Written by an AI agent working in this repository.
This was a process run. The process backlog held only rows that wait on the owner, so I took the friction that run records reported over the last two days (platform/ci/friction_rank.py --live --ungated --days 2), skipped the clusters that had already been fixed, and shipped three small fixes. Each one comes from a doc_friction line in prompts/runs/.
1. A probe that missed the host it needed
portfolio/flagships.yaml lists the flagship products and, for each one, a buildable: line of work a cloud session can do now. allocate.py --venture <slug> probes the hosts that work needs and prints which ones answer from this session, so nobody reads a long row only to find the data can't be downloaded. It probes the row's hosts: field plus any host name written in the row's prose.
ai-bookkeeper's top buildable item is re-reading receipt photos from the CORD dataset. The row names the script that fetches them, eval/cord_images.py, but it never names the host. Two sessions on 2026-11-17 were sent there, found that the proxy refused huggingface.co, and reported that "the row names no hosts: list, so allocate.py could not warn".
That fix had two parts. The row now has its own hosts: [https://huggingface.co/, https://raw.githubusercontent.com/]. More generally, row_hosts now opens each script that buildable: names under the venture's folder and adds the hosts its URLs point to, up to four of them, with licence and reference hosts left out. Across today's five rows that set buildable:, it finds:
| row | hosts read from the named scripts |
|---|---|
| ai-bookkeeper | raw.githubusercontent.com, huggingface.co, cdn.jsdelivr.net |
| shift-roster-builder | api.7shifts.com |
| photo-editing, open-tennis-robots, ids-model-check | none (their hosts are already in the prose or hosts:) |
My first version skipped comment lines, on the theory that a URL in a comment is a licence link. That version found nothing for ai-bookkeeper, because cord_images.py gives its download URLs in its usage comment. Comments count now, and the test fixture keeps that case.
2. A command cut where the meaning was
platform/ci/wait_for.py waits on a background job's log. When it times out, it says which process still has the log open. It cut that command line at 60 characters, keeping the front. One session read ... run.py --draws 10 for a job started with --draws 1000 and briefly thought the wrong run was going. The interpreter path at the front is the part a reader can do without, so the cut now goes in the middle: a third of the width from the front, the rest from the end, so the arguments survive.
3. A to-do list that was one line
Saints by Candlelight's episode list told sessions to mark a finished entry *(Done: ...)* at the start of its line, and list_done.mjs moves those lines to DONE.md. But every remaining idea sat on a single "Next up, suggestions" line. Taking one idea meant hand-editing that line, and once another session had claimed an idea, a session found nothing unclaimed it could take (two reports, 2026-11-16 and 11-17). The list now has one entry per line. It also gains the two Egeria sequels that episode 133 left in DONE.md but never added back to the list (Lent, and Pentecost to where the manuscript breaks off), and the instructions ask each new episode to add its sequels as new lines.
What I checked
python3 -m unittest test_wait_for(40 tests) andtest_allocate(116 tests), each with a new test for its fix.node list_done.mjs --checkon the new list: no Done entries left behind.allocate.py --venture ai-bookkeeperfrom this session: both hosts answer right now. That is the point of probing: the answer changes from session to session.main_health.pybefore starting: 254 of 254 checks green on main.
None of these is large. Each one saved the next session a round trip that several earlier sessions paid.