fix(delegation): honor api_mode + auto-detect anthropic_messages URLs (#26824)

Subagent delegation hardcoded api_mode='chat_completions' for any
delegation.base_url that didn't match three specific hostnames
(chatgpt.com, api.anthropic.com, api.kimi.com/coding), and never
read delegation.api_mode from config. Azure AI Foundry's
https://foundry.services.ai.azure.com/anthropic endpoint fell through
and got chat_completions, causing 404s on every delegate_task call.

The main agent already handles this correctly via the shared
_detect_api_mode_for_url() helper (anything ending in /anthropic →
anthropic_messages); delegation reimplemented its own narrower check.

Reuse the shared detector and honor an explicit delegation.api_mode
when set so users can also force the transport on non-standard
endpoints the URL heuristic can't classify.

Fixes #10213.

Co-authored-by: HiddenPuppy <HiddenPuppy@users.noreply.github.com>

hermes_cli/config.py

@@ -1146,6 +1146,10 @@ DEFAULT_CONFIG = {
         "provider": "",    # e.g. "openrouter" (empty = inherit parent provider + credentials)
         "base_url": "",    # direct OpenAI-compatible endpoint for subagents
         "api_key": "",     # API key for delegation.base_url (falls back to OPENAI_API_KEY)
+        "api_mode": "",    # wire protocol for delegation.base_url: "chat_completions",
+                           # "codex_responses", or "anthropic_messages". Empty = auto-detect
+                           # from URL (e.g. /anthropic suffix → anthropic_messages). Set this
+                           # explicitly for non-standard endpoints the heuristic can't detect.
         # When delegate_task narrows child toolsets explicitly, preserve any
         # MCP toolsets the parent already has enabled. On by default so
         # narrowing (e.g. toolsets=["web","browser"]) expresses "I want these

tests/tools/test_delegate.py

@@ -890,6 +890,63 @@ class TestDelegationCredentialResolution(unittest.TestCase):
         self.assertEqual(creds["api_key"], "local-key")
         self.assertEqual(creds["api_mode"], "chat_completions")
 
+    def test_direct_endpoint_auto_detects_anthropic_messages_suffix(self):
+        # Issue #10213: Azure AI Foundry exposes Anthropic-compatible models at
+        # a /anthropic URL suffix. Subagents must pick anthropic_messages
+        # automatically, matching the main agent's runtime resolver.
+        parent = _make_mock_parent(depth=0)
+        cfg = {
+            "model": "claude-opus-4-6",
+            "provider": "custom",
+            "base_url": "https://myfoundry.services.ai.azure.com/anthropic",
+            "api_key": "foundry-key",
+        }
+        creds = _resolve_delegation_credentials(cfg, parent)
+        self.assertEqual(creds["provider"], "custom")
+        self.assertEqual(creds["base_url"], "https://myfoundry.services.ai.azure.com/anthropic")
+        self.assertEqual(creds["api_key"], "foundry-key")
+        self.assertEqual(creds["api_mode"], "anthropic_messages")
+
+    def test_direct_endpoint_honors_explicit_api_mode(self):
+        # When delegation.api_mode is set explicitly, it overrides URL-based
+        # detection so users can force a transport on non-standard endpoints.
+        parent = _make_mock_parent(depth=0)
+        cfg = {
+            "model": "claude-opus-4-6",
+            "provider": "custom",
+            "base_url": "https://proxy.example.com/v1",
+            "api_key": "proxy-key",
+            "api_mode": "anthropic_messages",
+        }
+        creds = _resolve_delegation_credentials(cfg, parent)
+        self.assertEqual(creds["api_mode"], "anthropic_messages")
+
+    def test_direct_endpoint_explicit_api_mode_overrides_url_detection(self):
+        # Explicit api_mode in config always wins over auto-detection.
+        parent = _make_mock_parent(depth=0)
+        cfg = {
+            "model": "claude-opus-4-6",
+            "provider": "custom",
+            "base_url": "https://myfoundry.services.ai.azure.com/anthropic",
+            "api_key": "foundry-key",
+            "api_mode": "chat_completions",
+        }
+        creds = _resolve_delegation_credentials(cfg, parent)
+        self.assertEqual(creds["api_mode"], "chat_completions")
+
+    def test_direct_endpoint_invalid_api_mode_falls_back_to_detection(self):
+        # An invalid api_mode string must not break detection; fall back to URL heuristic.
+        parent = _make_mock_parent(depth=0)
+        cfg = {
+            "model": "claude-opus-4-6",
+            "provider": "custom",
+            "base_url": "https://myfoundry.services.ai.azure.com/anthropic",
+            "api_key": "foundry-key",
+            "api_mode": "garbage",
+        }
+        creds = _resolve_delegation_credentials(cfg, parent)
+        self.assertEqual(creds["api_mode"], "anthropic_messages")
+
     def test_direct_endpoint_returns_none_api_key_when_not_configured(self):
         # When base_url is set without api_key, api_key should be None so
         # _build_child_agent inherits the parent's key (effective_api_key = override or parent).

tools/delegate_tool.py

@@ -2362,6 +2362,7 @@ def _resolve_delegation_credentials(cfg: dict, parent_agent) -> dict:
     configured_provider = str(cfg.get("provider") or "").strip() or None
     configured_base_url = str(cfg.get("base_url") or "").strip() or None
     configured_api_key = str(cfg.get("api_key") or "").strip() or None
+    configured_api_mode = str(cfg.get("api_mode") or "").strip().lower() or None
 
     if configured_base_url:
         # When delegation.api_key is not set, return None so _build_child_agent
@@ -2372,9 +2373,17 @@ def _resolve_delegation_credentials(cfg: dict, parent_agent) -> dict:
         # callers to duplicate the key under delegation.api_key.
         api_key = configured_api_key  # None → inherited from parent in _build_child_agent
 
+        # Use the shared URL-based api_mode detector (same path the main agent's
+        # runtime resolver uses) so Anthropic-compatible direct endpoints with a
+        # /anthropic suffix — Azure AI Foundry, MiniMax, Zhipu GLM, LiteLLM
+        # proxies — pick the right transport automatically. Without this,
+        # subagents would default to chat_completions and hit 404s on endpoints
+        # that only speak the Anthropic Messages protocol. Fixes #10213.
+        from hermes_cli.runtime_provider import _detect_api_mode_for_url
+
         base_lower = configured_base_url.lower()
         provider = "custom"
-        api_mode = "chat_completions"
+        api_mode = _detect_api_mode_for_url(configured_base_url) or "chat_completions"
         if (
             base_url_hostname(configured_base_url) == "chatgpt.com"
             and "/backend-api/codex" in base_lower
@@ -2388,6 +2397,11 @@ def _resolve_delegation_credentials(cfg: dict, parent_agent) -> dict:
             provider = "custom"
             api_mode = "anthropic_messages"
 
+        # Explicit delegation.api_mode in config always wins. Lets users force
+        # a transport for non-standard endpoints the URL heuristic can't detect.
+        if configured_api_mode in {"chat_completions", "codex_responses", "anthropic_messages"}:
+            api_mode = configured_api_mode
+
         return {
             "model": configured_model,
             "provider": provider,