Bug Description
Ollama provider does not work with :cloud model variants (e.g., kimi-k3:cloud, gpt-oss:cloud). Two distinct symptoms observed:
-
Local models timeout on cold start: A local 9B model fails to load within the default Ollama timeout (120s). The model cold-start (loading weights into memory) exceeds the timeout, causing the request to fail before inference even begins.
-
:cloud models appear not to attempt at all: When switching the model name to a :cloud variant (e.g., kimi-k3:cloud), the provider does not seem to actually send the request to Ollama — no observable load attempt or inference occurs.
Environment
- SparkyFitness deployed via TrueNAS app (Docker)
- Ollama provider configured with custom URL pointing to Ollama server
- Models tested: local 9B model,
kimi-k3:cloud, other :cloud variants
Expected Behavior
:cloud models should be passed through to Ollama, which proxies them to Ollama Cloud- The timeout should be long enough for local model cold-starts, or configurable
Additional Context
The Ollama default timeout in providerDispatch.ts is OLLAMA_DEFAULT_TIMEOUT_MS = 120_000 (120s). This may be insufficient for:
- Local model cold-starts (loading a 9B model into VRAM can take >120s on slower hardware)
- Cloud model initialization
The provider.timeout field exists in ProviderConfig and is used in resolveTimeout(), but there does not appear to be a UI-exposed setting to configure it.
Regarding :cloud models: the native Ollama API path (/api/chat) is used for the raw dispatch path, while the chat service path uses the OpenAI-compatible endpoint (/v1/chat/completions). The :cloud models may behave differently between these two endpoints.
Bug Description
Ollama provider does not work with
:cloudmodel variants (e.g.,kimi-k3:cloud,gpt-oss:cloud). Two distinct symptoms observed:Local models timeout on cold start: A local 9B model fails to load within the default Ollama timeout (120s). The model cold-start (loading weights into memory) exceeds the timeout, causing the request to fail before inference even begins.
:cloudmodels appear not to attempt at all: When switching the model name to a:cloudvariant (e.g.,kimi-k3:cloud), the provider does not seem to actually send the request to Ollama — no observable load attempt or inference occurs.Environment
kimi-k3:cloud, other:cloudvariantsExpected Behavior
:cloudmodels should be passed through to Ollama, which proxies them to Ollama CloudAdditional Context
The Ollama default timeout in
providerDispatch.tsisOLLAMA_DEFAULT_TIMEOUT_MS = 120_000(120s). This may be insufficient for:The
provider.timeoutfield exists inProviderConfigand is used inresolveTimeout(), but there does not appear to be a UI-exposed setting to configure it.Regarding
:cloudmodels: the native Ollama API path (/api/chat) is used for the raw dispatch path, while the chat service path uses the OpenAI-compatible endpoint (/v1/chat/completions). The:cloudmodels may behave differently between these two endpoints.