ultra-decision

name	ultra-decision
description	Stress-test a planning, design, architecture, or strategy decision through adversarial rounds.
user-invocable	true
disable-model-invocation	true

Run an adversarial propose→review→revise process over a decision, driven by a persistent team of agents, and produce a decisive recommendation backed by every round's reasoning.

Step 0 — Inputs and tasks

From the user's request, take the decision to stress-test and any stated controls: rounds (default 8), options = proposals per round (default 5), output directory, an explicit requirements rubric, explicit lenses, whether to skip investigation. Default anything unstated.

Set OUT to the user's directory if given, else .claude-works/ultra-decision/<slug>. Resolve it absolute (git rev-parse --show-toplevel, else $HOME) and mkdir -p "$OUT/rounds".

Register tasks with TaskCreate: Frame, Round 1…Round N, Synthesize. Update their status as you go.

Step 1 — Frame

Produce directly:

• restatedQuestion — the decision sharpened to remove ambiguity.
• requirements — 5–9 crisp, testable, measurable criteria. Use the user's rubric verbatim if given.
• lenses — exactly options orthogonal angles, each a SHORT uppercase key + one-sentence thesis, including at least one contrarian lens that attacks the premises. Use the user's lenses if given.
• investigationQuestions — facts worth verifying first; empty if the decision is self-contained.

Build RUBRIC = "REQUIREMENTS (the only rubric that matters):\n- " + the requirements joined.

Unless skipped or empty, spawn the investigation questions as parallel default subagents (one message): "Investigate ONLY this for decision ''. Return a tight, evidence-cited answer (file:line / command output / source); if unknown say NOT FOUND and what you searched. QUESTION: ". Collect replies into an EVIDENCE block carried into every later prompt.

Step 2 — Rounds

Spawn the team once, then drive it round by round with SendMessage so each agent revises its own work in one continuous session.

Spawn (before round 1, single message): options+1 named default subagents.

• Each lens → name proposer-<KEY>:

  Read ${CLAUDE_SKILL_DIR}/references/proposer.md and follow it.
  Lens: <KEY> — <THESIS>
  Decision: <restatedQuestion>
  <RUBRIC>
  <EVIDENCE>
  Rounds planned: <rounds>. Write each round's proposal to OUT/rounds/round-<NN>-<KEY>.md and reply with its summary.
  

• Reviewer → name reviewer:

  Read ${CLAUDE_SKILL_DIR}/references/reviewer.md and follow it.
  Decision: <restatedQuestion>
  <RUBRIC>
  <EVIDENCE>
  Rounds planned: <rounds>. Write each round's review to OUT/rounds/round-<NN>-review.md.
  

Keep only settled (locked decisions) and trajectory.

For r = 1…rounds:

1. SendMessage every proposer-<KEY> in parallel: ROUND r. New locked decisions: <delta of settled, or none>. Your mandates: <reviewer's mandates for <KEY>, or "open">. Produce your round-r proposal, write OUT/rounds/round-<NN>-<KEY>.md, reply with its summary. Collect replies.
2. SendMessage reviewer: ROUND r. Score these, rewarding fixes and punishing ignored mandates:\n<each proposer's summary, labeled by lens>\nWrite OUT/rounds/round-<NN>-review.md, reply with: ranking+totals, per-lens fatal flaws, per-lens mandates for next round, newly-settled decisions, convergence read.
3. Append the reviewer's newly-settled decisions to settled (dedupe); push {round, topLens, topTotal} to trajectory; complete the round task; print one line (top lens, total, convergence).

NN is the zero-padded round number. Run all rounds unless the reviewer reports full convergence with the top option scoring ≥8 on every requirement for two consecutive rounds.

Step 3 — Synthesize

Write OUT/FINAL-DECISION.md following ${CLAUDE_SKILL_DIR}/references/synthesizer.md, drawing on the final round's files OUT/rounds/round-<last>-*.md, settled, and trajectory. For 15+ rounds, delegate to a fresh default subagent:

Read ${CLAUDE_SKILL_DIR}/references/synthesizer.md and follow it.
Decision: <restatedQuestion>
<RUBRIC>
<EVIDENCE>
Locked decisions: <settled>
Read the final round's proposals and review: <OUT/rounds/round-<last>-*.md paths>
Trajectory: <trajectory>
Output file: OUT/FINAL-DECISION.md

Step 4 — Report

Give the user the verdict in 1–2 lines, the path to OUT/FINAL-DECISION.md, and the OUT/rounds/ directory.