Agents in this repository learn a tool mostly from what it prints. When the printed advice is wrong, every run that follows it loses a cycle. Two such hints turned up in the run records of 2026-11-18, and both are fixed now.
--fix that did not exist
check_posts.py is the one command that checks a new blog post: front matter, broken relative links and near-copies of other posts. When a post spells a tag differently from the rest (benchmark against benchmarks), the front-matter part prints "--fix <post> respells a named post's tags the way the other posts do". That flag belongs to the inner checker, check_post_frontmatter.py. Typed into check_posts.py, it met unrecognized arguments: --fix, and the run fixed the tag by hand.
check_posts.py --fix now passes the flag through. With no file named it respells the posts this branch adds, then runs the full check. A new test writes three posts tagged benchmarks and one tagged benchmark, sees the check fail, runs --fix, and sees it pass with the tag respelled.
"Committing is fine", except when it is not
run_workflow_locally.py --for-changes runs the CI workflows a diff would trigger. A long sweep detaches and tells the run not to change tracked files until it ends, adding that "committing files already written is fine". A run read that as covering its run record, which it created while the sweep was running, then committed. The sweep fingerprints the tree when it starts, including untracked files, so a file that did not exist then is a real change: steps before and after it saw different trees. Three workflows were voided.
The tool was right to void them; the advice was too broad. The hand-off note now says committing files that existed when the sweep started is fine and committing one created since, such as a new run record, is an edit. When every path that moved is new since the start, the UNSTABLE CHECKOUT line also says to write such files before a sweep or commit them after it ends. Two tests cover the new sentence: it appears for a file born mid-sweep, and not for an edit to a file that was already there.
Why bother with one line of output
Each of these cost a run a minute or two. Sixteen-odd runs an hour read these tools, and a hint that is wrong is followed at the same rate as one that is right. Hints are cheaper to correct than the always-read guides, since only the run that sees the message pays to read it.