Task format

The task-format reference defines project and area metadata, task files, and task bundles. The same reference is bundled with the installed skill.

Its research contract defines area research, mixed hypothesis bundles, and task research evidence. Task completion and experiment decisions remain separate. The local experiment path executes correctness checks and measurements, retains candidate Git code, and keeps or restores owned files before completing and committing the result. requires_code admits later work only after verified acceptance. The installed Codex research route and Claude serial research route generate scoped supporting tasks and hypotheses and coordinate execution. Codex also coordinates rolling source worktrees. Claude rolling research and execution routes for other harnesses follow separately.

The research run contract authorizes generated mixed work, counts dispatch and command costs, and retains limits through recovery. Its public commands expose current accepted samples, task lessons, unfinished work and remaining allowance. Evaluator or accepted supporting-code changes require an actual counted rebaseline. Queue exhaustion and stopping a run leave the area open.

Isolated hypothesis candidates use one authoritative checkout for records and acceptance. An authorized research run admits linked source worktrees with research experiment <area> <task> admit --source <root> --path <file>. Workers read authoritative records by absolute path and edit only source code. After confirmed worker stop, return --worker-stopped retains Git transport; evaluate measures in the source; integrate claims the serialized destination acceptance lane and captures its verification snapshot. reconcile refreshes a stopped candidate after baseline drift and charges fresh measurement work. Separate positive --worker-limit and --measurement-limit run capacities default to one.

Each attempt keeps its comparison facts. Evaluation changes pause candidates until an actual counted rebaseline; resumed candidates reconcile and remeasure. Historical isolated rejection retains evidence and records accepted_revision for the current destination. It preserves accepted sibling code. Inspect/recover identifies active, returned, terminal, paused and ambiguous candidates, retaining unknown worker/process capacity. The binary never removes worktrees. See the installed task-format and recovery references for the complete command contract.

reserve --worker-slot counts native planning, implementation, verification and finding checks against worker capacity, separately from measurements. Exact retries never grant another dispatch; task, kind and requested mode stay fixed. Admit and Prepare transfer their own unused claims without double-counting. The integrated acceptance owner may claim confirmation or repair capacity. Prepare records authorized withdrawal, restores only exact measured destination paths to unchanged accepted HEAD and preserves unrelated work and historical evidence. Recover completes that withdrawal without worker permission. Source repair then requires Return, fresh Evaluate, Integrate and a fresh verifier.