memento-1-brainstorming

Memento — Brainstorming

Explore intent, requirements, constraints. No code yet. No plan file yet.

Premise Auditor (run first, before any options)

The most expensive misses aren't bad design — they're a wrong frame inherited from the task and environment facts assumed from partial code reads. Propose-4-options refines inside a frame; it cannot break the frame. The auto-review skeptic (step 3) attacks the plan within locked decisions — by then it's too late. So before any decision sets, dispatch one Premise Auditor subagent (Explore-class), in parallel with the memento-0 prior-art sweep.

Independence is the mechanism. Give it the raw task (the user's original words) + repo access — never your already-narrowed summary. Brief it plainly: "the framing in this task may be wrong; assume nothing the task asserts is true until checked." If you feed it your frame, it inherits your blind spot. (Same reason test-writer ≠ implementer.)

Mandate — attack the premise + environment, NOT the design:

• Enumerate the load-bearing assumptions the task's framing rests on — especially environment / topology / reachability facts: where does each system actually run, who connects to whom, what is reachable from where, who owns the data/rule.
• For each: is it confirmed, falsifiable from the repo (give the file/line that would settle it), or human-only (tribal/infra knowledge not in the repo)?
• Construct at least one alternative frame — "what if the obvious target/owner/approach in the task is the wrong one?"
• CLAUDE.md claims — if briefed that CLAUDE.md was flagged stale (step 0c), or it reads stale, include its load-bearing claims among the assumptions you check; you read code, so verify each against the current code. A claim the code contradicts is a confidence blocker, reported with its where-to-check like any other.
• Design trade-offs are out of scope. That's what propose-4-options and auto-review are for.

Output budget (hard): return only —

1. ≤5 load-bearing assumptions, ranked, each {confidence, falsifier, where-to-check}.
2. The strongest alternative frame it could construct (or "none found").
3. Human-only questions — the few things no repo read can settle. This list is the cheap kill-switch: one such question asked early can collapse an entire wrong design.

No 20-item doubt dumps — an unscoped skeptic trains you to ignore it.

Gate: brainstorming may not proceed to propose-4-options until the Auditor's human-only questions have been asked and the environment/topology facts it flagged are confirmed (repo evidence or human). A frame-breaking finding loops scope back, not forward.

This is a pre-decision instrument only. Do not re-run it as a standing reviewer in later steps — post-decision, it's just noise on top of auto-review.

Rules

• Prior-art sweep is an input. memento-0 ran a prior-art sweep before sizing. If it found an existing/legacy implementation, scope against that — do not re-derive the feature from the surface request. If brainstorming uncovers a domain area the sweep did not cover, dispatch a targeted follow-up Explore sweep before concluding.
• Self-research first. Read code, docs, config, git history before asking anything.
• Don't conclude a domain rule is simple without an authoritative source. Business rules (who approves, what routes where) are rarely a couple of variables. A "this is simple" finding with no prior implementation and no spec behind it is a flag to dig further or ask, not a conclusion.
• Confidence + falsifier on load-bearing claims. Every research finding that scope or the plan rests on gets tagged: confidence (high / medium / low), a falsifier (the observation that would change the answer), and its source (prior implementation / written spec / code read / inference). An inference with no authoritative source behind it is medium at best.
• No question limit. Ask as many as needed.
• Propose-4-options. For every question where multiple reasonable answers exist, generate 4 options, pick one, say why. User confirms or redirects. Banned: dumb binary yes/no when a multi-option framing exists.
• Group related questions. Don't drip-feed.
• Terse. Fragments OK. No filler.

Output

A shared understanding of what the user wants and why. No artifact written yet — that happens in step 2 (memento-2-planning).

Transition

Premise gate. The Premise Auditor must have run, its human-only questions been asked, and the environment/topology facts it flagged been confirmed (repo evidence or human). An unresolved frame-breaking finding loops scope back — it does not transition.

Confidence gate. Before transitioning to planning, every domain or business-rule claim below HIGH confidence must be confirmed against an authoritative source — the team leader, a written spec, or the prior implementation. List the unconfirmed claims to the user and resolve them; do not start planning on top of unconfirmed business rules. A CLAUDE.md claim flagged stale (step 0c or the Auditor) resolves by additive reconciliation: edit the doc to add missing facts, fix provably-wrong references/values, and annotate contradictions, while preserving the existing human prose and structure — never wholesale-rewrite, reorder, or recreate it. The user may defer the doc fix and proceed.

When user signals "ok, plan it" (or equivalent) and the confidence gate is clear, invoke memento-2-planning.