opsx-exttasklisted

# Autonomous Task Workflow Drives a single OpenSpec change end-to-end. The whole autonomous pipeline — explore, generate artifacts, self-review, implement, test, verify — runs inside ONE deterministic Workflow. This skill does only what a workflow physically cannot: check preconditions, and handle the parts that need a human or that spawn subagents (the review gate, and archive + commit). An A/B on this design measured it ~1.8× faster wall-clock-to-done than driving each phase from the main thread, because it collapses the many main-thread round-trips (one per phase, plus one per self-review and verify loop iteration) into a single hand-off. The gap widens with the number of phases and loop iterations. ## Why the split is exactly here A Workflow runs in the background, returns one result, and its leaf subagents have the `Skill` tool but NOT the `Agent`/`Task` tool. That fixes the boundary: - **Inside the workflow** (Phases 1–2f): explore, generate artifacts, self-review, implement in waves, test, verify. The workflow invokes the `openspec-explore`, `openspec-ff-change`, and `openspec-verify-change` skills directly — all three run non-interactively (explore has no mandatory prompt; ff/verify are given the change name/description so they never ask). It does NOT use `openspec-new-change` (halts for the user) or `openspec-apply-change` (serial) — `new` is subsumed by `ff`, and `apply` is replaced by script-driven wave fan-out. - **Outside the workflow** (this skill): the hum