devils-advocate

Sub-agent contract

/contract

When a proposal — a plan, fix, decomposition, scoping, or named recommendation — will drive user decisions, file edits, or commits, do NOT act on it as-given. Run a devils-advocate critique wave before acting, and use the recommendation as input to the decision.

Wave 2 — Parallel critics (3 fixed lenses + 1 conditional, independent):

1. Extract the proposal (the approach being critiqued) and the goal (what the proposal is trying to accomplish). Both should be plain prose. Do NOT include the original proposer's reasoning or evidence — critics must invent alternatives without anchoring on the chosen path.
2. Dispatch 3 critics in parallel. Default subagent_type: "research-agent" (mechanically locked to Read/Grep/Glob/WebFetch/WebSearch — cannot Edit/Write/commit). Each critic receives ONLY the proposal + goal + ONE lens:
   • pragmatist — cheapest-path. "What is the cheapest approach that solves the goal? Argue why the proposal may be over-engineered."
   • paranoid — safest-path. "What could go wrong with the proposal? Propose a safer alternative with narrower blast radius."
   • architect — right-level. "Is the proposal addressing the right abstraction level? Propose an alternative one level up (systemic fix) or down (targeted fix)."
3. Each critic returns {lens, alternative, tradeoff, strength} where strength ∈ {weak, medium, strong} reflects the critic's confidence that its alternative beats the original.
4. Conditional 4th critic — steelman. Fires only when the proposal is externally-authored: an inherited plan, someone else's PR, an external review, or third-party text the user pasted in. It does not fire on a proposal the agent authored itself this session — there the proposer's reasoning is already in main context and strengthening is a no-op that taxes the hottest call path. When it fires, dispatch it in the same parallel wave as the other three (never before them), on the same research-agent base, receiving ONLY the proposal + goal + the steelman lens — the closed input set of step 2 is unchanged, and no critic ever sees another critic's output.
   • steelman — strongest-version. "Restate this proposal as its strongest defensible version. Fill in assumptions its author left implicit, supply the evidence that would best support it, and drop claims too weak to defend. Do not critique it and do not propose an alternative."
5. The steelman returns {strengthened_original, gaps_filled, weak_claims_dropped} — deliberately not the {lens, alternative, tradeoff, strength} shape. It is an annotation on the original, never a competing option, so it does not enter the ranking as a 5th candidate and does not carry a strength score.

Invariant (why steelman sits inside Wave 2, not before it): a steelman is by construction the proposer's reasoning and evidence, reconstructed and amplified — the single most anchoring artifact obtainable. Routing it upstream of the other critics would hand them a hardened target and violate step 1's prohibition, converting invention into rebuttal. Keeping it a peer in the parallel wave preserves critic independence, which is what makes convergence (Wave 3.5) and dissent (Wave 3) informative at all. Never promote it to a pre-wave.

Wave 3 — Synthesis (sequential, single agent):

1. Dispatch one synthesis agent (same research-agent base). Input: original proposal + goal + all 3 critic outputs, plus the steelman annotation when Wave 2 produced one.
2. Rank all 4 options (original + 3 alternatives) along: cost (implementation + ongoing), risk (blast radius + reversibility), scope-fit (how cleanly it solves the stated goal, no more), goal-fit (how well it addresses the underlying intent, not just the surface goal). When a steelman annotation is present, score the original at its strengthened form — the point is to beat the proposal at its best, not to win against a version its author would disown. The candidate count stays 4: the annotation upgrades how original is judged, it does not add an option.
3. Recommend ONE top choice with a one-paragraph rationale.
4. Flag dissent = true when ≥2 critics returned strong alternatives disagreeing with the recommendation — signals the synthesizer is overruling well-argued dissent, so confidence is low. Include a dissent_note summarizing the strongest counter-argument.