Goals

How it works

Each goal turn is a complete run. When it finishes, a separate evaluator, a fresh model completion with no tools, reads the conversation and answers one question: is the condition met? It returns yes or no with a short reason.

• Yes: the goal completes and the loop stops.
• No: the reason is fed into the next turn as guidance, so each turn starts knowing exactly what was still missing.

Separating the judge from the worker is the whole trick. A worker grading itself says "done" early; an evaluator that can only read the record does not.

Setting a goal

1. Type the condition

   In the composer: /goal the report exists and cites three sources. Write conditions as observable outcomes, things the evaluator can check from the conversation, rather than instructions.

2. Watch the goal bar

   A slim bar docks above the composer: ◎ Goal · condition · elapsed, with pause, edit, clear, and expand controls. It stays visible across screens while the goal is active.

3. Let it work

   Turns run one after another, each a full run with the same approval gates as everything else. A goal never weakens permissions: in manual mode every turn's tool calls still ask.

The same engine is available headlessly:

anymo goal "the report exists and cites three sources" --workspace anymo-demo
anymo goal                  # bare: show condition, elapsed, turns, token spend, latest reason
anymo goal pause            # honored after the current turn
anymo goal resume           # counters reset, condition kept
anymo goal clear            # aliases: stop|off|reset|none|cancel

Options: --provider mock|http, --permission auto|manual|plan, --reasoning low|medium|high|xhigh|max, --max-turns <n>.

Pause, resume, replace, clear

• One goal per session. Setting a new one replaces the old.
• Pause stops after the current turn finishes and is evaluated; a turn is never killed halfway. Paused goals show "Paused goal" in the bar with the same controls.
• Resume starts a fresh driving session: the condition is kept, but turn, token, and elapsed counters reset.
• Clear removes the goal entirely; stop, off, reset, none, and cancel all mean clear.

Bounding long goals

Conditions can be up to 4,000 characters, and a good long-running goal states its own stop rules in the condition: "…or stop after 20 turns or 2 hours". Combined with --max-turns and the evaluator's turn-by-turn feedback, a goal cannot become an unbounded loop. Scheduled goal loops in Automations add hard stop clauses: max turns, max duration, and max cost.

Checking on a goal

Expand the goal bar, or run bare anymo goal, for the full status: the condition, elapsed time, turns evaluated, token spend, and the evaluator's latest reason. The reasons read like a progress report of what is still missing, because that is exactly what they are.

Frequently asked questions

I paused and it kept going for a minute. Why?

Pause and clear are honored at turn boundaries. The current turn finishes and gets evaluated, then the loop stops. This is deliberate: killing a turn mid-flight would leave half-applied work.

What model judges the condition?

A fresh, cheap completion against your configured provider, with no tools. It sees the conversation, not the worker's internal state.

Goal or automation?

A goal drives turns now, until done. An automation runs on a schedule, and can carry its own goal condition for "keep trying nightly until this is true".