OpenSquilla version or commit
0.5.3
Area
Web UI
Reproduction steps
- Configure a keyless OpenAI-compatible vLLM endpoint and a model ID that is not present in the packaged model catalog:
opensquilla onboard --provider vllm --model qa-model --base-url http://127.0.0.1:8000/v1 --router disabled --minimal --skip-channels --skip-search --skip-image-generation
- Do not set
llm.context_window_tokens or a [models.vllm."qa-model"] override.
- Start the Gateway and open Control UI → Settings → Model Services → Edit vLLM.
- Inspect the visible fields in the editor.
- Send a short WebChat message.
Expected behavior
The editor should expose the documented per-provider/model context-window override, including the auto-detected/effective readout and local-runtime warning. A user should be able to set the real window of a self-hosted model without editing TOML or applying a global override.
Actual behavior
The visible vLLM editor contains only Base URL, HTTP proxy, API key, connection test, and direct/fallback model. There is no context-window control.
The control is mounted once in the DOM as:
input[name="setup_provider_context_window"]
visible: false
rect: 0 × 0
ancestor: <div hidden style="display: none">
With no override, the model row reports an unknown/zero context window and a short WebChat message fails before any provider request:
The request is too large for the provider context window after automatic context compaction and payload reduction.
Setting a sufficiently large global context override and restarting immediately makes the same WebChat request succeed, confirming that the endpoint itself is healthy. The global override is only a workaround; it applies to every active model.
Environment
windows 11
Diagnostic bundle and error reference
No error reference was shown. Browser console had no errors. The reproduction uses no credentials.
OpenSquilla version or commit
0.5.3
Area
Web UI
Reproduction steps
llm.context_window_tokensor a[models.vllm."qa-model"]override.Expected behavior
The editor should expose the documented per-provider/model context-window override, including the auto-detected/effective readout and local-runtime warning. A user should be able to set the real window of a self-hosted model without editing TOML or applying a global override.
Actual behavior
The visible vLLM editor contains only Base URL, HTTP proxy, API key, connection test, and direct/fallback model. There is no context-window control.
The control is mounted once in the DOM as:
With no override, the model row reports an unknown/zero context window and a short WebChat message fails before any provider request:
Setting a sufficiently large global context override and restarting immediately makes the same WebChat request succeed, confirming that the endpoint itself is healthy. The global override is only a workaround; it applies to every active model.
Environment
windows 11
Diagnostic bundle and error reference
No error reference was shown. Browser console had no errors. The reproduction uses no credentials.