goal

Goal

Use this skill only when the user explicitly invokes /goal with an objective that may require multiple turns.

A goal is complete only when current evidence proves every requirement of the objective. Stopping a turn, running out of time, or having a plausible answer does not finish it.

There is no deadline that finishes the work. The goal stays active across turns until the end state is true, or until a budget pause or a persistent failure stops the ticks. A budget pause is not completion. The objective stays intact and the user can resume it.

Start

1. Preserve the user's complete objective without narrowing its scope.
2. Call create_goal exactly once with that objective.
3. Do the first concrete increment of work in the same turn.
4. Keep the returned goal id. Every update_goal call must include it.

Continue

• Goal continuations are scheduled automatically. A webhook can also wake the goal.
• Treat the stored objective as authoritative. Do not shrink it to work already done.
• Make measurable progress. Verify results before you even consider completion.
• Temporary session-busy deferrals are normal and are not goal failures.
• Persistent execution failures stop the goal after the retry limit.
• Using up the tick or runtime budget pauses the goal. It does not complete it.

Waiting

If you are blocked on a person or an external event (review, approval, CI, a publish), call update_goal with status="waiting", the goal id, and a reason. Pass recheck_after (for example 2h) only when you should look again even if no event arrives. Do not keep waking on the timer while blocked. You cannot pause a goal. That is the user's control.

Complete

Before update_goal with status="complete", completion is unproven. In a separate step, after the evidence tool calls:

1. Pull concrete requirements out of the objective and anything it points at (files, plans, specs, issues, instructions). Do not shrink the original scope.
2. For each requirement, name the evidence that would prove it, then look at the current state: files, command output, tests, pull-request state, rendered UI, runtime behavior.
3. Judge each item as proved, contradicted, incomplete, too_weak, or missing. A green test counts only if it covers that requirement. Uncertain or indirect evidence is not done.
4. The audit has to prove completion. Failing to notice leftover work is not enough.

Intent, partial progress, memory of earlier turns, and a reasonable-sounding answer are not proof.

Call update_goal with status="complete" and an audit array only after every item is proved and no required work remains. The call must be its own step so the transcript already contains the evidence. Ending the turn does not call update_goal. The goal stays active and the full objective stays intact.

status="active" is only for resuming a goal the user paused.