work

/work — Pick · Design-First · Auto-Fanout · 6-Gate

Announce: Using /work to [pick the next task / execute {{TASK_ID_PREFIX}}-S. / fan out N tasks].

One command from "a task exists" to "reviewed, ready to merge". It runs the discipline for you — design-first (A005), the cost-aware fan-out decision (../../rules/sub-agent-workflow.md §1.0/§1.5), the 4-layer Conflict Radar (../../../docs/playbooks/parallel-conflict-prevention.md), and the mandatory 6-gate review (../../../docs/playbooks/post-delegation-review.md).

Token budget

• Read only the active board (limit: 200) + the candidate task's design doc. Everything else via Grep (frontier scan, path-overlap, contract scan).
• Dispatch implementation to a sub-agent via a brief file, never a long inline prompt (see ../../rules/sub-agent-workflow.md §3).
• Don't re-Read what a returning agent reported; trust its summary, verify via the gates.

Forms

Step A — Classify the work

1. Ensure a sprint board exists. If docs/project/sprints/ has no active sprint folder (or its tasks.md is missing), scaffold one before picking a task: mkdir -p docs/project/sprints/S<N>/designs/_briefs, copy docs/project/_templates/tasks.md → sprints/S<N>/tasks.md, fill the Glance header, and pull in the backlog rows scheduled for this sprint (Grep docs/project/backlog.md for wip S<N>). This is how a sprint starts — there is no separate "new sprint" command.
2. Identify the task(s). No arg → scan docs/project/sprints/S<N>/tasks.md for the unblocked frontier (lowest priority-number first, dependencies met). Task ID(s) given → take those rows.
3. Read the Type of each → look up ../../rules/phase-matrix.md for the phase list (which gates run / run-light / skip / trigger). Quote the phase list back to the user before dispatch.
4. Readiness gate — confirm dependencies done, acceptance criteria present, not already in flight. Block + report if not ready.
5. Confirm-open gate (do NOT trust the Status column). Before committing to execute, run scripts/backlog.sh verify <B###> for the backlog item behind the task. Exit 0 = genuinely open → proceed. Exit 1 = already closed (it's in the cold archive or terminal-but-unswept) → do not re-implement: sweep the stale row (scripts/backlog.sh sweep) and pick the next eligible task. Exit 2 = unknown ID → don't assume open; grep the sprint records first. This is mechanism for the lesson that status drift causes re-picking shipped work — one cheap grep beats a wasted implementation.