/next

Advance the active Goals goal.

1. Run goals next --agent claude and read the current phase, its goal, and acceptance criteria. Work only the current phase.
2. Assess the phase before building (PACERS — see /goals-problem-solving). Break it into sub-problems, hunt the assumptions you'll depend on, and list the open questions. Record them so the user can trace the why: goals assess assume "I'm assuming X" --building "…" --toward "…" --depends --phase <PHASE> for each, and goals assess breakdown --file breakdown.json. Keep each statement plain enough for a high-school reader.
3. Make the change, then prove it by execution, not description. First invert — try to break it: how could it fail (boundaries/signs, time/locale, empty/huge, concurrency, storage or network failing)? Fix each, guard it with a check, or — if it's a premise you're relying on — record it as a load-bearing assumption. Then for each acceptance criterion AND each load-bearing assumption, write a runnable check that fails if it's wrong (a check that can't fail proves nothing). List them in the evidence verifications (each covers a criterion or an assumption id). Run goals phase evidence <PHASE> --file <evidence.json>, then goals phase verify <PHASE> — the engine runs your checks and records the real results; you cannot pass one yourself. Fix until every automated check passes.
4. Run goals phase review <PHASE>; fix anything it flags; when it passes, run goals phase accept <PHASE>.
5. Run goals check. If it says you're waiting on the user, ask one plain-language question using its wording and stop. Otherwise continue to the next phase.

End every turn with the goal pin (~4 lines), sourced from goals check, so the dashboard link and status stay on screen:

---
📌 **<goal objective>** · Phase <current>/<total> (<n> accepted)
Waiting on: <"you — <one-line ask>" or "nothing — agent is working">
Dashboard: <file:// link from goals check, click to open>   ·   Next: `goals next --agent claude`