Provider health checks can time out on thinking-enabled models

[szafranski]

Problem

Provider health checks use a short prompt and a 30-second timeout, but the probe is initialized like a full agent session. Thinking-enabled models can spend the timeout on reasoning and tool context instead of returning the health-check response.

A working provider was reported unhealthy after 30.13 seconds even though regular requests succeeded. Disabling thinking alone was insufficient because the probe still advertised 10 agent tools.

Expected behavior

A health check should test provider reachability and basic generation without advertising agent tools or enabling thinking.

Proposed fix

Build the health-check engine with thinking disabled and an empty tool policy. Keep this configuration scoped to probes so normal conversations remain unchanged.

Validation

With both changes applied, the same live probe completed successfully in 18.77 seconds within the existing timeout. Advertised tools dropped from 10 to 0 and input tokens dropped from approximately 2,850 to 449.

Targeted tests, formatting, clippy, diff checks, the debug build, and the full workspace pre-push gate pass.

[szafranski]

There is also a broader reason to keep this probe lightweight: more providers now run models locally, including setups where model weights are loaded from SSD on demand. Cold loading and initialization can consume a significant part of a 30-second budget even when the provider is healthy.\n\nIncreasing the probe timeout to 60 seconds could be considered as an additional safeguard for slow cold starts. It should complement, rather than replace, disabling thinking and tools for the health check, since avoiding unnecessary probe work improves both latency and reliability.