Local Runner Troubleshooting

This page covers common issues when running kew against a local OpenAI-compatible endpoint.

Context window too small

Symptom: The run fails with an HTTP 400 context_length_exceeded / "exceeds the available context size".

Fix: The model was loaded with too small a context (e.g. the default 4096 tokens); reload it with a larger context (16k+), since kew's prompt plus the files the agent reads can exceed 4k quickly.

No model loaded

Symptom: Requests return an empty models list or a model-not-found error.

Fix: Load/pull a model into your local server first, and confirm the served model id matches the lemonade/<model-id> in kew.toml.

Wrong port / base_url

Symptom: Connection refused.

Fix: Confirm the server's port and set [agent.lemonade] base_url to match (Lemonade defaults to http://localhost:13305/v1).

Slow first token

Symptom: Long pause before output.

Note: Local inference is wall-clock-bound, not a hang; larger models are slower on CPU/GPU without an NPU.