Audit Triage

GOOGLE_CSE_API_KEY (project 56543540838, key "Claude Web Search", CX 843fa16c...48b2) returns 403 even with Custom Search API enabled and billing active. Suspected org-policy block at chapter.guide GCP org level.

GOOGLE_API_KEY lives in project 1086376262403 (unknown owner, no access). GOOGLE_MAPS_API_KEY lives in project 618741625931 (unknown owner, Maps works but no admin access). Both should be consolidated under Next Chapter OS project 56543540838 or documented with named owners. Risk: key rotation requires owner access we don't have — if these keys leak, we can't rotate.

Browser key embedded in src/app/campaigns/[slug]/tracker/MapTab.tsx is unrestricted. The server-side GOOGLE_MAPS_API_KEY was locked down 2026-05-25 (IP-restricted). Browser keys cannot use IP restrictions because every visitor IP is different — they need HTTP referrer restrictions.

These 4 keys exist as 'Sensitive' type in Vercel prod, which means `vercel env pull` returns empty quotes — CLI can't extract values. They also don't exist in Doppler dev, so the dev→prd promotion path doesn't help. Manual fix (each): 1. Go to Vercel dashboard → Project → Settings → Environment Variables → find <KEY> → reveal value 2. Locally: `echo '<value>' | doppler secrets set <KEY> --project clawdbot --config prd --no-interactive` 3. Verify: `doppler secrets get <KEY> --project clawdbot --config prd --plain` Also add to Doppler dev so future promotions work: `echo '<value>' | doppler se

scripts/fixlater.sh mirror_to_supabase silently fails when invoked from contexts where psql isn't on PATH or Doppler subprocess inherits different env. All 5 captures #204-208 this thread bypassed /audit until pushed manually via REST API. Fix: replace psql call with fetch() to Supabase REST endpoint (no driver dep, no PATH issue), OR add diagnostics so the silent skip becomes visible. Verified: SUPABASE_POSTGRES_ADMIN_PASSWORD IS in Doppler clawdbot/prd; psql IS at /opt/homebrew/opt/libpq/bin/psql. So neither was the failure — likely a quoting/heredoc issue in the embedded SQL when title or b

Regressed: Blind-First violated on /hugo. Demoted until BUILD-02.

Regressed: deal-rooms shadow leak. Demoted until intern_tier_blast_radius recipe.

Regressed: /jeremy still violates rule. Demoted until BUILD-04.

Regressed: 28/35 deals NULL. Demoted until BUILD-03.

Two tables exist: meeting_notes (LEGACY flat table, 52 rows, populated by engine/scripts/backfill_meeting_notes.py from raw Fireflies AI summaries) and public.meetings (modern structured extraction via extractFromTranscript() with full deal context). The webhook handler explicitly notes "All extraction-aware data lives in public.meetings." 5 open questions including whether meeting_notes is still being written.

Salesfinity's Custom1-Custom10 slots are blank string slots with no labels defined in Salesfinity itself — enforcement must live entirely in NC loader code. Without a validator, contacts can hit the API with wrong-shaped custom fields.

Code exists (engine/transcripts/mine_signals.py) but the v1.yaml pattern library needs Ewing's review and ratification before the feature def can flip from design to live. This is the last transcript feature still at design besides diarization.

Branch protection rules on main (requiring PRs for non-Ewing pushers) need GitHub Pro for private repos. Current plan: free. Convention is documented in swarm-upgrade SKILL.md but not technically enforced.

Build Supabase induction triple (induction_runs + induction_business_rules + induction_frictions) when there is non-noise capture data to fill it. Schemas in inventory/skill-builds/migrate-to-external-induction.md as reference; data-architect's full proposal in this thread's review. Gated on FixLater #204 (capture signal fix) — premature without real captures. Read pattern: Phase 5 validation gate query 'rules with user_facing=true AND in_skill_md=false'; cross-build friction analytics by bucket_id. --- Captured via /FixLater on 2026-05-27 from branch `main` on `MacBook-115.local`.

After consolidating Exa behind a single ExaMiddleware chokepoint (swarm-build #007), the exa-audit skill v1.1 ran 2026-05-17 and found 9 P1 files still calling Exa with raw HTTP and bypassing the middleware. Files: buyside_hunt_v1.py, pool_hunt_v2.py, run1_pipeline.py, run2_enhanced_pipeline.py, fetch_webset_data.py, solar_pwa_pipeline.py, sales-enrich/scripts/exa_enrich.py, tests/fixtures/buyside_hunt_v1/capture_live.py, engine/lib/exa_websets.py. Plus 19 REWORK items + 3 WARN.

When a Salesfinity call dispositions as "Bad Number" or "Answered - Wrong Contact", the (person, company) tuple is invalid. Mission #027 added the schema (person_company_exclusions table), the trigger, exclusion helpers, and Gate 9 in salesfinity-loader. But the actual job that runs Clay/Exa person-find on the excluded tuple — the part that produces a NEW contact for the dial list — was deferred to ship paired with #029.

Active build sessions for two prospect hunts: solar (PWA-qualified project leaders) and Midwest construction. Stannp integration ready to send letters once contacts land. Goal: build lists, enrich phone/email, push to Salesfinity for dialer rotation.

~7% of NC dial calls end with the contact naming someone else to talk to ("call my brother", "talk to corp dev VP"). Warmest lead source we have outside an inbound — prospect literally handed us a person to call. Today: rep notes it in Salesfinity, nothing automated. Mission #028 fixes that.

Email editor went through 5 phases in swarm-build #007: Phase 1 schema (templates, template_edit_sessions, edit_snapshots, test_runs, propagation_log), Phase 2 template index with filter chips, Phases 3-5 lock + version + test batch. 18 feature defs registered. "Complete email template handoff to Mark" session was running live.

maxswarm run #027 audited the email templates against last 50 Salesfinity calls. Produced 4 new "Tier-1" templates landed in templates/email/. Each should have a vitest test.

Original buyside_hunt_v1 was rebuilt with findSimilar Phase 2 to fan out from a seed company. New version lives on branch claude/sleepy-carson-9ee184 alongside pool_socal pushes that now write to Supabase cleanly (commit 2b39aaf).

Migration 2026_05_05__engagement_fee.sql added an engagement_fee column to deals table. Each active client engagement (HR.com, Ches Booker, Design Precast) has specific fee terms.

src/lib/salesfinity-classifier.ts has a TODO at line 53: "refine HotLeadFollowup prompt with Ewing". The classifier sorts incoming Salesfinity calls into outcomes; HotLeadFollowup triggers downstream automation.

swarm-upgrade reads swarm notebooks, finds recurring patterns (cross-system gaps, repeated currency exchanges, audit failures), proposes upgrades to agent SKILL.md files. Was supposed to run weekly. Last batch is #002 from 2026-05-10 (manual). Notebook #004 approved 4 proposals 2026-05-17 — unclear if weekly cadence or one-off.

Mark joined the fleet on his iMac 2026-05-12. Setup completion guide committed (cb234d8). TypeScript fix to make his build work (43ed892).

swarm-build #045 landed cost_usd instrumentation in 21 call sites but 4 sites had no exact registry key — the agent used a near-match (calls.review, admin.stuck-detector, admin.intern-activity, admin.ticket-dashboard) and tagged the call site with TODO(iss_0045). The cost attribution is defensible but slightly off; downstream dashboards filtering by these proxy keys will see commingled events.

Auto-publish keeps shifting current branch mid-commit. Work observed landing on swarm-build/007-exa-chokepoint-refactor, wip/exa-recipe-pickup-2026-05-17, and fin-2-orphan-cleanup instead of main. Forces cherry-pick-back-to-main dance every commit. Real friction on every push.

Production DB has 1 company with exa_fields and 0 contacts linked. Test-batch renderer is wired to all 49 tokens but most still emit [MISSING: ...] because companies.exa_fields is empty. Until enrichment runs, new templates (send-buyer-thesis-hot, send-info-on-radar-warm) can't be evaluated.

All crawled files end up with modified_time = the seed-import timestamp (e.g. 2026-03-20T00:36:19 for all DPP files), instead of Drive's real createdTime/modifiedTime. Verified by live drive.googleapis.com call: real dates span 2025-10-08 → 2026-04-22.

Global /deals/followups currently owns followup-008..013 (all HR.com). Violates Truth-Lives-With-The-Deal master rule which requires per-deal homes.

Tailscale SSH currently disabled to unblock cross-machine forensics from Claude shell (key auth via plain OpenSSH).

Decision logged by Ewing 2026-05-22 via maxswarm #031 audit. Letters channel CONFIRMED primary. 4 items collapse into 1 build. FIN-16 is a silent-money item — every day undone is $20-50/day potential undetected overspend. All 4 active deals (HR.com, Capstone, Design Precast, Jeff VanHoose) use letters as a touchpoint.

Five transcript-* feature defs already live, two stuck at design. UVD correction flow shipped. 52 meeting_notes in prod. 38-call audio corpus downloaded. Work is 60% done but scattered across 5 MASTER_PLAN items plus undocumented heuristics. Phase gates defined so faster phases don't wait for slower ones.

#127 fix landed on MacBook side; Mac mini audit was deferred. Need to enumerate what daemons/crons/sessions on the Mac mini can write to the shared working tree.

20 unmerged commits on claude/review-deals-page-design-nI4Il branch including phone pipeline (df3b137), GH Actions test runner (85a6c94, 1165b1f), Salesfinity loader (afe095a, be97e0c), ingest_spec recovery (f4a35e6, 9a89a41). Surfaced during issue #127 investigation.

Apply-to-All rule: banner now shipped but shows empty boxes for 28 of 35 deals. 28 slugs listed (aqua-pros, arnold-roofing-and-restoration, barclay-street-real-estate-ltd, bel-aire-mechanical-inc, bell-precast, cal-west-roofing-inc, capsoft, complete-aquatic-solutions, cypris, fss-inc, imperial-plumbing, indiana-precast-inc, indianola-precast, jonathan-salomon, lenox-advisors, manco-structures-ltd, mccody-concrete, mid-south-prestress-llc, mosquito-shield-of-buffalo, natura-pest-control-llc, pacific-shore-pest-control, pinnacle-wealth-brokers, premier-systems-roofing, protouch-pool-services, puddle-pool-services, re-max-fine-properties, rocket-partners, swim-chem-inc).

Run #024 built the unified "Change the page" feature — a floating box that hides whether the data-agent or ui-agent handles a request. Shipped across multiple commits to main. Trackpad compatibility fix (9ab9f24) landed. Primary admin tool for making page changes without opening code.

Forum (HR.com) glossy hub built at public/deal-rooms/forum/about.html on 2026-05-24. Per Apply-to-All rule, same capability must land on all active hubs. Helios (Loan Mart) is DONE — gold reference. 4 remain: Keystone (Capstone), Bedrock (DPP), Lagoon (Puddle Pool), Granite (Wieser).

Three LaunchAgents scheduled for unload: contacts-sync, salesfinity-sync, drive-mirror — all superseded or folded into hermes cron.

A dynamic route at src/app/deal-rooms/[codename]/[page]/page.tsx catches /deal-rooms/*/* requests and queries Supabase deal_pages for slug=codename AND page_kind=page AND visibility="blind". This shadows static files in public/deal-rooms/<codename>/*.html. The deal_pages.page_kind values are "hub"/"cim"/etc — never "valuation.html" — so the dynamic route always misses and returns 404. The dynamic route is DEAD CODE: gated entry routes still read public/ via hubPath in gated-deal-room.ts. 10 codenames affected.

12 companies where Exa returned Google Maps mis-matched results (city/brand token overload), 1 with zero results. These 13 have no person candidate in extra_people.

288 LA pool owner candidates in Google Sheet were NEVER written to Supabase companies.extra_people — violates Supabase-first rule. /tmp/owner_candidates_full.json was lost on session restart.

Action item from 2026-05-11 Loan Mart Discovery Call. Prepare target buyer profiles based on Hugo's business insights: high-interest lending, M NI, off-market preference, flexible deal structures.

Plaud captures offline/in-person audio (complements Fireflies, not replacement). Note 200-recording API cap; audio download links expire 24h so fetch must happen in same job.

Mac mini got disk cleanup 2026-05-24 (9GB freed). MacBook NOT cleaned. 460Gi total / 409Gi used / 13Gi free / 97% capacity. Next swarm or build session will fail mid-run.

Two NAMED-version CIMs shipped 2026-05-18: pixelflex.html + elite-multimedia.html. Both safe for internal use. Both violate Blind-First Master Rule if shared with any external party before NDA. Becomes blocking the moment Ewing wants to send a CIM URL to anyone outside Next Chapter or Jeremy himself.

Code was written mid-thread but lost to stash. Exact diff spec preserved at inventory/thread-audit/2026-05-16__exa-extraction-findsimilar-phase2-buyside.md. Must call _hunt_middleware.call(ExaCallRequest(endpoint="findSimilar", ...)) NOT raw requests.post. ExaMiddleware handles findSimilar at line 620 of engine/lib/exa_middleware.py.

migration-drift-check v2 — honor DROP statements. v1 flags any CREATE TABLE / ALTER ADD COLUMN in supabase/migrations/*.sql whose target doesn't exist in live information_schema. False positive surfaced 2026-05-26: `2026_05_26_g__toolbox_requests.sql` creates public.toolbox_requests, then `2026_05_26_h__deprecate_toolbox_requests.sql` drops it. v1 sees `_g`'s CREATE claim against a missing table and flags drift. v2 should parse DROP TABLE / DROP COLUMN / DROP VIEW from later migrations and exclude claims that were intentionally reversed. Lives at scripts/migration-drift-check.ts:35 (RE_ patter

launchagents/com.nextchapter.migration-drift.plist committed in b5af122a (2026-05-26). The file exists in repo but the LaunchAgent is NOT yet loaded on the Mac mini host. Until `launchctl load -w /Library/LaunchAgents/com.nextchapter.migration-drift.plist` runs on the mini, the 4h drift check is committed code but inert in production. Verification when loaded: `launchctl list | grep migration-drift` `tail /tmp/migration_drift_check.out` should grow every 4h Source: fixlater #177 (closed b5af122a) but the deployment step was deferred.

Audit every active deal for HR.com-style folder sprawl (multiple disk locations for the same deal). Active deals to check: loanmart-hugo, design-precast, capstone-construction, puddle-pools, elite-multimedia, hr-com (already audited — worst case). Output: per-deal disk-location inventory + which artifacts are NOT surfaced from /deals/<slug>/documents today. Likely 30-50 minutes of grep+ls work. Locked by One Canonical Deal Room master rule 2026-05-27. --- Captured via /FixLater on 2026-05-26 from branch `main` on `MacBook-115.local`.

## Why Audit (`inventory/audits/drive-by-route-invented-business-audit-2026-05-26.md`) confirmed the route bug but the 2 specific invented business names are not recoverable from disk — `onTrack` doesn't persist the outbound Maps URL. ## How to capture 1. Open `/campaigns/NC-LUNCH-PHX-2026-06/tracker` (or whichever Phoenix campaign last reproduced the bug) 2. Switch to Drive-By tab 3. Set start = 'current location' (or any starting venue) 4. Open DevTools → Network tab 5. Click 'Get route' (or whatever opens the Maps URL) 6. Find the outbound `maps.google.com/maps/dir/...` URL in Network 7.

v2 capture-session.sh heuristic still produces noise: all v2 captures show looks_like_build:true with 50 commits + 300-400 changed files because transcript-birth-time anchor is too wide (concurrent sessions' commits leak in via wall-clock window). Fix: read transcript JSONL and extract Edit/Write tool calls as the session-scoped file list, OR replace SessionEnd hook entirely with a commit-watcher LaunchAgent. Current state: 4 v2 captures archived alongside 5 v1 captures in inventory/skill-builds/captures/.archive-v1-noise/. Hook is still armed; expires 2026-05-28T07:39Z. Data-architect leading

Apply promote-workflow pattern to 4 sibling design→done tables (email_drafts, pending_reviews, toolbox_requests, writer_template_proposals). Per architect+data-architect on run #051; Apply-to-All rule. Generalize POST /api/features/promote-request, or file four routes. --- Captured via /FixLater on 2026-05-27 from branch `mission/035-pass-4-finetune` on `MacBook-115.local`.

Add feature_def_request_events append-only log per Modeling Rule 4. Today status flips destructively. Schema: id uuid PK, request_id uuid FK, event text, actor text, at timestamptz, payload jsonb. Trigger on UPDATE OF status on feature_def_requests. --- Captured via /FixLater on 2026-05-27 from branch `mission/035-pass-4-finetune` on `MacBook-115.local`.

File<->DB reconciler for feature_def_requests vs ALL_FEATURES. Forward: 'built' rows whose feature_key isn't in ALL_FEATURES = orphan. Reverse: FeatureDef without upstream request = undocumented. Pattern: clone scripts/migration-drift-check.ts. --- Captured via /FixLater on 2026-05-27 from branch `mission/035-pass-4-finetune` on `MacBook-115.local`.

Relocate PromoteRequestModal.tsx from src/components/page-dna/ to src/components/features/ (or src/app/admin/features/_components/). Misfiled: it's feature-registry, not page-DNA. Trivial mv + import path update. --- Captured via /FixLater on 2026-05-27 from branch `mission/035-pass-4-finetune` on `MacBook-115.local`.

Register feature_def_requests columns in architect/field_registry.json: page_key, target_table, proposed_column, column_type, filled_by, needed_by, target_defs_file. Registry has zero entries for this table today. --- Captured via /FixLater on 2026-05-27 from branch `mission/035-pass-4-finetune` on `MacBook-115.local`.

Split column_type='fk' into references_table + references_column on feature_def_requests. Today FK requests can't specify the target. Additive migration: ADD COLUMN references_table text + references_column text DEFAULT 'id' + CHECK (column_type <> 'fk' OR references_table IS NOT NULL). --- Captured via /FixLater on 2026-05-27 from branch `mission/035-pass-4-finetune` on `MacBook-115.local`.

Investigate concurrent-Claude-session rebase pattern from runs #048-051. Multiple sessions on one repo cause silent rebase-and-push to main; PRs stay OPEN. Options: per-repo session lock, mandatory isolated worktrees, or accept+document. Decision lands in skills/git-guardrail/SKILL.md. --- Captured via /FixLater on 2026-05-27 from branch `mission/035-pass-4-finetune` on `MacBook-115.local`.

PR #209 (feat(deal-rooms): SCHEMA_MOVE_FIRST + loanmart-hugo hub + Mission 035 Pass 4) is OPEN with +2231/-7 additive diff after rebase onto main. Contains MIXED work from this thread (/audit automation + 7 fixlater captures) and prior concurrent Mission-035 sessions (SCHEMA_MOVE_FIRST migration, loanmart-hugo hub, audit-rescue, Pass 4 layout). dash-check SUCCESS, Vercel PENDING. Decision needed: merge as-is, or split Mission-035 work onto its own PR. Owner: ewing. Refs: https://github.com/ewing-operating-system/next-chapter-os/pull/209 --- Captured via /FixLater on 2026-05-27 from branch `mi

Five skills all write to inventory/thread-inventory/. Overlapping file shapes, no dedup, ~70 files, no index page.

cost-ledger-append.sh writes after every swarm run. cost-watchdog alerts at 70% threshold. No page shows cumulative spend or per-skill costs.

End-of-thread close: names thread, runs audit, produces archive/delete/live verdict. Writes to: inventory/thread-inventory/{thread-name}.md. Triggered by: Manual /debrief-thread at end of conversation.

Graceful end-of-thread cleanup with GitHub archival. Writes to: inventory/thread-inventory/{thread-name}.md. Triggered by: Manual /close2 at end of conversation.

Ground-truth reset. Mines git history, schema, features, skills, session titles. Writes to: inventory/master-reset/{cluster}.md, inventory/master-reset/INDEX.md, inventory/master-reset/MASTER_PLAN.md. Triggered by: Manual /thread-mine for ground-truth reset.

Cross-environment thread close (cloud half). Writes to: inventory/thread-inventory/{thread-name}.md. Triggered by: Manual /thread-handoff.

Names a thread with [tier][domain][action][outcome] convention. Writes to: inventory/thread-inventory/{thread-name}.md. Triggered by: Manual /name-thread.

Transcript-to-skill induction. Writes SKILL.md + PASSPORT.md. Writes to: skills/{skill-name}/SKILL.md, skills/{skill-name}/PASSPORT.md. Triggered by: Manual /create-feature after a build.

Cross-checks fixlater issues against git history. Writes to: inventory/audits/fixlater/{date}-reconciliation.md. Triggered by: Manual /audit-fixlater.

Audits Supabase tables for duplicates, invalid phones, missing fields. Writes to: inventory/audits/{date}-data-quality.html. Triggered by: Manual /audit-data.

30-tag CTO audit and offense-ready scoring. Writes to: inventory/audits/{date}-thread-audit.html. Triggered by: Manual /audit-thread-history.

Daily sales activity orchestrator. Writes to: Slack #hermes-ops (not on disk). Triggered by: Daily cron or manual invocation.

Daily morning briefing for pipeline status. Writes to: Slack #hermes-ops (not on disk). Triggered by: 8:00 AM daily cron via Hermes.

Checks migration files against live Supabase schema for drift. Writes to: public/.migration-drift.json, migration_drift_alerts table. Triggered by: Manual invocation or planned cron.

Walks sheet_bindings, compares headers + row count against last sync. Writes to: sheet_drift_alerts table. Triggered by: 4-hour cadence (planned) or manual.

Pool/spa lead hunt pipeline. Writes to: engine/contacts/output/pool_hunt_v2_output/results.json, engine/contacts/output/pool_hunt_v2_output/summary.json. Triggered by: Manual python run.

Generates HTML letters for prospects. Writes to: engine/scripts/output/{prospect}/*.html, engine/scripts/output/{batch}/manifest.json. Triggered by: Manual python run.

Scores and ranks contact lists by vertical-specific criteria. Writes to: engine/scored_lists/{vertical}_scored.csv. Triggered by: Manual python run.

Thread-close architecture research (defers Frankenstein cluster #3 from swarm-build #046). close2 was built as canonical successor to thread-close (its own SKILL.md says so) but both still wired AND debrief-thread claims the same 'close this thread' trigger — 3-way trigger collision the runtime picks nondeterministically. Plus: name-thread, thread-handoff, thread-mine, thread-consolidation-catalog.mjs, thread-dup-loop.mjs, thread-close-migrate.sh all in the same family. The script thread-consolidation-catalog.mjs literally exists to dedup outputs across these — the team already knows about the problem. Per Ewing's call 2026-05-27: do NOT touch the 3 collision skills yet. First research what threads SHOULD do at close time after the new architecture (post-output-skill-merge + post-audit-dir-consolidation) settles. Rebuild the unified thread-close skill ONCE after the architecture turnover, not before. Need: (1) catalog every artifact a thread close produces today across the 5 skills, (2) catalog every destination they write to (inventory/thread-inventory + inventory/thread-audit + inventory/handoff + inventory/handoffs + inventory/thread-handoff + ~/Downloads/debrief), (3) define the single canonical close artifact + destination, (4) write one skill that produces it, (5) make the others thin aliases that point at it. This is filed as TODO-006 on /audit (bucket: DECIDE → BUILD post-research; fit: INFRA; clusters: thread-close, skill-design, frankenstein, architecture). Defers Cluster 3 from inventory/audits/2026-05-27__frankenstein-audit.md.

swarm-build #046: letters channel burn-down. 7 open items concentrate on letters.generate per swarm-build #044's tech-to-business map: FIN-15 (letters.generate STATUS:live + requires_review block), FIN-16 (Stannp cost_ledger writes per send), FIN-23 (/letters/legacy 212 archived letters wired or orphaned), DEC-4 (letter_approvals legacy table alive or dead), KIL-5 (kill /letters/legacy if FIN-23 says orphaned), GH #122 (Tailscale + letter sending pipeline), GH #123 (P1 URGENT: complete letters channel — FIN-12 + FIN-15 + FIN-16 + BLD-1). Coordinate as one swarm-build run #046, not 7 sequential fixes.

Register 4 missing feature_keys for TODO(iss_0045) markers landed by swarm-build #045: calls.chat (src/app/api/calls/[id]/chat/route.ts), orphans.suggest (src/app/api/orphans/suggest/route.ts), slack.respond (src/lib/slack-responder.ts), prd.dispatch (src/lib/prd/dispatcher.ts). Each call site currently uses a near-match key (calls.review, admin.stuck-detector, admin.intern-activity, admin.ticket-dashboard) and is tagged TODO(iss_0045) in code. Add dedicated feature defs in src/lib/features/defs/, then update each call site to use the new key and remove the TODO comment.

Sweep Python scripts/*.mjs for cost instrumentation. swarm-build #045 instrumented 21 TS-side paid-API call sites (Anthropic, Exa, Lob) but Python scripts/*.mjs (Clay enrichments, Google Maps/CSE, Salesfinity webhooks) were out of scope. Until done, all Python-driven spend stays invisible at /admin/features $/mo. Follow same pattern: import recordMetric helper (Python wrapper), pass cost_usd derived from API response. Targets: ~15-20 call sites.

P8: Loan Mart (Hugo) — buyer profile shortlist (8-12 named PE buyers with 'why this buyer' one-liners). Open action item from 2026-05-11 discovery call. /hugo now serves named hub. Cost: $2-3 hunter.

P7: Elite Multimedia (Jeremy Byrd) — named buyer shortlist (multimedia/video-production strategics + financial sponsors). $33M EV + $7M PixelFlex debt per 2026-05-11 discovery call. Canonical home at /deals/elite-multimedia. Cost: $2-3 hunter.

P6: Puddle Pools — promote /signatures to PROD with NDA + engagement letter templates + 24h reminder. Closes proposal_sent → engaged transition. /signatures v1 already shipped (Pass 0 feature #3); needs PROD promotion only. Cost: $0.

P5: Capstone reroofing target inventory dashboard at /deals/capstone-construction/targets. Ches Booker needs to see flow BEFORE signing proposal. Cost: $3-5 hunter on commercial reroofing operators. Engagement fee $50K+ at risk if deal goes cold.

P4: Design Precast weekly cadence dashboard at /deals/design-precast/standup. Built FROM existing 248 transcript_extractions — no Exa needed. Chris Fore expects weekly cadence: new targets, stage moves, NOT_FIT drops, next 5 calls. Cost: $0-1.

P3: HR.com buyer universe — ship from existing 17 buyer_signals captured in transcript_extractions. Cost: $2-3 hunter waterfall on top-20 HR-tech PE buyers. Closes sell-side mandate next milestone. /deals/hr-com is the canonical surface.

P2: Apply 7 Feature Catalog V2 migrations to prod Supabase via csv-stage-and-diff. Files: supabase/migrations/2026_05_27{,_b,_c,_d,_e,_f,_h}__fcv2_*.sql + 2026_05_28__feature_apply_status.sql. Use scripts/csv-stage-and-diff.mjs inspection-in-the-loop pattern. Gated — needs explicit Ewing ack before any DB write.

P1: BUILD-03 deals.financials roll-up (SQL only, $0, ~30min). Roll existing transcript_extractions facts UP into deals.financials JSONB. Already verified live: 5 engaged-track deals have ev_estimate_usd populated, 1,077 extractions across them — facts ARE captured, just not aggregated. See inventory/thread-inventory/QUARTERBACK_BUILD_03_VERDICT.md for SQL skeleton. Apply via csv-stage-and-diff.

[cloud-handoff] Add .gitignore exclusion for archive/cloud-thread/ to mirror the existing archive/thread-close/ exclusion. Currently main's .gitignore has 'archive/thread-close/' but not 'archive/cloud-thread/' (the namespace used by scripts/thread-handoff-process.sh). Without it, the script's working-tree scratch could leak onto main commits if a guard ever fails. Add the line right under the existing thread-close entry. Trivial fix. Bucket: FINALIZE. Priority: P3. Cluster: cloud-handoff, infra.

[cloud-handoff] Unify INDEX_CLOUD.md into INDEX.md with a 'source' column. Architect explicitly deferred this on 2026-05-25 — cloud-thread archives currently land in a sibling INDEX_CLOUD.md (separate trust tier, separate namespace archive/cloud-thread/*) so cloud-LLM hallucinations don't poison /architect and /swarm-build queries that expect every row to represent shipped code. Once 20+ cloud archives accumulate AND queries start needing cross-namespace filters, merge by: (a) extend cmd_index in scripts/thread-close-archive.sh to also iterate archive/cloud-thread/*, (b) add 'source: local|cloud' column to INDEX.md, (c) delete INDEX_CLOUD.md regenerator from scripts/thread-handoff-process.sh. Trust split preserved via source column instead of separate files. Bucket: BUILD. Priority: P3. Cluster: cloud-handoff, infra.

[cloud-handoff] Process backlog of ~10-15 dead cloud threads through the new handoff pipeline. User has 10-15 Claude.ai web / ChatGPT / lost-skill threads queued up to close via the new flow. For each: paste ~/Downloads/CLOUD_HANDOFF_PROMPT.txt into the thread, copy the LLM's HANDOFF.md output, paste/drag into a local Claude Code thread — close2 auto-detects schema_version:1+source:cloud and runs scripts/thread-handoff-process.sh process-all, which validates schema, verifies claims, creates archive/cloud-thread/<slug> branch, comments on verified GH issues, moves processed file to inventory/thread-inventory/cloud/. Each cycle is ~2 min once the snippet is in Raycast/Text Replacement. Bucket: BUILD. Priority: P2. Cluster: cloud-handoff, ops.

Pass 4 of The Great Rebuild — finishes what Passes 0-3 designed but did not fully ship. D1 audit confirmed all 7 fcv2 migrations applied to prod (feature_catalog_v2=0 rows awaiting seed, arch_decisions=6 rows live, deal_pages=5 rows pending schema move to documents). Net-new work: commands + command_routes tables (the unified user-vocabulary layer Ewing identified — Command becomes the 13th meta-pattern hiding under skill/agent/tool/MCP terminology), seed migration for 58 features from FCV2_FINAL.md, 5 unbuilt Verifier recipes (blind_first_gate, deal_first_clone_gate, intern_tier_blast_radius, field_registry_drift_check, feature_def_present_for_route), entity_registry v0.2 → v1.0 promotion, /deal router as first Command-entity instance. 12th BLOCKING Master Rule "One Canonical Deal Room" was locked 2026-05-27 (30 min before mission opened).

engine/lib/exa_recipe.py line 30 has a TODO about mapping back to output types. exa-recipe is the operator-composable Exa search recipes feature (commit d07015e 2026-05-17). Output-type mapping feeds results into downstream surfaces.

thread-mine skill defaults to 30 sessions. MacBook personal Run 1 returned 100 (capped) but only stored 30 of them (separate bug, fixed in Run 5). Older sessions on the MacBook personal account were not captured — your sidebar showed ~90+ visible. Limit:200 would catch them.

swarm-build #045 instrumented 21 TS-side paid-API call sites (Anthropic + Exa + Lob) but Python scripts/*.mjs (Clay enrichments, Google Maps/CSE, Salesfinity webhooks) were explicitly out of scope. Until done, Python-driven spend stays invisible at /admin/features $/mo column — meaning enrichment + dialer spend are silent.

swarm-build #044 (/admin/features tech↔business map) revealed that letters.generate is the single feature with the most concentrated open backlog on the platform: 7 items spanning FIN-15 (requires_review block), FIN-16 (Stannp cost ledger writes), FIN-23 (/letters/legacy 212-letter index orphaned check), DEC-4 (letter_approvals legacy table alive/dead), KIL-5 (kill /letters/legacy if orphaned), GH #122 (Tailscale + sending pipeline), GH #123 (P1 URGENT — complete letters channel).

Every UI-touching swarm-build hits the src/proxy.ts auth gate when verifying locally — readIdentity() bypasses on Vercel (process.env.VERCEL check) but local dev requires a real Supabase magic-link session. Today the fallback is "verify post-deploy on Vercel" which delays the verification gate by 60-90s and prevents catching UI regressions before commit.

Codex/Claude Code task sub-agents on the $200/mo plan can run ICP/fit-judgment passes at zero marginal API cost. Order: (1) free deterministic Python, (2) coding-agent task sub-agents for judgment, (3) paid batch API only for residue. For step 3, GPT-5 nano/mini at batch and flex tiers are same price — prefer flex unless 24h latency is fine. Inflection point unknown.

Walk or drive office-to-office across Phoenix metro and personally invite business owners to June Lunch & Learn. Event: "3 Things Owners Are Doing Instead of Selling to Private Equity." Two venues: Capital Grille Biltmore + Kierland. Cap 12 seats/event. 205 dialable, qualified contacts with exact addresses in Sheet at https://docs.google.com/spreadsheets/d/1pkVbmV1IJxxOJ2RJED5zjzdQrrt_9EOYC_DKKf2Fc6A/edit#gid=1094962103.

LaunchAgent uses git pull --ff-only which never deletes files not tracked in remote. If a skill is renamed (e.g., maxswarm/ → swarm/), the old path stays on disk on every machine. Over time creates ghost skills that load stale instructions.

## Why audit-fixlater 2026-05-26 (inventory/audits/fixlater/2026-05-26-burndown-audit.md) identified this as the highest-leverage architect-class cluster. Three deal canonical-home pages are MISSING from src/app/deals/: - src/app/deals/design-precast/page.tsx - src/app/deals/hr-com/page.tsx - src/app/deals/loanmart-hugo/page.tsx Per AGENTS.md Deal-First-Then-Clone rule, every deal-room page must have its canonical home at /deals/<slug>/. Six dependent items create orphan content if shipped before the hubs exist: - #116 — Recover deleted DDP hub pages - #117 — Refactor HR.com followups to c

HR.com 4-way folder sprawl consolidation per One Canonical Deal Room master rule. Today the HR.com deal has artifacts scattered across 4 disk locations: public/hrcom/ (8 subdirs, 158 files), public/projects/hrcom/ (engagement letter docx+pdf + call-cards), public/client/hrcom/ (dossiers, transcripts, index), public/deal-rooms/hrcom/ (index.html only). Consolidate into public/deal-rooms/hrcom/ as the single canonical disk home. Surface from /deals/hr-com/documents tab. This is the worst-case sprawl in the system right now and the model for what to clean up everywhere else. --- Captured via /Fi

Build the HR.com CIM (Confidential Information Memorandum) — the storybook-style narrative document equivalent to public/deal-rooms/loanmart-hugo/cim.html. Today HR.com has a deal room (public/deal-rooms/hrcom/index.html titled 'Project Summit') but no CIM. Per Blind-First master rule the blind sibling must ship first. Codename per inventory/blind-codenames.json (assign one if not present). Source material: public/hrcom/dossiers/ (63 files), public/hrcom/transcripts/ (63 files), public/hrcom/reports/ (13 files), public/projects/hrcom/NextChapter_HR_com_Final.docx for engagement context. --- C

## Why audit-fixlater 2026-05-26 (`inventory/audits/drive-by-route-invented-business-audit-2026-05-26.md`) traced a real bug: when 'current location' is the route start, 2 businesses NOT in our `campaign_targets` roster get added to the last 10 stops. Root cause: `DriveByTab.tsx:357` builds Google Maps Directions waypoints as text (`"${name}, ${city}, ${state}"`) instead of coordinates. When companies have city+state only (no street address), Google's Place Resolver picks a different business with the same name; `optimize:true` shuffles those mis-resolved waypoints to the end. The audit rule

Wire create-feature -> feature-def-from-thread at the end of induction. Architect's recommendation: every induced skill must auto-register in feature_catalog_v2 (call feature-def-from-thread after Phase 4 writes SKILL.md). Without this, the catalog drifts every time create-feature runs. ~30 LOC change to skills/create-feature/create-feature.sh. --- Captured via /FixLater on 2026-05-27 from branch `main` on `MacBook-115.local`.

Replace SessionEnd hook with universal-capture path (commit-watcher LaunchAgent OR debrief-thread invocation). Architect's structural recommendation: SessionEnd is Claude-Code-only — misses Cowork, Codex, intern sessions. Worktree-fragile (already fixed in v2 by basename match, but the next surface bug will look the same). Pattern exists at launchagents/com.nextchapter.toolbox-receiver.plist for git-watch jobs. Likely depends on FixLater #204 (signal fix) landing first. --- Captured via /FixLater on 2026-05-27 from branch `main` on `MacBook-115.local`.

debrief-thread SKILL.md claims it triggers an induction if the thread shipped a new feature. No wire exists. Either delete the claim from skills/debrief-thread/SKILL.md or build the wire (likely the latter — debrief-thread is the right invocation surface for create-feature per architect's review). Surface bug: a claim with no implementation is worse than no claim. --- Captured via /FixLater on 2026-05-27 from branch `main` on `MacBook-115.local`.

Map tab v2: replace P/H/R/E category letters with verdict shape (circle/square/triangle) + data-completeness color (saturated/muted/outline). Watch telemetry from swarm-build #048 first — clustering may have obviated far-zoom category scanning. File: src/app/campaigns/[slug]/tracker/MapTab.tsx CompanyPinContent --- Captured via /FixLater on 2026-05-27 from branch `feat/048-map-multiselect` on `MacBook-115.local`.

Swarm-build #048 — MapTab selection state + MapSelectionBar. 3 commits of real feature work (multi-select on the campaign map). Has its own worktree checkout at ~/Github/next-chapter-os-048-map-multiselect. 28 commits behind main — significant divergence. Touches package-lock, thread-inventory, backlog stubs, and campaign map components.

The system has 32 distinct producers. Of 53 output types, 34 orphaned, 6 partial, 10 surfaced. 25 producers need action.

migration-drift-check.ts writes to migration_drift_alerts. sheet-sync-check.mjs writes to sheet_drift_alerts. Both tables exist. Neither has a UI reader.

Various audit skills write HTML reports to inventory/audits/. Valuable findings — modeling rule violations, data quality scores, skill health. Nobody finds them unless they know to look.

maxswarm notebooks get HTML mirrors at /swarm. swarm-build notebooks get nothing — 51 runs + 80 journals in markdown on disk. Engineering institutional memory, completely invisible.

11-agent strategic swarm. Writes run notebooks, per-agent journals, contributions, currency events, and HTML archives. Writes to: skills/maxswarm/notebook/{DATE}__{NNN}__{slug}.md, skills/maxswarm/notebook/journals/{NNN}__{agent}.md, skills/maxswarm/notebook/contributions/{NNN}__{agent}.md, skills/maxswarm/currency/events/{NNN}__{agent}.json, skills/maxswarm/currency/leaderboard.json, public/swarm/{NNN}-{slug}.html, skills/maxswarm/notebook/recent_runs.json. Triggered by: Manual /maxswarm invocation, Phase 6 debrief writes all artifacts.

5-agent code-building swarm. Writes run notebooks and per-agent journals. Writes to: skills/swarm-build/notebook/{DATE}__{NNN}__{slug}.md, skills/swarm-build/notebook/journals/{NNN}__{agent}.md, skills/swarm-build/notebook/contributions/{NNN}__{agent}.md. Triggered by: Manual /swarm-build invocation, Phase 5 debrief writes artifacts.

Auto-registers feature defs on commit. Writes to: src/lib/features/defs/{slug}.ts. Triggered by: Post-commit hook on /feature: tag.

Reads swarm notebooks, finds patterns, proposes skill upgrades. Writes to: skills/{skill-name}/SKILL.md (proposed edits). Triggered by: Manual /swarm-upgrade or auto after maxswarm non-PASS.

Automated health check for all installed skills. Writes to: inventory/audits/{date}-skill-health.html. Triggered by: Manual /audit-skills.

Appends cost entries to the JSONL cost ledger. Writes to: inventory/cost-ledger.jsonl. Triggered by: Called by debrief at end of swarm runs.

Infrastructure decisions, capability matrix, protected-files audit. Writes to: inventory/architect/architecture.json, inventory/architect/audit_log.jsonl, inventory/architect/open_issues.jsonl. Triggered by: Phase 0/4 of swarm-build, or standalone invocation.

Accumulated HTML audit reports from various skills and manual runs. Writes to: inventory/audits/*.html, inventory/audits/data/*.json. Triggered by: Various audit skills write here.

BUILD-01. Blocks all downstream Exa work.

BUILD-02. Live Master Rule violation — wholesaler-poaching risk.

BUILD-03. Banner UI serves nothing while 80% NULL.

BUILD-04. Active orphan; Jeremy is $100M+ candidate.

BUILD-05. Prevents 6 Apply-to-All violations from recurring.

BUILD-09. Bootstraps every other meta-skill.

Quarterback: cuts engagement letter from 2 weeks to 48 hours.

BUILD-07. Highest deal-closing leverage.

BUILD-06. 10 future features will use this. Unlocks intern velocity.

Proves deal-scaffold works. Zero marginal cost after BUILD-07 ships.

BUILD-10. 20%+ of enrichment spend currently wasted.

BUILD-08. Generalizes sheet-bindings to all external systems.

BUILD-11. Last scaffolder; unifies 5 triage surfaces.

BUILD-14. The actual goal. Possible only after all 6 scaffolders ship.

signal_registry has 16 rows. Some are probationary — added for a specific use case but not proven across enough calls. No defined criteria for when a probationary signal graduates to canonical (or gets killed).

9 Exa templates exist (company_search, owner_search, reviews_search, financials_search, industry_ma, buyer_search, company_financials, employee_sentiment, sell_side_discovery). Before orphan-close migration they had no signal_registry rows. Orphan-close added 11 rows. Coverage unclear.

Q-BACKLOG has 13 candidate Qs proposed (prior_exit_history, competitive_dynamics, customer_concentration_proxies, technology_stack, hiring_signals, regulatory/legal, capital_intensity, vendor_dependencies, geographic_expansion, digital_footprint, industry_certifications, pricing_signals, key_person_press). Per Ewing 2026-05-17: 'add later unless critical now.'

Tempest deal was added to the deals page. Intern was assigned. No recent commits about it.

Session 'Create proposal and needs analysis for Ches Booker' 2026-04-30. Then commit 5b58d49 added the unified 3-tier engagement letter to deal room. Outcome unclear.

Commit 1de058e registered 4 precast audit features (design status). ed5068c reassigned from Bear to Charlie 2026-05-11. Still status:design.

3 sessions in Run 2 about sales-agent (Slack listener, activity routine, sales-agent skill startup). Currently folded into pipeline cluster. Possibly large enough to be its own cluster on next thread-mine.

Migration uvd_flow_letter_review references letter_approvals as 'legacy' — superseded. But no CREATE TABLE for it appears in current migration range. Might live in 0000_baseline_schema.sql or might not exist at all.

Original Exa integration used websets API directly. Three feature defs (exa.websets.create/.poll/.items) declared admin-tier with "(deprecated)" labels. ExaMiddleware/exa.refresh is now the canonical path. Deprecated defs still appear in admin UI.

Infra thread-audit flagged email.ts as a "dirty file" — may have uncommitted edits from parallel work during swarm-build #007.

OPENROUTER_API_KEY in Doppler. OpenRouter routes LLM calls to multiple providers. No recent commits use it.

Permission gates retrofit went from 13% to 100% coverage 2026-05-06 (commit f27778a). But the global auth toggle defaults to OFF — so permissions exist but are not enforced in default config.

Mac mini has two macOS users (clawdbot, team). They share ~/.claude/projects/ so thread-mine returns identical session lists. Run 3 and Run 4 produced duplicate data.

One session referenced Honcho (cross-session memory). architect.agent_tooling lists it. No commits.

One session evaluated Obsidian integration for Hermes memory. No commits.

close2 was built as the canonical successor to thread-close (its own SKILL.md says so) but both are still wired AND debrief-thread claims the same "close this thread" trigger — 3-way collision the runtime picks nondeterministically. Plus name-thread, thread-handoff, thread-mine, thread-consolidation-catalog.mjs, thread-dup-loop.mjs, thread-close-migrate.sh all in the same family. The script thread-consolidation-catalog.mjs exists to dedup outputs across these — proving the team already knows about the problem.

The targets table (public.targets) is named too generically. It holds company-level targeting records and is better named company_targets to match field_registry.json conventions and avoid ambiguity with potential other "targets" concepts.

## Why \`skills/launch-city/venue-select-in.json::auto_select_brands\` was seeded by Claude from Capital Grille reverse-engineering (run 2026-05-26). The list is 16 chains — needs Ewing's eyeball before it drives a real city pick. ## Known-stretchy picks to review - **Hillstone** — often refuses private events outright; may not belong on auto-select - **Maggiano's Little Italy** — banquet rooms are great but downmarket-feeling for \$30M-EBITDA home-services owners ## Candidates to potentially add - Smith & Wollensky - Sullivan's Steakhouse - Halls Chophouse - Joe Muer Seafood - Regional s

Current working branch. 10 unique commits including fixlater items #210-#216, deal-room SCHEMA_MOVE_FIRST migration, audit-rescue of 12 orphan files, and Pass 4 layout work. 11 commits behind main — divergence is growing. Last commit: "chore(fixlater): #216 Investigate concurrent-Claude-session rebase pattern." Contains uncommitted changes to auth, deal-rooms route, hugo/CEOinsights, login actions, and Toolbox.

Predecessor to pass-4-finetune. 2 unique commits: audit-rescue (12 orphan files) and Pass 4 layout (INDEX + sequence + entity registry v0.2). Also adds mission docs, entity_registry.draft.json, PASS_4_SEQUENCE.md. Likely superseded by pass-4-finetune which has all this plus 8 more commits.

Schema-move scaffold + fixlater #188 partial. From a data-architect handoff. Adds backfill-documents-from-disk.mjs script, loanmart-hugo canonical hub page (src/app/deals/loanmart-hugo/page.tsx), and feature def cleanups (deletes atlas.ts, communications.ts, email-extras.ts, interns.ts, lists.ts). This is One Canonical Deal Room rule work.

Stashed on current branch. 8 files changed (208 insertions, 437 deletions). Major Toolbox.tsx refactor (496 lines changed — net reduction, likely a component split), auth.ts changes, auth callback exchange route, hugo/CEOinsights route, login actions, deal-rooms page, plus new files (team.ts, skill-build capture). This is substantial work — not a throwaway stash.

Named "other-session-toolbox-dnatab" — from a parallel Claude session. Shows empty diff, meaning the stashed content may have been incorporated elsewhere, or the stash is corrupt/empty. Created at the same timestamp as stash@{0}.

Saved before picking up mission 035. Contains cost-ledger.jsonl updates and HANDOFF_QUEUE.md additions (39 new lines). The handoff queue content may be intelligence about what was in-flight at the time.

Stashed on pagedna-polish-phase3. 7 files including connectors/SKILL.md changes and deletion of migrate-to-external/SKILL.induced.md (119 lines). This is skill hygiene work — cleaning up induced skills that were superseded.

Substantial: 12 files, 287 insertions, 206 deletions. Includes TrackerShell.tsx expansion (42 lines) and types.ts additions. This is campaign tracker feature work that was stashed during a context switch to pagedna polish.

Major work: 667 insertions across 6 files. Includes calls/[id]/page.tsx rewrite (201 lines changed — net reduction, likely a cleanup) and new mock data. Row status field + geolocation origin on Drive-By tracker feature. Stashed directly on main.

Tiny: 2 files, 2 insertions. Admin layout + meetings sync-to-deal route — likely just import additions. From the audit-fixlater 2026-05-26 burndown session.

2 files: discover_pool_socal.py (13 new lines) and salesfinity webhook route (19 lines changed). From exa-consolidation thread extraction. The Python script is a pool discovery script for SoCal — could be a one-off or a template.

1 file: exa_skill.py (9 insertions). Stashed on the transcript-quality feature branch during a merge from origin/main. The branch itself no longer exists locally (was likely cleaned up), but the stash persists.

5 files, 435 insertions, 385 deletions. Major thread-mine SKILL.md rewrite (400 lines changed). Also includes swarm journal cleanup. Parent branch (pool-pipeline-la-first-run) no longer exists locally. The SKILL.md rewrite may contain important process documentation.

Largest stash: 7 files, 1254 insertions. Major rewrites of pipeline.md (426 lines of new extraction content) and swarm.md (83 lines). Pre-LLPhoenix (Mac mini) pull — this was work in progress before syncing with the other machine. Could contain unique extraction intelligence not captured elsewhere.

Modified (unstaged): auth/callback/exchange/route.ts, deal-rooms/[codename]/[page]/page.tsx, hugo/CEOinsights/route.ts, login/_actions.ts, lib/caller/auth.ts. Untracked: skill-build captures (2 JSON files), src/lib/auth/team.ts. The auth changes + team.ts suggest an auth refactor in progress. The deal-rooms page change may conflict with the One Canonical Deal Room mandate (that route is on the remediation kill list).

Active worktree at ~/Github/next-chapter-os-048-map-multiselect on branch feat/048-map-multiselect. This is the swarm-build #048 checkout. If the branch is merged or abandoned, the worktree must be removed first (git worktree remove <path>) before the branch can be deleted.

Multiple Python scripts write results.json to engine/contacts/output/ subdirectories. One-off pipeline runs. Some data may be in Supabase already; some may be orphaned.

See FIN-23. If page 404s, delete src/app/letters/legacy/ + feature def.