fix(moa): route Copilot slots by target model

hermes_cli/models.py

@@ -3649,19 +3649,11 @@ def copilot_model_api_mode(
     if _should_use_copilot_responses_api(normalized):
         return "codex_responses"
 
-    # Secondary: check catalog for non-GPT-5 models (Claude via /v1/messages, etc.)
-    if catalog:
-        catalog_entry = next((item for item in catalog if item.get("id") == normalized), None)
-        if isinstance(catalog_entry, dict):
-            supported_endpoints = {
-                str(endpoint).strip()
-                for endpoint in (catalog_entry.get("supported_endpoints") or [])
-                if str(endpoint).strip()
-            }
-            # For non-GPT-5 models, check if they only support messages API
-            if "/v1/messages" in supported_endpoints and "/chat/completions" not in supported_endpoints:
-                return "anthropic_messages"
-
+    # Copilot's Claude models are exposed through its OpenAI-compatible chat
+    # endpoint, not through Hermes' native Anthropic adapter. The live catalog may
+    # advertise /v1/messages, but the Copilot token/header scheme is handled by
+    # the OpenAI client path; selecting anthropic_messages would send the wrong
+    # auth/wire shape. Keep non-GPT Copilot slots on chat_completions.
     return "chat_completions"
 
 

hermes_cli/runtime_provider.py

@@ -317,13 +317,24 @@ def _provider_supports_explicit_api_mode(provider: Optional[str], configured_pro
     return normalized_configured == normalized_provider
 
 
-def _copilot_runtime_api_mode(model_cfg: Dict[str, Any], api_key: str) -> str:
+def _copilot_runtime_api_mode(
+    model_cfg: Dict[str, Any],
+    api_key: str,
+    *,
+    target_model: Optional[str] = None,
+) -> str:
     configured_provider = str(model_cfg.get("provider") or "").strip().lower()
     configured_mode = _parse_api_mode(model_cfg.get("api_mode"))
     if configured_mode and _provider_supports_explicit_api_mode("copilot", configured_provider):
         return configured_mode
 
-    model_name = str(model_cfg.get("default") or "").strip()
+    # Use the model being resolved for this runtime, not the persisted global
+    # default. MoA slots, fallback models, and mid-session model switches all
+    # resolve credentials for a target model that can differ from config.yaml's
+    # model.default. If we derive Copilot api_mode from the stale default, a
+    # Claude/Gemini MoA slot can inherit codex_responses from a GPT-5 default and
+    # fail with "model ... does not support Responses API".
+    model_name = str(target_model or model_cfg.get("default") or "").strip()
     if not model_name:
         return "chat_completions"
 
@@ -449,7 +460,11 @@ def _resolve_runtime_from_pool_entry(
         api_mode = "chat_completions"
         base_url = _nous_inference_base_url_override() or base_url
     elif provider == "copilot":
-        api_mode = _copilot_runtime_api_mode(model_cfg, getattr(entry, "runtime_api_key", ""))
+        api_mode = _copilot_runtime_api_mode(
+            model_cfg,
+            getattr(entry, "runtime_api_key", ""),
+            target_model=effective_model,
+        )
         base_url = base_url or PROVIDER_REGISTRY["copilot"].inference_base_url
     elif provider == "azure-foundry":
         # Azure Foundry: read api_mode and base_url from config

tests/hermes_cli/test_copilot_model_api_mode.py

@@ -0,0 +1,23 @@
+"""Tests for Copilot model API-mode routing."""
+
+from __future__ import annotations
+
+
+def test_copilot_claude_stays_on_chat_completions_even_if_catalog_lists_messages():
+    from hermes_cli.models import copilot_model_api_mode
+
+    catalog = [
+        {
+            "id": "claude-opus-4.8",
+            "supported_endpoints": ["/v1/messages"],
+        }
+    ]
+
+    assert copilot_model_api_mode("claude-opus-4.8", catalog=catalog) == "chat_completions"
+
+
+def test_copilot_gpt5_still_uses_responses_api():
+    from hermes_cli.models import copilot_model_api_mode
+
+    assert copilot_model_api_mode("gpt-5.5", catalog=[]) == "codex_responses"
+    assert copilot_model_api_mode("gpt-5-mini", catalog=[]) == "chat_completions"

tests/hermes_cli/test_copilot_runtime_api_mode.py

@@ -0,0 +1,40 @@
+"""Regression tests for Copilot runtime api_mode resolution."""
+
+from __future__ import annotations
+
+
+def test_copilot_runtime_api_mode_uses_target_model_over_stale_config_default(monkeypatch):
+    """MoA/fallback slots must derive Copilot api_mode from the slot model.
+
+    Repro: user's main/default Copilot model is GPT-5.5 (Responses API), but a
+    MoA reference slot is Copilot Claude Opus (Chat Completions). Runtime
+    resolution previously looked only at model.default and returned
+    codex_responses for the Claude slot, causing Copilot to reject it with
+    "model ... does not support Responses API".
+    """
+    from hermes_cli import runtime_provider as rp
+
+    monkeypatch.setattr(
+        "hermes_cli.models.copilot_model_api_mode",
+        lambda model, api_key=None: "codex_responses" if str(model).startswith("gpt-5") else "chat_completions",
+    )
+
+    assert rp._copilot_runtime_api_mode(
+        {"provider": "copilot", "default": "gpt-5.5"},
+        "token",
+        target_model="claude-opus-4.8",
+    ) == "chat_completions"
+
+
+def test_copilot_runtime_api_mode_still_uses_default_without_target(monkeypatch):
+    from hermes_cli import runtime_provider as rp
+
+    monkeypatch.setattr(
+        "hermes_cli.models.copilot_model_api_mode",
+        lambda model, api_key=None: "codex_responses" if str(model).startswith("gpt-5") else "chat_completions",
+    )
+
+    assert rp._copilot_runtime_api_mode(
+        {"provider": "copilot", "default": "gpt-5.5"},
+        "token",
+    ) == "codex_responses"