Goal-Driven Autonomy

Set a goal with --goal so the run loop checks the target when the agent tries to stop. The current code has Goal Store, Goal Gate, and a 12 re-entry cap. Full independent LLM Judge integration is still pending.

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 Gate calls an evaluator
5. If goal not met: injects synthetic message, agent continues
6. If goal met or impossible: stops normally

Goal Judge

The intended evaluator returns { ok, impossible, reason }.

• ok=true clears the goal and stops
• impossible=true clears the goal and stops
• Otherwise the loop injects a <system-reminder> and continues

In the current core runner, the evaluator is still a placeholder returning Judge not yet integrated. The --goal option exists, but the final judgment quality depends on the pending Judge wiring.

Re-entry Control

Role	Max Re-entries
Main Agent	12

Task Gate

taskGate() can force continuation when non-terminal tasks exist. The current runner still passes an empty task list, so wiring real todowrite state remains pending.