diff --git a/hermes_cli/models.py b/hermes_cli/models.py index 9b2e3c8463..eb9bea8ee8 100644 --- a/hermes_cli/models.py +++ b/hermes_cli/models.py @@ -1536,14 +1536,8 @@ def cached_provider_model_ids( fp = _credential_fingerprint(normalized) entry = cache.get(normalized) now = time.time() - # Even a credential-matched fresh disk cache predates the current account's - # entitlement check. Revalidate gated entries before offering them again. - cached_models = entry.get("models") if isinstance(entry, dict) else None - gated_cache = isinstance(cached_models, list) and any( - _model_requires_account_discovery(normalized, model) for model in cached_models - ) - if not force_refresh and not gated_cache and _cache_entry_valid(entry, fp, allow_empty=is_ollama): + if not force_refresh and _cache_entry_valid(entry, fp, allow_empty=is_ollama): age = now - entry["at"] if age < ttl_seconds: return list(entry["models"]) @@ -1570,7 +1564,9 @@ def cached_provider_model_ids( if same_creds and isinstance(entry.get("models"), list) and entry["models"]: return list(entry["models"]) return [] - # Live returned nothing: a stale same-fingerprint entry beats an empty result. + # Live returned nothing: a stale same-fingerprint entry beats an empty result — minus account-gated + # models, which only a successful discovery may advertise (the entry itself is untouched, so the + # next successful fetch restores them). if _cache_entry_valid(entry, fp): return [model for model in entry["models"] if not _model_requires_account_discovery(normalized, model)] return [] diff --git a/tests/agent/test_astra_baseline_runtime.py b/tests/agent/test_astra_baseline_runtime.py index b3591ad11a..b1fad0e133 100644 --- a/tests/agent/test_astra_baseline_runtime.py +++ b/tests/agent/test_astra_baseline_runtime.py @@ -77,7 +77,6 @@ def test_astra_900k_opt_in_preserves_live_limits_and_wire_contract(monkeypatch, kwargs = ResponsesApiTransport().build_kwargs(model="gpt-6-astra-900k", **params) assert kwargs["model"] == "gpt-6-astra" assert kwargs == ResponsesApiTransport().build_kwargs(model="gpt-6-astra", **params) - assert "prompt_cache_options" not in kwargs @pytest.mark.parametrize("provider,model", [ @@ -99,7 +98,13 @@ def test_picker_revalidates_cached_astra_and_never_injects_saved_entitlement(mon return list(live) monkeypatch.setattr(models, "provider_model_ids", discover) - assert models.cached_provider_model_ids(provider) == ["gpt-5.6-sol"] + # Only live discovery writes Astra into a same-credential entry, so a fresh entry IS the + # entitlement record: served without a round-trip (a per-open revalidation defeated the cache). + assert model in models.cached_provider_model_ids(provider) + assert calls == [] + # After the entry ages past the fresh window, a failed refresh must not resurrect Astra from it. + monkeypatch.setattr(models, "_PROVIDER_MODELS_STALE_SERVE_MAX", 0) + assert models.cached_provider_model_ids(provider, ttl_seconds=0) == ["gpt-5.6-sol"] assert calls == [provider] row = {"slug": provider, "is_current": True, "models": ["gpt-5.6-sol"], "total_models": 1} assert model not in _finalize_picker_rows([row], {}, model)[0]["models"]