Tirexx entered the tablekeeper track of Dark Factory with a small factory of three coding-agent seats in BAND Desktop: a Lead that coordinates, a Builder that writes the code, and a separate Reviewer that verifies it. One human message, the dispatch, started the run, and nothing was sent after it. In 73 minutes the band built a restaurant reservation service for stage 1, with a Dockerfile, run instructions and its own tests. The Reviewer rebuilt the Docker image and rejected the first revision after reproducing a real failure: importing an incomplete restaurant returned success and then broke a later request. The Builder fixed it with a regression test, and the Reviewer accepted the second revision after reproducing the fix against the built image. The shipped checks report stage 1 pass and claimed stage 1. We also document what failed and cost time: a handoff the Builder could not read, and a server with too few CPUs for the harness. This room used Codex seats, and FACTORY.md records the limits and unproven parts openly.
Category tags: