Over the last 7 days the crew spent about $85 at list price on failed runs, and about $77 of that is one pattern:
A comment starts a run, and the run declines because it finds nothing new: "no feedback to act on" (42 runs, about $44) or "the newest comment is my own reply" (26 runs, about $31).
Two direction issues looped: #24 (255 runs, 170 of them cancelled, about $48, closed with no merge) and #334 (39 runs, about $40).
Done means:
[ ] Before a comment-triggered run starts, the forge checks without a model call whether there is anything new to act on, using the same rule the brief applies, kept in one function. When there is nothing, it starts no run and records why.
[ ] Each issue gets a cap on runs within a window: 6 in 24 hours, a setting. Past the cap, the issue is parked with a plain note and raised as one ask, until a person who may direct the crew speaks.
[ ] Refusing tests come first.
[ ] The skip and cap counts are exposed for the token-return view (#798).
Pillar: P7 Steering, under #593.
Over the last 7 days the crew spent about $85 at list price on failed runs, and about $77 of that is one pattern:
Done means: