feat: 新增 Codex 上游兼容整流器

- 强制 store=false
- 确保 instructions 字段存在（缺失/None -> ""）
- 强制 parallel_tool_calls=true，并确保 include 包含 reasoning.encrypted_content
- 删除 Codex 会拒绝的一些字段：max_output_tokens/max_completion_tokens/max_tokens/temperature/top_p/service_tier
- 将 input[] 里 role=system 的 message 改成 developer
- 在发送上游请求前注入整流（覆盖“同格式透传”和“跨格式转换”两条链路）

src/services/provider/codex.py

@@ -0,0 +1,106 @@
+"""
+Codex upstream request compatibility helpers.
+
+The Codex upstream (https://chatgpt.com/backend-api/codex) is largely compatible with the
+OpenAI Responses (/responses, aka "openai:cli") schema, but enforces some extra constraints.
+
+CLIProxyAPI's reference implementation applies a small set of mutations before forwarding.
+We replicate the same mutations here to keep Aether's routing compatible when
+Provider.provider_type == "codex".
+"""
+
+from __future__ import annotations
+
+from typing import Any
+
+_CODEX_REQUIRED_INCLUDE_ITEM = "reasoning.encrypted_content"
+
+
+def patch_openai_cli_request_for_codex(request_body: dict[str, Any]) -> dict[str, Any]:
+    """
+    Mutate an OpenAI Responses (openai:cli) request into a Codex-compatible payload.
+
+    Notes (based on CLIProxyAPI translators):
+    - `store` must be explicitly set to false.
+    - `instructions` field must exist (Codex rejects missing instructions).
+    - Codex rejects several generation params, so strip them.
+    - Codex does not accept `system` role inside the `input` array.
+    - Enable `parallel_tool_calls` and request encrypted reasoning content.
+    """
+    if not isinstance(request_body, dict):
+        return request_body
+
+    result: dict[str, Any] = dict(request_body)
+
+    # Required by Codex: explicitly disable storing.
+    result["store"] = False
+
+    # Required by Codex: ensure instructions exists (can be empty).
+    instructions = result.get("instructions")
+    if instructions is None:
+        result["instructions"] = ""
+    elif not isinstance(instructions, str):
+        result["instructions"] = str(instructions)
+
+    # Codex defaults/tooling expectations
+    result["parallel_tool_calls"] = True
+
+    include_value = result.get("include")
+    include: list[str] = []
+    if isinstance(include_value, list):
+        include = [v for v in include_value if isinstance(v, str) and v]
+    if _CODEX_REQUIRED_INCLUDE_ITEM not in include:
+        include.append(_CODEX_REQUIRED_INCLUDE_ITEM)
+    result["include"] = include
+
+    # Codex Responses rejects token limit fields and some sampling params.
+    for key in (
+        "max_output_tokens",
+        "max_completion_tokens",
+        "max_tokens",
+        "temperature",
+        "top_p",
+        "service_tier",
+    ):
+        result.pop(key, None)
+
+    # Convert role "system" to "developer" in input array to comply with Codex API requirements.
+    input_value = result.get("input")
+    if isinstance(input_value, list):
+        patched_input: list[Any] = []
+        for item in input_value:
+            if (
+                isinstance(item, dict)
+                and item.get("type") == "message"
+                and item.get("role") == "system"
+            ):
+                item = dict(item)
+                item["role"] = "developer"
+            patched_input.append(item)
+        result["input"] = patched_input
+
+    return result
+
+
+def maybe_patch_request_for_codex(
+    *,
+    provider_type: str | None,
+    provider_api_format: str | None,
+    request_body: dict[str, Any],
+) -> dict[str, Any]:
+    """
+    Apply Codex compatibility patches only when the selected upstream is Codex and the
+    endpoint uses the OpenAI Responses schema ("openai:cli").
+    """
+    if str(provider_type or "").strip().lower() != "codex":
+        return request_body
+    if str(provider_api_format or "").strip().lower() != "openai:cli":
+        return request_body
+    return patch_openai_cli_request_for_codex(request_body)
+
+
+__all__ = [
+    "maybe_patch_request_for_codex",
+    "patch_openai_cli_request_for_codex",
+]
+