An agent needs three answers. Is the server answering? Is the expected model loaded? Can that model complete a small request? Run these checks after starting a local server.
1. Check health and identify the model
API=http://127.0.0.1:8080/v1
curl -fsS --max-time 5 "$API/health"
curl -fsS --max-time 5 "$API/models"The v8 text server returned {"status":"ok","mode":"live","inference":true}. Its route is /v1/health. Some dedicated audio services use /health. A successful health check says the live endpoint is up. It can still be busy, have the wrong model, or fail to render a chat prompt.
Find your configured ID in data[].id. If present, read cke_context_length, cke_default_max_output_tokens and the model's advertised reasoning options. On the tested Qwen2 server these were a 1,024-token window and a default 512-token output allowance.
2. Send one small request
curl -fsS --max-time 90 "$API/responses" \
-H 'Content-Type: application/json' \
-d '{"model":"cke-local","input":"What is the capital of France?","temperature":0,"max_output_tokens":24}'Read status, output_text (or the message content in output) and usage. Our Qwen2 request completed with “The capital of France is Paris.” The formatted input consumed 47 tokens and the reply used seven. A client can show that result separately from a green health indicator.
3. Check what a tool-using client can rely on
- Bundle identity. GET
/v1/cke/loaded-identity. On an identified bundle compareserving_identityandsession_library_sha256with the expected build and retain theserver_instance_id. An older bundle can serve text and return 404 here. An identity-bound harness then lacks the evidence it requires. - Native progress. GET
/v1/cke/native-progresswhen available. This reports completed native work. A missing route means progress is unavailable. Record it as unknown. In experimental batch mode the optional observation route is/v1/cke/batch-observation. - Tool use. A client also needs a declared tool protocol and a real round trip. For a file-read task, verify that the harness read the intended file and that the model used the returned value. Compare tested harnesses.
If a check fails
- Connection refused. Inspect the server terminal. A download or compile can still be running. Poll the inexpensive health GET with a bounded deadline.
- 400 context_length_exceeded or 413 prompt_bytes_limit_exceeded. Shorten the formatted conversation, shrink the output allowance, or choose a suitable build and byte limit. Sending the same request again keeps the same problem.
- 422 template_render_failed. Check that weights, tokenizer, template and serving package match. An ordinary GET can succeed while every chat request fails here.
- 429 busy. Honor Retry-After when supplied. A generation request that timed out on the client may still be running. Check its response ID if you received one before starting another.
- 501 on tool use. Check the declared serving variant and protocol. Plain chat and tool calling have separate acceptance tests.