopAsyncIteration:
break
await on_content_delta(text)
response = await stream.get_final_message()
response = await asyncio.wait_for(
stream.get_final_message(),
timeout=idle_timeout_s,
)
return self._parse_response(response)
except asyncio.TimeoutError:
return LLMResponse(
content=(
f"Error calling LLM: stream stalled for more than "
f"{idle_timeout_s} seconds"
),
finish_reason="error",
error_kind="timeout",
)
except Exception as e:
return LLMResponse(content=f"Error calling LLM: {e}", finish_reason="error")
return self._handle_error(e)
def get_default_model(self) -> str:
return self.default_model
nanobot/providers/azure_openai_provider.py
+99 -227
View File
@@ -1,31 +1,36 @@
"""Azure OpenAI provider implementation with API version 2024-10-21."""
"""Azure OpenAI provider using the OpenAI SDK Responses API.
Uses ``AsyncOpenAI`` pointed at ``https://{endpoint}/openai/v1/`` which
routes to the Responses API (``/responses``). Reuses shared conversion
helpers from :mod:`nanobot.providers.openai_responses`.
"""
from __future__ import annotations
import json
import uuid
from collections.abc import Awaitable, Callable
from typing import Any
from urllib.parse import urljoin
import httpx
import json_repair
from openai import AsyncOpenAI
from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
_AZURE_MSG_KEYS = frozenset({"role", "content", "tool_calls", "tool_call_id", "name"})
from nanobot.providers.base import LLMProvider, LLMResponse
from nanobot.providers.openai_responses import (
consume_sdk_stream,
convert_messages,
convert_tools,
parse_response_output,
)
class AzureOpenAIProvider(LLMProvider):
"""
Azure OpenAI provider with API version 2024-10-21 compliance.
"""Azure OpenAI provider backed by the Responses API.
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
- Uses the OpenAI Python SDK (``AsyncOpenAI``) with
``base_url = {endpoint}/openai/v1/``
- Calls ``client.responses.create()`` (Responses API)
- Reuses shared message/tool/SSE conversion from
``openai_responses``
"""
def __init__(
@@ -36,40 +41,29 @@ class AzureOpenAIProvider(LLMProvider):
):
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 += '/'
# Normalise: ensure trailing slash
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"
# SDK client targeting the Azure Responses API endpoint
base_url = f"{api_base.rstrip('/')}/openai/v1/"
self._client = AsyncOpenAI(
api_key=api_key,
base_url=base_url,
default_headers={"x-session-affinity": uuid.uuid4().hex},
max_retries=0,
)
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
}
# ------------------------------------------------------------------
# Helpers
# ------------------------------------------------------------------
@staticmethod
def _supports_temperature(
@@ -82,38 +76,56 @@ class AzureOpenAIProvider(LLMProvider):
name = deployment_name.lower()
return not any(token in name for token in ("gpt-5", "o1", "o3", "o4"))
def _prepare_request_payload(
def _build_body(
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,
tool_choice: str | dict[str, Any] | None = None,
tools: list[dict[str, Any]] | None,
model: str | None,
max_tokens: int,
temperature: float,
reasoning_effort: str | None,
tool_choice: str | dict[str, Any] | None,
) -> dict[str, Any]:
"""Prepare the request payload with Azure OpenAI 2024-10-21 compliance."""
payload: dict[str, Any] = {
"messages": self._enforce_role_alternation(
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
"""Build the Responses API request body from Chat-Completions-style args."""
deployment = model or self.default_model
instructions, input_items = convert_messages(self._sanitize_empty_content(messages))
body: dict[str, Any] = {
"model": deployment,
"instructions": instructions or None,
"input": input_items,
"max_output_tokens": max(1, max_tokens),
"store": False,
"stream": False,
}
if self._supports_temperature(deployment_name, reasoning_effort):
payload["temperature"] = temperature
if self._supports_temperature(deployment, reasoning_effort):
body["temperature"] = temperature
if reasoning_effort:
payload["reasoning_effort"] = reasoning_effort
body["reasoning"] = {"effort": reasoning_effort}
body["include"] = ["reasoning.encrypted_content"]
if tools:
payload["tools"] = tools
payload["tool_choice"] = tool_choice or "auto"
body["tools"] = convert_tools(tools)
body["tool_choice"] = tool_choice or "auto"
return payload
return body
@staticmethod
def _handle_error(e: Exception) -> LLMResponse:
response = getattr(e, "response", None)
body = getattr(e, "body", None) or getattr(response, "text", None)
body_text = str(body).strip() if body is not None else ""
msg = f"Error: {body_text[:500]}" if body_text else f"Error calling Azure OpenAI: {e}"
retry_after = LLMProvider._extract_retry_after_from_headers(getattr(response, "headers", None))
if retry_after is None:
retry_after = LLMProvider._extract_retry_after(msg)
return LLMResponse(content=msg, finish_reason="error", retry_after=retry_after)
# ------------------------------------------------------------------
# Public API
# ------------------------------------------------------------------
async def chat(
self,
@@ -125,92 +137,15 @@ class AzureOpenAIProvider(LLMProvider):
reasoning_effort: str | None = None,
tool_choice: str | dict[str, Any] | 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,
tool_choice=tool_choice,
body = self._build_body(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
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)
response = await self._client.responses.create(**body)
return parse_response_output(response)
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",
)
return self._handle_error(e)
async def chat_stream(
self,
@@ -223,89 +158,26 @@ class AzureOpenAIProvider(LLMProvider):
tool_choice: str | dict[str, Any] | None = None,
on_content_delta: Callable[[str], Awaitable[None]] | None = None,
) -> LLMResponse:
"""Stream a chat completion via Azure OpenAI SSE."""
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, tool_choice=tool_choice,
body = self._build_body(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
payload["stream"] = True
body["stream"] = True
try:
async with httpx.AsyncClient(timeout=60.0, verify=True) as client:
async with client.stream("POST", url, headers=headers, json=payload) as response:
if response.status_code != 200:
text = await response.aread()
return LLMResponse(
content=f"Azure OpenAI API Error {response.status_code}: {text.decode('utf-8', 'ignore')}",
finish_reason="error",
)
return await self._consume_stream(response, on_content_delta)
except Exception as e:
return LLMResponse(content=f"Error calling Azure OpenAI: {repr(e)}", finish_reason="error")
async def _consume_stream(
self,
response: httpx.Response,
on_content_delta: Callable[[str], Awaitable[None]] | None,
) -> LLMResponse:
"""Parse Azure OpenAI SSE stream into an LLMResponse."""
content_parts: list[str] = []
tool_call_buffers: dict[int, dict[str, str]] = {}
finish_reason = "stop"
async for line in response.aiter_lines():
if not line.startswith("data: "):
continue
data = line[6:].strip()
if data == "[DONE]":
break
try:
chunk = json.loads(data)
except Exception:
continue
choices = chunk.get("choices") or []
if not choices:
continue
choice = choices[0]
if choice.get("finish_reason"):
finish_reason = choice["finish_reason"]
delta = choice.get("delta") or {}
text = delta.get("content")
if text:
content_parts.append(text)
if on_content_delta:
await on_content_delta(text)
for tc in delta.get("tool_calls") or []:
idx = tc.get("index", 0)
buf = tool_call_buffers.setdefault(idx, {"id": "", "name": "", "arguments": ""})
if tc.get("id"):
buf["id"] = tc["id"]
fn = tc.get("function") or {}
if fn.get("name"):
buf["name"] = fn["name"]
if fn.get("arguments"):
buf["arguments"] += fn["arguments"]
tool_calls = [
ToolCallRequest(
id=buf["id"], name=buf["name"],
arguments=json_repair.loads(buf["arguments"]) if buf["arguments"] else {},
stream = await self._client.responses.create(**body)
content, tool_calls, finish_reason, usage, reasoning_content = (
await consume_sdk_stream(stream, on_content_delta)
)
for buf in tool_call_buffers.values()
]
return LLMResponse(
content="".join(content_parts) or None,
tool_calls=tool_calls,
finish_reason=finish_reason,
)
return LLMResponse(
content=content or None,
tool_calls=tool_calls,
finish_reason=finish_reason,
usage=usage,
reasoning_content=reasoning_content,
)
except Exception as e:
return self._handle_error(e)
def get_default_model(self) -> str:
"""Get the default model (also used as default deployment name)."""
return self.default_model
return self.default_model
nanobot/providers/base.py
+350 -51
View File
@@ -2,13 +2,18 @@
import asyncio
import json
import re
from abc import ABC, abstractmethod
from collections.abc import Awaitable, Callable
from dataclasses import dataclass, field
from datetime import datetime, timezone
from email.utils import parsedate_to_datetime
from typing import Any
from loguru import logger
from nanobot.utils.helpers import image_placeholder_text
@dataclass
class ToolCallRequest:
@@ -46,9 +51,17 @@ class LLMResponse:
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.
retry_after: float | None = None # Provider supplied retry wait in seconds.
reasoning_content: str | None = None # Kimi, DeepSeek-R1, MiMo etc.
thinking_blocks: list[dict] | None = None # Anthropic extended thinking
# Structured error metadata used by retry policy when finish_reason == "error".
error_status_code: int | None = None
error_kind: str | None = None # e.g. "timeout", "connection"
error_type: str | None = None # Provider/type semantic, e.g. insufficient_quota.
error_code: str | None = None # Provider/code semantic, e.g. rate_limit_exceeded.
error_retry_after_s: float | None = None
error_should_retry: bool | None = None
@property
def has_tool_calls(self) -> bool:
"""Check if response contains tool calls."""
@@ -57,13 +70,7 @@ class LLMResponse:
@dataclass(frozen=True)
class GenerationSettings:
"""Default generation parameters for LLM calls.
Stored on the provider so every call site inherits the same defaults
without having to pass temperature / max_tokens / reasoning_effort
through every layer. Individual call sites can still override by
passing explicit keyword arguments to chat() / chat_with_retry().
"""
"""Default generation settings."""
temperature: float = 0.7
max_tokens: int = 4096
@@ -71,14 +78,12 @@ class GenerationSettings:
class LLMProvider(ABC):
"""
Abstract base class for LLM providers.
Implementations should handle the specifics of each provider's API
while maintaining a consistent interface.
"""
"""Base class for LLM providers."""
_CHAT_RETRY_DELAYS = (1, 2, 4)
_PERSISTENT_MAX_DELAY = 60
_PERSISTENT_IDENTICAL_ERROR_LIMIT = 10
_RETRY_HEARTBEAT_CHUNK = 30
_TRANSIENT_ERROR_MARKERS = (
"429",
"rate limit",
@@ -93,6 +98,52 @@ class LLMProvider(ABC):
"server error",
"temporarily unavailable",
)
_RETRYABLE_STATUS_CODES = frozenset({408, 409, 429})
_TRANSIENT_ERROR_KINDS = frozenset({"timeout", "connection"})
_NON_RETRYABLE_429_ERROR_TOKENS = frozenset({
"insufficient_quota",
"quota_exceeded",
"quota_exhausted",
"billing_hard_limit_reached",
"insufficient_balance",
"credit_balance_too_low",
"billing_not_active",
"payment_required",
})
_RETRYABLE_429_ERROR_TOKENS = frozenset({
"rate_limit_exceeded",
"rate_limit_error",
"too_many_requests",
"request_limit_exceeded",
"requests_limit_exceeded",
"overloaded_error",
})
_NON_RETRYABLE_429_TEXT_MARKERS = (
"insufficient_quota",
"insufficient quota",
"quota exceeded",
"quota exhausted",
"billing hard limit",
"billing_hard_limit_reached",
"billing not active",
"insufficient balance",
"insufficient_balance",
"credit balance too low",
"payment required",
"out of credits",
"out of quota",
"exceeded your current quota",
)
_RETRYABLE_429_TEXT_MARKERS = (
"rate limit",
"rate_limit",
"too many requests",
"retry after",
"try again in",
"temporarily unavailable",
"overloaded",
"concurrency limit",
)
_SENTINEL = object()
@@ -150,6 +201,38 @@ class LLMProvider(ABC):
result.append(msg)
return result
@staticmethod
def _tool_name(tool: dict[str, Any]) -> str:
"""Extract tool name from either OpenAI or Anthropic-style tool schemas."""
name = tool.get("name")
if isinstance(name, str):
return name
fn = tool.get("function")
if isinstance(fn, dict):
fname = fn.get("name")
if isinstance(fname, str):
return fname
return ""
@classmethod
def _tool_cache_marker_indices(cls, tools: list[dict[str, Any]]) -> list[int]:
"""Return cache marker indices: builtin/MCP boundary and tail index."""
if not tools:
return []
tail_idx = len(tools) - 1
last_builtin_idx: int | None = None
for i in range(tail_idx, -1, -1):
if not cls._tool_name(tools[i]).startswith("mcp_"):
last_builtin_idx = i
break
ordered_unique: list[int] = []
for idx in (last_builtin_idx, tail_idx):
if idx is not None and idx not in ordered_unique:
ordered_unique.append(idx)
return ordered_unique
@staticmethod
def _sanitize_request_messages(
messages: list[dict[str, Any]],
@@ -177,7 +260,7 @@ class LLMProvider(ABC):
) -> LLMResponse:
"""
Send a chat completion request.
Args:
messages: List of message dicts with 'role' and 'content'.
tools: Optional list of tool definitions.
@@ -185,7 +268,7 @@ class LLMProvider(ABC):
max_tokens: Maximum tokens in response.
temperature: Sampling temperature.
tool_choice: Tool selection strategy ("auto", "required", or specific tool dict).
Returns:
LLMResponse with content and/or tool calls.
"""
@@ -196,6 +279,80 @@ class LLMProvider(ABC):
err = (content or "").lower()
return any(marker in err for marker in cls._TRANSIENT_ERROR_MARKERS)
@classmethod
def _is_transient_response(cls, response: LLMResponse) -> bool:
"""Prefer structured error metadata, fallback to text markers for legacy providers."""
if response.error_should_retry is not None:
return bool(response.error_should_retry)
if response.error_status_code is not None:
status = int(response.error_status_code)
if status == 429:
return cls._is_retryable_429_response(response)
if status in cls._RETRYABLE_STATUS_CODES or status >= 500:
return True
kind = (response.error_kind or "").strip().lower()
if kind in cls._TRANSIENT_ERROR_KINDS:
return True
return cls._is_transient_error(response.content)
@staticmethod
def _normalize_error_token(value: Any) -> str | None:
if value is None:
return None
token = str(value).strip().lower()
return token or None
@classmethod
def _extract_error_type_code(cls, payload: Any) -> tuple[str | None, str | None]:
data: dict[str, Any] | None = None
if isinstance(payload, dict):
data = payload
elif isinstance(payload, str):
text = payload.strip()
if text:
try:
parsed = json.loads(text)
except Exception:
parsed = None
if isinstance(parsed, dict):
data = parsed
if not isinstance(data, dict):
return None, None
error_obj = data.get("error")
type_value = data.get("type")
code_value = data.get("code")
if isinstance(error_obj, dict):
type_value = error_obj.get("type") or type_value
code_value = error_obj.get("code") or code_value
return cls._normalize_error_token(type_value), cls._normalize_error_token(code_value)
@classmethod
def _is_retryable_429_response(cls, response: LLMResponse) -> bool:
type_token = cls._normalize_error_token(response.error_type)
code_token = cls._normalize_error_token(response.error_code)
semantic_tokens = {
token for token in (type_token, code_token)
if token is not None
}
if any(token in cls._NON_RETRYABLE_429_ERROR_TOKENS for token in semantic_tokens):
return False
content = (response.content or "").lower()
if any(marker in content for marker in cls._NON_RETRYABLE_429_TEXT_MARKERS):
return False
if any(token in cls._RETRYABLE_429_ERROR_TOKENS for token in semantic_tokens):
return True
if any(marker in content for marker in cls._RETRYABLE_429_TEXT_MARKERS):
return True
# Unknown 429 defaults to WAIT+retry.
return True
@staticmethod
def _enforce_role_alternation(messages: list[dict[str, Any]]) -> list[dict[str, Any]]:
"""Merge consecutive same-role messages and drop trailing assistant messages.
@@ -244,7 +401,7 @@ class LLMProvider(ABC):
for b in content:
if isinstance(b, dict) and b.get("type") == "image_url":
path = (b.get("_meta") or {}).get("path", "")
placeholder = f"[image: {path}]" if path else "[image omitted]"
placeholder = image_placeholder_text(path, empty="[image omitted]")
new_content.append({"type": "text", "text": placeholder})
found = True
else:
@@ -309,6 +466,8 @@ class LLMProvider(ABC):
reasoning_effort: object = _SENTINEL,
tool_choice: str | dict[str, Any] | None = None,
on_content_delta: Callable[[str], Awaitable[None]] | None = None,
retry_mode: str = "standard",
on_retry_wait: Callable[[str], Awaitable[None]] | None = None,
) -> LLMResponse:
"""Call chat_stream() with retry on transient provider failures."""
if max_tokens is self._SENTINEL:
@@ -324,28 +483,13 @@ class LLMProvider(ABC):
reasoning_effort=reasoning_effort, tool_choice=tool_choice,
on_content_delta=on_content_delta,
)
for attempt, delay in enumerate(self._CHAT_RETRY_DELAYS, start=1):
response = await self._safe_chat_stream(**kw)
if response.finish_reason != "error":
return response
if not self._is_transient_error(response.content):
stripped = self._strip_image_content(messages)
if stripped is not None:
logger.warning("Non-transient LLM error with image content, retrying without images")
return await self._safe_chat_stream(**{**kw, "messages": stripped})
return response
logger.warning(
"LLM transient error (attempt {}/{}), retrying in {}s: {}",
attempt, len(self._CHAT_RETRY_DELAYS), delay,
(response.content or "")[:120].lower(),
)
await asyncio.sleep(delay)
return await self._safe_chat_stream(**kw)
return await self._run_with_retry(
self._safe_chat_stream,
kw,
messages,
retry_mode=retry_mode,
on_retry_wait=on_retry_wait,
)
async def chat_with_retry(
self,
@@ -356,6 +500,8 @@ class LLMProvider(ABC):
temperature: object = _SENTINEL,
reasoning_effort: object = _SENTINEL,
tool_choice: str | dict[str, Any] | None = None,
retry_mode: str = "standard",
on_retry_wait: Callable[[str], Awaitable[None]] | None = None,
) -> LLMResponse:
"""Call chat() with retry on transient provider failures.
@@ -375,28 +521,181 @@ class LLMProvider(ABC):
max_tokens=max_tokens, temperature=temperature,
reasoning_effort=reasoning_effort, tool_choice=tool_choice,
)
return await self._run_with_retry(
self._safe_chat,
kw,
messages,
retry_mode=retry_mode,
on_retry_wait=on_retry_wait,
)
for attempt, delay in enumerate(self._CHAT_RETRY_DELAYS, start=1):
response = await self._safe_chat(**kw)
@classmethod
def _extract_retry_after(cls, content: str | None) -> float | None:
text = (content or "").lower()
patterns = (
r"retry after\s+(\d+(?:\.\d+)?)\s*(ms|milliseconds|s|sec|secs|seconds|m|min|minutes)?",
r"try again in\s+(\d+(?:\.\d+)?)\s*(ms|milliseconds|s|sec|secs|seconds|m|min|minutes)",
r"wait\s+(\d+(?:\.\d+)?)\s*(ms|milliseconds|s|sec|secs|seconds|m|min|minutes)\s*before retry",
r"retry[_-]?after[\"'\s:=]+(\d+(?:\.\d+)?)",
)
for idx, pattern in enumerate(patterns):
match = re.search(pattern, text)
if not match:
continue
value = float(match.group(1))
unit = match.group(2) if idx < 3 else "s"
return cls._to_retry_seconds(value, unit)
return None
@classmethod
def _to_retry_seconds(cls, value: float, unit: str | None = None) -> float:
normalized_unit = (unit or "s").lower()
if normalized_unit in {"ms", "milliseconds"}:
return max(0.1, value / 1000.0)
if normalized_unit in {"m", "min", "minutes"}:
return max(0.1, value * 60.0)
return max(0.1, value)
@classmethod
def _extract_retry_after_from_headers(cls, headers: Any) -> float | None:
if not headers:
return None
def _header_value(name: str) -> Any:
if hasattr(headers, "get"):
value = headers.get(name) or headers.get(name.title())
if value is not None:
return value
if isinstance(headers, dict):
for key, value in headers.items():
if isinstance(key, str) and key.lower() == name.lower():
return value
return None
try:
retry_ms = _header_value("retry-after-ms")
if retry_ms is not None:
value = float(retry_ms) / 1000.0
if value > 0:
return value
except (TypeError, ValueError):
pass
retry_after = _header_value("retry-after")
if retry_after is None:
return None
retry_after_text = str(retry_after).strip()
if not retry_after_text:
return None
if re.fullmatch(r"\d+(?:\.\d+)?", retry_after_text):
return cls._to_retry_seconds(float(retry_after_text), "s")
try:
retry_at = parsedate_to_datetime(retry_after_text)
except Exception:
return None
if retry_at.tzinfo is None:
retry_at = retry_at.replace(tzinfo=timezone.utc)
remaining = (retry_at - datetime.now(retry_at.tzinfo)).total_seconds()
return max(0.1, remaining)
@classmethod
def _extract_retry_after_from_response(cls, response: LLMResponse) -> float | None:
if response.error_retry_after_s is not None and response.error_retry_after_s > 0:
return response.error_retry_after_s
if response.retry_after is not None and response.retry_after > 0:
return response.retry_after
return cls._extract_retry_after(response.content)
async def _sleep_with_heartbeat(
self,
delay: float,
*,
attempt: int,
persistent: bool,
on_retry_wait: Callable[[str], Awaitable[None]] | None = None,
) -> None:
remaining = max(0.0, delay)
while remaining > 0:
if on_retry_wait:
kind = "persistent retry" if persistent else "retry"
await on_retry_wait(
f"Model request failed, {kind} in {max(1, int(round(remaining)))}s "
f"(attempt {attempt})."
)
chunk = min(remaining, self._RETRY_HEARTBEAT_CHUNK)
await asyncio.sleep(chunk)
remaining -= chunk
async def _run_with_retry(
self,
call: Callable[..., Awaitable[LLMResponse]],
kw: dict[str, Any],
original_messages: list[dict[str, Any]],
*,
retry_mode: str,
on_retry_wait: Callable[[str], Awaitable[None]] | None,
) -> LLMResponse:
attempt = 0
delays = list(self._CHAT_RETRY_DELAYS)
persistent = retry_mode == "persistent"
last_response: LLMResponse | None = None
last_error_key: str | None = None
identical_error_count = 0
while True:
attempt += 1
response = await call(**kw)
if response.finish_reason != "error":
return response
last_response = response
error_key = ((response.content or "").strip().lower() or None)
if error_key and error_key == last_error_key:
identical_error_count += 1
else:
last_error_key = error_key
identical_error_count = 1 if error_key else 0
if not self._is_transient_error(response.content):
stripped = self._strip_image_content(messages)
if stripped is not None:
logger.warning("Non-transient LLM error with image content, retrying without images")
return await self._safe_chat(**{**kw, "messages": stripped})
if not self._is_transient_response(response):
stripped = self._strip_image_content(original_messages)
if stripped is not None and stripped != kw["messages"]:
logger.warning(
"Non-transient LLM error with image content, retrying without images"
)
retry_kw = dict(kw)
retry_kw["messages"] = stripped
return await call(**retry_kw)
return response
if persistent and identical_error_count >= self._PERSISTENT_IDENTICAL_ERROR_LIMIT:
logger.warning(
"Stopping persistent retry after {} identical transient errors: {}",
identical_error_count,
(response.content or "")[:120].lower(),
)
return response
if not persistent and attempt > len(delays):
break
base_delay = delays[min(attempt - 1, len(delays) - 1)]
delay = self._extract_retry_after_from_response(response) or base_delay
if persistent:
delay = min(delay, self._PERSISTENT_MAX_DELAY)
logger.warning(
"LLM transient error (attempt {}/{}), retrying in {}s: {}",
attempt, len(self._CHAT_RETRY_DELAYS), delay,
"LLM transient error (attempt {}{}), retrying in {}s: {}",
attempt,
"+" if persistent and attempt > len(delays) else f"/{len(delays)}",
int(round(delay)),
(response.content or "")[:120].lower(),
)
await asyncio.sleep(delay)
await self._sleep_with_heartbeat(
delay,
attempt=attempt,
persistent=persistent,
on_retry_wait=on_retry_wait,
)
return await self._safe_chat(**kw)
return last_response if last_response is not None else await call(**kw)
@abstractmethod
def get_default_model(self) -> str:
nanobot/providers/github_copilot_provider.py
+257
View File
@@ -0,0 +1,257 @@
"""GitHub Copilot OAuth-backed provider."""
from __future__ import annotations
import time
import webbrowser
from collections.abc import Callable
import httpx
from oauth_cli_kit.models import OAuthToken
from oauth_cli_kit.storage import FileTokenStorage
from nanobot.providers.openai_compat_provider import OpenAICompatProvider
DEFAULT_GITHUB_DEVICE_CODE_URL = "https://github.com/login/device/code"
DEFAULT_GITHUB_ACCESS_TOKEN_URL = "https://github.com/login/oauth/access_token"
DEFAULT_GITHUB_USER_URL = "https://api.github.com/user"
DEFAULT_COPILOT_TOKEN_URL = "https://api.github.com/copilot_internal/v2/token"
DEFAULT_COPILOT_BASE_URL = "https://api.githubcopilot.com"
GITHUB_COPILOT_CLIENT_ID = "Iv1.b507a08c87ecfe98"
GITHUB_COPILOT_SCOPE = "read:user"
TOKEN_FILENAME = "github-copilot.json"
TOKEN_APP_NAME = "nanobot"
USER_AGENT = "nanobot/0.1"
EDITOR_VERSION = "vscode/1.99.0"
EDITOR_PLUGIN_VERSION = "copilot-chat/0.26.0"
_EXPIRY_SKEW_SECONDS = 60
_LONG_LIVED_TOKEN_SECONDS = 315360000
def _storage() -> FileTokenStorage:
return FileTokenStorage(
token_filename=TOKEN_FILENAME,
app_name=TOKEN_APP_NAME,
import_codex_cli=False,
)
def _copilot_headers(token: str) -> dict[str, str]:
return {
"Authorization": f"token {token}",
"Accept": "application/json",
"User-Agent": USER_AGENT,
"Editor-Version": EDITOR_VERSION,
"Editor-Plugin-Version": EDITOR_PLUGIN_VERSION,
}
def _load_github_token() -> OAuthToken | None:
token = _storage().load()
if not token or not token.access:
return None
return token
def get_github_copilot_login_status() -> OAuthToken | None:
"""Return the persisted GitHub OAuth token if available."""
return _load_github_token()
def login_github_copilot(
print_fn: Callable[[str], None] | None = None,
prompt_fn: Callable[[str], str] | None = None,
) -> OAuthToken:
"""Run GitHub device flow and persist the GitHub OAuth token used for Copilot."""
del prompt_fn
printer = print_fn or print
timeout = httpx.Timeout(20.0, connect=20.0)
with httpx.Client(timeout=timeout, follow_redirects=True, trust_env=True) as client:
response = client.post(
DEFAULT_GITHUB_DEVICE_CODE_URL,
headers={"Accept": "application/json", "User-Agent": USER_AGENT},
data={"client_id": GITHUB_COPILOT_CLIENT_ID, "scope": GITHUB_COPILOT_SCOPE},
)
response.raise_for_status()
payload = response.json()
device_code = str(payload["device_code"])
user_code = str(payload["user_code"])
verify_url = str(payload.get("verification_uri") or payload.get("verification_uri_complete") or "")
verify_complete = str(payload.get("verification_uri_complete") or verify_url)
interval = max(1, int(payload.get("interval") or 5))
expires_in = int(payload.get("expires_in") or 900)
printer(f"Open: {verify_url}")
printer(f"Code: {user_code}")
if verify_complete:
try:
webbrowser.open(verify_complete)
except Exception:
pass
deadline = time.time() + expires_in
current_interval = interval
access_token = None
token_expires_in = _LONG_LIVED_TOKEN_SECONDS
while time.time() < deadline:
poll = client.post(
DEFAULT_GITHUB_ACCESS_TOKEN_URL,
headers={"Accept": "application/json", "User-Agent": USER_AGENT},
data={
"client_id": GITHUB_COPILOT_CLIENT_ID,
"device_code": device_code,
"grant_type": "urn:ietf:params:oauth:grant-type:device_code",
},
)
poll.raise_for_status()
poll_payload = poll.json()
access_token = poll_payload.get("access_token")
if access_token:
token_expires_in = int(poll_payload.get("expires_in") or _LONG_LIVED_TOKEN_SECONDS)
break
error = poll_payload.get("error")
if error == "authorization_pending":
time.sleep(current_interval)
continue
if error == "slow_down":
current_interval += 5
time.sleep(current_interval)
continue
if error == "expired_token":
raise RuntimeError("GitHub device code expired. Please run login again.")
if error == "access_denied":
raise RuntimeError("GitHub device flow was denied.")
if error:
desc = poll_payload.get("error_description") or error
raise RuntimeError(str(desc))
time.sleep(current_interval)
else:
raise RuntimeError("GitHub device flow timed out.")
user = client.get(
DEFAULT_GITHUB_USER_URL,
headers={
"Authorization": f"Bearer {access_token}",
"Accept": "application/vnd.github+json",
"User-Agent": USER_AGENT,
},
)
user.raise_for_status()
user_payload = user.json()
account_id = user_payload.get("login") or str(user_payload.get("id") or "") or None
expires_ms = int((time.time() + token_expires_in) * 1000)
token = OAuthToken(
access=str(access_token),
refresh="",
expires=expires_ms,
account_id=str(account_id) if account_id else None,
)
_storage().save(token)
return token
class GitHubCopilotProvider(OpenAICompatProvider):
"""Provider that exchanges a stored GitHub OAuth token for Copilot access tokens."""
def __init__(self, default_model: str = "github-copilot/gpt-4.1"):
from nanobot.providers.registry import find_by_name
self._copilot_access_token: str | None = None
self._copilot_expires_at: float = 0.0
super().__init__(
api_key="no-key",
api_base=DEFAULT_COPILOT_BASE_URL,
default_model=default_model,
extra_headers={
"Editor-Version": EDITOR_VERSION,
"Editor-Plugin-Version": EDITOR_PLUGIN_VERSION,
"User-Agent": USER_AGENT,
},
spec=find_by_name("github_copilot"),
)
async def _get_copilot_access_token(self) -> str:
now = time.time()
if self._copilot_access_token and now < self._copilot_expires_at - _EXPIRY_SKEW_SECONDS:
return self._copilot_access_token
github_token = _load_github_token()
if not github_token or not github_token.access:
raise RuntimeError("GitHub Copilot is not logged in. Run: nanobot provider login github-copilot")
timeout = httpx.Timeout(20.0, connect=20.0)
async with httpx.AsyncClient(timeout=timeout, follow_redirects=True, trust_env=True) as client:
response = await client.get(
DEFAULT_COPILOT_TOKEN_URL,
headers=_copilot_headers(github_token.access),
)
response.raise_for_status()
payload = response.json()
token = payload.get("token")
if not token:
raise RuntimeError("GitHub Copilot token exchange returned no token.")
expires_at = payload.get("expires_at")
if isinstance(expires_at, (int, float)):
self._copilot_expires_at = float(expires_at)
else:
refresh_in = payload.get("refresh_in") or 1500
self._copilot_expires_at = time.time() + int(refresh_in)
self._copilot_access_token = str(token)
return self._copilot_access_token
async def _refresh_client_api_key(self) -> str:
token = await self._get_copilot_access_token()
self.api_key = token
self._client.api_key = token
return token
async def chat(
self,
messages: list[dict[str, object]],
tools: list[dict[str, object]] | None = None,
model: str | None = None,
max_tokens: int = 4096,
temperature: float = 0.7,
reasoning_effort: str | None = None,
tool_choice: str | dict[str, object] | None = None,
):
await self._refresh_client_api_key()
return await super().chat(
messages=messages,
tools=tools,
model=model,
max_tokens=max_tokens,
temperature=temperature,
reasoning_effort=reasoning_effort,
tool_choice=tool_choice,
)
async def chat_stream(
self,
messages: list[dict[str, object]],
tools: list[dict[str, object]] | None = None,
model: str | None = None,
max_tokens: int = 4096,
temperature: float = 0.7,
reasoning_effort: str | None = None,
tool_choice: str | dict[str, object] | None = None,
on_content_delta: Callable[[str], None] | None = None,
):
await self._refresh_client_api_key()
return await super().chat_stream(
messages=messages,
tools=tools,
model=model,
max_tokens=max_tokens,
temperature=temperature,
reasoning_effort=reasoning_effort,
tool_choice=tool_choice,
on_content_delta=on_content_delta,
)
nanobot/providers/openai_codex_provider.py
+23 -185
View File
@@ -6,13 +6,18 @@ import asyncio
import hashlib
import json
from collections.abc import Awaitable, Callable
from typing import Any, AsyncGenerator
from typing import Any
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
from nanobot.providers.openai_responses import (
consume_sse,
convert_messages,
convert_tools,
)
DEFAULT_CODEX_URL = "https://chatgpt.com/backend-api/codex/responses"
DEFAULT_ORIGINATOR = "nanobot"
@@ -36,7 +41,7 @@ class OpenAICodexProvider(LLMProvider):
) -> LLMResponse:
"""Shared request logic for both chat() and chat_stream()."""
model = model or self.default_model
system_prompt, input_items = _convert_messages(messages)
system_prompt, input_items = convert_messages(messages)
token = await asyncio.to_thread(get_codex_token)
headers = _build_headers(token.account_id, token.access)
@@ -56,7 +61,7 @@ class OpenAICodexProvider(LLMProvider):
if reasoning_effort:
body["reasoning"] = {"effort": reasoning_effort}
if tools:
body["tools"] = _convert_tools(tools)
body["tools"] = convert_tools(tools)
try:
try:
@@ -74,7 +79,9 @@ class OpenAICodexProvider(LLMProvider):
)
return LLMResponse(content=content, tool_calls=tool_calls, finish_reason=finish_reason)
except Exception as e:
return LLMResponse(content=f"Error calling Codex: {e}", finish_reason="error")
msg = f"Error calling Codex: {e}"
retry_after = getattr(e, "retry_after", None) or self._extract_retry_after(msg)
return LLMResponse(content=msg, finish_reason="error", retry_after=retry_after)
async def chat(
self, messages: list[dict[str, Any]], tools: list[dict[str, Any]] | None = None,
@@ -115,6 +122,12 @@ def _build_headers(account_id: str, token: str) -> dict[str, str]:
}
class _CodexHTTPError(RuntimeError):
def __init__(self, message: str, retry_after: float | None = None):
super().__init__(message)
self.retry_after = retry_after
async def _request_codex(
url: str,
headers: dict[str, str],
@@ -126,97 +139,12 @@ async def _request_codex(
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, on_content_delta)
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":
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}",
})
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"))
input_items.append({
"type": "function_call",
"id": item_id or f"fc_{idx}",
"call_id": call_id or f"call_{idx}",
"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})
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
retry_after = LLMProvider._extract_retry_after_from_headers(response.headers)
raise _CodexHTTPError(
_friendly_error(response.status_code, text.decode("utf-8", "ignore")),
retry_after=retry_after,
)
return await consume_sse(response, on_content_delta)
def _prompt_cache_key(messages: list[dict[str, Any]]) -> str:
@@ -224,96 +152,6 @@ def _prompt_cache_key(messages: list[dict[str, Any]]) -> str:
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,
on_content_delta: Callable[[str], Awaitable[None]] | None = None,
) -> 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":
delta_text = event.get("delta") or ""
content += delta_text
if on_content_delta and delta_text:
await on_content_delta(delta_text)
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."
nanobot/providers/openai_compat_provider.py
+371 -25
View File
@@ -2,7 +2,9 @@
from __future__ import annotations
import asyncio
import hashlib
import importlib.util
import os
import secrets
import string
@@ -11,9 +13,25 @@ from collections.abc import Awaitable, Callable
from typing import TYPE_CHECKING, Any
import json_repair
from openai import AsyncOpenAI
if os.environ.get("LANGFUSE_SECRET_KEY") and importlib.util.find_spec("langfuse"):
from langfuse.openai import AsyncOpenAI
else:
if os.environ.get("LANGFUSE_SECRET_KEY"):
import logging
logging.getLogger(__name__).warning(
"LANGFUSE_SECRET_KEY is set but langfuse is not installed; "
"install with `pip install langfuse` to enable tracing"
)
from openai import AsyncOpenAI
from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest
from nanobot.providers.openai_responses import (
consume_sdk_stream,
convert_messages,
convert_tools,
parse_response_output,
)
if TYPE_CHECKING:
from nanobot.providers.registry import ProviderSpec
@@ -101,6 +119,14 @@ def _uses_openrouter_attribution(spec: "ProviderSpec | None", api_base: str | No
return bool(api_base and "openrouter" in api_base.lower())
def _is_direct_openai_base(api_base: str | None) -> bool:
"""Return True for direct OpenAI endpoints, not generic OpenAI-compatible gateways."""
if not api_base:
return True
normalized = api_base.strip().lower().rstrip("/")
return "api.openai.com" in normalized and "openrouter" not in normalized
class OpenAICompatProvider(LLMProvider):
"""Unified provider for all OpenAI-compatible APIs.
@@ -125,6 +151,7 @@ class OpenAICompatProvider(LLMProvider):
self._setup_env(api_key, api_base)
effective_base = api_base or (spec.default_api_base if spec else None) or None
self._effective_base = effective_base
default_headers = {"x-session-affinity": uuid.uuid4().hex}
if _uses_openrouter_attribution(spec, effective_base):
default_headers.update(_DEFAULT_OPENROUTER_HEADERS)
@@ -135,6 +162,7 @@ class OpenAICompatProvider(LLMProvider):
api_key=api_key or "no-key",
base_url=effective_base,
default_headers=default_headers,
max_retries=0,
)
def _setup_env(self, api_key: str, api_base: str | None) -> None:
@@ -151,8 +179,9 @@ class OpenAICompatProvider(LLMProvider):
resolved = env_val.replace("{api_key}", api_key).replace("{api_base}", effective_base)
os.environ.setdefault(env_name, resolved)
@staticmethod
@classmethod
def _apply_cache_control(
cls,
messages: list[dict[str, Any]],
tools: list[dict[str, Any]] | None,
) -> tuple[list[dict[str, Any]], list[dict[str, Any]] | None]:
@@ -180,7 +209,8 @@ class OpenAICompatProvider(LLMProvider):
new_tools = tools
if tools:
new_tools = list(tools)
new_tools[-1] = {**new_tools[-1], "cache_control": cache_marker}
for idx in cls._tool_cache_marker_indices(new_tools):
new_tools[idx] = {**new_tools[idx], "cache_control": cache_marker}
return new_messages, new_tools
@staticmethod
@@ -221,6 +251,21 @@ class OpenAICompatProvider(LLMProvider):
# Build kwargs
# ------------------------------------------------------------------
@staticmethod
def _supports_temperature(
model_name: str,
reasoning_effort: str | None = None,
) -> bool:
"""Return True when the model accepts a temperature parameter.
GPT-5 family and reasoning models (o1/o3/o4) reject temperature
when reasoning_effort is set to anything other than ``"none"``.
"""
if reasoning_effort and reasoning_effort.lower() != "none":
return False
name = model_name.lower()
return not any(token in name for token in ("gpt-5", "o1", "o3", "o4"))
def _build_kwargs(
self,
messages: list[dict[str, Any]],
@@ -235,7 +280,9 @@ class OpenAICompatProvider(LLMProvider):
spec = self._spec
if spec and spec.supports_prompt_caching:
messages, tools = self._apply_cache_control(messages, tools)
model_name = model or self.default_model
if any(model_name.lower().startswith(k) for k in ("anthropic/", "claude")):
messages, tools = self._apply_cache_control(messages, tools)
if spec and spec.strip_model_prefix:
model_name = model_name.split("/")[-1]
@@ -243,9 +290,13 @@ class OpenAICompatProvider(LLMProvider):
kwargs: dict[str, Any] = {
"model": model_name,
"messages": self._sanitize_messages(self._sanitize_empty_content(messages)),
"temperature": temperature,
}
# GPT-5 and reasoning models (o1/o3/o4) reject temperature when
# reasoning_effort is active. Only include it when safe.
if self._supports_temperature(model_name, reasoning_effort):
kwargs["temperature"] = temperature
if spec and getattr(spec, "supports_max_completion_tokens", False):
kwargs["max_completion_tokens"] = max(1, max_tokens)
else:
@@ -261,12 +312,112 @@ class OpenAICompatProvider(LLMProvider):
if reasoning_effort:
kwargs["reasoning_effort"] = reasoning_effort
# Provider-specific thinking parameters.
# Only sent when reasoning_effort is explicitly configured so that
# the provider default is preserved otherwise.
if spec and reasoning_effort is not None:
thinking_enabled = reasoning_effort.lower() != "minimal"
extra: dict[str, Any] | None = None
if spec.name == "dashscope":
extra = {"enable_thinking": thinking_enabled}
elif spec.name in (
"volcengine", "volcengine_coding_plan",
"byteplus", "byteplus_coding_plan",
):
extra = {
"thinking": {"type": "enabled" if thinking_enabled else "disabled"}
}
if extra:
kwargs.setdefault("extra_body", {}).update(extra)
if tools:
kwargs["tools"] = tools
kwargs["tool_choice"] = tool_choice or "auto"
return kwargs
def _should_use_responses_api(
self,
model: str | None,
reasoning_effort: str | None,
) -> bool:
"""Use Responses API only for direct OpenAI requests that benefit from it."""
if self._spec and self._spec.name != "openai":
return False
if not _is_direct_openai_base(self._effective_base):
return False
model_name = (model or self.default_model).lower()
if reasoning_effort and reasoning_effort.lower() != "none":
return True
return any(token in model_name for token in ("gpt-5", "o1", "o3", "o4"))
@staticmethod
def _should_fallback_from_responses_error(e: Exception) -> bool:
"""Fallback only for likely Responses API compatibility errors."""
response = getattr(e, "response", None)
status_code = getattr(e, "status_code", None)
if status_code is None and response is not None:
status_code = getattr(response, "status_code", None)
if status_code not in {400, 404, 422}:
return False
body = (
getattr(e, "body", None)
or getattr(e, "doc", None)
or getattr(response, "text", None)
)
body_text = str(body).lower() if body is not None else ""
compatibility_markers = (
"responses",
"response api",
"max_output_tokens",
"instructions",
"previous_response",
"unsupported",
"not supported",
"unknown parameter",
"unrecognized request argument",
)
return any(marker in body_text for marker in compatibility_markers)
def _build_responses_body(
self,
messages: list[dict[str, Any]],
tools: list[dict[str, Any]] | None,
model: str | None,
max_tokens: int,
temperature: float,
reasoning_effort: str | None,
tool_choice: str | dict[str, Any] | None,
) -> dict[str, Any]:
"""Build a Responses API body for direct OpenAI requests."""
model_name = model or self.default_model
sanitized_messages = self._sanitize_messages(self._sanitize_empty_content(messages))
instructions, input_items = convert_messages(sanitized_messages)
body: dict[str, Any] = {
"model": model_name,
"instructions": instructions or None,
"input": input_items,
"max_output_tokens": max(1, max_tokens),
"store": False,
"stream": False,
}
if self._supports_temperature(model_name, reasoning_effort):
body["temperature"] = temperature
if reasoning_effort and reasoning_effort.lower() != "none":
body["reasoning"] = {"effort": reasoning_effort}
body["include"] = ["reasoning.encrypted_content"]
if tools:
body["tools"] = convert_tools(tools)
body["tool_choice"] = tool_choice or "auto"
return body
# ------------------------------------------------------------------
# Response parsing
# ------------------------------------------------------------------
@@ -308,6 +459,13 @@ class OpenAICompatProvider(LLMProvider):
@classmethod
def _extract_usage(cls, response: Any) -> dict[str, int]:
"""Extract token usage from an OpenAI-compatible response.
Handles both dict-based (raw JSON) and object-based (SDK Pydantic)
responses. Provider-specific ``cached_tokens`` fields are normalised
under a single key; see the priority chain inside for details.
"""
# --- resolve usage object ---
usage_obj = None
response_map = cls._maybe_mapping(response)
if response_map is not None:
@@ -317,19 +475,53 @@ class OpenAICompatProvider(LLMProvider):
usage_map = cls._maybe_mapping(usage_obj)
if usage_map is not None:
return {
result = {
"prompt_tokens": int(usage_map.get("prompt_tokens") or 0),
"completion_tokens": int(usage_map.get("completion_tokens") or 0),
"total_tokens": int(usage_map.get("total_tokens") or 0),
}
if usage_obj:
return {
elif usage_obj:
result = {
"prompt_tokens": getattr(usage_obj, "prompt_tokens", 0) or 0,
"completion_tokens": getattr(usage_obj, "completion_tokens", 0) or 0,
"total_tokens": getattr(usage_obj, "total_tokens", 0) or 0,
}
return {}
else:
return {}
# --- cached_tokens (normalised across providers) ---
# Try nested paths first (dict), fall back to attribute (SDK object).
# Priority order ensures the most specific field wins.
for path in (
("prompt_tokens_details", "cached_tokens"), # OpenAI/Zhipu/MiniMax/Qwen/Mistral/xAI
("cached_tokens",), # StepFun/Moonshot (top-level)
("prompt_cache_hit_tokens",), # DeepSeek/SiliconFlow
):
cached = cls._get_nested_int(usage_map, path)
if not cached and usage_obj:
cached = cls._get_nested_int(usage_obj, path)
if cached:
result["cached_tokens"] = cached
break
return result
@staticmethod
def _get_nested_int(obj: Any, path: tuple[str, ...]) -> int:
"""Drill into *obj* by *path* segments and return an ``int`` value.
Supports both dict-key access and attribute access so it works
uniformly with raw JSON dicts **and** SDK Pydantic models.
"""
current = obj
for segment in path:
if current is None:
return 0
if isinstance(current, dict):
current = current.get(segment)
else:
current = getattr(current, segment, None)
return int(current or 0) if current is not None else 0
def _parse(self, response: Any) -> LLMResponse:
if isinstance(response, str):
@@ -342,9 +534,13 @@ class OpenAICompatProvider(LLMProvider):
content = self._extract_text_content(
response_map.get("content") or response_map.get("output_text")
)
reasoning_content = self._extract_text_content(
response_map.get("reasoning_content")
)
if content is not None:
return LLMResponse(
content=content,
reasoning_content=reasoning_content,
finish_reason=str(response_map.get("finish_reason") or "stop"),
usage=self._extract_usage(response_map),
)
@@ -356,7 +552,12 @@ class OpenAICompatProvider(LLMProvider):
finish_reason = str(choice0.get("finish_reason") or "stop")
raw_tool_calls: list[Any] = []
# StepFun Plan: fallback to reasoning field when content is empty
if not content and msg0.get("reasoning"):
content = self._extract_text_content(msg0.get("reasoning"))
reasoning_content = msg0.get("reasoning_content")
if not reasoning_content and msg0.get("reasoning"):
reasoning_content = self._extract_text_content(msg0.get("reasoning"))
for ch in choices:
ch_map = self._maybe_mapping(ch) or {}
m = self._maybe_mapping(ch_map.get("message")) or {}
@@ -412,6 +613,8 @@ class OpenAICompatProvider(LLMProvider):
finish_reason = ch.finish_reason
if not content and m.content:
content = m.content
if not content and getattr(m, "reasoning", None):
content = m.reasoning
tool_calls = []
for tc in raw_tool_calls:
@@ -428,17 +631,22 @@ class OpenAICompatProvider(LLMProvider):
function_provider_specific_fields=fn_prov,
))
reasoning_content = getattr(msg, "reasoning_content", None) or None
if not reasoning_content and getattr(msg, "reasoning", None):
reasoning_content = msg.reasoning
return LLMResponse(
content=content,
tool_calls=tool_calls,
finish_reason=finish_reason or "stop",
usage=self._extract_usage(response),
reasoning_content=getattr(msg, "reasoning_content", None) or None,
reasoning_content=reasoning_content,
)
@classmethod
def _parse_chunks(cls, chunks: list[Any]) -> LLMResponse:
content_parts: list[str] = []
reasoning_parts: list[str] = []
tc_bufs: dict[int, dict[str, Any]] = {}
finish_reason = "stop"
usage: dict[str, int] = {}
@@ -492,6 +700,11 @@ class OpenAICompatProvider(LLMProvider):
text = cls._extract_text_content(delta.get("content"))
if text:
content_parts.append(text)
text = cls._extract_text_content(delta.get("reasoning_content"))
if not text:
text = cls._extract_text_content(delta.get("reasoning"))
if text:
reasoning_parts.append(text)
for idx, tc in enumerate(delta.get("tool_calls") or []):
_accum_tc(tc, idx)
usage = cls._extract_usage(chunk_map) or usage
@@ -506,6 +719,12 @@ class OpenAICompatProvider(LLMProvider):
delta = choice.delta
if delta and delta.content:
content_parts.append(delta.content)
if delta:
reasoning = getattr(delta, "reasoning_content", None)
if not reasoning:
reasoning = getattr(delta, "reasoning", None)
if reasoning:
reasoning_parts.append(reasoning)
for tc in (delta.tool_calls or []) if delta else []:
_accum_tc(tc, getattr(tc, "index", 0))
@@ -524,13 +743,76 @@ class OpenAICompatProvider(LLMProvider):
],
finish_reason=finish_reason,
usage=usage,
reasoning_content="".join(reasoning_parts) or None,
)
@classmethod
def _extract_error_metadata(cls, e: Exception) -> dict[str, Any]:
response = getattr(e, "response", None)
headers = getattr(response, "headers", None)
payload = (
getattr(e, "body", None)
or getattr(e, "doc", None)
or getattr(response, "text", None)
)
if payload is None and response is not None:
response_json = getattr(response, "json", None)
if callable(response_json):
try:
payload = response_json()
except Exception:
payload = None
error_type, error_code = LLMProvider._extract_error_type_code(payload)
status_code = getattr(e, "status_code", None)
if status_code is None and response is not None:
status_code = getattr(response, "status_code", None)
should_retry: bool | None = None
if headers is not None:
raw = headers.get("x-should-retry")
if isinstance(raw, str):
lowered = raw.strip().lower()
if lowered == "true":
should_retry = True
elif lowered == "false":
should_retry = False
error_kind: str | None = None
error_name = e.__class__.__name__.lower()
if "timeout" in error_name:
error_kind = "timeout"
elif "connection" in error_name:
error_kind = "connection"
return {
"error_status_code": int(status_code) if status_code is not None else None,
"error_kind": error_kind,
"error_type": error_type,
"error_code": error_code,
"error_retry_after_s": cls._extract_retry_after_from_headers(headers),
"error_should_retry": should_retry,
}
@staticmethod
def _handle_error(e: Exception) -> LLMResponse:
body = getattr(e, "doc", None) or getattr(getattr(e, "response", None), "text", None)
msg = f"Error: {body.strip()[:500]}" if body and body.strip() else f"Error calling LLM: {e}"
return LLMResponse(content=msg, finish_reason="error")
body = (
getattr(e, "doc", None)
or getattr(e, "body", None)
or getattr(getattr(e, "response", None), "text", None)
)
body_text = body if isinstance(body, str) else str(body) if body is not None else ""
msg = f"Error: {body_text.strip()[:500]}" if body_text.strip() else f"Error calling LLM: {e}"
response = getattr(e, "response", None)
retry_after = LLMProvider._extract_retry_after_from_headers(getattr(response, "headers", None))
if retry_after is None:
retry_after = LLMProvider._extract_retry_after(msg)
return LLMResponse(
content=msg,
finish_reason="error",
retry_after=retry_after,
**OpenAICompatProvider._extract_error_metadata(e),
)
# ------------------------------------------------------------------
# Public API
@@ -546,11 +828,22 @@ class OpenAICompatProvider(LLMProvider):
reasoning_effort: str | None = None,
tool_choice: str | dict[str, Any] | None = None,
) -> LLMResponse:
kwargs = self._build_kwargs(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
try:
if self._should_use_responses_api(model, reasoning_effort):
try:
body = self._build_responses_body(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
return parse_response_output(await self._client.responses.create(**body))
except Exception as responses_error:
if not self._should_fallback_from_responses_error(responses_error):
raise
kwargs = self._build_kwargs(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
return self._parse(await self._client.chat.completions.create(**kwargs))
except Exception as e:
return self._handle_error(e)
@@ -566,22 +859,75 @@ class OpenAICompatProvider(LLMProvider):
tool_choice: str | dict[str, Any] | None = None,
on_content_delta: Callable[[str], Awaitable[None]] | None = None,
) -> LLMResponse:
kwargs = self._build_kwargs(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
kwargs["stream"] = True
kwargs["stream_options"] = {"include_usage": True}
idle_timeout_s = int(os.environ.get("NANOBOT_STREAM_IDLE_TIMEOUT_S", "90"))
try:
if self._should_use_responses_api(model, reasoning_effort):
try:
body = self._build_responses_body(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
body["stream"] = True
stream = await self._client.responses.create(**body)
async def _timed_stream():
stream_iter = stream.__aiter__()
while True:
try:
yield await asyncio.wait_for(
stream_iter.__anext__(),
timeout=idle_timeout_s,
)
except StopAsyncIteration:
break
content, tool_calls, finish_reason, usage, reasoning_content = await consume_sdk_stream(
_timed_stream(),
on_content_delta,
)
return LLMResponse(
content=content or None,
tool_calls=tool_calls,
finish_reason=finish_reason,
usage=usage,
reasoning_content=reasoning_content,
)
except Exception as responses_error:
if not self._should_fallback_from_responses_error(responses_error):
raise
kwargs = self._build_kwargs(
messages, tools, model, max_tokens, temperature,
reasoning_effort, tool_choice,
)
kwargs["stream"] = True
kwargs["stream_options"] = {"include_usage": True}
stream = await self._client.chat.completions.create(**kwargs)
chunks: list[Any] = []
async for chunk in stream:
stream_iter = stream.__aiter__()
while True:
try:
chunk = await asyncio.wait_for(
stream_iter.__anext__(),
timeout=idle_timeout_s,
)
except StopAsyncIteration:
break
chunks.append(chunk)
if on_content_delta and chunk.choices:
text = getattr(chunk.choices[0].delta, "content", None)
if text:
await on_content_delta(text)
return self._parse_chunks(chunks)
except asyncio.TimeoutError:
return LLMResponse(
content=(
f"Error calling LLM: stream stalled for more than "
f"{idle_timeout_s} seconds"
),
finish_reason="error",
error_kind="timeout",
)
except Exception as e:
return self._handle_error(e)
nanobot/providers/openai_responses/__init__.py
+29
View File
@@ -0,0 +1,29 @@
"""Shared helpers for OpenAI Responses API providers (Codex, Azure OpenAI)."""
from nanobot.providers.openai_responses.converters import (
convert_messages,
convert_tools,
convert_user_message,
split_tool_call_id,
)
from nanobot.providers.openai_responses.parsing import (
FINISH_REASON_MAP,
consume_sdk_stream,
consume_sse,
iter_sse,
map_finish_reason,
parse_response_output,
)
__all__ = [
"convert_messages",
"convert_tools",
"convert_user_message",
"split_tool_call_id",
"iter_sse",
"consume_sse",
"consume_sdk_stream",
"map_finish_reason",
"parse_response_output",
"FINISH_REASON_MAP",
]
nanobot/providers/openai_responses/converters.py
+110
View File
@@ -0,0 +1,110 @@
"""Convert Chat Completions messages/tools to Responses API format."""
from __future__ import annotations
import json
from typing import Any
def convert_messages(messages: list[dict[str, Any]]) -> tuple[str, list[dict[str, Any]]]:
"""Convert Chat Completions messages to Responses API input items.
Returns ``(system_prompt, input_items)`` where *system_prompt* is extracted
from any ``system`` role message and *input_items* is the Responses API
``input`` array.
"""
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":
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}",
})
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"))
input_items.append({
"type": "function_call",
"id": item_id or f"fc_{idx}",
"call_id": call_id or f"call_{idx}",
"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})
return system_prompt, input_items
def convert_user_message(content: Any) -> dict[str, Any]:
"""Convert a user message's content to Responses API format.
Handles plain strings, ``text`` blocks -> ``input_text``, and
``image_url`` blocks -> ``input_image``.
"""
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 convert_tools(tools: list[dict[str, Any]]) -> list[dict[str, Any]]:
"""Convert OpenAI function-calling tool schema to Responses API 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 split_tool_call_id(tool_call_id: Any) -> tuple[str, str | None]:
"""Split a compound ``call_id|item_id`` string.
Returns ``(call_id, item_id)`` where *item_id* may be ``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
nanobot/providers/openai_responses/parsing.py
+297
View File
@@ -0,0 +1,297 @@
"""Parse Responses API SSE streams and SDK response objects."""
from __future__ import annotations
import json
from collections.abc import Awaitable, Callable
from typing import Any, AsyncGenerator
import httpx
import json_repair
from loguru import logger
from nanobot.providers.base import LLMResponse, ToolCallRequest
FINISH_REASON_MAP = {
"completed": "stop",
"incomplete": "length",
"failed": "error",
"cancelled": "error",
}
def map_finish_reason(status: str | None) -> str:
"""Map a Responses API status string to a Chat-Completions-style finish_reason."""
return FINISH_REASON_MAP.get(status or "completed", "stop")
async def iter_sse(response: httpx.Response) -> AsyncGenerator[dict[str, Any], None]:
"""Yield parsed JSON events from a Responses API SSE stream."""
buffer: list[str] = []
def _flush() -> dict[str, Any] | None:
data_lines = [l[5:].strip() for l in buffer if l.startswith("data:")]
buffer.clear()
if not data_lines:
return None
data = "\n".join(data_lines).strip()
if not data or data == "[DONE]":
return None
try:
return json.loads(data)
except Exception:
logger.warning("Failed to parse SSE event JSON: {}", data[:200])
return None
async for line in response.aiter_lines():
if line == "":
if buffer:
event = _flush()
if event is not None:
yield event
continue
buffer.append(line)
# Flush any remaining buffer at EOF (#10)
if buffer:
event = _flush()
if event is not None:
yield event
async def consume_sse(
response: httpx.Response,
on_content_delta: Callable[[str], Awaitable[None]] | None = None,
) -> tuple[str, list[ToolCallRequest], str]:
"""Consume a Responses API SSE stream into ``(content, tool_calls, finish_reason)``."""
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":
delta_text = event.get("delta") or ""
content += delta_text
if on_content_delta and delta_text:
await on_content_delta(delta_text)
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:
logger.warning(
"Failed to parse tool call arguments for '{}': {}",
buf.get("name") or item.get("name"),
args_raw[:200],
)
args = json_repair.loads(args_raw)
if not isinstance(args, dict):
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") or "",
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"}:
detail = event.get("error") or event.get("message") or event
raise RuntimeError(f"Response failed: {str(detail)[:500]}")
return content, tool_calls, finish_reason
def parse_response_output(response: Any) -> LLMResponse:
"""Parse an SDK ``Response`` object into an ``LLMResponse``."""
if not isinstance(response, dict):
dump = getattr(response, "model_dump", None)
response = dump() if callable(dump) else vars(response)
output = response.get("output") or []
content_parts: list[str] = []
tool_calls: list[ToolCallRequest] = []
reasoning_content: str | None = None
for item in output:
if not isinstance(item, dict):
dump = getattr(item, "model_dump", None)
item = dump() if callable(dump) else vars(item)
item_type = item.get("type")
if item_type == "message":
for block in item.get("content") or []:
if not isinstance(block, dict):
dump = getattr(block, "model_dump", None)
block = dump() if callable(dump) else vars(block)
if block.get("type") == "output_text":
content_parts.append(block.get("text") or "")
elif item_type == "reasoning":
for s in item.get("summary") or []:
if not isinstance(s, dict):
dump = getattr(s, "model_dump", None)
s = dump() if callable(dump) else vars(s)
if s.get("type") == "summary_text" and s.get("text"):
reasoning_content = (reasoning_content or "") + s["text"]
elif item_type == "function_call":
call_id = item.get("call_id") or ""
item_id = item.get("id") or "fc_0"
args_raw = item.get("arguments") or "{}"
try:
args = json.loads(args_raw) if isinstance(args_raw, str) else args_raw
except Exception:
logger.warning(
"Failed to parse tool call arguments for '{}': {}",
item.get("name"),
str(args_raw)[:200],
)
args = json_repair.loads(args_raw) if isinstance(args_raw, str) else args_raw
if not isinstance(args, dict):
args = {"raw": args_raw}
tool_calls.append(ToolCallRequest(
id=f"{call_id}|{item_id}",
name=item.get("name") or "",
arguments=args if isinstance(args, dict) else {},
))
usage_raw = response.get("usage") or {}
if not isinstance(usage_raw, dict):
dump = getattr(usage_raw, "model_dump", None)
usage_raw = dump() if callable(dump) else vars(usage_raw)
usage = {}
if usage_raw:
usage = {
"prompt_tokens": int(usage_raw.get("input_tokens") or 0),
"completion_tokens": int(usage_raw.get("output_tokens") or 0),
"total_tokens": int(usage_raw.get("total_tokens") or 0),
}
status = response.get("status")
finish_reason = map_finish_reason(status)
return LLMResponse(
content="".join(content_parts) or None,
tool_calls=tool_calls,
finish_reason=finish_reason,
usage=usage,
reasoning_content=reasoning_content if isinstance(reasoning_content, str) else None,
)
async def consume_sdk_stream(
stream: Any,
on_content_delta: Callable[[str], Awaitable[None]] | None = None,
) -> tuple[str, list[ToolCallRequest], str, dict[str, int], str | None]:
"""Consume an SDK async stream from ``client.responses.create(stream=True)``."""
content = ""
tool_calls: list[ToolCallRequest] = []
tool_call_buffers: dict[str, dict[str, Any]] = {}
finish_reason = "stop"
usage: dict[str, int] = {}
reasoning_content: str | None = None
async for event in stream:
event_type = getattr(event, "type", None)
if event_type == "response.output_item.added":
item = getattr(event, "item", None)
if item and getattr(item, "type", None) == "function_call":
call_id = getattr(item, "call_id", None)
if not call_id:
continue
tool_call_buffers[call_id] = {
"id": getattr(item, "id", None) or "fc_0",
"name": getattr(item, "name", None),
"arguments": getattr(item, "arguments", None) or "",
}
elif event_type == "response.output_text.delta":
delta_text = getattr(event, "delta", "") or ""
content += delta_text
if on_content_delta and delta_text:
await on_content_delta(delta_text)
elif event_type == "response.function_call_arguments.delta":
call_id = getattr(event, "call_id", None)
if call_id and call_id in tool_call_buffers:
tool_call_buffers[call_id]["arguments"] += getattr(event, "delta", "") or ""
elif event_type == "response.function_call_arguments.done":
call_id = getattr(event, "call_id", None)
if call_id and call_id in tool_call_buffers:
tool_call_buffers[call_id]["arguments"] = getattr(event, "arguments", "") or ""
elif event_type == "response.output_item.done":
item = getattr(event, "item", None)
if item and getattr(item, "type", None) == "function_call":
call_id = getattr(item, "call_id", None)
if not call_id:
continue
buf = tool_call_buffers.get(call_id) or {}
args_raw = buf.get("arguments") or getattr(item, "arguments", None) or "{}"
try:
args = json.loads(args_raw)
except Exception:
logger.warning(
"Failed to parse tool call arguments for '{}': {}",
buf.get("name") or getattr(item, "name", None),
str(args_raw)[:200],
)
args = json_repair.loads(args_raw)
if not isinstance(args, dict):
args = {"raw": args_raw}
tool_calls.append(
ToolCallRequest(
id=f"{call_id}|{buf.get('id') or getattr(item, 'id', None) or 'fc_0'}",
name=buf.get("name") or getattr(item, "name", None) or "",
arguments=args,
)
)
elif event_type == "response.completed":
resp = getattr(event, "response", None)
status = getattr(resp, "status", None) if resp else None
finish_reason = map_finish_reason(status)
if resp:
usage_obj = getattr(resp, "usage", None)
if usage_obj:
usage = {
"prompt_tokens": int(getattr(usage_obj, "input_tokens", 0) or 0),
"completion_tokens": int(getattr(usage_obj, "output_tokens", 0) or 0),
"total_tokens": int(getattr(usage_obj, "total_tokens", 0) or 0),
}
for out_item in getattr(resp, "output", None) or []:
if getattr(out_item, "type", None) == "reasoning":
for s in getattr(out_item, "summary", None) or []:
if getattr(s, "type", None) == "summary_text":
text = getattr(s, "text", None)
if text:
reasoning_content = (reasoning_content or "") + text
elif event_type in {"error", "response.failed"}:
detail = getattr(event, "error", None) or getattr(event, "message", None) or event
raise RuntimeError(f"Response failed: {str(detail)[:500]}")
return content, tool_calls, finish_reason, usage, reasoning_content
nanobot/providers/registry.py
+22 -2
View File
@@ -34,7 +34,7 @@ class ProviderSpec:
display_name: str = "" # shown in `nanobot status`
# which provider implementation to use
# "openai_compat" | "anthropic" | "azure_openai" | "openai_codex"
# "openai_compat" | "anthropic" | "azure_openai" | "openai_codex" | "github_copilot"
backend: str = "openai_compat"
# extra env vars, e.g. (("ZHIPUAI_API_KEY", "{api_key}"),)
@@ -200,6 +200,7 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
env_key="OPENAI_API_KEY",
display_name="OpenAI",
backend="openai_compat",
supports_max_completion_tokens=True,
),
# OpenAI Codex: OAuth-based, dedicated provider
ProviderSpec(
@@ -218,8 +219,9 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
keywords=("github_copilot", "copilot"),
env_key="",
display_name="Github Copilot",
backend="openai_compat",
backend="github_copilot",
default_api_base="https://api.githubcopilot.com",
strip_model_prefix=True,
is_oauth=True,
),
# DeepSeek: OpenAI-compatible at api.deepseek.com
@@ -296,6 +298,15 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
backend="openai_compat",
default_api_base="https://api.stepfun.com/v1",
),
# Xiaomi MIMO (小米): OpenAI-compatible API
ProviderSpec(
name="xiaomi_mimo",
keywords=("xiaomi_mimo", "mimo"),
env_key="XIAOMIMIMO_API_KEY",
display_name="Xiaomi MIMO",
backend="openai_compat",
default_api_base="https://api.xiaomimimo.com/v1",
),
# === Local deployment (matched by config key, NOT by api_base) =========
# vLLM / any OpenAI-compatible local server
ProviderSpec(
@@ -338,6 +349,15 @@ PROVIDERS: tuple[ProviderSpec, ...] = (
backend="openai_compat",
default_api_base="https://api.groq.com/openai/v1",
),
# Qianfan (百度千帆): OpenAI-compatible API
ProviderSpec(
name="qianfan",
keywords=("qianfan", "ernie"),
env_key="QIANFAN_API_KEY",
display_name="Qianfan",
backend="openai_compat",
default_api_base="https://qianfan.baidubce.com/v2"
),
)
nanobot/providers/transcription.py
+31 -1
View File
@@ -1,4 +1,4 @@
"""Voice transcription provider using Groq."""
"""Voice transcription providers (Groq and OpenAI Whisper)."""
import os
from pathlib import Path
@@ -7,6 +7,36 @@ import httpx
from loguru import logger
class OpenAITranscriptionProvider:
"""Voice transcription provider using OpenAI's Whisper API."""
def __init__(self, api_key: str | None = None):
self.api_key = api_key or os.environ.get("OPENAI_API_KEY")
self.api_url = "https://api.openai.com/v1/audio/transcriptions"
async def transcribe(self, file_path: str | Path) -> str:
if not self.api_key:
logger.warning("OpenAI 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-1")}
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()
return response.json().get("text", "")
except Exception as e:
logger.error("OpenAI transcription error: {}", e)
return ""
class GroqTranscriptionProvider:
"""
Voice transcription provider using Groq's Whisper API.