fix(openai): keep Astra rules on eligible routes

(cherry picked from commit f92fb9d9964e6e8ea118035c548aae812054f504)

agent/auxiliary_client.py

@@ -1380,8 +1380,13 @@ class _CodexCompletionsAdapter:
             if isinstance(reasoning_cfg, dict) and reasoning_cfg.get("enabled") is not False:
                 # Truthy-only: Codex 400s on e.g. {"effort": null}, so falsy → default. Shared
                 # per-model clamp with the main transport ("max" is gpt-5.6-only; "minimal"/"ultra" rejected).
-                from agent.reasoning_effort import clamp_effort, codex_supported_efforts
-                effort = clamp_effort(reasoning_cfg.get("effort") or "medium", codex_supported_efforts(model))
+                from agent.reasoning_effort import clamp_effort
+                from agent.transports.codex import _codex_efforts_for_route
+                is_codex_backend = base_url_host_matches(host, "chatgpt.com") and "/backend-api/codex" in host.lower()
+                effort = clamp_effort(
+                    reasoning_cfg.get("effort") or "medium",
+                    _codex_efforts_for_route(model, host, is_codex_backend=is_codex_backend),
+                )
                 resp_kwargs["reasoning"] = {"effort": effort, "summary": "auto"}
                 resp_kwargs["include"] = ["reasoning.encrypted_content"]
         tools = kwargs.get("tools")

agent/transports/codex.py

@@ -11,7 +11,8 @@ import re
 from typing import Any, Callable, Optional
 
 from agent.reasoning_effort import (
-    ACTUAL_RELAY_EFFORTS, CODEX_ASTRA_EFFORTS, XAI_GROK46_EFFORTS, XAI_LEGACY_EFFORTS, clamp_effort,
+    ACTUAL_RELAY_EFFORTS, CODEX_ASTRA_EFFORTS, CODEX_LEGACY_EFFORTS,
+    XAI_GROK46_EFFORTS, XAI_LEGACY_EFFORTS, clamp_effort,
     # Same declared vocabulary + shared clamp as the main Codex transport (agent.reasoning_effort):
     # per-model — "max" is gpt-5.6-only, "minimal"/"ultra" always rejected (live-verified, #68365).
     codex_supported_efforts,
@@ -226,7 +227,9 @@ def _resolve_reasoning(model: str, params: dict[str, Any]) -> tuple[Any, bool]:
         declared = _profile_declared_efforts(params.get("provider"), model, params.get("base_url"))
         if declared is not None and not declared:
             reasoning_enabled = False
-        supported = declared or codex_supported_efforts(model)
+        supported = declared or _codex_efforts_for_route(
+            model, params.get("base_url"), is_codex_backend=params.get("is_codex_backend") is True
+        )
     return clamp_effort(reasoning_effort, supported), reasoning_enabled
 
 
@@ -275,6 +278,15 @@ def _is_official_openai_responses_route(model: Any, base_url: Any) -> bool:
     return base_url_host_matches(str(base_url or ""), "api.openai.com")
 
 
+def _codex_efforts_for_route(model: Any, base_url: Any, *, is_codex_backend: bool = False) -> tuple[str, ...]:
+    """Keep Astra's new vocabulary off unrelated Responses-compatible endpoints."""
+    if _is_astra_model(model) and not (
+        is_codex_backend or _is_official_openai_responses_route(model, base_url)
+    ):
+        return CODEX_LEGACY_EFFORTS
+    return codex_supported_efforts(str(model or ""))
+
+
 def _sanitize_astra_request_kwargs(kwargs: dict[str, Any], model: Any, base_url: Any) -> None:
     """Apply Astra's model-specific restrictions after all request overrides are merged."""
     if not _is_astra_model(model):

tests/agent/test_auxiliary_client.py

@@ -3197,6 +3197,18 @@ class TestCodexAdapterPromptCacheKey:
         assert captured["prompt_cache_options"] == {"ttl": "30m"}
         assert "prompt_cache_retention" not in captured
 
+    def test_astra_auxiliary_proxy_keeps_legacy_effort_contract(self):
+        adapter, captured = self._build_adapter(
+            base_url="https://responses.example.com/v1",
+            model="gpt-6-astra",
+        )
+        adapter.create(
+            messages=[{"role": "user", "content": "hi"}],
+            extra_body={"reasoning": {"effort": "none"}},
+        )
+        assert captured["reasoning"]["effort"] == "none"
+        assert "prompt_cache_options" not in captured
+
     def test_codex_backend_forwards_auxiliary_service_tier(self):
         adapter, captured = self._build_adapter(
             base_url="https://chatgpt.com/backend-api/codex",

tests/agent/transports/test_codex_transport.py

@@ -96,10 +96,12 @@ class TestCodexBuildKwargs:
             messages=[{"role": "user", "content": "Hi"}],
             tools=[],
             base_url="https://responses.example.com/v1",
+            reasoning_config={"enabled": True, "effort": "none"},
             request_overrides={"prompt_cache_options": {"ttl": "30m"}},
         )
 
         assert "prompt_cache_options" not in kw
+        assert kw["reasoning"]["effort"] == "none"
 
     def test_900k_context_variant_suffix_stripped_on_wire(self, transport):
         """``-900k`` large-context picker variants are Hermes-side aliases —