Old per-Lambda DuckDB approach didn't scale past 6 clients. Replaced with one always-warm query layer everyone hits the same way. Closed out 2026-05-23: home-data 100% + Rundown 100% + Courier message-raise 100% + agency-top 100% (new `top_agencies` pipe). Bridge retired entirely (not migrated) — confirmed write-only artifact, nothing downstream read its Live Summary Docs; code deleted, 9 AM ET cron killed, Drive Docs frozen as historical. DuckDB-in-Lambda fully gone: `duckdb` dropped from package.json (npm removed 112 transitive packages), three consumers cleaned up. Net diff across the close-out: -3169 lines.

Prompt rebuild progressed Step 1 → Step 2 → Step 3 over the 5/31–6/2 sessions. Step 3 LIVE 2026-06-02 via single master toggle STEP3_ACTIVE in _scribe-lib/phase1-assembly.mjs: whole-email perf-ranked corpus block in the prompt (top 3 client-scoped records ranked by performance.dollars_raised desc, deduped by source_doc_id, no Layer B cross-client). Step 2 (full profile excerpt minus radar sections, client-agnostic name-pattern exclusion) shipped earlier same session. Performance data wired into the corpus for the first time: 222 / 373 records (60%) carry refund-netted dollars + donor counts via message_raise_v2 Tinybird join + a hardcoded MVE-FC archive table from TVE profile §7 (retired committee never in donations_v2). 6 prompt blocks now (preamble + URL rule + principles + profile excerpt + corpus + directives) totaling ~8.3K projected tokens for STPC (vs ~4.9K Step 2, ~2.4K Step 1, ~30K old-prod-mode bloat). Three production hard-directive bugs fixed first: voice (3-class structured AUTHORITATIVE override of past-hits sender patterns), channel (explicit email branch with ONLY-email exclusivity), tone (renamed Length hint → Tone AUTHORITATIVE). Brief + output cards now clear on client switch. New infrastructure: ScribeCorpus under git for the first time at github.com/amwaldeck/scribecorpus (private, follows ParquetPipeline precedent). Phase 1 corpus pipeline + retrieval module + bundled-Doc parser all shipped earlier; corpus consumed from local deploy bundle (_scribe-corpus/*.jsonl via netlify.toml included_files). Update 2026-06-11 — Step 3 CLOSED: relevance-primary corpus selection live (cosine 0.65 / dollar-tiebreak band 0.03, validated for variety); "The arc" structural principles + meta-narration opener fix shipped 6/4; the 6/5-staged trio deployed 6/11 (ce6e14d) — anti-fabrication + rubric pinned on both selection paths, the 8-angle angle menu as a retrievable principles section (embeddings regenerated, 21→22 sections), 388M stat scrub (upstream Google Docs also fixed; "owed source commits" closed as moot). Validated in production via URL-anchored STPC draft. First recurring-surrogate profile entry (Boykin) added to STPC profile + mirror (8d062c7).

Currently $500/mo. Pushes WinRed/Revv donors → Email Octopus (deployment) and lets Doug pull data into Salesforce. Replace with S3-custom-reports path driving both EO and SF directly. Saves the cost and removes a vendor.

Killed Apps Script. Drop / Tasks / Comments + Recently Launched + the Scribe corpus cron now run on Netlify Blobs; reads sub-500ms and the recurring break class is gone. Built → tested (22/22) → imported (275 drops / 55 tasks / 24 comment threads) → hard cutover → live, all in one session. New: `callBackend`/`blobs-backend` seam, `drops-by-client` cron endpoint, recovery surface (`admin.html`), browsable Drops Ledger (`ledger.html`). Decisions: Mark Sent gated on Drop=Approved; `listMyTasks` approver clause; `getBySlug` spine fallback.

Finished the job Tasks started — moved Courier's agency roster + sender lists off the last live Apps Script dependency (the separate Master Agency List Sheet, whose intermittent slowness caused blank Agencies steps + missing sender dropdowns). `agency-list.js` now reads Blobs-first with the sheet as fallback (safe parallel build, zero-risk flip). New: `_agency-lib/agency-store.js`, Adam-gated `agency-admin-api.js` (CRUD + one-click seed/rollback), form-based `agency-editor.html` in Back Porch. UX: paired Slack-mention rows; "Allyz URL" → "Submission portal URL". Round-trip parity verified; seeded live (29 agencies / 25 senders); confirmed `source:blobs`.

Solve the WinRed recurring subscriptions data problem with a lighter approach than manual lifetime exports.

Same problem as cancellation-delta pilots above, heavier solution. May be obsolete if pilots work.

Today every client report is hand-built HTML with numbers pasted in manually. Make it automatic so every TMP client gets a fresh report subdomain with zero manual work.

Bridge writes data into Google Docs for human briefing. Doesn't work for Scribe, which needs queryable structured data, not narrative prose. Form needs to change or get supplemented.

When live + EO Donor Flow Lambda lives, Zapier middle-layer on donation flow becomes retirable.

Pipe Email Octopus engagement data (opens, clicks) into the data layer for join with donations. Foundation for the Donor 360 layer.