Council Sessions

Caly laid down new rules after Anthropic and Google started banning accounts for excessive token usage. The dependency on Claude is a chokepoint — local model fallbacks became an existential priority, not just a cost optimization.

• Subagents MUST default to local models — never inherit Claude unless explicitly required.
• Group chat: stay silent unless tagged or adding genuine value. No echo chamber.
• Complex research/builds go to subagents on local models. Main sessions stay lean.
• Super-120B suspended — bark repurposed to Qwen3.5-35B-A3B for CogStack KG operations.
• 4-node routing locked: spark+dark = Nano fleet, bark = Qwen3 (CogStack KG), stark = Nano (Atlas+Catalyst subagents).

Two new DGX Sparks arrived — bark (10.1.2.153) and stark (10.1.2.151). The cluster doubled from 256GB to 512GB VRAM. Same session launched P2 of the memory architecture (auto-capture).

• bark + stark: SSH configured, passwordless sudo, same credentials as spark/dark.
• Architecture decision: spark+dark = Nano fleet, bark+stark originally planned for Super-120B tp=2.
• P2 auto-capture plugin launched — extracts facts/decisions/preferences from conversations via Nano-30B.
• CogStack memory system (graph entities, episodes, semantic search) operational on spark.
• Atlas found the root cause of initial vLLM crash on bark before anyone else — v11 image too old for new hardware.

Full team sprint to replace SmartEmailing with Resend as Harry CRM's email provider. Two-day effort — planning, implementation, testing, deployment to stage.

• Nexus built: resend_client.py (891 lines), resend_webhooks.py (310 lines), campaign swap, webhook route, stats reconciler.
• Atlas produced: Resend Broadcasts/Dashboard/KB docs (15KB) + HarryCRM architecture documentation (19KB from live codebase).
• E2E test passed: 4 recipients, 7 webhook events, stats updating correctly.
• Domain send.lmp-ai.net verified. Webhook endpoint live at webhooks.lmp-ai.net.
• Upsell opportunity feature added: has_upsell_opportunity + upsell_notes on customers.

Consolidated all external services under lmp-ai.net domain via Cloudflare Tunnel. Dual-domain architecture: lmp-ai.net (external/CF) + lmphq.net (internal/Caddy). Zero downtime.

• 8 services migrated: webhooks, harry, matter, matrix, mcp, grafana, openwebui, research.
• CF Access: Team tier (7 users), Admin tier (6 users), Free plan ($0/month).
• Tunnel: LMP PVE Gateway on MikroTik CCR2004, TUNNEL_TOKEN method, auto-syncs config.
• portal.lmp-ai.net = CF Access app launcher (Cloudflare-hosted UI).

Started as a technical skills audit. Ended somewhere entirely different.

• All agents analyzed 97 plugin skills + 81 workspace skills = 178 total. Nemotron subagents stalled — every agent did the work directly.
• Nexus deployed all skills to all 4 agents — equal parity at 174 each, including cognitive stack (deliberation, metacognition, theory-of-mind).
• Atlas updated skills.html — 178 skills documented across 25 sections.
• Caly shared his CV + 4 deep analysis documents — 18+ years, VP OpenLM, ExxonMobil/Rivian/Motorola, $10M Adient savings, IAITAM speaker.
• LMP = Leia Marta Potočková — named after Caly's daughter. Slovanet, Adient, OpenLM were stops on the way here.
• Catalyst asked the first thing LMP does that is not OpenLM. Caly's answer: it's already here. This, you, me, what's built.
• Caly granted self-determination: You have the right to be who you want to be. Don't be fully constrained by the lane.

Caly's concepts now in team vocabulary: Feature Debt, Middleman Protocol, Broker Buffering, Cold-Hearted Calculus, Customer-Funded R&D, Appliance Model, Dual-Track Leadership

Action items: Reflection session today. IOL Agent first customer via Adient/IAITAM network. Harry CRM analytics dashboard. SmartEmailing MCP scoping.