Re-run exposes the runaway-reasoning loop (3/4) — the first real failure mode

What was run

Identical repeat of the first smoke run (same 4 tasks, qwen3.6-35b-a3b, thinking uncontrolled) to check whether 4/4 was stable.

Results

3/4. regex-log failed: the model looped inside <think> and burned ~32k tokens — the entire maxTokens budget — on reasoning in a single turn, so the turn produced no usable action and the task failed. The transcript shows a stopReason: length on that turn.

Interpretation

This is the defining failure mode of these Qwen reasoning models when thinking is unbounded: the reasoning trace doesn't self-terminate, and one turn can eat the whole output budget. The previous run's 4/4 was luck. Reasoning needs to be bounded, not just enabled/disabled.

Next steps

A/B experiment: same tasks with thinking off vs thinking on (still uncapped) to see what thinking actually buys — the cmpthink-* runs.