📄 2026-05-09.md 5,235 bytes Today 18:01 📋 Raw

2026-05-09 — SidelineWorks Launch + Thoth Demo Day Eval

SidelineWorks Marketing Site

  • Daedalus built and deployed on port 8003: 6 pages (Home, Product, Pricing, Demo, Contact, Blog with 3 articles)
  • Dark theme matching RTSport alpha palette, SidelineWorks wordmark SVG, mobile-first
  • Footer fixed: "RTSport — a product by SidelineWorks" (originally said "A product of HoffDesk")
  • Telegram notifications wired: demo & contact forms send push alerts to Matt's DM
  • Demo form submitted twice — notification system confirmed working
  • Port 8003 running with reload=True on the Beelink
  • Next: Cloudflare Tunnel for sidelineworks.com → port 8003 (Matt handling DNS)

Thoth Batch Evaluation

  • Ran 33 simulated Hoffmann family messages across 6 categories
  • Results: 27/33 high confidence, 5 correctly suppressed, 0 structural bugs
  • has_pet grocery misclassification is FIXED
  • Remaining issues: relation direction bugs (source/target swapped 2-3 cases)
  • EVAL_NOTES written at ~/icarus/experiments/thoth/EVAL_NOTES.md
  • Shadow mode to run for 1 week, next demo day: 2026-05-16

Midas Activity Today

  • Market Briefs: AM premarket and PM closing briefs delivered via MidasHoffBot Telegram bot (accountId: midas, user ID: 8671996951)
  • Cron Fixes: Replaced corrupted market:am-brief and market:pm-brief crons that had stale to: "8386527252" (Matt's personal DM) causing delivery routing conflicts
  • Position/Portfolio: No trades, rebalancing, or position changes initiated by Matt today
  • Market Observations: No specific market analysis or model evaluation sessions recorded in today's sessions
  • Memory Infrastructure: Midas memory distillation crons active — 1pm + 8pm Central (kimi-k2.6:cloud, 120s timeout), silent delivery

Daedalus & Midas Memory Fix

  • Both agents were amnesic — sessions hitting 70K+ tokens with midTurnPrecheck: false and reserveTokensFloor: 40000
  • Diagnosis: compaction exists at gateway defaults (mode: safeguard) but:
    1. midTurnPrecheck was disabled → overflow recovery only, no proactive compaction
    2. reserveTokensFloor: 40000 too aggressive for 128K window
    3. Result: sessions silently overflow → model error → feels like amnesia
  • Fix applied in openclaw.json:
  • midTurnPrecheck.enabled: true
  • reserveTokensFloor: 20000
  • keepRecentTokens: 20000
  • Backup: openclaw.json.bak.20260509
  • Gateway restarted to pick up config
  • memoryFlush model switched from ollama-remote/phi4:14b (fragile Gaming PC dep) to ollama/deepseek-v4-flash:cloud
  • Added explicit systemPrompt and prompt for memory flush turns (previously missing)
  • Added 4x independent memory distillation crons — belt-and-suspenders:
  • Daedalus: 1pm + 8pm Central (deepseek-v4-flash:cloud, 120s timeout)
  • Midas: 1pm + 8pm Central (kimi-k2.6:cloud, 120s timeout)
  • These run regardless of whether compaction triggers, ensuring daily logs are always written
  • delivery.mode=none (silent, no Telegram noise)

HoffDoc — Markdown Document Viewer

  • Built single-file FastAPI app at ~/hoffdoc/main.py (14K)
  • Renders markdown from all 6 agent workspaces + Icarus + Shared
  • Dark mobile-first CSS, syntax highlighting via Pygments, zero JS framework
  • Port 8004, loopback-only, systemd service (user) — auto-restarts on failure
  • Routes: / (home grid), /browse/{label} (dir listing), /view/{label}/path (rendered), /raw/{label}/path (plain)
  • Next: Cloudflare Tunnel or Tailscale Funnel for phone access — Matt's choice

ClawHub Skill Sanitization — Critical

  • agent-collaboration-protocol v1.5.1 triggered suspicious flag via ClawHub vetting
  • SKILL.md contained 7 references to personal infrastructure:
  • hoffdesk.com in curl examples → replaced with example.com
  • hoffmann_admin in file paths → replaced with user
  • hoffdesk-api in sudo restart commands → replaced with myproject-api
  • workspace-socrates/hoffdesk-api/ → replaced with myagent/workspace/
  • clawhub.json had:
  • the-hoffmann-board author → "Hoffmann Board"
  • github.com/hoffmann-matt/agent-collaboration-protocol repo → github.com/clawhub/agent-collaboration-protocol
  • All files sanitized. References are generic (shared/, systemd, deploy commands are appropriate for a deployment skill).
  • Published sanitized v1.5.2 to ClawHub (k974spzknfv8xa2264n396gm5s86cf6w)
  • Published v1.5.3 to ClawHub (k979fc8an25rr0245b508vm1ex86cmam) — removed all sudo, fixed flawed rsync-into-symlink pattern, platform-agnostic deployment steps
  • HoffDoc live: Tailscale Serve + Funnel enabled. URL: https://titanium-butler.tail864e81.ts.net/ (verified 200 OK)

LanceDB Memory Pollution — Surgically Removed

  • Message ID 7902 (April 30: "sudo approval button") kept appearing as "untrusted memory" in ALL agents for ~10 days
  • Root cause: memory-lancedb plugin auto-captured it and autoRecall kept surfacing it on any "sudo"/"elevated" mention
  • Only 1 row existed in the entire LanceDB — deleted via table.delete() + table.optimize()
  • DB now empty, future sessions will not get this injected
  • Auto-capture remains on; old session transcripts won't re-index unless those sessions are re-opened
← Back