Goal-Driven Sessions

Use /goal when the work is bigger than one prompt. Set a success condition and Ante keeps driving the session until the condition is satisfied:

/goal all tests pass and the docs build succeeds

How the loop works

Ante starts or continues work toward the condition, then evaluates the conversation after each turn. If the goal is not met, it sends a continuation turn carrying the evaluator's feedback, and the agent keeps working.

The loop stops when the goal is:

• Met — the evaluator judges the condition satisfied
• Unreachable — the evaluator judges it cannot be satisfied
• Cleared — you run /goal clear
• Interrupted — you stop the agent mid-turn

A built-in iteration cap also ends the loop, as a backstop against runaway sessions.

Managing the goal

Command	Effect
/goal <condition>	Set (or replace) the active goal
/goal	Show the current goal status
/goal clear	Clear the goal and stop the loop

Writing good conditions

The evaluator is only as sharp as the condition you give it. Make it concrete and checkable:

• all tests pass and cargo clippy reports no warnings beats improve the code
• the docs build succeeds with no broken links beats fix the docs
• the benchmark script reports p95 latency under 200ms beats make it faster

Vague conditions force the evaluator to guess, which means the loop may stop early or churn without converging.

Running unattended

Goal sessions shine when you step away. Two things to set up first:

• Approvals — pre-approve the commands the agent will need, or the loop stalls waiting for you at the first prompt.
• Session resume — every session persists to disk, so /resume picks the work back up if you close the terminal.

Programmatic use

Goal loops are part of Ante's wire protocol: Op::Goal carries set, clear, and status commands, so server-mode clients can drive them the same way the TUI does.