Debate

Analyze and debate the topic from the user request.

Start by running Phase 0 to sharpen the topic (three questions in one call). Use the grill-me skill when it is installed (upstream mattpocock/skills); if unavailable, ask the Phase 0 context questions inline. A well-formed topic makes a better debate.

Process

Phase 0: Context questions

Ask these three questions in a single call:

1. Decision type: Feature proposal, Architecture decision, Technology choice, or Process change
2. Context scope: Isolated (clean-room, no codebase) or Full context (codebase-aware)
3. Knowledge source: LLM knowledge only (fast) or Research with subagents (thorough, slower)

Phase 1: Decomposition

Break the topic into 4-6 investigation angles tailored to the decision type. Each type has its own angle set (user need, implementation complexity, simpler alternatives, reversibility, operational complexity, exit strategy, etc.).

Phase 2: Research (conditional)

If "Research with subagents":

• Launch 3-5 parallel subagents, one per angle
• Full context: use scout agents -- they examine local code
• Isolated: use task agents -- they must not reference local code or conversation history

If "LLM knowledge only": skip to Phase 3.

Phase 3: Main analysis

Synthesize into structured sections:

• Problem Validation -- is the problem real and worth solving?
• Pros -- with evidence strength (strong / moderate / weak)
• Cons -- with severity (blocker / major / minor)
• Tradeoffs -- what you gain vs. what you give up
• Alternatives -- "Do nothing" is always first; "Simplest viable approach" is always second
• Overengineering Assessment -- answer these 5 questions:
  1. Would doing nothing solve the problem adequately?
  2. What is the simplest thing that could possibly work?
  3. Which part of this solution is solving a problem we don't have yet?
  4. If we had to ship this in 48 hours, what would we cut?
  5. How hard is this to undo if we're wrong?
• Reversibility -- one-way door, two-way door, or reversible with cost

Phase 4: Devil's advocate

Launch a single adversarial-challenger subagent with ONLY the finished Phase 3 analysis (never raw research). It must:

• Challenge every pro
• Deepen every con
• Check for biases: survivorship, sunk cost, herd mentality, optimism, complexity, resume-driven
• Identify unstated assumptions
• Name the single strongest argument against the proposal

If that agent is not installed, run this critique yourself as a separate pass, still using only the finished Phase 3 analysis as input.

Phase 5: Synthesis

Merge the main analysis with the devil's advocate critique:

• Incorporate valid criticisms; note deflected ones with reasoning
• Calibrate confidence: High (75-95%), Medium (40-74%), Low (10-39%)
• Produce a conditional verdict: "This makes sense IF... It does NOT make sense IF..."

Then offer interactive debate rounds, capped at 3. Each round must revise a verdict item or state why it is unchanged. After round 3: "We've explored this from three additional angles. Here's where things stand. Want to continue or call it?"

Phase 6: Save

Save the report to research/debate-<slug>.md relative to the project root. Only skip if the user explicitly declines.

Workflow turbo-path (optional)

The prose Process above is the default. When the user asks for orchestration or dynamic workflows, the research fan-out and devil's advocate become a single task batch instead of sequential launches. Same phases, same outputs -- only the orchestration moves onto the task wire.

Shape:

• Phase 0-1 stay in the main thread -- context questions and angle decomposition need the user.
• Phase 2 (research): one task item per angle, isolated: true when Isolated.
  • scout when scope is Full context; task when Isolated (and instruct it NOT to reference local code or history).
  • Per-angle agents: medium effort (breadth, not depth).
  • Barrier on all angles before synthesis.
• Phase 3 synthesis in the main thread (or one task at high effort).
• Phase 4 (devil's advocate): one adversarial-challenger spawn given ONLY the finished Phase 3 analysis (never the raw research) -- preserving the structural isolation rule below.
• Phase 5-6 (merge, verdict, save) in the main thread.

adversarial-challenger resolves to the existing agent definition -- do not duplicate it.

Rules

• YAGNI is the default stance. Burden of proof is on complexity.
• "Do nothing" is mandatory and must be taken seriously.
• Devil's advocate is structurally isolated from raw research to prevent shared reasoning biases.
• Verdict is ALWAYS conditional, never binary.
• Overengineering assessment must be substantive, not perfunctory.