---
name: backlog-autolab
description: Execute approved research backlog tasks autonomously in isolated worktrees, using native subagents when available and reporting a digest for human review.
---

Before acting, read [runtime guidance](../../RUNTIME.md) and [backlog-run](../backlog-run/SKILL.md).
Use this workflow only when the user requests autonomous execution.

1. Resolve named items, or the Ready queue in Work/README.md order.
   Eligible items are Ready, or Active with `> Autonomous: allowed`; resolve the existing work branch's item state before using the checkout's stale path.
   Exclude terminal items, `(human)` next tasks, malformed/legacy routing and `needs-human` hand-offs.
2. Report the queue, requested caps, required permissions and missing capabilities.
   Preserve current permission settings; a preflight is not blanket approval.
   Default to one item in flight and three tasks per item; honor user-selected lower limits.
   Serialize all Wolfram work on the shared kernel.
3. Claim each item in its own `work/<Item>` worktree outside cloud synchronization, following backlog-run.
   A busy branch is held by another session; do not reuse or force-remove its tree.
   A Ready item moves to Active with the autonomy marker in that branch only.
4. With native agent tools available, dispatch one fresh worker per task with the absolute worktree path, item, exact next task and backlog-run SKILL.md path.
   The prompt must state autonomous mode, one task only, no other checkout, no merge/push/publication, preserve human edits, and commit work plus bookkeeping on the work branch.
   Ask for task, commit, outcome and a one-line note in the final report.
   Inherit model/effort unless explicitly selected in accepted task routing and supported by the host.
5. On completion verify: clean tree, HEAD advanced, item still present, the exact task moved to `### Done`, no unexpected open task changed, no `needs-human` hand-off.
   Stop that item on any failed check, permission denial, `(human)` task, terminal state or configured cap.
   Do not continue after an uncertain tool outcome without checking current state.
6. Report per-item branch, commits, checks, stop reason and unresolved decisions in a digest under Work/Runs/.
   Keep dirty worktrees and name them; clean worktrees may be removed without force after the digest.
   Human review and explicit merge remain the integration gate.

If native subagents are unavailable, use [the bounded CLI workflow](../backlog-run-scheduled/SKILL.md) when local Codex exists; otherwise report the limit and offer backlog-run for one task.
Do not claim background workers survive the conversation unless the host guarantees it.