From e7bb67332d758c7093e58efa5caa30fef96051b2 Mon Sep 17 00:00:00 2001
From: Gille <4317663+helix4u@users.noreply.github.com>
Date: Sat, 27 Jun 2026 14:31:24 -0600
Subject: [PATCH] fix(moa): preserve Codex slot routing

---
 agent/moa_loop.py                     |  7 +++++++
 tests/run_agent/test_moa_loop_mode.py | 26 ++++++++++++++++++++++++++
 2 files changed, 33 insertions(+)

diff --git a/agent/moa_loop.py b/agent/moa_loop.py
index bab36fb4010..ed566325eeb 100644
--- a/agent/moa_loop.py
+++ b/agent/moa_loop.py
@@ -55,6 +55,13 @@ def _slot_runtime(slot: dict[str, str]) -> dict[str, Any]:
         from hermes_cli.runtime_provider import resolve_runtime_provider
 
         rt = resolve_runtime_provider(requested=provider, target_model=model)
+        resolved_provider = str(rt.get("provider") or provider).strip().lower()
+        # call_llm treats an explicit base_url as a custom endpoint. That is
+        # correct for ordinary OpenAI-compatible targets, but wrong for OAuth /
+        # adapter-backed providers whose provider branch adds auth headers and
+        # request-shape adapters. Keep those providers identified by name.
+        if resolved_provider in {"openai-codex", "xai-oauth"}:
+            return out
         # Pass the resolved endpoint through so call_llm builds the request for
         # the provider's actual API surface instead of auto-detecting. base_url
         # routes call_llm to the right adapter (incl. anthropic_messages mode);
diff --git a/tests/run_agent/test_moa_loop_mode.py b/tests/run_agent/test_moa_loop_mode.py
index 2bf5624df69..c05dd3b267f 100644
--- a/tests/run_agent/test_moa_loop_mode.py
+++ b/tests/run_agent/test_moa_loop_mode.py
@@ -172,6 +172,32 @@ def test_moa_slots_routed_through_resolve_runtime_provider(monkeypatch):
     assert rt["api_key"] == "key-for-minimax"
 
 
+def test_moa_codex_slot_preserves_provider_identity(monkeypatch):
+    """Codex slots must not become custom chat-completions endpoints.
+
+    _resolve_task_provider_model treats any explicit base_url as provider=custom.
+    For openai-codex that bypasses the Codex auxiliary branch, losing the
+    Cloudflare headers and Responses adapter required for chatgpt.com/backend-api/codex.
+    """
+    from agent import moa_loop
+
+    def fake_resolve(*, requested, target_model=None):
+        return {
+            "provider": requested,
+            "api_mode": "codex_responses",
+            "base_url": "https://chatgpt.com/backend-api/codex",
+            "api_key": "codex-oauth-token",
+        }
+
+    monkeypatch.setattr(
+        "hermes_cli.runtime_provider.resolve_runtime_provider", fake_resolve
+    )
+
+    rt = moa_loop._slot_runtime({"provider": "openai-codex", "model": "gpt-5.5"})
+
+    assert rt == {"provider": "openai-codex", "model": "gpt-5.5"}
+
+
 def test_moa_slot_runtime_falls_back_on_resolution_error(monkeypatch):
     """A slot whose provider can't be resolved still attempts the call with the
     bare provider/model rather than aborting the whole MoA turn."""