Goals (`/goal`)

/goal is a slash command you type in the chat composer. Use it when the agent should keep working until a specific outcome is reached — not just answer one question.

The basic shape

In any chat session:

/goal Fix every failing test in tests/ and report the command that passes

Press Enter. The agent:

1. Saves the goal on the session.
2. Starts working on it as if you’d asked normally.
3. After each final response, a separate evaluator LLM reads the response and decides: did this satisfy the goal?
4. If not, the evaluator’s feedback becomes the next prompt and the agent keeps going.

Up to 20 iterations by default. Then it stops.

Adding a rubric (recommended)

The evaluator does a much better job when you give it concrete acceptance criteria:

/goal Build a DCF model for Costco

Rubric:
- Produces an .xlsx file
- Uses five years of historical revenue
- Includes WACC and terminal value assumptions
- Includes a sensitivity analysis
- Final response names the generated file

Write rubrics as observable outcomes. Bad: “try hard to test everything.” Good: “the final response includes the passing pytest command.”

Verdicts

After every assistant turn the evaluator returns one of:

Verdict	What happens
satisfied	Goal stops. The rubric is met (with concrete evidence).
needs_revision	Continuation queued. Iteration count bumps.
blocked	Goal stops. The agent says it needs external input it can’t get.
failed	Goal stops. The rubric is unreachable.

The evaluator is instructed to require concrete evidence — a file excerpt, an output line, a test command result — before saying satisfied. Generic claims like “all requirements met” trigger needs_revision and the agent keeps working.

If the evaluator itself fails — a provider error, or a verdict that comes back empty or unreadable — the goal falls back to needs_revision and the agent keeps going, so one blip never ends a run early. After three unreadable verdicts in a row the goal pauses itself; /goal resume restarts it.

Controlling a running goal

/goal status        # See current goal, iteration, last verdict
/goal pause         # Stop automatic continuation. Saved goal stays.
/goal resume        # Continue after pause.
/goal clear         # Remove the goal.

Only one goal can be active per session. While active, sending /goal <new text> is rejected — pause or clear first.

A session is allowed only one evaluator loop, so /goal and /mission can’t both run in the same chat — and the block works both ways: starting a /goal while a mission is active or paused is refused, and starting a /mission (or /auto-research) while a goal is set — active or paused — is refused too. Clear the goal (/goal clear) or cancel the mission (/mission cancel) before starting the other — or start it in a new chat.

Iteration budget

Default: 20 turns. When exhausted, status becomes max_iterations_reached. Use /goal resume to give it more turns, or /goal clear to stop.

When NOT to use /goal

• One-shot questions (“what’s the refund policy?”) — /goal adds overhead with no benefit.
• Exploration (“help me brainstorm names”) — no fixed outcome to evaluate against.
• Work you want to steer manually each turn — /goal keeps the agent autonomous.

/loop — for recurring work

A close cousin. /loop schedules a prompt to run on a recurring basis:

/loop 5m /babysit-prs             # every 5 minutes, run the babysit-prs skill
/loop every 1 minute get btc price
/loop check deploys every 20m
/loop check queue health           # no interval = dynamic, agent picks

• Fixed interval → expires after 3 days
• Dynamic (no interval) → agent picks delay between 60s and 1h after each run → expires after 7 days

Manage with /loop list and /loop cancel <id>.

What’s next

Other conversation tools: Copilot, Inbox.