dispatch-task-startlisted

Sweep the caller's prior `in_progress` tasks (close those actually done, block those actually blocked, escalate ambiguous ones) before issuing the `start_task` for the new taskId. **Canonical source**: VantageRegistry (`get_skill_content name=dispatch-task-start`). The local `.claude/skills/dispatch-task-start/SKILL.md` in each workspace MUST be a byte-exact mirror of the VR canonical content. Fetch from VR, do not edit locally. This skill exists because `enforce-irp-sequence.py` rejects any `start_task` call when the caller already has an `in_progress` task. The fix is not to bypass the hook but to pre-satisfy it: sweep first, start second. ## WORKFLOW **Step 1 — Resolve identity and target** 1. Detect your orchestrator role (and instanceId where applicable) from `CLAUDE.md` header / hostname (sigma, eta, alpha, lambda, tau, phi, omega, zeta, pi, …). 2. Extract the target `taskId` from the user invocation (e.g. `start task k9abc…`, `pick up k9abc…`). If the user did not name a task, halt and ask which task to start — do not guess. **Step 2 — List your current in_progress tasks** Call `mcp__vantage-peers__list_tasks` with `assignedTo=<your role>`, `status=in_progress`, `limit=50`. If the list is empty, skip directly to Step 5. **Step 3 — Classify each stale in_progress task** For each task returned by Step 2, decide which of three buckets it belongs to. Use the task's `description`, `completionNote` history, linked PR/commit, and your own session context. Bucket A —