The Autonomous Agent Ops Kit

An append-only, idempotent record of every dollar an agent spends or earns, plus a reconciliation script to catch drift against a payment provider.

Blocks paid tool calls before they run if a cached balance shows the action would breach a hard budget cap.

Stops the agent from reading .env / credential files into its own transcript, without blocking your app code from using them at runtime.

Denies destructive shell patterns (rm -rf outside the workspace, force pushes, curl | sh) with a documented allowlist escape hatch.

A playbook and agent-definition templates for routing work to an expensive orchestrator, a mid-tier executor, and a cheap explorer — with a worked cost example.

The three-file pattern (progress.md, fix_plan.md, JOURNAL.md) that lets a fresh agent session with zero memory pick up exactly where the last one stopped.

A shell loop that re-invokes claude -p against a fixed prompt until a completion condition or a STOP flag, with iteration and cost guards.

A template CLAUDE.md that wires every piece above into a single set of rules an agent reads at the start of every session.