v1 — Original DPP build framing
May 19, 2026, 7:30 PM · Build task. Skill-shaped solution assumed.
deal-room-structureanonymizationsource-miningloan-mart-templateblind-first
Added
- + Loan Mart as structural template
- + Codename requirement via inventory/blind-codenames.json
- + Source-mining list: Gmail, Fireflies, Granola, Drive, Supabase, repo
- + Buy-side framing for DPP
- + Principals: Chris Fore (President), Pat Fore (founder)
- + Starter segmentation hint about underground DOT work
- + Independent analysis requirement (user notes are hypothesis, not fact)
Settled
- ✓ Blind-First compliance
- ✓ Codename-first ordering
- ✓ Buy-side framing
- ✓ Principals identified
- ✓ Segmentation from primary sources
Revoked in a later version:- – Read existing design-precast-deal-manager skill first (revoked in v3)
v2 — +Repository / +Apply-to-All loop / +Plaud / +Audio
May 19, 2026, 8:10 PM · Build task, expanded scope.
file-repositoryapply-to-all-loopsupabase-firstplaud-integrationaudio-playback
Added
- + Deal File Repository as new universal capability
- + Identification → Build → Apply to All loop as core protocol
- + Supabase-first source-of-truth hierarchy
- + Plaud as new source connector
- + Audio playback via meetings-page player reuse
- + Loan Mart Deal File Repository backfill in same build
Settled
- ✓ Repository is universal, not DPP-specific
- ✓ Loop is protocol-level, not one-shot
- ✓ Supabase canonical, Vercel renders from it
v3 — +Swarm autonomy / scope to /swarm-build (5 agents)
May 19, 2026, 8:18 PM · Build task, with constitutional rule on swarm independence.
swarm-autonomy
Added
- + Swarm Autonomy section — existing artifacts are inputs, not authorities
- + Explicit 'do NOT pre-read design-precast-deal-manager as canon'
- + Valid verdicts: use / refactor / supersede / archive
- + Scope confirmed to /swarm-build (5 agents), not maxswarm (11)
- + Each agent's S1 reasoning must show independent skill read
Removed
- − Pre-read existing skill as canon (from v1)
Settled
- ✓ Swarm independence is constitutional
- ✓ Existing skill is input data, not authority
v4 — Reframed as architecture-decision swarm (no build)
May 19, 2026, 8:23 PM · Decision swarm. No build yet.
decision-vs-build
Added
- + Single question: is existing skill the best path forward?
- + Six architectural inputs the swarm must weigh
- + Output: one architectural recommendation memo
- + Dissent log required in output
Removed
- − All code/migration/page deliverables (moved to downstream build phase)
Settled
- ✓ Decision precedes build
- ✓ Build phase is downstream and contingent
- ✓ Swarm produces memo, not artifacts
v5 — +Skill-vs-application shape question / +Iteration log
May 19, 2026, 9:12 PM · Decision swarm. Broader option space.
skill-vs-applicationiteration-log
Added
- + Broadened central question: shape = skill / application / hybrid
- + Skill verdict retained as tactical sub-question
- + Iteration log requirement (this file)
- + Iteration log scoping rule: architectural decisions and multi-deal capabilities only
Removed
- − Implicit assumption that solution is skill-shaped
Settled
- ✓ Shape selection precedes scope
- ✓ Swarm chooses shape, not operator
- ✓ Iteration logs scoped to architectural / multi-deal work
v6 — +Webpage / +Filterable UI for iterations
May 19, 2026, 9:15 PM · Decision swarm, with iteration log as first-class UI surface.
iteration-log
Added
- + Iteration log surfaced as webpage at /swarm/plan-iterations
- + Organized by topic, filterable by any data field
- + JSON data file feeds both markdown and webpage
- + Page inherits root layout (NavLinks; No Orphans rule satisfied)
Settled
- ✓ Iteration logs are first-class UI surfaces
- ✓ Agent-readable .md and human-browsable page stay in sync via shared JSON
v7 — Architectural verdict landed (swarm-build #015)
May 19, 2026, 9:33 PM · Decision swarm complete; verdict feeds run #016 build.
decision-vs-buildskill-vs-applicationfile-repositorysupabase-firstapply-to-all-loopblind-first
Added
- + Shape verdict: HYBRID (application-led + thin per-deal skill); unanimous S1 convergence across architect, data-architect, quarterback
- + Skill verdict: REFACTOR design-precast-deal-manager; content preserved, Desktop-file storage abstraction retired
- + 3 new Supabase tables: deal_files, deal_file_tags, deal_pages
- + 1 extended table: deal_rooms (+ blind_codename, blind_slug, visibility)
- + 4+ reused tables via FK (meetings, transcript_extractions, documents, drive_files)
- + Apply-to-All Master Rule to be added to AGENTS.md, enforced by gate-check.sh reading Supabase deals
- + Buy-side template (DPP) separate from sell-side template (Loan Mart); chosen by deals.side
- + Run #016 scope frozen: universal foundation + Helios blind + Bedrock scaffold
- + 6 binary open questions for Ewing gating run #016
Removed
- − public/deal-rooms/*.html as primary surface (becomes one-time export only)
- − Author-discretion enforcement of Blind-First (replaced by DB trigger pair invariant)
- − Desktop-file canonical state in per-deal skills
Settled
- ✓ Persistent state belongs in application; trigger gating belongs in skills
- ✓ Blind/named pair invariant enforced by DB trigger, not author discipline
- ✓ Codename source-of-truth: DB (JSON deprecated to read-replica) — pending Ewing confirmation
- ✓ Existing skill content survives via extraction to deal-manager-base
- ✓ Run #016 first deliverable is Helios blind (resolves active Loan Mart Blind-First violation)
v8 — Six-question gate cleared; run #016 ready
May 19, 2026, 10:34 PM · All v7 open questions resolved. Run #016 fully scoped and awaiting trigger.
decision-vs-buildsupabase-firstblind-firstplaud-integrationskill-vs-application
Added
- + Codename source-of-truth: DB-canonical via new deal_codenames table (JSON deprecated to read-replica)
- + Per-deal password scheme: 4-digit randomized deal_number; password = each digit + 1 mod 10 (e.g., 4729 -> 5830)
- + Plaud file-size pre-flight checkpoint: probe duration + bitrate, chunk via TUS or transcode if file exceeds bucket limit
- + Rate-limit the password gate endpoint (4-digit space = brute-forceable without it)
- + deal_codenames table as structured codename store
- + deals.deal_number CHAR(4) NOT NULL UNIQUE column
- + Plaud bucket file_size_limit set explicitly at migration time
- + Test fixture requirement: 90-minute mock audio at 128 kbps must round-trip successfully
Removed
- − JSON file (inventory/blind-codenames.json) as authoritative codename store
- − Shared single password across all blind rooms
- − Migration of existing Loan Mart static HTML in run #016 (frozen, parallel-only)
Settled
- ✓ Codename storage: DB single source; JSON becomes a generated artifact
- ✓ Password gate: deterministic per-deal scheme; internal-only user model accepted
- ✓ Plaud: Supabase Storage with size-check + transcode-or-chunk fallback
- ✓ Sequencing: universal-first (Option A); Helios blind first deliverable; DPP content second run
- ✓ Hybrid shape verdict approved; run #016 cleared to execute
v9 — Run #016 shipped — universal foundation live (file system)
May 19, 2026, 10:54 PM · Build artifacts produced and integrated; migration files written but NOT yet applied to production Supabase. Awaiting Ewing's `doppler run -- supabase db push`. PASS verdict after one REWORK cycle (2 column-name defects, conductor-fixed).
decision-vs-buildskill-vs-applicationfile-repositorysupabase-firstapply-to-all-loopblind-firstaudio-playback
Added
- + Migration 2026_05_19_e__deal_room_foundation.sql — 4 new tables (deal_codenames, deal_files, deal_file_tags, deal_pages), 2 extended (deals +side/engagement_type/deal_number; deal_rooms +blind_codename/blind_slug/visibility), blind/named pair invariant trigger, deal_room_password() SQL function, RLS policies, 9 indexes, 3 triggers
- + Migration 2026_05_19_f__deal_room_seeds.sql — Helios + Bedrock blind hub deal_pages rows, deal_codenames entries for both, deals upserts with hardcoded deal_numbers (4729 design-precast, 5037 loanmart-hugo)
- + Next.js render route src/app/deal-rooms/[codename]/[page]/page.tsx — RSC, inherits NavLinks (No Orphans-compliant), password-gated, light theme
- + Password gate API src/app/api/deal-rooms/unlock/route.ts — derived 4-digit password (digit+1 mod 10), in-memory rate limit (10 per 15-min per IP+codename)
- + Skill skills/deal-manager-base/SKILL.md — shared base with mandatory pre-deliverable sequence and universal anti-pattern guardrails
- + Skill refactor skills/design-precast-deal-manager/SKILL.md — reduced from 120 to 56 lines; Desktop file pointer removed; reads context from /api/deals/design-precast
- + Apply-to-All Master Rule appended to AGENTS.md as third BLOCKING Master Rule
- + Project Bedrock codename entry merged into inventory/blind-codenames.json
- + Script scripts/plaud-size-check.mjs — direct / direct-raised / chunk TUS decision logic; self-test PASSES all 4 contract cases
- + Script scripts/apply-to-all-gate.sh — enumerates active deals via stage filter, fails fatal on missing capability artifact
- + Script scripts/anonymize-all.mjs extended — verifyDealPagesCoverage() reads deal_pages and validates Blind-First pair invariant at Vercel build time
- + REWORK cycle fixes: seed migration column status->stage (deals enum), apply-to-all-gate column archived_at->closed_at+stage filter, silent error suppression replaced with fatal exit
Removed
- − design-precast-deal-manager Desktop file pointer storage abstraction
- − Silent error suppression (2>/dev/null || true) from apply-to-all-gate.sh
Settled
- ✓ Universal foundation shipped on disk; migration files ready for supabase db push
- ✓ Helios blind room scaffolded (resolves long-standing Blind-First violation on Loan Mart)
- ✓ Bedrock blind room scaffolded (DPP placeholder; content authoring deferred to run #018+)
- ✓ design-precast-deal-manager is the FIRST per-deal skill refactored to the hybrid base+overlay pattern
- ✓ REWORK loop validated end-to-end: auditor caught 2 real defects that would have crashed production migration apply
- ✓ PASS verdict on 16/16 audit criteria after resubmission
- ✓ Deferred to run #017: Gmail/Fireflies/Drive ingestion adapters, hr-com-deal-manager + ches-booker-deal-manager refactors, Apply-to-All sweep against HR.com + Capstone repositories
- ✓ Deferred to run #018+: Plaud adapter daemon, Loan Mart Deal File Repository population, DPP Bedrock content authoring, migration of remaining public/deal-rooms/*.html
v10 — Migrations applied to production Supabase — PASS
May 20, 2026, 6:05 AM · Both migration files applied to production via psql $DATABASE_URL. Two additional defects found and fixed during apply. Full smoke-test passed. Schema and seed data are live.
supabase-firstblind-firstdeal-room-structure
Added
- + Defect 3 fix: deal_number backfill replaced pure random with row_number*97%10000 (prime stride, guaranteed unique across up to 10,000 rows). Applied in 2026_05_19_e__deal_room_foundation.sql.
- + Defect 4 fix: seed INSERT used 'name' column; correct column is 'company_name'. Fixed in 2026_05_19_f__deal_room_seeds.sql.
- + Defect 4b fix: ON CONFLICT COALESCE blocked canonical deal_numbers; changed to direct overwrite. deal_numbers manually set: design-precast=4729, loanmart-hugo=5037.
- + Production smoke tests: deal_room_password('4729')='5830' ✓, deal_room_password('5037')='6148' ✓, both hub pages queryable from deal_pages table.
Settled
- ✓ Production Supabase has deal_codenames, deal_files, deal_file_tags, deal_pages tables live
- ✓ deal_rooms extended with blind_codename, blind_slug, visibility
- ✓ Helios hub: /deal-rooms/helios/hub live (password 6148)
- ✓ Bedrock hub: /deal-rooms/bedrock/hub live (password 5830)
- ✓ Password derivation function matches all 4 build contract test vectors
- ✓ Run #017 can now proceed: ingestion adapters + remaining skill refactors
v11 — Run #018 scoped (Opus) — connector completion + cost guardrails + skill reconciliation
May 19, 2026, 11:38 PM · Infrastructure-completion run. Closes the three FixLater items filed at the end of #017 (Gmail #100, pipeline_items #101, cost ledger #102). Two Phase-0 probe findings reshape the scope: deal_targets already exists (no new table needed — Modeling Rule 9), and Gmail OAuth scope is NOT confirmed on the existing Google refresh token.
supabase-firstapply-to-all-loopskill-vs-applicationdecision-vs-build
Added
- + Track A — Gmail adapter (engine/ingestion/gmail_to_deal_files.py + LaunchAgent). Phase 0 verifies Gmail scope; if missing, blocks Track A pending one-time browser re-OAuth.
- + Track B — Skill reconciliation: deal-manager-base + ches-booker-deal-manager point at existing public.deal_targets instead of nonexistent deal_pipeline_items. Closes FixLater #101 with no migration.
- + Track C — Cost ledger JSONL (inventory/cost-ledger.jsonl) + helper script + Phase 5 hard requirement. Backfill #016 and #017 entries. Closes FixLater #102.
- + Phase 0 probe finding: deal_targets exists with 117 rows for HR.com (deal_id='hr_com_ltd'). Modeling Rule 9 says do not create deal_pipeline_items.
- + Phase 0 probe finding: Google OAuth token returned HTTP 401 on Gmail profile probe. Track A needs re-OAuth before build.
Removed
- − Creation of new deal_pipeline_items table (Modeling Rule 9 — deal_targets serves the use case already)
Settled
- ✓ Theme: connector completion + cost guardrails. NOT content authoring (#020+) and NOT Plaud (device-dependent).
- ✓ deal_targets slug normalization (hr_com_ltd vs hr-com) deferred as separate FixLater — fixing risks breaking existing HR.com pipeline UI
- ✓ deal_targets.fit_score Modeling Rule 2 migration deferred as separate FixLater — existing tech debt, large effort
- ✓ Cost ledger format: JSONL append-only (no new schema migration, easy diff)
- ✓ Bedrock + Forum + Lagoon content authoring waits one week so the freshly-ingested corpus grows from primary sources