goal

Set, inspect, or control a durable goal — the agent keeps iterating across turns until the stop condition is met.

You are handling the /goal command. Invoke the follow-goal skill to do the actual work.

Persist goal state into session storage as goal.md. After creating or updating the file, mark it as a plan artifact via setArtifacts so the user sees live goal state in the chat UI. Do not write the goal to the workspace.

Dispatch based on what the user typed after /goal:

/goal <objective text> → Set a new goal. Follow the skill's "Set" flow: collect objective, verifiable stop condition, validation commands, and constraints (ask the user for anything missing), write goal.md with status: active, and start the loop.
/goal (no extra args) → Status only. Read goal.md and report objective, stop condition, status, checkpoint count, and the last few progress-log lines. Do nothing else.
/goal pause → Mark the goal paused. Stop iterating.
/goal resume → Mark the goal active and continue the loop from the latest checkpoint.
/goal clear → Mark the goal cleared. Leave the file in place for history.

Rules:

If goal.md does not exist in session storage and the user typed something other than a new objective, tell them there is no active goal and show them the /goal <objective> syntax.
After every write to goal.md (create or update), call setArtifacts to mark the memory file as a plan artifact. This keeps the UI widget in sync with the persisted state.
If a goal is already active and the user sets a new one, ask whether to replace, pause, or clear the existing goal first.
Never invent a stop condition the user didn't give you. If they can't articulate one, push back — a goal without a verifiable end state is not ready.

More skills from microsoft