Skip to content

The agent that grows with you

fix(providers): one guarded api_mode_for_model for resolution, /model and resume

Tekniumfbb1b417e2bc
diff --git a/agent/auxiliary_client.py b/agent/auxiliary_client.py
index 20a05e7faa3..3ffe40dbcf2 100644
--- a/agent/auxiliary_client.py
+++ b/agent/auxiliary_client.py
@@ -5300,10 +5300,10 @@ def _resolve_api_key_branch(req: _ResolveRequest, pconfig: Any, resolve_creds: C
return _route_client(req, client, final_model)
headers = _endpoint_default_headers(base_url, provider, is_vision=req.is_vision, xai=True)
client = _create_openai_client(api_key=api_key, base_url=base_url, **({"default_headers": headers} if headers else {}))
- # A profile serving this model only on the Responses API: route call_llm() through responses.stream().
+ from agent.provider_route_hooks import model_api_mode # a Responses-only model: call_llm() streams responses
from providers import get_provider_profile
profile = get_provider_profile(provider) if final_model and not req.raw_codex else None
- if profile is not None and profile.api_mode_for_model(final_model) == "codex_responses":
+ if profile is not None and model_api_mode(profile, final_model) == "codex_responses":
logger.debug("resolve_provider_client: %s model %s needs Responses API — wrapping", provider, final_model)
client = CodexAuxiliaryClient(client, final_model)
# api_mode handling for any API-key provider (direct OpenAI + codex model) and Anthropic-wire
diff --git a/agent/provider_route_hooks.py b/agent/provider_route_hooks.py
index 9a92b4a55fe..819b9e44442 100644
--- a/agent/provider_route_hooks.py
+++ b/agent/provider_route_hooks.py
@@ -137,6 +137,25 @@ def ask_on_error_path(profile: Any, hook: str, fallback: Any, **kwargs: Any) ->
_ERROR_PATH_WARNED: set[tuple] = set()
+def model_api_mode(profile: Any, model: str, *, api_key: str | None = None) -> str | None:
+ """*profile*'s ``api_mode_for_model`` answer for *model* (None = no per-model wire). The ONE guard for
+ resolution, ``/model``, resume and the agent: a raising hook (a live catalog fetch, a lost plugin host)
+ answers the profile's declared ``api_mode``, logged with its traceback once per provider per process,
+ so a model pick never fails where runtime resolution succeeded."""
+ try:
+ return profile.api_mode_for_model(model, api_key=api_key)
+ except Exception: # health: allow BLE001 -- plugin hook boundary; logged with traceback once below
+ name = getattr(profile, "name", "?")
+ if name not in _API_MODE_WARNED:
+ _API_MODE_WARNED.add(name)
+ logger.warning("provider %s: api_mode_for_model failed; using its declared api_mode", name,
+ exc_info=True)
+ return profile.api_mode or "chat_completions"
+
+
+_API_MODE_WARNED: set[str] = set()
+
+
def route_answer(hook: str, *, provider: str = "", scope: Any = None, **kwargs: Any) -> Any:
"""First truthy answer of *hook* across ``route_profiles``; None when no profile claims the endpoint.
Failures follow ``ask_profile``'s policy (*scope* = the agent, when there is one)."""
diff --git a/hermes_cli/model_switch.py b/hermes_cli/model_switch.py
index 94035151a01..077f3e62fd1 100644
--- a/hermes_cli/model_switch.py
+++ b/hermes_cli/model_switch.py
@@ -1707,7 +1707,8 @@ def model_derived_api_mode(provider: str, model: str, api_key: str = "") -> Opti
from providers import get_provider_profile
profile = get_provider_profile(key)
if profile is not None:
- return profile.api_mode_for_model(model, api_key=api_key or "")
+ from agent.provider_route_hooks import model_api_mode
+ return model_api_mode(profile, model, api_key=api_key or "")
return override(key, model, api_key) if override is not None else None
diff --git a/hermes_cli/runtime_provider.py b/hermes_cli/runtime_provider.py
index 8bbf1146149..da99b41c201 100644
--- a/hermes_cli/runtime_provider.py
+++ b/hermes_cli/runtime_provider.py
@@ -275,6 +275,7 @@ def _profile_runtime_api_mode(provider: str, model_cfg: dict[str, Any], api_key:
target_model: Optional[str] = None) -> str:
"""api_mode for a provider that routes per model: its hard per-model wire wins over a stale persisted
``codex_responses`` (pure check, no network); otherwise the profile decides with the live key."""
+ from agent.provider_route_hooks import model_api_mode
profile = _per_model_profile(provider)
configured_mode = _configured_api_mode(provider, model_cfg)
# Use the model being resolved, not the persisted default: a Claude MoA slot inheriting
@@ -285,19 +286,13 @@ def _profile_runtime_api_mode(provider: str, model_cfg: dict[str, Any], api_key:
# switch to a chat-only model) must not override the provider's per-model requirement: the
# Responses request silently succeeds but returns no reasoning content (#46527).
if (configured_mode == "codex_responses" and model_name and profile is not None
- and profile.api_mode_for_model(model_name) == "chat_completions"):
+ and model_api_mode(profile, model_name) == "chat_completions"):
return "chat_completions"
return configured_mode
if profile is None:
return "chat_completions"
declared = profile.api_mode or "chat_completions" # the hook's None means "keep the static api_mode"
- if not model_name:
- return declared
- try:
- return profile.api_mode_for_model(model_name, api_key=api_key) or declared
- except Exception: # plugin hook (live catalog fetch): degrade to the declared wire, never fail resolution
- logger.debug("%s api_mode_for_model(%s) failed", provider, model_name, exc_info=True)
- return declared
+ return (model_api_mode(profile, model_name, api_key=api_key) or declared) if model_name else declared
def _azure_inferred_api_mode(effective_model: str, api_mode: str) -> str:
diff --git a/run_agent.py b/run_agent.py
index eea433714c1..021d3db0c0f 100644
--- a/run_agent.py
+++ b/run_agent.py
@@ -650,7 +650,8 @@ class AIAgent(
# A provider serving several wires behind one endpoint decides per model (pure, no I/O).
from providers import get_provider_profile
profile = get_provider_profile(normalized_provider) if normalized_provider else None
- derived = profile.api_mode_for_model(model) if profile is not None else None
+ from agent.provider_route_hooks import model_api_mode
+ derived = model_api_mode(profile, model) if profile is not None else None
if derived is not None:
return derived == "codex_responses"
return AIAgent._model_requires_responses_api(model)
diff --git a/tests/hermes_cli/test_runtime_api_mode_profile_default.py b/tests/hermes_cli/test_runtime_api_mode_profile_default.py
index e22be961202..1a35c18154a 100644
--- a/tests/hermes_cli/test_runtime_api_mode_profile_default.py
+++ b/tests/hermes_cli/test_runtime_api_mode_profile_default.py
@@ -38,3 +38,13 @@ def test_hook_none_keeps_the_profiles_declared_api_mode(responses_default, model
from hermes_cli.runtime_provider import resolve_runtime_provider
assert resolve_runtime_provider(requested="respdefault", target_model=model)["api_mode"] == api_mode
+
+
+def test_a_model_pick_with_a_raising_hook_gets_the_wire_resolution_gave(responses_default):
+ """``/model`` and session resume re-derive the wire through ``model_derived_api_mode``: a raising hook
+ degrades there exactly as in runtime resolution instead of failing the switch."""
+ from hermes_cli.model_switch import model_derived_api_mode
+ from hermes_cli.runtime_provider import resolve_runtime_provider
+
+ resolved = resolve_runtime_provider(requested="respdefault", target_model="broken-z")["api_mode"]
+ assert model_derived_api_mode("respdefault", "broken-z", "sk-fake") == resolved == "codex_responses"