fix(models): a model the current provider serves never re-routes to OpenRouter

detect_provider_for_model() walked static catalogs and then the OpenRouter
catalog. Providers whose live catalog outruns the static list (Codex
early-access ids, Nous Portal slugs, Ollama Cloud) had nothing to stop the
ladder, so `/model gpt-6-astra` from a Codex session, or `/model glm-5.3-flash`
from a Nous-sub-only session, silently rebuilt the agent on metered OpenRouter
(or on a vendor the user has no key for).

The current provider's disk-cached live catalog (cached_provider_model_ids, 1h
TTL + SWR) is now consulted first: an exact or bare-name hit stays put and
resolves to the provider's own spelling. Fixes the whole class at the shared
choke point, covering /model, ACP, oneshot and the dashboard.

Refs #97487 (sj-unit72 diagnosed the ollama-cloud instance of this).

hermes_cli/models.py

@@ -936,6 +936,15 @@ def detect_provider_for_model(
     if not name:
         return None
 
+    # The current provider's LIVE catalog outranks every static guess: a model it already serves
+    # (Codex early-access ids, Portal-only slugs, Ollama Cloud models absent from _PROVIDER_MODELS)
+    # must never re-route the session to another vendor or to metered OpenRouter.
+    from hermes_cli.models_detect import current_provider_catalog_match
+
+    served = current_provider_catalog_match(name, current_provider)
+    if served is not None:
+        return (current_provider, served) if served != name else None
+
     static_match = detect_static_provider_for_model(name, current_provider)
     if static_match:
         return static_match

hermes_cli/models_detect.py

@@ -0,0 +1,39 @@
+"""Live-catalog guard for ``detect_provider_for_model``.
+
+Split out of ``hermes_cli.models``. The detection ladder there consults static catalogs, then the
+OpenRouter catalog. Providers whose static list lags their live catalog (Codex accounts with
+early-access models, Nous Portal, Ollama Cloud) have no static entry to stop the ladder, so a bare
+name the CURRENT provider already serves fell through to OpenRouter and the session was silently
+rebuilt on a metered aggregator (#97487, WolframRvnwlf's $100 Astra incident).
+"""
+
+from __future__ import annotations
+
+from typing import Optional
+
+# Providers where the ladder's own OpenRouter step is the answer, or where there is no catalog to
+# consult; ``custom`` endpoints are never auto-switched away from (handled upstream).
+_SKIP = frozenset({"", "auto", "openrouter", "custom"})
+
+
+def current_provider_catalog_match(model_name: str, current_provider: str) -> Optional[str]:
+    """Return the current provider's own spelling of *model_name* when its live (disk-cached)
+    catalog serves it — exact id, or the bare part after ``vendor/`` — else ``None``.
+
+    Goes through :func:`hermes_cli.models.cached_provider_model_ids` (1h TTL, stale-while-
+    revalidate) so a model switch does not block on a cold ``/v1/models`` round-trip in the
+    common case; a fetch failure yields an empty catalog and the ladder continues unchanged."""
+    from hermes_cli.models import cached_provider_model_ids, normalize_provider
+
+    provider = (current_provider or "").strip().lower()
+    if provider in _SKIP or provider.startswith("custom:") or normalize_provider(provider) in _SKIP:
+        return None
+    wanted = (model_name or "").strip().lower()
+    if not wanted:
+        return None
+    try:
+        catalog = cached_provider_model_ids(provider)
+    except Exception:
+        return None
+    return next((mid for mid in catalog if mid.lower() == wanted), None) or next(
+        (mid for mid in catalog if "/" in mid and mid.split("/", 1)[1].lower() == wanted), None)

tests/hermes_cli/test_models_detect_live_catalog.py

@@ -0,0 +1,40 @@
+"""``detect_provider_for_model`` must not re-route a model the CURRENT provider already serves.
+
+Regression for the OpenRouter hijack (#97487, the ``/model gpt-6-astra`` incident): a bare name
+absent from the static ``_PROVIDER_MODELS`` list but present in the current provider's live catalog
+fell through to the OpenRouter lookup and silently rebuilt the session on a metered aggregator.
+"""
+
+from __future__ import annotations
+
+import pytest
+
+from hermes_cli import models
+
+
+@pytest.fixture
+def live_catalog(monkeypatch):
+    """Pin the disk-cached live catalog per provider and make the OpenRouter lookup loud."""
+    catalogs: dict[str, list[str]] = {}
+    monkeypatch.setattr(models, "cached_provider_model_ids", lambda provider, **_: list(catalogs.get(provider, [])))
+    monkeypatch.setattr(models, "_find_openrouter_slug", lambda name: f"vendor/{name}")
+    return catalogs
+
+
+class TestCurrentProviderCatalogWins:
+    @pytest.mark.parametrize("provider,model", [
+        ("openai-codex", "gpt-6-astra"),       # early-access id, static Codex list lags
+        ("nous", "some-portal-only-model"),    # Portal serves it, static snapshot doesn't
+        ("ollama-cloud", "glm-5.3-flash"),     # static list points at zai; OpenRouter has zai/…
+    ])
+    def test_served_model_stays_on_current_provider(self, live_catalog, provider, model):
+        live_catalog[provider] = [model]
+        assert models.detect_provider_for_model(model, provider) is None
+
+    def test_bare_name_resolves_to_current_providers_full_slug(self, live_catalog):
+        live_catalog["nous"] = ["zai/glm-5.3-flash"]
+        assert models.detect_provider_for_model("glm-5.3-flash", "nous") == ("nous", "zai/glm-5.3-flash")
+
+    def test_unserved_model_still_walks_the_ladder(self, live_catalog):
+        live_catalog["nous"] = ["hermes-4-405b"]
+        assert models.detect_provider_for_model("no-such-model", "nous") == ("openrouter", "vendor/no-such-model")