advisor-checkpoint

Advisor Checkpoint

Spend the expensive model only at a task's few load-bearing moments; let a cheaper/faster model handle the rest.

When to consult

• Before committing to a non-obvious approach. Orientation (reading files, finding the shape of the problem) doesn't need it; the moment you're about to write code, edit a plan, or declare an interpretation does.
• When stuck — an error is recurring, an approach isn't converging, results don't fit what you expected.
• Before declaring a non-trivial task done. Make the deliverable durable first (commit, save, write the file) — the consult adds latency, and a durable result survives an interruption where an unwritten one doesn't.

Skip it on short, reactive steps where the next action is dictated by what you just read — the value is highest on the first consult, before an approach crystallizes, not on every turn.

In an interactive session

Dispatch an Agent call at a checkpoint above, with model set to a stronger tier than the one currently running (opus or fable), given the plan or diff so far and a specific question — not "review everything," but "does this approach have a flaw" or "which constraint breaks the tie between X and Y."

Treat the response as advice, not a verdict:

• If you act on it and it fails empirically, or you have primary-source evidence it's wrong (the file says X, the test says Y), adapt — don't re-litigate an empirical result against advice.
• If you already have evidence pointing one way and the advice points another, don't silently switch. Surface the conflict in one more consult — "I found X, you suggest Y, which constraint wins?" — rather than picking a side alone.
• A passing self-check is not evidence the advice was wrong; it's evidence your check doesn't test what the advice was about.

This is the same posture as the security-review skill's verify-before-report principle — a second opinion is only useful if you actually weigh it, not just collect it.