Decision records

Answer

One record per big technical call, each with its context, the decision, what follows from it and the date, so no session re-decides a settled thing without new facts.

Today

DR-1. A realistic prototype, cut to one flow

• Date: 15 Oct 2026. memory brew-prototype-repo
• Context: the old feed/ repo had grown Do it now, measurement, sending and Arata-as-code. Param: "we are building a realistic prototype".
• Decision: a new repo brew/ with one flow, read, Plays, feed, jam, share the brief. Cut: Do it now, writing to accounts, sending to customers, measurement, audit Plays, the free cap. Dummies are tagged in three places. brew/SCOPE.md brew/CLAUDE.md
• Consequences: feed/ archived (tag archive-2026-10-15); old surfaces retired; leftover code from cut features remains (see Tech debt).

DR-2. Read-only connectors, built by us

• Date: 16 Oct 2026. memory brew-real-connectors
• Context: real brand data was the biggest blocker; Composio was proposed.
• Decision: no vendor. Revive our own connectors, read-only scopes only; Brew never writes to an account. brew/SCOPE.md brew/connectors/shopify.js:13
• Consequences: each platform's approval gates us directly (Google verified 21 Oct; Shopify submitted 21 Oct; Meta blocked on business verification). Do not propose vendors again without new facts.

DR-3. Gemini Flash while building, Pro for Plays

• Date: 16 Oct 2026 (Flash), 20 Oct 2026 (Pro for Plays). memory brew-gemini-provider memory brew-pro-switch-and-flush
• Context: Anthropic credit at $0; Google credits about $20,763 of $25,000 left, ending 2028-05-19.
• Decision: AI_PROVIDER=gemini; gemini-3.5-flash for everything except Plays, which use GEMINI_MOVES_MODEL=gemini-3.1-pro-preview. AI_PROVIDER=anthropic switches back. brew/moves-engine.js:17 brew/deploy.sh:39
• Consequences: Plays are shorter and more generic than Opus (20 Oct: median 2,250 vs 5,450 words); the weekly rewrite pot is $40 on Pro, $10 on Flash. The launch model is still open.

DR-4. Deploys never wait

• Date: 16 Oct 2026, commit 522662c. memory brew-pro-switch-and-flush
• Context: a deploy guard waited for background work; Param: "continuous deployment is part of development".
• Decision: remove the guard; SIGTERM stops new work and the next instance resumes. brew/CLAUDE.md
• Consequences: every background run must be resumable and never pay twice; never re-add a wait or gap.

DR-5. Every Play is kept; make each cheaper

• Date: 15 Oct 2026. memory keep-all-moves
• Context: a "planner picks about 10 Plays" option was proposed to save cost.
• Decision: keep all Plays; cut tokens, caching and repeat evidence calls; PDFs only when opened. memory brew-lean-pipeline
• Consequences: cost work targets per-Play spend and the rewrite pot.

DR-6. Sloosh makes finals, quote first

• Date: 16 Oct 2026. memory brew-sloosh-finals
• Decision: drafts stay on Gemini; Sloosh makes go-live finals only behind Make the final, after a quote and a yes; chat never spends; Gemini fallback while the workspace cannot pay. brew/SCOPE.md
• Consequences: finals keys written before paying; production runs with Sloosh off (25 Oct). git 0fc7cfc

DR-7. One version, then Redo it

• Date: 19 Oct 2026. memory brew-one-version
• Decision: never several options to pick from; one version and Redo. A/B decisions and test batches are not options.
• Consequences: MAKE_N 1; saved "3 versions" text heals on read.

DR-8. Plays start on the first feed open

• Date: 22 Oct 2026. memory brew-feed-open-moves
• Decision: a brand's first read writes Plays only after someone opens the feed after onboarding (feedGate). brew/SCOPE.md
• Consequences: corrections made while waiting reach the brief first; two reminder emails.

DR-9. Protected connections

• Date: 22 Oct 2026 (rule), 25 Oct 2026 (in code, rev 00710). memory never-disconnect-arata-meta memory brew-prod-backend
• Context: Arata is the client going live; its Meta account act_2942894532542334 drives the private rewrite of its Plays.
• Decision: never disconnect, re-select or reconnect Arata's Meta or Shopify. PROTECTED_CONNECTIONS (default arata-in) refuses those actions unless a root admin forces it; the vault refuses to drop the token. brew/connectors/protected.js
• Consequences: disconnect tests run on Cozy; Arata's vault is shared between staging and production rather than reconnected.

DR-10. Postgres over the bucket

• Date: 16 Oct 2026 (not yet, not Supabase), 25 Oct 2026 (Cloud SQL, approved and live). memory brew-storage memory brew-prod-backend
• Context: whole-file writes to gs://brew-prototype were last-write-wins and safe only with exactly one instance.
• Decision: Cloud SQL Postgres in asia-south1 on Google credits, not Supabase. First step: every structured object as a row (objects), big documents in parts (doc_parts), binaries in the bucket; the bucket mirrored as the way back. brew/db/migrations/001_objects.sql brew/db/migrations/004_doc_parts.sql
• Consequences: more than one instance became possible; per-domain tables are still to come; the bucket-era tools (reset-jams.sh, reset-brand.sh) no longer work on Postgres sites; since 26 Oct db/reset-brand.mjs and db/delete-brand.mjs replace the brand ones, which now exit 3 unless STORAGE=gcs brew/tools/reset-brand.sh:11-19; brew/db/reset-brand.mjs:1-11.

DR-11. Web and worker roles with one lease holder

• Date: 25 Oct 2026 (Phase 2, rev 00707). memory brew-prod-backend
• Decision: ROLE=web|worker; one worker holds a Postgres advisory-lock lease and alone runs timers, resumes and paid watchers; web instances ask the worker through a Postgres job queue. brew/roles.js:1-9 brew/jobs.js:1-8
• Consequences: production runs 2 to 10 web instances and exactly 1 worker until per-brand build leases exist.

DR-12. Production on feed.joinsquareup.com, staging on brew

• Date: 25 Oct 2026. memory brew-prod-backend
• Decision: production is brew-web plus brew-worker on its own database, bucket and identity; staging is shopos-feed; production is released only by promoting staging's exact image. brew/.wt-prod/docs/production.md
• Consequences: two sites to keep in step for sign-in callbacks and Slack; a moved-brand rule decides each brand's home.

DR-13. Nothing dummy or test reaches production

• Date: 25 Oct 2026. memory prod-clean-only
• Decision: production never runs demo, mock, dev sign-in, open-site or simulated connectors; a clean gate in every promote refuses any finding. Four kept exceptions decided one by one. brew/db/prod-clean.mjs:1-6
• Consequences: test brands stay on staging; partner copies exclude testing data.

DR-14. Every partner's home is production

• Date: 25 Oct 2026, 20:40 IST, commit 119dc8a (it reversed deb7d90 from 20:11). git 119dc8a, deb7d90
• Decision: Arata, Nutrova, Mana, PUMA India and Scrubroom onboard on production; staging lists all five in MOVED_TO_PROD; production's STAGING_HOME is empty. brew/deploy.sh:39 brew/tools/prod-env.sh:9
• Consequences: staging is for testing only; the reverse list stays in code, empty.

DR-15. Retire Jam v2's own screen

• Date: 25 Oct 2026, commit 9ee9d14. git 9ee9d14 memory brew-one-stage
• Context: v2 was locked behind ?v=2 on 22 Oct; on 23 Oct Param ruled no v2 jam inside v1; two jams confused people.
• Decision: remove ?v=2, window.Jam2 and public/jam2/jam2.js; every card opens v1's jam, whose Set-up drives the jam2/ Studio engine.
• Consequences: jam2.css, scene.js and the /api/jam2 routes stay; v2-only routes remain mounted until shared links are checked.

DR-16. No Gmail scope for briefs

• Date: 23 Oct 2026. memory brew-brief-mode
• Decision: briefs open in Gmail by a compose page with the files downloaded, or as an .eml draft; no gmail.compose, which is a restricted scope needing a yearly assessment.
• Consequences: Google verification stays at two scopes.

DR-17. Brew stays independent of the sister company

• Date: 24 Oct 2026. memory brew-qa-round2-2026-10-24
• Decision: no new ties to the sister company or its studio, Sloosh; existing ones (Send brief to our team, Sloosh finals, the staff domains) stay.
• Consequences: new features never depend on those services.

Tomorrow

How we do it

1. A decision is big if it changes storage, roles, a provider, scope, access or what reaches production.
2. Write the record the same day with context, decision, consequences and date, each sourced; add it to content/company-and-strategy/decisions-log.md too.
3. A record is replaced only by a newer record that names it; never edited away.

Numbers

Not applicable: decisions are records, not metrics.

Open

1. Which model writes Plays at launch? Proposal: Opus with Anthropic credit, after a 5-Play blind test against Pro (about $1).
2. Should DR-17 also cover Sloosh finals on production? Proposal: yes, keep finals on Gemini on production until Param says otherwise.