Model errors

Model selection is validated when agent configuration is saved, but provider availability can also change at runtime. Treat configuration failures and temporary provider failures as different recovery paths.

Symptom#

Saving or running an agent fails because no active model adapter exists or the selected provider does not complete the request. Test Chat may stop before producing an assistant response.

Cause#

llm_unavailable means the model adapter or upstream model request is unavailable. A model listed in configuration is not a promise that its provider is healthy at every moment, and unsupported temperature or capability settings can be rejected before runtime.

Verify#

Open the agent Model settings and confirm the selected model and supported options. Check whether the same published version fails in a new Test Chat. Record the model identifier, agent and version identifiers, UTC time, and error code. Do not copy prompts, customer messages, credentials, or provider responses into logs used for escalation.

Recovery#

If the setting is unsupported, choose an available model or remove the incompatible option, save the draft, and test again before publishing. If a previously working version reports a transient provider failure, retry once after a short delay. Do not loop indefinitely or silently substitute a different model in production.

Escalation#

Escalate when two clean tests of the same version fail, or no listed model can be saved. Include safe configuration metadata and whether the failure occurred at save time or runtime. See Models, Test Chat errors, and Support.

Error code index#

• llm_unavailable — the configured model adapter or provider request is unavailable.

Publishing does not repair a provider outage. Validate the draft in Test Chat, then publish only after the intended model succeeds.