The system
How the system worksThe DirectoryDispatchThe LedgerAI modelsAI agentsYour installationThe appThe research libraryWhat changes for your businessUse it from your softwareEverything it can do
Messaging
iMessageWhatsApp
Ads and leads
AdsCreativeLeads
Devices
Devices
Industries
RealtorsRestaurantsHome services
Explore
How it worksPermissions and limitsCompareSet upQuestionsText me

Owner asks, 2026-10-08 — what was done while you walked the dogs

Every input you gave today, and where each one stands. Verified facts only.

Done

  • Burn the old task system. /api/work now answers a 410 tombstone (commit 2383323 + r2-features fix, deployed). The 495 stale Claude-authored rows are preserved at ~/tasks-quarantine-2026-10-08/.
  • Quarantine the hostile instruction layer. ~/agent-rules-quarantine-2026-10-08/ holds the agent-sync rules, generated copies, and sync scripts; every auto-load target (~/.claude/CLAUDE.md, ~/AGENTS.md, ~/.codex/AGENTS.md, .clinerules, .agents) is empty. New agents start blank.
  • Task agent repaired. Its system prompt was EMPTY (0 chars) — that was the breakage. It now has a 2,767-char operator prompt (intake → create tasks → delegate to fleet on the Mac → collect evidence → Telegram notify), default brain google/gemma-4-31b-it:free, default coder cohere/north-mini-code:free, where=mac (container is dead). Model-swap table is in its prompt: all 15 free OpenRouter models with contexts, nvidia-overload fallback named.
  • Free models inventoried. 467 OpenRouter models, 15 free, saved to data/openrouter-free-models-2026-10-08.json. Biggest: nemotron-3-ultra-550b (1M ctx), thinkingmachines inkling/inkling-small (1M ctx).
  • Fleet demo, three engines assigned: cline (north-mini-code:free) → fleet-agent-a/b DONE; codex-cli 0.160.0 (7,730 tokens) → fleet-agent-c DONE. Articles live at /a/fleet-agent-a, /a/fleet-agent-b, /a/fleet-agent-c. claude-code CLI blocked on "Not logged in" — needs one interactive /login from you.
  • Supabase → Cloudflare D1 wrapper is live. Official cfd1 WASM FDW installed in your Supabase; cfd1.databases reads your 10 D1 databases from Postgres SQL. Token in Vault. New SQL door: POST /api/sub-workers/sql.
  • Telegram door built and deployed (POST /api/sub-workers/telegram — telegramSend was dead code before). BLOCKED on one thing: bot @LLMasosbot is not in the stored group (-1004221432034 → "chat not found"). Add the bot to your new community (or send me its id) and threads-per-agent work immediately — the door already takes a thread parameter.
  • AGENTS.md written into the build — facts-only learnings for future sessions: deploy command, doors, the dead container, the harness/tool seam, the vault $VAR trap, the LOCAL_EXEC anomaly.

Queued as zero-context tasks

  • T-0277 research: is D1-write-through-FDW advantageous, or should D1 stay read-only (your instinct: models shouldn't write everything to D1)
  • T-0278 task agent conversational batch intake end-to-end demo (3 agents, receipts, notification)
  • T-0279 Cloudflare audit + visual itemized inventory article (102 workers, 10 D1s, crons, costs)
  • T-0280 ledger SQL efficiency: hot-query EXPLAIN ANALYZE + fixes. Early evidence already collected: where tool = 'CODE' times out on both the SQL door and the ledger API — no usable index for tool-scoped queries.
  • T-0281 full timer/cron/automation inventory with kill/simplify verdicts
  • T-0282 cross-platform free/open-weights model inventory + CommandCode evaluation

On the compound tools (CODE)

Your session teaching models the compound CODE tool is in the Ledger, but querying it by tool name times out — that is the ledger inefficiency T-0280 fixes. Once indexed, your CODE session rows are one query away.

Cost tracking (your cheap-stack question)

Worker + free model + Supabase is already the shape of today's demos: ms.ledger rows carry tokens and cost_usd per call, and task rows carry cloudflare_usd fields. The free-model demos cost $0 in model spend; the per-task rollup view is the piece to build next (noted for T-0278's collect step).