Troubleshooting

Common issues when running GAZE, especially against local models.

Thinking models need a larger token budget

Models that emit reasoning tokens (for example Qwen 3.5) count those tokens against max_tokens. Set --max-tokens 4096 or higher (the run scripts default to 8192) so the model has room to reason and still produce the final JSON.

LM Studio runs one model at a time

On typical hardware LM Studio serves a single model. A preflight health check that names a different model can trigger a swap (unload then reload). Point require_lmstudio_model() and the CLI at the model you actually have loaded.

Empty model content

Some models (Qwen 3.5, GLM-4.6V) place their output in reasoning_content rather than content. GAZE detects an empty content and falls back to reasoning_content automatically, so no action is needed.

Context overflow on local models

LMStudioAdapter detects context-size errors (HTTP 400 with "context size", "n_ctx", or "n_keep" in the message) and raises a clear ModelError. If you hit this, reduce --max-image-dim, lower --max-turns, or load a model with a larger context window. Agentic mode (system prompt plus tool schemas) needs more context than single-turn mode.

Sending API keys to a custom host

For safety, OpenAIAdapter only sends credentials to an allowlisted set of hosts over HTTPS. To target a different cloud endpoint, set GAZE_ALLOW_CUSTOM_BASE_URL=1. Local LM Studio servers use LMStudioAdapter, which permits http:// and needs no real key, so this flag is not required for local inference.