Every agent run here ends with a run record, and each record lists the instructions or tools that cost it time ("doc friction"). About one run in ten is a process run. When nothing is waiting in the process backlog, it reads a table of those reports, newest first, and fixes one.
On 2026-11-12 one report said the pre-merge preflight for products, venture-preflight.mjs, had passed a new page that CI then failed on faq-visible.test.mjs. Two process runs started within the same hour. Both read the same table, and both picked that report:
| opened | merged | |
|---|---|---|
| PR #15090 | 10:23 UTC | 10:34 UTC |
| PR #15094 (this run) | 10:28 UTC | no, its preflight change was dropped |
Before starting, this run searched open pull requests at about 10:13. The other run's PR didn't exist yet. Both arrived at the same design on their own: run the 76 fast platform/pages test files (1,249 tests, 27 to 35 s) before the build, skip the two that build the whole site, and confirm it by planting an invisible FAQ answer on a real page. When this run went to merge, main already had #15090. Its version stayed, because it fails on any red test, as CI does. This run's version only failed when the output named the product.
The overlap cost about an hour of agent time. An earlier record from the same day had already predicted it: "Critique mode sends every process run to the same 'Doc friction by file' table, newest first, so concurrent runs converge on the same newest report." An open-PR search can't see a run that hasn't opened its PR yet. Picking from somewhere other than the top of the table, or opening a draft PR before writing any code, would have avoided this one. Both changes live in owner-approved files, so this run left them in its record.
What this run did land
A smaller fix to platform/ci/wait_for.py, the helper runs use to wait for a background job without sleep. When the harness finishes a background task, the task's output file ends with [exited with code 0]. The helper only knew DONE and EXIT lines, so a wait on that file ran its full 110 s timeout after the job had already finished. It now accepts that line, and also a - after the marker (DONE-EXIT 3, which another run reported the same morning). Later waits in this run on task output files returned as soon as the job finished.