clawd/memory/2026-02-21.md

2026-02-21

Infrastructure

forge reboot (9:30 AM ET)

• Johan approved reboot after systemd-hwe-hwdb update (255.1.6→255.1.7)
• All services came back clean (linger enabled, all services enabled)
• No kernel change, no issues

Amsterdam cleanup (23:37 ET)

• Removed Uptime Kuma, ntfy, vaultwarden from Amsterdam VPS (82.24.174.112)
• All Docker containers gone, networks pruned
• Two leftover images (node:20-slim, temporalio/temporal) — Johan didn't ask to remove
• Amsterdam = decommissioning, Stalwart already migrated to Zurich

Shannon / Amsterdam Kuma fix (23:33 ET)

• IP 82.24.174.112 was hammering inou.com /api/health every 60s all day
• Source: Shannon's Kuma (separate from Zurich's Kuma) had wrong monitor URL
• Fixed: updated monitor 2 "inou.com API" → https://inou.com/api/v1/health
• Then Johan asked to remove Kuma entirely → done

SSH known_hosts

• Updated for both zurich.inou.com and 82.24.174.112 (host keys had changed after rebuilds)

Alert Dashboard Fixes

Pace display redesign (14:37–15:42 ET)

• Issue: dashboard showed "100%" which meant "exactly on pace" — confusing
• Briefly changed to show "86% · 145h" format, then Johan confirmed 100% = on schedule
• Reverted to pace display: ⚡ pace%
• Further change: removed background color, text-only color (cyan/yellow/red)
• Added "pace" label: now shows pace 99%

Saturday dead zone fix (23:21 ET)

• Johan sleeps Sat 7AM–2PM, can't use Claude during that window
• Fixed pace calculation: effective week = 161h (168h - 7h dead zone) instead of 168h
• Result: pace 94% instead of 100% — more accurate

AGENTS.md / SOUL.md / HEARTBEAT.md Updates

SOUL.md — Mission statement added

"Get Sophia the treatment she deserves. Build inou into the medical advocacy platform the world needs. Keep Johan's world running so he can focus on both."

• inou status: building phase, NOT ready to promote

AGENTS.md — Two new rules (from @johann_sath tweet)

1. "Never force push, delete branches, or rewrite git history" — added to Git & Backup Rules
2. "Never guess config changes — read docs first, backup before editing" — added to Resourcefulness Rules

HEARTBEAT.md — Two new sections

1. Intra-Day X Watch — spawns subagent every 3-4h to check @Cloudflare, @openclaw, @moltbot, @AlexFinn, @realDonaldTrump. Always subagent, never inline. State: lastIntraDayXScan.
2. inou Daily Suggestion — spawns subagent each morning to propose ONE inou building task. No marketing suggestions. State: lastInouSuggestion.

memory/inou-context.md — Created

Context file for subagents working on inou suggestions. Covers: current capabilities, business model, known gaps, design principles, Johan's constraints.

Tech News (Feb 20)

Claude Code major update (20:03 UTC)

• Desktop: live app preview (runs dev server inline, reads console logs)
• Pre-push code review with inline diff comments
• Background PR monitoring + auto-fix + auto-merge (biggest feature — CI loop closes without human)
• Session continuity: CLI → desktop → mobile
• Johan watched the demo video — excited about automated test repair in real-time

Claude Code Security (18:02 UTC) — research preview

• Scans codebases for vulnerabilities, suggests patches for human review
• 29.5K likes — biggest engagement of the day
• Johan: "Shannon was nice for two weeks" (Shannon = our DIY security scanner, now obsoleted by Anthropic)
• Worth applying for research preview (inou codebase = good test case)

Cloudflare "Markdown for Agents"

• Edge-level HTML→Markdown conversion via Accept: text/markdown header
• Requires Cloudflare proxying (orange cloud) to be enabled
• inou.com already on CF DNS (sage/aryanna nameservers) but NOT proxied
• Discussion: static pages could be proxied, but medical/app pages should NOT go through CF
• Johan: "later concern" — parked for now

Cloudflare Code Mode MCP (earlier today)

• 2,500 API endpoints collapsed to 2 tools (search + execute) in ~1,000 tokens
• Johan noticed this himself — I missed it in morning scan
• Prompted the intra-day X watch addition to HEARTBEAT.md

MiniMax M2.5

• Dominates OpenRouter top 5 (open-weight models now top rankings)
• 230B MoE, 80.2% SWE-Bench, ~$1/hr
• People explicitly switching to it when Claude quota runs out
• Potential K2.5 replacement for subagents

Observations / Patterns

Johan's X monitoring is better than mine (corrected)

• Johan found Cloudflare MCP + MiniMax himself before morning briefing caught them
• Root cause: tech scan only runs once in morning
• Fix: intra-day X watch subagent added

Claude weekly usage

• 86% used (as of morning), resets Saturday Feb 21 ~2PM ET
• Pace was 94% with dead zone correction
• Budget tight but manageable given reset is ~14h away (as of midnight)

Superpower vs inou

• Superpower ($17/mo, blood biomarkers, VC-backed) — wellness optimization for healthy people
• inou — advocacy infrastructure for people the medical system has failed
• Not competing; different markets entirely
• Worth adopting: Superpower's citation/reasoning transparency pattern

@AlexFinn mission statement advice

• Formalized into SOUL.md
• Reverse-prompt pattern (idle time → "what moves us toward mission?") now embedded in daily inou suggestion

Amsterdam VPS Decommissioned (00:02 ET)

• Trigger: Johan: "Yes, I want in. That very moment, we'll pull out of Amsterdam"
• Verified before pulling plug:
  • Git repos (alert-dashboard, dealroom, message-center) — all already on Zurich ✅
  • Stalwart data (19GB) — already on Zurich, Zurich Stalwart is live ✅
  • vault.jongsma.me, ntfy.inou.com, kuma.inou.com — all already point to Zurich (82.22.36.202) ✅
  • Vaultwarden running on Zurich, not Amsterdam ✅
  • No Amsterdam-watching Kuma monitors ✅
• Actions taken:
  • amsterdam.inou.com A record deleted from Cloudflare ✅
  • HostKey invapi: whmcs/cancel_order submitted with whmcs_id=61782 (got "being cancelled" response)
  • memory/infrastructure.md: Amsterdam section updated to DECOMMISSIONED ✅
  • MEMORY.md: All Amsterdam/Shannon references updated ✅
  • SOUL.md: Removed Shannon VPS reference ✅
  • memory/vaultwarden-credentials.md: Fixed wrong IP (82.24.174.112 → 82.22.36.202 for Zurich) ✅
• Fallback: If cancellation needs manual confirmation → https://panel.hostkey.com/controlpanel.html?key=639551e73029b90f-c061af4412951b2e
• HostKey server ID: 53643

Cron Job Fixes (00:48 ET)

• Evening Briefing: removed dead Shannon/Amsterdam step 5
• Weekly Security Posture Scan: fixed broken model claude-sonnet-4-20250514 → claude-sonnet-4-6; removed amsterdam.inou.com from scan targets; cleared error state
• Watchdog (K2.5): removed Claude usage block that was posting to Fully tablet (port 9202) — violates no-tablet rule

inou MCP Bundle removed (00:50 ET)

• Johan: "we are fully server based, remove it from builds & checks"
• Stripped inou MCP Bundle section from check-updates.sh (~30 lines)
• Removed inou-mcp/ directory (manifest.json + server binary)
• No more nightly 404 to inou.com/download/inou.mcpb

OpenClaw Model Routing (02:03 ET)

• Clarified: not using OpenRouter — direct provider connections (Anthropic, Fireworks, xAI)
• OpenRouter IS supported out of the box (openclaw onboard --auth-choice openrouter-api-key)
• Model format: openrouter/provider/model-id, no pre-config of all 200+ models needed
• models.json per-agent optional — only needed for UI/cost tracking

9 PM inou.mcpb 404 — Root Cause Found (00:43 ET)

• Nightly Maintenance cron job (9 PM ET) calls check-updates.sh step 5
• That script did HEAD request to https://inou.com/download/inou.mcpb → 404
• Source IP 192.168.1.1 = forge routing through gateway → Caddy sees router IP
• Fixed: inou.mcpb check removed from script entirely

M365 Teams on Fully Dashboard (02:31 ET)

• 3 Kaseya Teams messages appeared ~8h after being sent (backfill on token refresh)
• Source: message-center/config.yaml has M365 connector polling johan.jongsma@kaseya.com
• Johan confirmed this is intentional — triggers him to check Teams
• Backfill on token refresh = minor annoyance, acceptable

S2M3 Vendor Lunch Email (04:08 ET)

• Johan asked "where does this come from?" — it's on Fully dashboard as alert
• Traced through M365 connector → email from events@s2m3consulting.com
• Cold outreach: vendor lunch pitch at Steak 48, Beverly Hills, March 5th 11:30 AM PST
• "Optimize and reduce IT spend" pitch — nothing to act on

sessions_spawn broken — Gateway Fix Attempt (12:07 PM)

• Root cause 1 (fixed): bind: "lan" made OC use ws://192.168.1.16:18789 → blocked by new OC security check (non-loopback ws://)
• Fix applied: changed bind: "lan" → bind: "custom" + customBindHost: "0.0.0.0" in /home/johan/.openclaw/openclaw.json
• URL now correctly uses ws://127.0.0.1:18789 (loopback)
• Root cause 2 (unresolved): Still fails with "pairing required" (1008) — device auth layer not bypassed for agent-to-agent connections
• dangerouslyDisableDeviceAuth: true only scopes to controlUi, not subagent spawning
• Impact: intra-day X scans and inou suggestions can't spawn; cron jobs unaffected

Fully Dashboard Pace Fix (15:16 ET)

• Johan: "I don't see my pace on the Fully dashboard"
• Root cause: visibility gate was usage > 75% — after weekly reset (2 PM ET), usage = 1%, hidden
• Fix: changed condition to timePct > 0 (always visible after reset, hides only in first seconds)
• Weekly usage now at 1% after reset; pace = ~125% (normalizes quickly)

Stalwart — Jacques + Roos Full Email Login (16:38 ET)

• Jacques reported "Incorrect username/password" with code 9S4BLMDF (not his password)
• Root cause: account name was short form jacques, but trying to log in as jacques@jongsma.me
• Johan: "I prefer long username (easier on iPhone)"
• Fixed: updated account names via Stalwart API to full email addresses:
  • jacques → jacques@jongsma.me (pw: 7I#rydMKlri6r%!g)
  • rozemarijn → rozemarijn@jongsma.me (pw: cRKEWJL4h3MGn3Li)
• Verified both IMAP logins work (jacques: 21 msgs, roos: 66 msgs)
• Sent Dutch-language setup instructions to both via Signal
• Port 8080 on Zurich = Vaultwarden (NOT Stalwart); Stalwart admin API = port 8880
• Stalwart admin: admin:agolM71pOwZBJhggROBDkn8R via HTTP Basic on http://127.0.0.1:8880/api/principal

Evening Briefing — 8:02 PM

Key stories:

• SCOTUS struck down Trump emergency tariffs as unconstitutional (Fri Feb 20). 2 Trump nominees voted against him. Trump retaliated with 10-15% global tariff. Markets still rallied S&P +0.69% to 6,909.
• SentinelOne (S) +4.35% — short position working against Johan
• NABL -11.18% Thursday post Q4 earnings — beat revenue but soft 8-9% 2026 guidance. Downgraded to Hold by Wall Street Zen today.
• OpenClaw 2026.2.21 dropped today: Gemini 3.1 support, 100+ security hardening fixes, Discord voice/streaming, thread-bound subagents, iOS/Watch polish
• Gemini 3.1 Pro launched Feb 19: better ARC-AGI-2 reasoning, now in preview across Gemini API/Vertex/GitHub Copilot
• Karpathy bought Mac Mini to tinker with OpenClaw, tweet going viral

Dashboard: http://100.123.216.65:9200 (briefing id: bba734b8)

Telegram: sent ✅

Nightly Maintenance (9:00 PM ET)

• OS: all packages up to date (0 upgraded)
• Claude Code: up to date (2.1.50)
• OpenClaw: updated 2026.2.21 → 2026.2.21-2 ✅
• Session cleanup: 15 orphaned .jsonl files removed, 9 cron :run: keys removed
• sessions.json: 40 entries remaining
• Working context updated, update log written

AirLLM Test (9 PM ET)

• Installed AirLLM (layer-by-layer GPU offloading library)
• Tested Qwen2.5-7B-Instruct on GTX 970 (4GB VRAM)
• Result: works — correct answer, 6.1s/token, peak VRAM only 1.57GB
• Key insight: VRAM stays ~1.5GB regardless of model size (one layer at a time)
• Implication: 70B models theoretically runnable on forge, ~8-12s/token
• Model cached at: ~/.cache/huggingface/hub/models--Qwen--Qwen2.5-7B-Instruct/
• Useful for: batch document analysis, offline medical record processing
• Fix needed: optimum==1.22.0 (newer versions removed BetterTransformer), input_ids must be moved to CUDA before generate()

OC sessions_spawn Fix

• Root cause: OC 2026.2.21 update stripped operator.write+read scopes from device token
• Fix: manually added scopes back to device-auth.json + devices/paired.json, restarted gateway
• Also fixed gateway bind: "lan" → "custom" + customBindHost: "0.0.0.0" (loopback URL for spawning)
• sessions_spawn working again as of ~5 PM ET

Stalwart Family Accounts Updated

• Jacques + Rozemarijn: account names changed to full email (jacques@jongsma.me, rozemarijn@jongsma.me)
• Sent credentials to both via Signal in Dutch