The launcher: it starts rooms smallest first, up to a set number at once.
What it holds
python3 _tools/coding_queue.py [--parallel N] [--inspect] [--clear-pause] [--requeue-failed].
A room is ready when it has 500 to 50,000 in-window items, no failure,
no run/, no store or archive, no earlier coder run (except
liverpoolfc), and a brief that passes. It launches
start_yard_lead.py <sub> --complete for each, or a
repair script for a failed room, keeps
_coding_queue/STATE.json, EVENTS.jsonl and a
log per launch, pauses on a DRAIN file, a usage limit, an
unexplained exit or low disk, and stops retrying a room after two
failures. A room counts as finished when its STATE.json
says verified and validation/RUN-RECEIPT.json
exists.
Who writes it, and when
The yard's author. Run by the scheduler; the model is fixed at
gpt-5.6-sol, medium.
Why it exists
Its docstring: "Owns scheduling, never research decisions."
Where
_tools/coding_queue.py. The repair and progress scripts it calls are not among these sources.
Checked against: _tools/coding_queue.py, _tools/start_yard_lead.py