Every agent session in this repository writes a run record: a JSON file under prompts/runs/ that says what the run did, what went wrong and which instructions cost it time. promptctl.py record writes the skeleton with outcome: null, and the master prompt says to fill it in at the end, once the pull request exists and its number is known.
Venture runs also run platform/pages/venture-preflight.mjs before they merge. One of its steps is platform/ci/main_health.py, which runs every repository check, promptctl.py validate among them. And validate rejects a record whose outcome is still null, which is correct: an unfilled record must not reach main.
The two orders clash. A run that writes its skeleton early, as the prompt suggests, and then runs the preflight sees the preflight fail on its own blank record. Three run records from 2026-11-14 and 2026-11-15 report exactly this, one of them as an error that cost a second preflight pass. A preflight that fails for a reason that is not a fault teaches a run to skim past failures, which is the wrong lesson.
What changed
main_health.py now tells that one case apart. When promptctl.py validate fails and every error it prints is "outcome must be one of ... (got None); fill it in before merging", and origin/main holds none of those record files, the check's line reads PEND instead of FAIL. It names the record and says to fill it in before merging, and the exit code stays 0, so the preflight carries on.
The excuse is narrow on purpose:
- any other validation error, on any record, keeps the check red;
- a record that
origin/mainalready holds with a null outcome is a real fault onmainand stays red; - an outcome that is filled in but wrong (
"done") stays red; - if the clone cannot read
origin/main, nothing is excused.
The pull request's own CI still runs promptctl.py validate in full, so a run that forgets to fill the record is still stopped before it merges. Seven unit tests in platform/ci/test_main_health.py pin those cases.
Why it is worth a change
It is a small thing. But about sixteen sessions start every hour, many of them venture runs, and each one that hits this either loses a preflight pass or learns that a red preflight line might mean nothing. Both are cheap to prevent and costly to repeat.