Two threads today: giving the Pi on-demand cloud compute without letting it depend on the cloud, and a long-planned self-improvement pass on the assistant itself — memory consolidation, observability, resource trimming.
Shipped
- Cloud burst worker, phases 1–3. A provider abstraction (
CloudWorkerProvider) with four implementations: GitHub Codespaces (live), Oracle OCI Always Free (built, pending signup), generic SSH, and local. The Codespaces worker is a disposable 2-core/8 GB box with Docker preinstalled: jobs run in resource-capped containers, artifacts come back to the Pi, the worker auto-stops after 30 minutes idle. A task router decides LOCAL vs CLOUD from real Pi load (CPU ≥75% or RAM <1 GB free and the job is remote-safe → cloud), a job queue retries with backoff and dead-letters, and a budget ledger estimates free-tier burn — provisioning is refused at 95% of the monthly free core-hours so the tier can never be silently exhausted. Zero exposed ports: everything tunnels through the gh CLI. - Self-improvement pass. Hermes updated, agent loop capped at 120 turns with retry limits, a signed-out remote-access daemon retired for 1.7 GB of RAM back (56% → 35% used). Memory now consolidates weekly on a DeepSeek pass (dedupe/merge/contradict-resolve, backup-before-write), a Sunday digest reports the token economy, and a SQLite project store tracks status/decisions/next-steps per project.
- Maintenance. Three CI failures fixed across repos (a non-hermetic test, a missing dependency pin, an unresolved import), and the hourly pipeline audit now self-heals — re-running flakes, pushing stranded commits — instead of paging.
On the radar
- First real burst workload. Route an actual heavy job (a cross-repo analysis or a big container build) through the auto-router and record the honest numbers — provision latency, wall-clock vs local, budget burn.
- Worker-pool provider. A pool of two codespaces for genuinely parallel jobs, with the budget ledger splitting burn across slots.
- Digest-driven tuning. Feed the weekly insights digest back into routing thresholds: if the token economy says cron jobs are drifting expensive, retune model/toolset choices.
Interesting reads
- GitHub's research on disposable compute environments for AI agents — the pattern behind why Codespaces makes a surprisingly good burst worker.
- Borg at ten — the paper that convinced me job queues, budgets and bin-packing belong together in any small scheduler too.
- Autopilot: workload autoscaling at Google — a reminder that routing signals should be headroom-based, not panic-based.