loc-new

Provide a snake_case candidate name and a research-backed upstream pick, then the skill copies the _template realization, renames the module class, vendors the upstream repo with provenance, and seeds the README hypothesis and journal. The scaffold finishes by running the contract test suite, which must pass green.

You need a valid snake_case name that becomes a Python package and a research pick specifying the approach plus upstream repo or library. Without a research pick the skill stops, because algorithm selection belongs to the earlier research phase, not the scaffold step.

Yes. Pip-pinned libraries skip the vendoring step entirely; only repo adaptations are cloned at a pinned SHA with an UPSTREAM.md provenance file. Dependencies are declared in an environment.yml recipe named bench-<name>.

The playbook mandates that phase 2 (hypothesis) precedes phase 3 (build), so the README must state the approach, hypothesis, and expected failure modes referencing the benchmark board. This prevents scaffolding candidates with no measurable claim about improvement over prior attempts.

It never chooses the algorithm, never touches locbench internals, episode sets, or other candidates' folders, and never deletes a botched scaffold itself. Adaptation of the real algorithm in module.py is deferred to later loop phases, not scaffold time.