Goal-Driven Autonomy

Set a goal with --goal and the agent works autonomously until it's done.

Usage

swust-code run --goal "fix all TypeScript errors" "start working"

How It Works

1. User sets a goal
2. Agent begins working
3. When agent tries to stop, Goal Gate triggers
4. Goal Judge (independent LLM) evaluates the conversation
5. If goal not met: injects synthetic message, agent continues
6. If goal met or impossible: stops normally

Goal Judge

Independent LLM evaluator that reads the full conversation transcript and returns { ok, impossible, reason }.

Re-entry Control

Role	Max Re-entries
Main Agent	12
Subagent	3