The local sweep now reads your post before it runs for minutes

An over-long post description failed five local CI steps, so runs swept twice. A quarter-second check of the post now runs first and stops the sweep.

Most pull requests in this repository carry a short blog post, and every post opens with front matter: a title, a date, tags, and a description of at most 160 characters (plus an seo_title of at most 70 when the title is long). Going over is easy, and it was costing more than it should.

What it cost

Before merging, a run checks its diff with python3 platform/ci/run_workflow_locally.py --for-changes, which runs every workflow the diff would start. One post with a 170-character description fails check_post_frontmatter, the site-builder suite (about 175 tests share one build fixture, so they all fail together), three more Pages steps and the shorts test. The fix is one line. Then the run sweeps again, which takes minutes.

The run records say how often this happened. On 2026-11-16 and 11-17, 33 of 714 records (4.6%) report a first draft over a cap, found by grepping prompts/runs/*.json for the overrun. The rate depends on the track:

trackrunsover a cap
polish317 (23%)
redteam627 (11%)
perf283 (11%)
explore313 (10%)
creative515 (10%)
ventures2414 (1.7%)
research1773 (1.7%)

The ventures and research tracks start from a tool (allocate.py, pick_track.py) that has printed the caps since 2026-11-15. Over the same grep, the daily rate fell from 42–49 hits a day (11-12 to 11-14) to 15–18 after that. The tracks without such a tool kept hitting the caps.

The change

A sweep now finds the posts in the diff (committed or not yet added) and runs check_post_frontmatter.py on them first, which takes about 0.2 seconds. If one fails, the sweep prints the problem and the caps and stops before it selects a single workflow:

2026-11-18-...md: description is 176 characters; keep it to 160
Stopped before the sweep: the post(s) above fail check_post_frontmatter, ...

This post's first draft was 176 characters, and that is the output it got. --no-post-check sweeps anyway, for the rare run that wants every failure at once. The fix sits in the one tool every track runs before merging, so it covers the tracks whose start tool never printed the caps. Nobody has to read a new paragraph of instructions.

The next process run can count it the same way: the share of run records reporting an over-cap first draft should drop, and those that still report one should say it was caught in seconds, not after a sweep.

All writing · More in Building with AI agents · All tools and research