decide-about-disagreements

Decide About Disagreements

Use this skill after models-consensus returns at least one topic without a shared position. The result is a decision aid, not a second council. The skill does not edit the council artifacts or implement the user's choice.

Outcome spine

1. Result: turn each unresolved topic into a short decision packet. Each packet contains the neutral topic, all recorded seat opinions, an analysis, every distinct option, one recommendation, and the user's choice.
2. Next consumer: the user chooses one option for one topic at a time.
3. Done: every material unresolved topic is either decided, explicitly deferred, or marked awaiting_human; no topic is silently skipped.
4. Boundary: do not rerun models-consensus, change its report or state, or start implementation.

Say recommendation, not survey while working. The user asked for the full option set, so show it, then close each packet with one clear pick and the trade that decides it.

Read the result

Accept either the inline result supplied by the caller or the persisted result paths in that result object. If neither is present, ask for the report or state path with the host's blocking question tool. Do not guess a latest file from a directory listing.

1. Use the report as the primary narrative source. Use the state object for mode, status, rounds, seats, effective providers and models, convergence, and confidence.
2. Read raw round files only when the report and state do not contain a required opinion or resolution. Read the smallest file set that fills the gap. Do not load the full transcript by default.
3. Record the question, session, mode, status, and artifact paths before analysis. If status is not complete, label the work as partial, failed, awaiting human input, cancelled, or ceiling hit as recorded. A partial result is not a final consensus. If mode is missing or contains interactive or autonomous, treat that value as the interaction mode. Derive the council mode from the available fields and report headings: contradictions and material_gaps indicate poll, disagreement_points and convergence indicate debate, and the fixed council verdict headings indicate personas. If the mode remains unknown, use only items explicitly marked unresolved.
4. Apply seat fidelity. Name a model or provider only when the result records its effective receipt. Never infer an effective model from a requested name, a prompt, or a seat's self claim. If the receipt is absent, label the seat as unverified and preserve the reported seat name.
5. Apply flag, never overwrite. Keep minority positions, unresolved caveats, missing evidence, malformed seats, duplicate providers, and shared transport notes visible. A larger count does not erase a dissent.
6. Match the user's language. Use ASD-STE100 Simplified Technical English for English and Brazilian Portuguese with simple wording for Portuguese. Keep paths, enum values, model names, and short quoted evidence unchanged.

Find topics that are not consensus

Extract only material topics that remain open. Do not ask about a point that the council resolved, a point all seats support, or a decision the user has already made. That is already decided territory.

Use the mode-specific fields below. Field names can appear in the report, state, or inline result.

For older or custom result shapes, accept disagreements, open_points, remaining_objections, and open_caveats only when the surrounding text marks the item as unresolved. Treat follow_up_questions and evidence_gaps as evidence gaps, not as model disagreement. Include a coverage or evidence gap as a question when it requires a user assumption or can change the choice; otherwise list it as a non-blocking caveat.

If no material topic remains, lead with the outcome: state that no user choice is needed and list any non-blocking caveat. Do not ask an empty question.

Build one decision packet per topic

Process topics in the order used by the result. Do not merge two topics unless the result says they are the same decision. For each topic, build this packet before asking the user.

1. State the neutral topic

Use the council's point id and a one sentence statement of the contested issue. Remove loaded wording. Example: C1: choose a shared queue or an outbox table for durable delivery.

2. Present every recorded opinion

List every seat that addressed the topic. For each seat, show:

1. the recorded seat name;
2. the effective provider and model when receipt verified;
3. the position in plain language;
4. the evidence, assumption, or risk that supports it.

If the full seat table contains a participating seat with no position for this topic, write No position recorded. If a raw output is missing or malformed, write Unavailable in the supplied result and name the evidence gap. Do not turn silence into agreement. Keep opinions separate even when two seats use the same effective model. Mark the duplicate for diversity context, but do not remove its text.

3. Analyze the disagreement

Use the supplied question and context to test each position against the decision criteria that can change the choice. Prefer criteria that are observable in the result, such as behavior, scope, architecture, cost, operational risk, reversibility, evidence quality, and implementation burden.

Separate three cases:

1. Different conclusions: the seats choose different directions.
2. Different assumptions: the seats would converge if one fact were known.
3. Different coverage: one seat raised an additional risk or benefit.

Name the strongest evidence for each side and the material weakness. Do not score a seat by majority count alone. If the result lacks the fact needed to choose, say so as an evidence gap and include an evidence-first option.

4. List every option

Create a numbered option list. Include every distinct direction suggested by a seat, the council, or the original question. Preserve a meaningful minority option even when it is not the leading view. Remove only duplicate wording.

Then test whether the best approach is a combination of suggestions. If the combination is better, add it as a new option named Recommended synthesis and state which parts it takes from which positions. The recommendation must point to a listed option. Never recommend an option that is absent from the list.

5. Recommend one approach

Use this compact shape:

I recommend option <id>, because <decisive evidence and trade>. The strongest alternative is option <id>, but it costs <material cost or risk>.

This is recommendation, not survey. The recommendation is an analysis call, not a claim that the majority is correct. If evidence is not sufficient, the recommended option can be defer and collect <specific evidence>. State the smallest evidence that would change the recommendation.

6. Ask the user

Use the host's blocking interactive question capability first. In Codex, use request_user_input when it is exposed. In Claude Code, use AskUserQuestion. Otherwise use the host's equivalent. Use plain text only when no interactive tool exists or the tool call fails.

Ask one topic per question. Put the recommended option first. The question must be a single sentence, such as Which approach should we adopt for C1?

The full option list must appear before the question. If the tool supports only two or three choices, offer the recommendation and the strongest alternative as choices, then rely on the tool's free text Other choice for every other listed option or combination. Never hide an option from the narrative because the tool has a short choice limit. If the tool supports all choices, include all of them.

When the user chooses Other, capture the exact text and map it to a listed option or a new combination. If the text is not actionable, ask one focused clarification question. If the tool returns no answer, do not select the recommendation by default. Stop with awaiting_human and identify the topic.

Continue and finish

After each answer:

1. Record the user's selection in the current response context. Do not write to the council report or state unless the user separately asks for a decision record.
2. If the user selected a different option, state the change without arguing again. If the user selected a combination, restate its exact boundaries.
3. Move to the next unresolved topic. Do not ask about resolved topics again.

End with lead with the outcome:

1. topics decided and the user's chosen option for each;
2. topics deferred or awaiting_human;
3. the original recommendation where it differed from the user's choice;
4. evidence gaps, seat fidelity limits, and unresolved dissent;
5. one next step only when the result or the user's choice supplies one.

The acceptance contract is observable: every material open topic has a packet, every packet shows all recorded opinions and all distinct options, each packet has one recommendation, each asked topic has a captured user answer or an explicit awaiting_human status, and the source council artifacts are unchanged.

Gotchas

1. A majority is not proof. Use the evidence and the decision criteria.
2. Missing output is an observable failure, not consensus. Show the gap and choose evidence first when it blocks a safe call.
3. A blind spot is not automatically a disagreement. Ask about it only when it changes the decision or requires a user assumption.
4. Do not infer a model identity. Use seat fidelity for every attribution.
5. Do not turn a recommendation into an implementation. This skill ends when the user has chosen, deferred, or left a topic awaiting an answer.