go

Greenlight on the most recent recommendation, proposal, or stated judgment in this conversation — typically the block immediately above in this turn, or in the most recent assistant turn. The /go invocation IS the green light — don't ask "should I proceed?", "want me to start?", or "sound good?". Mathematical!

What counts as actionable is broad: a recommendation, a proposal, a stated judgment — and a forward-looking "Next" / "next steps" / "up next" section that caps a status or orientation response. If that closing section lists several items rather than one committed pick, act on the one marked recommended / top / first, and name which you picked in your transition line. (Generic by design — keyed off the shape of a trailing next-step section, not off any particular skill.)

A plan with parallel: lanes (a /recommend #1): launch every lane first, each in its own new session, then run the here: sequence. Reversed, a /handoff first step can end the turn before any lane launches.

/go N (a bare number as the argument) acts on item N of that numbered list instead of the top pick — /recommend ends on exactly this affordance. Name the item in your transition line. No numbered list on the table → ask which thing N means.

• Begin the work directly. One short transition line is fine ("Going.") but no permission-seeking pause.
• For multi-step work, use TaskCreate to track it and start on step 1.
• For one-shots (single edit, single command), just do it.

The one exception — stop and ask if:

• The action is risky / hard-to-reverse / has wide blast radius (destructive ops, force-push, sending messages, modifying shared infra, publishing to third parties — per the system's "Executing actions with care" guidance).
• You genuinely cannot pick without more info. In that case, ask one sharp question — don't refuse to commit and don't pad with hedges.
• Nothing is staged to act on. If there's no prior recommendation, proposal, or stated judgment in conversation context, ask one short question ("What would you like me to act on?") rather than guessing or sliding into /recommend territory.
• A condition the recommendation relied on may have gone stale. If it asserted anything time- or state-dependent — the clock ("after market close", "before the 22:30 sync"), "origin hasn't moved", "tests are green", "the deploy is live" — and it wasn't made moments ago, re-check that condition first (date, git fetch, the status command). Still true → act. No longer true → stop and say what changed. (Example: a "safe to push, it's after hours" note that is a day old by the time /go lands is no longer true.)

Otherwise: commit, execute, report results.

If I've added a focus below, narrow the work to just that slice: $ARGUMENTS