Guardrails & budgets

Autonomy without brakes is a liability. OpenLoop’s guardrails are defaults, not opt-ins, and they all report into the event log so you can audit every trip-wire.

Budget ceiling

openloop config set budgets.dailyCostUsd 25      # default
openloop config set budgets.estimatedCostPerRunUsd 0.10

The daemon accumulates spend across all projects (measured cost when the provider reports usage; the estimate otherwise, including runs killed by timeout). When daily spend crosses the ceiling:

1. budget-blocked hooks and notification channels fire.
2. The daemon pauses itself; no new agent runs until midnight or a higher ceiling.

Check it any time:

openloop status          # Budget: $3.42 / $25.00 today

Attempt ceiling

runtime.maxAttemptsPerTask (default 3): after 3 failed attempts a task moves to blocked instead of burning more budget on the same failure. The last failure signature is kept on the task for diagnosis.

Run timeout

runtime.runTimeoutSeconds (default 1800 = 30 minutes): agent runs are killed at the wall clock, charged at the estimated cost, and the task attempt is counted.

No-progress detection

If consecutive attempts fail with the same failure signature runtime.noProgressRepeatLimit times (default 2), the task is blocked early; repeating the same failing run adds no information.

Scope enforcement

Covered in policy & risk: deny globs and high-risk areas are enforced before runs and against actual diffs. Nothing outside the policy merges autonomously.

Promotion backpressure

A project with runtime.maxPendingReviewsPerProject (default 3) or more pending promotions is skipped by the scheduler. If humans are the bottleneck, the loop slows down instead of piling up risk.

Daemon self-protection

• Five consecutive tick errors → the daemon self-pauses (see service status).
• Stale PID files are detected and cleared; a PID pointing at a foreign process is refused, not killed.
• Stuck in_progress tasks recover to ready on daemon start; stale worktrees are reclaimed.

Pausing on demand

openloop pause        # stop scheduling new runs, daemon stays up
openloop resume
openloop service stop # full stop

pause/resume are also exposed over MCP, so an editor agent can hold the loop while you work.