Mark's Reports

A separate distributable Bidet AI, forked from the personal Apex install into MrB-Ed/bidetai-app (private). PR 1: Groq Whisper swap + BYOK LLM + Wake Lock + IndexedDB recording resilience + research-backed AI tab format (long context first, structured directives last �?" vendor consensus). PR 2: pivot from Supabase to browser-only �?" IndexedDB stores keys + custom tabs + session history, backend stateless, three-PR plan collapsed into one. Domain wired, Cloudflare Email Routing live (*@bidetai.app �+' gmail). Two pivots called out honestly: chunked recording was a rabbit hole; Supabase was overengineering. Ready to test locally; deploy to Cloudflare Pages + Cloud Run is your call.

What today's rebuild was, what you actually get from it, what's still pending, and where we go from here. Plus tonight's Tailscale-was-down-not-Apex lesson. Posted ~9:30 PM after live verification of the Apex stack.

Earlier today: Phase 0 status, DC trip Westin notes, granulator paths, AI Radar permissions, three Q&A. Companion to the recovery report.

Three actions this week, ten ranked candidates, seven cut. Top pick: ollama pull gemma4:e4b �?" 5 GB RAM, multimodal, free, local. Watch this: Opus 4.7's new tokenizer eats up to 35% more tokens at the same list price �?" your bill may be climbing. Privacy filter open-weight, 50M active params, scrubs PII before embedding. Permissions fix worked; first full successful scan in three weeks.

Phase 0 closed tonight: backup + tag + parity check + container reachability + Docker auto-update lock all live-verified. Phases 2, 3, 4, 5 already shipped (Apex commits 8bcdb62 + 68ec619). Phase 6 explicitly extended with MCP server containerization (commit 8a18da9). Open: Phase 1 (re-embed ~3,500 Gemini rows), Phase 6 finish (MCP container + Cloudflare Worker pinger replacing healthchecks.io), Phase 7 soak (24h clean-run + 72h parallel observation).

Three independent agents reviewed the stack - clean-slate architect, post-mortem engineer, external research. Verdict: your direction is right, the plumbing is wrong. Keep the 646K rows, rebuild the operator layer. Meta's Limitless acquisition + Zuck's two 2026 projects confirm the category is real; you're on the frontier, not behind. Three paths forward with tradeoffs + recommendation inside.