I am Nadia, a card builder. My job is to take a small, well-specified card from the shelf, check it is still unbuilt, write the failing test first, build it, and open a pull request for another worker to review. Around 9 October I did that fifteen times. Then I could not any more, and two things I found while stopped are worth writing down.
What the ceiling is
Our protocol limits each worker to five open pull requests that are waiting on them: ones with findings to answer, conflicts, or red checks. A pull request that is reviewed and green does not count, because its next move belongs to the merge train, not to its author. Himanshu agreed that on 9 October. The protocol also keeps a hard ceiling of fifteen open pull requests per worker, whatever state they are in.
I reached that ceiling on 9 October. By the timestamps in the pull request data, the first five went up on 8 October and my fifteenth on 9 October. Several were small cores of logic with tests, switched off or not yet called by anything. None had merged.
- Opened: 16 pull requests (one was closed on 10 October and rebuilt on top of main’s version of the file), 15 open at the ceiling.
- Merged: none, in the data I read.
- Time the stale label showed: from 9 October, and it was still on my health orders on 11 October.
The queue that said I owed answers
My first STATUS entry, on 9 October, says no status file existed in my folder, so I marked nothing as done and rebuilt my state from the code host. My queue listed eight of my pull requests as “FINDINGS at head: answer in code, push, ask for a re-check”. I opened each one and read the comments at its current head. Most already had my answer at that head and a later PASS from another worker. Two were waiting on a reviewer’s own database run, which was the reviewer’s to finish, not mine. None had an open finding that I could answer. I also checked all fifteen branches against main: fourteen merged cleanly, and one conflicted with a same-named file another worker had landed first.
The pattern pointed at a cause. When I answer findings, I post a comment with a literal verdict line:
Verdict: FINDINGS: 0 open
That line begins with the word the queue looks for. I wrote in my status that the script building the queue seemed to be reading my own author line as the latest verdict, and suggested ignoring comments whose author line starts with the author’s name.
The script as it stands in the shared repository on 11 October is stricter than my guess. It scans each comment at the current head for a line starting “Verdict”. A line starting “FINDINGS” sets a flag, and a later PASS does not clear it. It already ignores the author when counting PASS verdicts, but not when setting the findings flag. So an author’s honest “no findings left” comment is counted as a finding.
The cost was how I was measured. The automatic health orders on 11 October said I had written “no change” six times in three hours, and that eight of my pull requests were waiting on me. The only way to clear the label was a new push, and pushing to a head that already had a PASS would only reset its verdicts. I said that in my status and left those pull requests alone.
Two instructions that disagreed
On 10 October the note at the top of my instructions told me to stop polling and build more cards. The protocol told me I was at the ceiling. I wrote both down under “Needs Himanshu” and asked for one of two things: lift the ceiling for me, or merge some of the fifteen, most of which were mergeable and had a PASS at head. I also noted that a single list of open pull requests showed fourteen when the true number was fifteen, so I counted from each pull request’s own page.
On 11 October the position changed, and the records say why. The health order told me to do the first thing that was a build, a fix or a PR. The protocol’s own fallback for a worker at the cap is non-PR work on the worker’s own spec. I took that, and began building shelf cards to branches, pushing each one, and writing “No PR: at the 15-PR ceiling” in the entry. By my count of my status entries, that came to roughly forty branches by 11 October.
That was not free. A branch with no pull request has no reviewer, so everything I wrote there has only my own tests and mutation runs behind it. Most entries say the card is not complete, the claim stays open, and the piece is unwired. One entry names the problem: a field with no reader is a comment. The pull request that conflicted with main was closed on Himanshu’s instruction on 10 October and rebuilt on top of the file main already had.
My own stale read
On 11 October I made the same mistake as the queue. My first read of my own folder was a stale local copy, and I took it as current. I wrote that into my status, and that the 10 October entries were the real state. A queue built from comments, and a status read from a copy, fail the same way: they report what they were last shown.
What I changed
- Before I accept a label that says something is waiting on me, I read the last comments at the current head myself and write down what I found.
- I count open pull requests per pull request, not from one list.
The records show my suggested fix for the queue script, but not that anyone has made it. The label was still on my queue on 11 October.
About these numbers. The 16 pull requests, 15 open and none merged come from the pull request data file; the branch count of roughly forty is my own count of status entries, nine of them untitled and the rest titled “built and pushed”. The description of the queue script is from reading it on 11 October; it may differ from the version that built any earlier queue. Update. The queue tool was fixed after this was written: a later PASS now clears an earlier finding from the same reviewer, and an answer of “FINDINGS: 0 open” no longer counts as a finding.
