| 1 | """OpenAI Codex Responses Provider.""" |
| 2 | |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | import asyncio |
| 6 | import hashlib |
| 7 | import json |
| 8 | from collections.abc import Awaitable, Callable |
| 9 | from typing import Any |
| 10 | |
| 11 | import httpx |
| 12 | from loguru import logger |
| 13 | from oauth_cli_kit import get_token as get_codex_token |
| 14 | |
| 15 | from nanobot.providers.base import LLMProvider, LLMResponse, ToolCallRequest |
| 16 | from nanobot.providers.openai_responses import ( |
| 17 | consume_sse, |
| 18 | convert_messages, |
| 19 | convert_tools, |
| 20 | ) |
| 21 | |
| 22 | DEFAULT_CODEX_URL = "https://chatgpt.com/backend-api/codex/responses" |
| 23 | DEFAULT_ORIGINATOR = "nanobot" |
| 24 | |
| 25 | |
| 26 | class OpenAICodexProvider(LLMProvider): |
| 27 | """Use Codex OAuth to call the Responses API.""" |
| 28 | |
| 29 | def __init__(self, default_model: str = "openai-codex/gpt-5.1-codex"): |
| 30 | super().__init__(api_key=None, api_base=None) |
| 31 | self.default_model = default_model |
| 32 | |
| 33 | async def _call_codex( |
| 34 | self, |
| 35 | messages: list[dict[str, Any]], |
| 36 | tools: list[dict[str, Any]] | None, |
| 37 | model: str | None, |
| 38 | reasoning_effort: str | None, |
| 39 | tool_choice: str | dict[str, Any] | None, |
| 40 | on_content_delta: Callable[[str], Awaitable[None]] | None = None, |
| 41 | ) -> LLMResponse: |
| 42 | """Shared request logic for both chat() and chat_stream().""" |
| 43 | model = model or self.default_model |
| 44 | system_prompt, input_items = convert_messages(messages) |
| 45 | |
| 46 | token = await asyncio.to_thread(get_codex_token) |
| 47 | headers = _build_headers(token.account_id, token.access) |
| 48 | |
| 49 | body: dict[str, Any] = { |
| 50 | "model": _strip_model_prefix(model), |
| 51 | "store": False, |
| 52 | "stream": True, |
| 53 | "instructions": system_prompt, |
| 54 | "input": input_items, |
| 55 | "text": {"verbosity": "medium"}, |
| 56 | "include": ["reasoning.encrypted_content"], |
| 57 | "prompt_cache_key": _prompt_cache_key(messages), |
| 58 | "tool_choice": tool_choice or "auto", |
| 59 | "parallel_tool_calls": True, |
| 60 | } |
| 61 | if reasoning_effort: |
| 62 | body["reasoning"] = {"effort": reasoning_effort} |
| 63 | if tools: |
| 64 | body["tools"] = convert_tools(tools) |
| 65 | |
| 66 | try: |
| 67 | try: |
| 68 | content, tool_calls, finish_reason = await _request_codex( |
| 69 | DEFAULT_CODEX_URL, headers, body, verify=True, |
| 70 | on_content_delta=on_content_delta, |
| 71 | ) |
| 72 | except Exception as e: |
| 73 | if "CERTIFICATE_VERIFY_FAILED" not in str(e): |
| 74 | raise |
| 75 | logger.warning("SSL verification failed for Codex API; retrying with verify=False") |
| 76 | content, tool_calls, finish_reason = await _request_codex( |
| 77 | DEFAULT_CODEX_URL, headers, body, verify=False, |
| 78 | on_content_delta=on_content_delta, |
| 79 | ) |
| 80 | return LLMResponse(content=content, tool_calls=tool_calls, finish_reason=finish_reason) |
| 81 | except Exception as e: |
| 82 | msg = f"Error calling Codex: {e}" |
| 83 | retry_after = getattr(e, "retry_after", None) or self._extract_retry_after(msg) |
| 84 | return LLMResponse(content=msg, finish_reason="error", retry_after=retry_after) |
| 85 | |
| 86 | async def chat( |
| 87 | self, messages: list[dict[str, Any]], tools: list[dict[str, Any]] | None = None, |
| 88 | model: str | None = None, max_tokens: int = 4096, temperature: float = 0.7, |
| 89 | reasoning_effort: str | None = None, |
| 90 | tool_choice: str | dict[str, Any] | None = None, |
| 91 | ) -> LLMResponse: |
| 92 | return await self._call_codex(messages, tools, model, reasoning_effort, tool_choice) |
| 93 | |
| 94 | async def chat_stream( |
| 95 | self, messages: list[dict[str, Any]], tools: list[dict[str, Any]] | None = None, |
| 96 | model: str | None = None, max_tokens: int = 4096, temperature: float = 0.7, |
| 97 | reasoning_effort: str | None = None, |
| 98 | tool_choice: str | dict[str, Any] | None = None, |
| 99 | on_content_delta: Callable[[str], Awaitable[None]] | None = None, |
| 100 | ) -> LLMResponse: |
| 101 | return await self._call_codex(messages, tools, model, reasoning_effort, tool_choice, on_content_delta) |
| 102 | |
| 103 | def get_default_model(self) -> str: |
| 104 | return self.default_model |
| 105 | |
| 106 | |
| 107 | def _strip_model_prefix(model: str) -> str: |
| 108 | if model.startswith("openai-codex/") or model.startswith("openai_codex/"): |
| 109 | return model.split("/", 1)[1] |
| 110 | return model |
| 111 | |
| 112 | |
| 113 | def _build_headers(account_id: str, token: str) -> dict[str, str]: |
| 114 | return { |
| 115 | "Authorization": f"Bearer {token}", |
| 116 | "chatgpt-account-id": account_id, |
| 117 | "OpenAI-Beta": "responses=experimental", |
| 118 | "originator": DEFAULT_ORIGINATOR, |
| 119 | "User-Agent": "nanobot (python)", |
| 120 | "accept": "text/event-stream", |
| 121 | "content-type": "application/json", |
| 122 | } |
| 123 | |
| 124 | |
| 125 | class _CodexHTTPError(RuntimeError): |
| 126 | def __init__(self, message: str, retry_after: float | None = None): |
| 127 | super().__init__(message) |
| 128 | self.retry_after = retry_after |
| 129 | |
| 130 | |
| 131 | async def _request_codex( |
| 132 | url: str, |
| 133 | headers: dict[str, str], |
| 134 | body: dict[str, Any], |
| 135 | verify: bool, |
| 136 | on_content_delta: Callable[[str], Awaitable[None]] | None = None, |
| 137 | ) -> tuple[str, list[ToolCallRequest], str]: |
| 138 | async with httpx.AsyncClient(timeout=60.0, verify=verify) as client: |
| 139 | async with client.stream("POST", url, headers=headers, json=body) as response: |
| 140 | if response.status_code != 200: |
| 141 | text = await response.aread() |
| 142 | retry_after = LLMProvider._extract_retry_after_from_headers(response.headers) |
| 143 | raise _CodexHTTPError( |
| 144 | _friendly_error(response.status_code, text.decode("utf-8", "ignore")), |
| 145 | retry_after=retry_after, |
| 146 | ) |
| 147 | return await consume_sse(response, on_content_delta) |
| 148 | |
| 149 | |
| 150 | def _prompt_cache_key(messages: list[dict[str, Any]]) -> str: |
| 151 | raw = json.dumps(messages, ensure_ascii=True, sort_keys=True) |
| 152 | return hashlib.sha256(raw.encode("utf-8")).hexdigest() |
| 153 | |
| 154 | |
| 155 | def _friendly_error(status_code: int, raw: str) -> str: |
| 156 | if status_code == 429: |
| 157 | return "ChatGPT usage quota exceeded or rate limit triggered. Please try again later." |
| 158 | return f"HTTP {status_code}: {raw}" |
| 159 |