fix: improve ACP type check and restore comment accuracy

- Use isinstance() with try/except import for CopilotACPClient check
  in _to_async_client instead of fragile __class__.__name__ string check
- Restore accurate comment: GPT-5.x models *require* (not 'often require')
  the Responses API on OpenAI/OpenRouter; ACP is the exception, not a
  softening of the requirement
- Add inline comment explaining the ACP exclusion rationale

agent/auxiliary_client.py

@@ -1223,8 +1223,12 @@ def _to_async_client(sync_client, model: str):
         return AsyncCodexAuxiliaryClient(sync_client), model
     if isinstance(sync_client, AnthropicAuxiliaryClient):
         return AsyncAnthropicAuxiliaryClient(sync_client), model
-    if sync_client.__class__.__name__ == "CopilotACPClient":
+    try:
-        return sync_client, model
+        from agent.copilot_acp_client import CopilotACPClient
+        if isinstance(sync_client, CopilotACPClient):
+            return sync_client, model
+    except ImportError:
+        pass
 
     async_kwargs = {
         "api_key": sync_client.api_key,

run_agent.py

@@ -705,10 +705,12 @@ class AIAgent:
         except Exception:
             pass
 
-        # GPT-5.x models often require the Responses API path — they are
+        # GPT-5.x models require the Responses API path — they are rejected
-        # rejected on /v1/chat/completions by both OpenAI and OpenRouter.
+        # on /v1/chat/completions by both OpenAI and OpenRouter.  Also
-        # Also auto-upgrade for direct OpenAI URLs (api.openai.com) since
+        # auto-upgrade for direct OpenAI URLs (api.openai.com) since all
         # newer tool-calling models prefer Responses there.
+        # ACP runtimes are excluded: CopilotACPClient handles its own
+        # routing and does not implement the Responses API surface.
         if (
             self.api_mode == "chat_completions"
             and self.provider != "copilot-acp"