fix(codex): clamp reasoning.effort max/ultra to xhigh (HTTP 400)

agent/transports/codex.py

@@ -221,10 +221,13 @@ def build_kwargs(
             elif reasoning_config.get("effort"):
                 reasoning_effort = reasoning_config["effort"]
 
-        _effort_clamp = {"minimal": "low"}
-        if "gpt-5.6" in (model or "").lower():
-            # Ultra is the Codex product tier; the Responses API wire value is max.
-            _effort_clamp["ultra"] = "max"
+        # Clamp Hermes-generic effort levels onto the wire vocabulary for this
+        # surface. Codex/OpenAI Responses accepts none|minimal|low|medium|high|xhigh
+        # (observed 2026-07: effort=max returns HTTP 400 invalid_value). xAI
+        # Responses tops out at high. "minimal" is always promoted to "low" for
+        # backends that never learned the minimal tier.
+        reasoning_effort = str(reasoning_effort or "medium").strip().lower()
+        _effort_clamp = {"minimal": "low", "max": "xhigh", "ultra": "xhigh"}
         if params.get("is_xai_responses", False):
             # xAI Responses tops out at high; keep generic stronger values usable.
             _effort_clamp.update({"xhigh": "high", "max": "high", "ultra": "high"})

tests/agent/transports/test_codex_transport.py

@@ -38,12 +38,33 @@ def test_convert_tools(self, transport):
 
 
 class TestCodexBuildKwargs:
+    @pytest.mark.parametrize("effort, wire_effort", [("max", "xhigh"), ("ultra", "xhigh")])
+    def test_extended_reasoning_efforts_use_api_wire_value(self, transport, effort, wire_effort):
+        kw = transport.build_kwargs(
+            model="gpt-5.6-sol",
+            messages=[{"role": "user", "content": "Hi"}],
+            tools=[],
+            reasoning_config={"enabled": True, "effort": effort},
+        )
+        assert kw.get("reasoning", {}).get("effort") == wire_effort
 
+    def test_max_effort_clamped_for_non_56_codex_models(self, transport):
+        """Config effort=max must not reach Codex wire as max (HTTP 400)."""
+        kw = transport.build_kwargs(
+            model="gpt-5.5",
+            messages=[{"role": "user", "content": "Hi"}],
+            tools=[],
+            reasoning_config={"enabled": True, "effort": "max"},
+        )
+        assert kw.get("reasoning", {}).get("effort") == "xhigh"
 
-
-
-
-
+    def test_reasoning_disabled(self, transport):
+        messages = [{"role": "user", "content": "Hi"}]
+        kw = transport.build_kwargs(
+            model="gpt-5.4", messages=messages, tools=[],
+            reasoning_config={"enabled": False},
+        )
+        assert "reasoning" not in kw or kw.get("include") == []
 
     def test_cache_key_is_content_addressed_not_session_id(self, transport):
         """prompt_cache_key is content-addressed from the static prefix

tests/run_agent/test_run_agent_codex_responses.py

@@ -323,8 +323,55 @@ def test_build_api_kwargs_mantle_sets_extended_prompt_cache_retention(monkeypatc
 
     assert kwargs["prompt_cache_retention"] == "24h"
 
+def test_build_api_kwargs_codex_preserves_supported_efforts(monkeypatch):
+    """Wire-supported effort levels pass through; Hermes max/ultra clamp to xhigh."""
+    _patch_agent_bootstrap(monkeypatch)
+
+    for effort in ("low", "medium", "high", "xhigh"):
+        agent = run_agent.AIAgent(
+            model="gpt-5-codex",
+            base_url="https://chatgpt.com/backend-api/codex",
+            api_key="codex-token",
+            quiet_mode=True,
+            max_iterations=4,
+            skip_context_files=True,
+            skip_memory=True,
+            reasoning_config={"enabled": True, "effort": effort},
+        )
+        agent._cleanup_task_resources = lambda task_id: None
+        agent._persist_session = lambda messages, history=None: None
+        agent._save_trajectory = lambda messages, user_message, completed: None
+
+        kwargs = agent._build_api_kwargs(
+            [
+                {"role": "system", "content": "sys"},
+                {"role": "user", "content": "hi"},
+            ]
+        )
+        assert kwargs["reasoning"]["effort"] == effort, f"{effort} should pass through unchanged"
 
+    for effort, wire in (("max", "xhigh"), ("ultra", "xhigh")):
+        agent = run_agent.AIAgent(
+            model="gpt-5-codex",
+            base_url="https://chatgpt.com/backend-api/codex",
+            api_key="codex-token",
+            quiet_mode=True,
+            max_iterations=4,
+            skip_context_files=True,
+            skip_memory=True,
+            reasoning_config={"enabled": True, "effort": effort},
+        )
+        agent._cleanup_task_resources = lambda task_id: None
+        agent._persist_session = lambda messages, history=None: None
+        agent._save_trajectory = lambda messages, user_message, completed: None
 
+        kwargs = agent._build_api_kwargs(
+            [
+                {"role": "system", "content": "sys"},
+                {"role": "user", "content": "hi"},
+            ]
+        )
+        assert kwargs["reasoning"]["effort"] == wire, f"{effort} should clamp to {wire}"
 
 
 
@@ -1740,4 +1787,3 @@ def test_duplicate_detection_uses_commentary_when_hidden_reasoning_changes(monke
 
 
 
-