fix: derive delegation api_mode from delegation model, not main model

tests/tools/test_delegate.py

@@ -703,6 +703,46 @@ def test_missing_config_keys_inherit_parent(self):
         self.assertIsNone(creds["model"])
         self.assertIsNone(creds["provider"])
 
+    @patch("hermes_cli.runtime_provider.resolve_runtime_provider")
+    def test_copilot_delegation_api_mode_uses_delegation_model(self, mock_resolve):
+        """When copilot resolves codex_responses (main model GPT-5+), delegation model
+        GPT-4.x should override to chat_completions.
+
+        Regression test for: delegation api_mode was derived from main model
+        (gpt-5.4 → codex_responses) instead of delegation model (gpt-4.1 →
+        chat_completions), causing HTTP 400 on subagent creation.
+        """
+        # Simulate: resolve_runtime_provider returns codex_responses because
+        # the main model is gpt-5.4 (GPT-5+ triggers Responses API)
+        mock_resolve.return_value = {
+            "provider": "copilot",
+            "base_url": "https://api.githubcopilot.com",
+            "api_key": "ghu_test_key_123",
+            "api_mode": "codex_responses",  # Wrong for gpt-4.1
+        }
+        parent = _make_mock_parent(depth=0)
+        cfg = {"model": "gpt-4.1", "provider": "copilot"}
+        creds = _resolve_delegation_credentials(cfg, parent)
+        # The fix should override api_mode to chat_completions for gpt-4.1
+        self.assertEqual(creds["api_mode"], "chat_completions")
+        self.assertEqual(creds["model"], "gpt-4.1")
+        self.assertEqual(creds["provider"], "copilot")
+
+    @patch("hermes_cli.runtime_provider.resolve_runtime_provider")
+    def test_copilot_delegation_api_mode_preserves_responses_for_gpt5(self, mock_resolve):
+        """When delegation model is also GPT-5+, codex_responses should be preserved."""
+        mock_resolve.return_value = {
+            "provider": "copilot",
+            "base_url": "https://api.githubcopilot.com",
+            "api_key": "ghu_test_key_123",
+            "api_mode": "codex_responses",
+        }
+        parent = _make_mock_parent(depth=0)
+        cfg = {"model": "gpt-5.4", "provider": "copilot"}
+        creds = _resolve_delegation_credentials(cfg, parent)
+        # GPT-5.4 should keep codex_responses
+        self.assertEqual(creds["api_mode"], "codex_responses")
+
 
 class TestDelegationProviderIntegration(unittest.TestCase):
     """Integration tests: delegation config → _run_single_child → AIAgent construction."""

tools/delegate_tool.py

@@ -801,12 +801,27 @@ def _resolve_delegation_credentials(cfg: dict, parent_agent) -> dict:
             f"Set the appropriate environment variable or run 'hermes auth'."
         )
 
+    api_mode = runtime.get("api_mode")
+
+    # Fix: when delegation uses copilot with a different model than the main
+    # agent, the api_mode resolved by resolve_runtime_provider is based on the
+    # main model (model.default in config).  We must re-derive it for the
+    # delegation model.  E.g. main model gpt-5.4 → codex_responses, but
+    # delegation model gpt-4.1 → chat_completions.  Without this, the
+    # subagent sends gpt-4.1 to the Responses API which returns HTTP 400.
+    if configured_model and configured_provider in ("copilot",):
+        try:
+            from hermes_cli.models import copilot_model_api_mode
+            api_mode = copilot_model_api_mode(configured_model, api_key=api_key)
+        except Exception:
+            api_mode = "chat_completions"
+
     return {
         "model": configured_model,
         "provider": runtime.get("provider"),
         "base_url": runtime.get("base_url"),
         "api_key": api_key,
-        "api_mode": runtime.get("api_mode"),
+        "api_mode": api_mode,
         "command": runtime.get("command"),
         "args": list(runtime.get("args") or []),
     }