acervo-ship-queue

Sequential queue runner for acervo ship. One model at a time — when an item finishes (success, failure, or died), record its outcome and launch the next pending item. Continue until the queue is exhausted. Built on top of [[acervo-download-ship]] — this skill does not reimplement the launcher; it calls the sibling's start.sh / check.sh and adds a queue layer.

Why this exists

The sibling acervo-download-ship skill handles one model at a time. Real workflows often involve shipping a whole family of checkpoints (every quant of a Qwen release, every variant of a fine-tune) — and these can each take minutes to over an hour. Manually re-invoking the single-model skill after each completion wastes user attention and is exactly what /loop dynamic mode was built to automate.

This skill:

• Accepts a list of model ids (inline in the user's message or via a file path).
• Persists queue state at /tmp/acervo-ship-queue.json with per-item status (pending, running, success, failed, died).
• On each invocation, finalizes any just-completed item (parsing the sibling's tracking file + log sentinels), then launches the next pending item via the sibling's start.sh.
• Skips and continues on failure — a busted model id, a transient HF 5xx, or a CHECK 6 mismatch does not block the rest of the queue. The failure is recorded and the runner moves to the next item.
• Plays cleanly with /loop dynamic mode: the agent fires once to kick off the queue, then re-fires on a backoff schedule until the queue is complete.
• Never runs two ships in parallel — the sibling's /tmp/acervo-ship.tracking lock is the single source of truth for "what is currently in flight."

Locating the bundled scripts and the sibling