fix(delegation): carry provider request_overrides through the base_url path (#65035)

tests/tools/test_delegate.py

@@ -827,6 +827,58 @@ def test_direct_endpoint_auto_detects_anthropic_messages_suffix(self):
         self.assertEqual(creds["api_mode"], "anthropic_messages")
 
 
+    @patch("hermes_cli.runtime_provider.resolve_runtime_provider")
+    def test_base_url_with_provider_carries_runtime_request_overrides(self, mock_resolve):
+        """#65035: the base_url short-circuit must not drop the configured
+        provider's request_overrides / max_output_tokens."""
+        mock_resolve.return_value = {
+            "provider": "custom",
+            "base_url": "https://provider-default.example/v1",
+            "api_key": "provider-key",
+            "api_mode": "chat_completions",
+            "request_overrides": {"extra_body": {"thinking": {"type": "disabled"}}},
+            "max_output_tokens": 8192,
+        }
+        parent = _make_mock_parent(depth=0)
+        cfg = {
+            "model": "mimo-v2.5-pro",
+            "provider": "mimo",
+            "base_url": "https://api.xiaomimimo.com/v1",
+            "api_key": "cfg-key",
+        }
+        creds = _resolve_delegation_credentials(cfg, parent)
+        # Explicitly configured endpoint + key still win over the runtime's.
+        self.assertEqual(creds["base_url"], "https://api.xiaomimimo.com/v1")
+        self.assertEqual(creds["api_key"], "cfg-key")
+        # The provider's request personality survives the short-circuit.
+        self.assertEqual(
+            creds["request_overrides"],
+            {"extra_body": {"thinking": {"type": "disabled"}}},
+        )
+        self.assertEqual(creds["max_output_tokens"], 8192)
+
+    def test_bare_base_url_returns_none_overrides(self):
+        """No provider alongside base_url → no overrides source; keys are
+        present but None (shape parity with the inherit-everything path)."""
+        parent = _make_mock_parent(depth=0)
+        cfg = {"model": "m", "provider": "", "base_url": "http://localhost:1234/v1", "api_key": "k"}
+        creds = _resolve_delegation_credentials(cfg, parent)
+        self.assertIsNone(creds["request_overrides"])
+        self.assertIsNone(creds["max_output_tokens"])
+
+    @patch("hermes_cli.runtime_provider.resolve_runtime_provider")
+    def test_base_url_survives_runtime_resolution_failure(self, mock_resolve):
+        """Best-effort: the explicit endpoint worked before this change even
+        when the provider can't resolve — a resolution failure must not
+        break it, only skip the overrides."""
+        mock_resolve.side_effect = RuntimeError("MIMO_API_KEY not set")
+        parent = _make_mock_parent(depth=0)
+        cfg = {"model": "m", "provider": "mimo", "base_url": "https://api.xiaomimimo.com/v1", "api_key": "k"}
+        creds = _resolve_delegation_credentials(cfg, parent)
+        self.assertEqual(creds["base_url"], "https://api.xiaomimimo.com/v1")
+        self.assertIsNone(creds["request_overrides"])
+        self.assertIsNone(creds["max_output_tokens"])
+
     @patch("hermes_cli.runtime_provider.resolve_runtime_provider")
     def test_provider_resolution_failure_raises_valueerror(self, mock_resolve):
         """When provider resolution fails, ValueError is raised with helpful message."""

tools/delegate_tool.py

@@ -4531,12 +4531,42 @@ def _resolve_delegation_credentials(cfg: dict, parent_agent) -> dict:
         if configured_api_mode in {"chat_completions", "codex_responses", "anthropic_messages"}:
             api_mode = configured_api_mode
 
+        # A provider configured ALONGSIDE base_url means the user wants that
+        # provider's request personality on an explicit endpoint. This
+        # short-circuit runs before the resolve_runtime_provider() call below,
+        # so without this block the runtime-carried request_overrides
+        # (extra_body / extra_headers, e.g. `thinking: {type: disabled}`) and
+        # max_output_tokens are silently dropped for subagents (#65035).
+        # Best-effort: the explicit endpoint worked before this change even
+        # when the provider can't resolve, so a resolution failure only skips
+        # the overrides — it must not fail the dispatch.
+        request_overrides = None
+        max_output_tokens = None
+        if configured_provider:
+            try:
+                from hermes_cli.runtime_provider import resolve_runtime_provider
+
+                runtime = resolve_runtime_provider(
+                    requested=configured_provider, target_model=configured_model
+                )
+                request_overrides = dict(runtime.get("request_overrides") or {}) or None
+                max_output_tokens = runtime.get("max_output_tokens")
+            except Exception as exc:
+                logger.debug(
+                    "delegation.base_url: runtime resolution for provider '%s' "
+                    "failed; proceeding without request_overrides: %s",
+                    configured_provider,
+                    exc,
+                )
+
         return {
             "model": configured_model,
             "provider": provider,
             "base_url": configured_base_url,
             "api_key": api_key,
             "api_mode": api_mode,
+            "request_overrides": request_overrides,
+            "max_output_tokens": max_output_tokens,
         }
 
     if not configured_provider: