doubt-pass

doubt-pass — break the plan before you build it

The cheapest bug is the one caught before the code exists. A fresh-context reviewer gets the plan and one job: find where it breaks, not whether it is nice. Bounded at three cycles, every finding classified, with a name for the case where doubt spins and lands nothing.

When to use

• Before a non-trivial architecture or design decision.
• The user asks for a plan to be stress-tested, holes poked, a second opinion — while it is still a plan, no diff yet.

Neighbours: /code-review needs a diff that already exists; whetstone:grill interviews you for requirements. This one sends a fresh agent at an artifact.

Protocol

Five steps, detailed in reference/doubt-protocol.md:

1. CLAIM — state the decision in one paragraph.
2. EXTRACT — strip it to an artifact plus a contract (inputs, outputs, invariants). Send those two; the reasoning trail stays behind, so the reviewer judges the design rather than your defence of it.
3. DOUBT — spawn whetstone-doubter (Agent, subagent_type: whetstone:whetstone-doubter) with the extracted artifact and contract as its whole input. The agent already carries the adversarial method — "find where this breaks" — and its word budget; reference/adversarial-prompt-template.md is the mission text to fill and send. Pass no model: this break mission runs on the agent's own opus. It returns DOUBT: FAILURES or NO FAILURE FOUND.
4. RECONCILE — tag every returned finding actionable or not-actionable, each on its own merits.
5. STOP — at the 3-cycle cap, or at the first cycle that returns nothing new. Fold actionable findings into the plan between cycles.

Subtraction pass

A second mission for the same agent asks what can go, where the doubt asks where the plan breaks. One cycle, no loop. Fill reference/subtraction-prompt-template.md and spawn whetstone-doubter with it. Run it beside the doubt cycle, in the same message, when both apply.

• Ladder — delete > simplify > optimise > automate. New code runs the finer ladder: skip it, reuse an existing pattern, stdlib, the native platform, an installed dependency, a one-liner, then minimal new code with one test when it has logic. "It may be done" is an allowed answer.
• Verdict — SUBTRACT: NOTHING TO CUT or SUBTRACT: <n> CUTS (n ≥ 1) as the first line, then one line per cut: the piece, the rung, and what still holds without it.
• Model — pass the session model's family as the Agent model. The advisorModel setting is for the advisor alone; this mission does not read it.
• Reconcile — tag each cut actionable or not-actionable, as in step 4. The operator decides a cut that removes something they asked for.

Doubt theater

Zero cumulative actionable findings across every cycle is its own result. Report it as one:

doubt theater — N findings raised, 0 actionable across M cycles. Escalate to a human or ship as-is.

Cycling a reviewer that lands nothing is motion. Naming it keeps it from reading as confidence.

Verification

Done = the loop stopped at ≤ 3 cycles, every finding carries an actionable / not-actionable tag, and the outcome is one of: plan amended with the actionable findings, or an explicit doubt theater note. The cap and the classification are the gate; "let me think about it more" is a different activity.

Dossier breadcrumb

In a repo with a live dossier ledger, record the verdict as one §S line through the dossier plugin's append tooling (the host session reminds when applicable; the first live row is the current dossier). No dossier → skip: this skill ships no hooks and no dossier dependency.