Stop a scheduled agent collect from tearing down its own schedule.
Improve runbook to more decisively prompt for signoff gates.
0.2.4 (08 September 2026)
Bound adaptive model connections: spawn workers with a fixed range (min 20, max the samples ramp’s top) and raise the ceiling only on genuine sustained saturation.
More accurate model context window detection, using the window the definition resolved at capture and checking it during smoke runs.
Ensure that tool call approval is not applied to llm_scanner answer() tool.
Prevent overlapping scheduled agent collects: a collect that outruns its interval holds a lock so the next scheduler fire is skipped rather than stacking a second agent on top.
0.2.3 (07 September 2026)
Improved handling of live scan results (fold periodically, cleanup local buffer).