tao-loop

What problem does it solve?

The tao-loop skill solves the need to safely run multiple autonomous coding iterations while stopping automatically when a judge metric indicates completion or when cost/iteration limits are reached.

Core Features & Use Cases

• Judge-gated termination: Stops when the judge’s score reaches the completion criteria (single-metric termination) rather than relying on fixed iteration counts alone.
• Budget-bounded looping: Enforces hard constraints for MAX_ITERS, MAX_COST_USD, and an external HARD_STOP, capturing breaches into the final LoopResult instead of crashing the caller.
• Iteration-level streaming and traceability: Produces a fully-populated LoopResult including judge_history and final state, with optional per-iteration events via on_event.

Use Case: When implementing a non-trivial feature (e.g., “implement auth improvements and tests”) where you want iterative progress with automatic verification—stop early if the judge says it’s done, otherwise continue until the budget or hard stop is hit.

Quick Start

Run tao-loop to implement the goal with judge checks every N iterations and strict max iterations/cost limits by executing a loop runner command like: python scripts/run_tao_loop.py --goal "implement X" --workspace /path --max-iters 10 --max-cost 1.50 --judge-every 2.