[FR] Add /goal mode for persistent work until acceptance criteria are met

Problem

When I ask Rowboat to complete a substantial task, the model or harness often stops after one turn even though the requested outcome has not been achieved.

Increasing the model-call limit helps, but it does not create persistent goal-completion behavior. I still have to notice the stop, determine what remains, and manually say “continue.”

Proposed behavior

Add a /goal command or Goal Mode that accepts:

• An objective
• Acceptance criteria
• A model-call, token, credit, or time budget
• Conditions that require user input or approval

Rowboat should checkpoint progress and automatically continue across turn boundaries until:

• The acceptance criteria are verified
• User input or permission is required
• The configured budget is exhausted
• An unrecoverable error occurs
• The user pauses or cancels the goal

The UI should show completed and remaining criteria, current activity, budget usage, and why the goal paused.

This should work with context compaction so long-running goals can resume without losing the objective or redoing completed work.

Example

/goal Audit this project, fix every confirmed high-severity issue, run the relevant tests, and stop when the tests pass or you need my decision.

Related

• #768: Configurable model-call limits
• #1094: Context budgeting and compaction
• #1096: Live visibility and controls for delegated workers