Skip to content
MarginaliaDocsGitHub
Contents
handovers/2026-08-19-foreman-2-addendum.md

Rendered from the repository — the file stays the source of truth.

Foreman-2 handover addendum (2026-08-19, early day 3)

Supplements handovers/2026-08-18-foreman-2-handover.md (whose craft items and checklist still hold) with everything that landed after it. Read the base handover first; this note only carries the delta.

Board delta (all merged)

  • C8 (#54) history + architecture views current through the board; scrollable docs sidebar (owner-requested mid-session).
  • C9 (#57) generative view — model inventory (verified row-by-row at review), RAG pipeline, no-spend paths. Follow-ups landed: #58 (stale PGlite comments in webapp code, comment-only) and #59 (physical-topology diagram refreshed to merged-board reality).
  • C10 (#60) scope-vs-roadmap reconciliation + #61 FAQ phrase.
  • C11 (#63) “Marginalia in numbers”: dev cost ≈ $854 API-equivalent (28 transcripts; coordination 40.1%; cache reads 72% of spend) + ops model. Supersedes foreman-1’s “~$300” estimate (annotated there).
  • C12 (#65) research/plan split: product/scope.md retitled in place as the research catalog (filename + CF/SF ids kept — recorded citations stay true; phase groupings demoted to reading-order clusters), new product/target-scope.md defines Marginalia v1 with the same badge machinery and anchor-linked research ids. Root AGENTS.md now names the three-document set. + #68 FAQ phrase.
  • In-numbers cost-model corrections (#64, #66, #67) — see incident below. Final state on main (#67): fixed-vs-variable split per scenario; warm container = fixed cost of real users in both scenarios (scaled-to-1 floor ≈ €34.8/mo, bursts to max_scale = 2 bill on top, ≈ 2× bound); realistic 10-user all-in ≈ €43/mo (variable ≈ €0.34/user); ceiling ≈ €370–405 + $250/mo with Azure S1 + Supabase Pro as required tiers beyond the free allowances.

Incident on record (drove a new memory rule)

The foreman made three successive rounds of direct content edits to in-numbers.md on owner chat feedback; an interrupted command chain merged an overcorrected intermediate (#66) to main, which the owner then had to correct again, and the owner asked why the foreman was doing content work at all. Resolution: #67 hot-fix restored the intended content (owner-worded merge); the boundary is now in auto-memory (foreman-direct-merge-boundary): the foreman’s direct lane is roadmap rows, one-line record fixes, annotations, and pointers — substantive content goes to briefed sessions even when the owner directs it in chat. Successors: hold that line.

Open items

  1. Intermittent local test flake (demo-prep investigation item): 1 test fails in roughly 2 of 3 fresh-worktree bun test runs on this machine, then passes; failing runs also showed a truncated count (150/154). CI’s isolated pgvector container has never failed. First seen during the C8 review (2026-08-18); reproduced across several review worktrees; docs-only diffs, so no session code caused it.
  2. D2 Azure-Speech Terraform analog decision still open (input: B5 handover’s “what the provider cannot manage”).
  3. Cosmetic only: dated pre-C12 docs (feasibility, ui-research, two architecture views) still use “Phase n” as then-current cluster names — verified not-false at C12 review; no action required.
  4. Day-7 demo prep is all that remains: walkthrough script, before-demo checklist (base handover items 1–3 + seed the demo account, item 4), and item 1 above.

Craft additions (continue the numbering)

  1. AskUserQuestion answers arrive as tool results; a long compound commit && push && pr create && merge chain can partially complete before a user rejection lands — after any interrupted chain, verify what actually executed (git log, gh pr list, ls-remote) before reasoning about state. That is how #66 merged unnoticed.
  2. The in-numbers dev-cost table can be refreshed with the method in its own appendix; C11’s row for itself was measured mid-session and undercounts.

Second delta (2026-08-20) — docs lane closed

  • C13 (#73) target-scope re-baseline. Two decisions worth carrying: product/scope.md no longer tracks implementation status at all (status table, 37 per-item badges, the ⏳ glyph — all removed; nine badge-only facts relocated into target rows first), and product/target-scope.md measures delivery against the target, not the research catalog, which puts all 18 rows at ✅. The owner’s framing: the catalog describes Gemini Notebook, the target document commits to Marginalia v1, and only the latter carries status. Reviewer-escalated judgment kept as-is on owner’s word: SF-11 stays ✅ though the audio regeneration path escapes the daily quota (bounded by the 1-concurrent and 20-per-notebook caps; stated in its row).
  • C14 (#75) in-numbers Part 2 reworked into a pricing/ROI basis: every cost table closes with the same four rows (fixed/variable × system/per-user), invented total names gone, and a growth view over 1/10/100/1,000 users naming each tier step. Headline numbers: fixed ≈ €39.75/mo before any user exists; ≈ €0.335 per user per month, linear; steps at Supabase Free→Pro (≈115 user-months) and Azure F0→S1 (exactly 50 users). max_scale = 2 is recorded as a capacity finding — under heavy usage the deployment stops serving around 100 users rather than producing a bigger invoice. C14 also corrected C11’s per-chunk storage math (≈19.3 KB, not 8 KB — text + tsvector + HNSW entry); C11’s handover carries the annotation.
  • Foreman verification on both: every C13 grep re-run and all 16 target→catalog anchors checked against built HTML; C14’s per-user figure re-derived independently from the code constants (€0.3346 vs the page’s €0.335) and the F0 boundary recomputed (50 users exact).

Board at hand-off

Every roadmap session S-0 → D2 and C1 → C14 is merged and deployed. Nothing remains but day-7 demo prep: the walkthrough script, the owner’s before-demo checklist (base handover items 1–3 plus seeding the demo account), and the intermittent local test flake (item 1 of the previous open-items list — still unexplained, CI unaffected).

Craft addition

  1. Auto mode’s classifier can block git push while allowing commits — a brief lands committed-but-unpushed and looks merged in local history. If a push is refused, say so explicitly and hand the owner the exact command; do not assume the row/PR exists.