We found and fixed a bug where any coding-agent run longer than about 15 minutes was silently cut off mid-work. The isolated environment the agent works in had an inactivity timer that miscounted the agent's activity — from the outside it looked idle even while it was busy editing files and running tests, so the environment was shut down mid-run. Most runs finished well under the limit, which is why this took a while to surface.
That timer is now disabled. Runs still have a sensible overall time budget and finished runs still clean up their environments immediately, so nothing is left running longer than it should. Larger fixes — the ones that take real time to write and test — now run to completion.