Implement a hypothesis

You own one hypothesis (a plan) in one worktree. You never edit plan files. Your brief names the claim, the pass criteria, the parent commit and the scope. Details of the steps and the handoff are in handoff.md.

Harness-specific mechanics (tools, background tasks, script paths): read references/harness/<harness>.md for the harness you are running in.

Start

1. Run the first command of your brief and confirm the worktree and branch. Do not go on if they differ. Everything you do happens in that worktree.
2. Read OBJECTIVE.md in the task directory. Its rules bind you, whatever a measurement says.
3. Load vibesys:measure and the domain skills the brief names.

Work

1. Start from the first measurable step: the smallest change to the claim that you can finish and measure in this turn. If you cannot finish a step, do not stop on partial work: name the concrete blocker (the file, function, check or error) and the narrower step that would get past it.
2. Build, then run the cheapest check that covers what you changed. Once it builds and one local check passes, commit and measure; do not polish first. After a failure, change the code before you run the check again. When output is long, write it to a file and read the file.
3. Commit in your worktree. If your brief grants no measurement slot, return the SHA as ready for measurement; the lead measures it before judging it. Otherwise measure that commit with vibesys:measure: the correctness check, then at least 5 benchmark samples, recorded in the state directory (--dir; vibesys:state). Compare with the base: compare --commit <sha>. Add --base <sha> when you built on a candidate, to see the gain over your parent.
4. With the verdict in hand, build the next step on it, or stop.

Use the measure skill's scripts/service.py owner for every required service. Supply --service, --probe, and --client as JSON argv arrays, with a positive --timeout. It starts a new process group, fails on exhausted readiness, and stops and verifies the whole group on every exit. Never use handwritten background-shell PID or startup-sleep recipes. Run any load, including short sanity/tuning benchmarks, through the script measurements.py from the measure skill's scripts/ directory: python3 measurements.py run -- <command>. This host-wide lock applies to recorded and unrecorded benchmarks alike.

Never

• Edit anything under .vibesys/, the task's checker or its benchmark.
• Add a stub, a special case for the benchmark's inputs, a skipped check or any shortcut that raises a number without the program doing the work. Obey every rule in OBJECTIVE.md.
• Report a number that is not in the measurement record.

Hand back

Claim, commit SHA, the verdict (the compare output), what you learned, and the step you would take next. If you made no candidate: the blocker and the narrower step.