[matchalatteguy]

Keep task claims and queue replenishment consistent across failures and competing local operators.

This 0.2.0 update atomically claims tasks and worker capacity, records launch and timeout failures instead of leaving tasks running, preserves intervening operator state changes, and applies each replenishment plan with its tasks, events, and cursor in one transaction. Invalid output bytes do not abort subsequent workers. Applied duplicate-only plans still advance the cursor so later cycles can recover. Dry-run, duplicate, and stale plans cannot silently create partial work.

The README leads with a real documentation audit that produces three reports and persists task outcomes. It explains synchronous dispatch and the distinct detached-worker role of Local Agent Task Runtime. The changelog records the explicit dry_run=False requirement when constructing a plan for application. Pinned CI checks Python 3.11–3.14, builds the package, and executes the installed demo outside the checkout.

Validation: 41 tests passed; Ruff passed; locked install and refreshed source/wheel builds passed; the isolated installed-wheel demo produced the documented three audit reports. A separate review reproduced and verified binary-output recovery and duplicate-only cursor advancement. Tests cover competing claims, capacity, timeout and launch recovery, stale plans, duplicate titles, and transactional rollback. This remains a synchronous single-machine tool; process trees, distributed leases, and automatic crash recovery are outside its scope.