What problem does it solve? AI coding sessions tend to drift: they batch multiple tasks, wander into unrelated workstreams, or declare work done without running real checks. This Skill enforces a strict one-unit-per-invocation discipline, so every piece of work is executed, verified against the project's real gates, recorded, and reviewable as a single auditable receipt. ## Core Features & Use Cases - One bounded unit per run: Selects exactly one eligible unit from the focused workstream — a planning decision (shape), a ready roadmap chunk behind real quality gates (build), or one stakeholder signal (listen) — then stops. - Evidence-backed receipts: Every invocation ends with a structured receipt naming the unit, the checks run with real exit codes, cross-referenced think/ship/roadmap/signal records, discoveries, and the next candidate. - Honest stopping conditions: Emits explicit receipts for no focus set, no ready work, failed required checks, or decisions that belong to a human — never fabricates progress. - Spec Kit integration: Automatically uses .specify/ artifacts (specs, acceptance scenarios, task lists, constitution) as the source of requirements and definition of done when present. - Use Case: A developer says "do the next unit" mid-project; the agent loads the current focus, picks the frontier's next chunk, implements it, runs cargo test and clippy with real exit codes, records the trace, and stops with a reviewable receipt. ## Quick Start Ask the agent to advance the current work by doing the next single unit in the focused workstream and report the receipt.
