/stuck-crewmate-recovery

Skill type: Agent-only
Upstream path: .agents/skills/stuck-crewmate-recovery/SKILL.md

Escalating recovery for an ordinary direct report that is missing or stuck.

What it does

This skill recovers normal ship and scout direct reports without discarding their work.

It starts with evidence - current task state, endpoint presence, status, steering inbox, and existing worktree - then escalates from a simple answer or corrective steer to interrupt, relaunch, and finally failure reporting.

When to reach for it

Load it when session start reports an ordinary direct report with a dead endpoint or missing window, or after stale wakes, looping output, repeated confusion, an unanswered steer, or an apparently unresponsive worker.

How it behaves

The skill treats a dead endpoint as a presence signal, not proof that the task's work is gone. It checks for authoritative validation or existing task work before attempting replacement.

A relaunch keeps the same task identity and existing local copy. It does not create a fresh generic task while the recorded worktree is still unaccounted for.

Secondmates are explicitly out of scope; use secondmate-provisioning for them.

Common questions

Is low context enough reason to relaunch a worker? No. Modern harnesses may compact and continue.

What if the worker is asking something already answered by its brief? Answer it with one concise steer before escalating.

It's working if

Existing commits and uncommitted work survive recovery.
Simple confusion is corrected before expensive relaunch.
A genuinely wedged worker can be replaced in the same task copy.
After repeated failed relaunches, Firstmate reports the preserved work and plain consequence instead of silently losing state.

Where it fits

This is the recovery ladder for ordinary direct reports. It sits beside harness-adapters and below the separate persistent secondmate recovery path.

Upstream: .agents/skills/stuck-crewmate-recovery/SKILL.md