Goal-Driven Autonomy

Set a target with --goal and the run loop checks it when the agent tries to stop. Since v0.4, Goal is a dedicated agent mode: when --agent is omitted, swust-code run --goal ... routes to the goal agent and injects goal-specific prompts and reminders.

Usage

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

Flow

1. User sets a goal with --goal
2. CLI selects the goal agent unless --agent is explicit
3. The agent runs normal tool calls
4. Goal Gate triggers when the agent is ready to stop
5. The runner builds a transcript and resolves the current model
6. Goal Judge calls LLM.generateObject() for { ok, impossible, reason }
7. If not satisfied, a synthetic reminder is published and the agent continues
8. If satisfied or impossible, the goal is cleared and the run stops

Goal Judge

The Judge uses a separate system prompt and is instructed to evaluate concrete evidence such as code changes, tests, file state, and tool output.

Behavior	Current implementation
Output	Verdict schema: ok, impossible, reason
Sampling	temperature: 0, maxTokens: 500
Context	User text, assistant text, reasoning, tool success/error results, shell output, compaction summaries
Failure policy	If Judge evaluation fails, it returns not-satisfied and continues

Re-entry Control

Role	Max Re-entries
Main Agent	12

After the cap is exceeded, the goal is cleared to prevent an infinite loop.

Task Gate

taskGate() can force continuation for non-terminal tasks. v0.4 also adds durable Task Registry and actor task binding, so task-oriented subagents can start, complete, or block tasks through the runtime.