Running Unattended Overnight Solves on Ollama Cloud Without Runaway Spend or Lost Work

How AgentAFK pairs pre-task Git rollback bundles with strict call-allocation reservations so an unattended overnight run never corrupts a working checkout or exceeds its ceiling.

Developers hesitate to leave an autonomous agent running overnight for two reasons: fear of waking up to a $400 API invoice from an infinite retry loop, and fear of waking up to a broken working tree where an unverified edit overwrote yesterday's working code. agentafk.si prevents both by combining fixed-subscription Ollama Cloud / local model bindings, pre-execution call reservations, and mandatory Git rollback bundles.

Four accounting quantities tracked separately

Following the Loop Engine model-accounting rule, AgentAFK never conflates capacity with spend. Every overnight run record tracks four distinct quantities:

• source_backed_capacity: The verified rate limit or subscription allowance of the configured endpoint (for example, Ollama Cloud or local 127.0.0.1:11434).
• selected_allocation: The maximum calls reserved for the specific task (max_calls = 24).
• total_authority: The queue-wide ceiling across all tasks scheduled for the night.
• provider_reported_usage: Observed prompt tokens, completion tokens, and wall-clock latency returned by each completed call.

Pre-mutation Git bundle snapshots

Before any task in the overnight queue is permitted to modify a file in its isolated worktree, the worker creates a self-contained Git bundle (git bundle create backup-<task_id>.bundle --all) and records its SHA-256 digest. If a task exhausts its call allocation without passing verifier.py, the worktree automatically restores to the pre-task bundle and saves the failed candidate diff as a separate diagnostic patch for morning review.