Skip to content

Budget & Guardrails

Autonomous AI coding sessions can consume significant resources quickly. Cub provides robust budget management and safety guardrails to help you maintain control over costs, iterations, and sensitive data.

Why Budget Management Matters

When running AI coding assistants autonomously, costs can spiral unexpectedly:

  • Token consumption: Each task iteration can use 50K-500K+ tokens
  • Iteration loops: A confused AI might retry the same failing task repeatedly
  • Session length: Long-running sessions accumulate costs over time

Cub's budget system provides multiple layers of protection:

+------------------+
|   Hard Limits    |  <-- Absolute stops (tokens, cost, iterations)
+------------------+
|  Warning Zones   |  <-- Alerts before hitting limits
+------------------+
|  Task Limits     |  <-- Per-task protections
+------------------+

Budget Types

Cub tracks multiple budget dimensions:

Budget Type What It Tracks Default Limit
Tokens Total tokens consumed None (unlimited)
Cost USD spent on API calls None (unlimited)
Tasks Number of tasks completed None (unlimited)
Iterations Loop cycles completed 100 per run
Task Iterations Retries per task 3 per task

Quick Configuration

Set budgets in your .cub.json or global config:

{
  "budget": {
    "max_tokens_per_task": 500000,
    "max_tasks_per_session": 10,
    "max_total_cost": 5.0
  },
  "guardrails": {
    "max_task_iterations": 3,
    "max_run_iterations": 50,
    "iteration_warning_threshold": 0.8
  }
}

Or via CLI flags:

# Set a $5 budget for this run
cub run --budget 5.0

# Set a token limit
cub run --budget-tokens 1000000

Budget Status

Monitor budget usage during a run:

# View current run status
cub status

# Watch live dashboard
cub run --monitor

The status shows:

Budget Status
+------------------+------------+
| Metric           | Value      |
+------------------+------------+
| Tokens Used      | 245,892    |
| Tokens Limit     | 500,000    |
| Cost             | $1.23      |
| Cost Limit       | $5.00      |
| Tasks Completed  | 3          |
+------------------+------------+

What Happens When Limits Are Reached

When a budget limit is hit:

  1. Current task completes - The AI finishes its current work
  2. Run stops gracefully - No abrupt termination
  3. Status logged - Budget exhaustion recorded in logs
  4. Summary displayed - Final statistics shown
[yellow]Budget exhausted. Stopping.[/yellow]

Run Summary
+------------------+------------+
| Duration         | 847.3s     |
| Iterations       | 12         |
| Tasks Completed  | 8          |
| Tokens Used      | 1,000,000  |
| Final Phase      | completed  |
+------------------+------------+

Guardrails Overview

Beyond budgets, guardrails provide additional safety:

Guardrail Purpose Default
Max Task Iterations Prevents infinite retry loops 3
Max Run Iterations Caps total loop cycles 50
Warning Threshold Alerts at percentage of limit 80%
Secret Patterns Redacts sensitive data in logs Enabled

Next Steps

  • Token Management


    Deep dive into token tracking across different harnesses.

    Tokens

  • Iteration Limits


    Configure task and run iteration limits.

    Limits

  • Secret Redaction


    Protect sensitive data in logs and output.

    Secrets