council

name	council
description	Convene a bounded multi-model, multi-perspective deliberation to produce a recommendation with confidence, preserved dissent, and a transcript. Use on "council", "panel", "deliberate", "multi-agent review", "advisory council", "get diverse perspectives", or "discuss this with specialists".

Council

Convene a bounded multi-model, multi-perspective deliberation when a decision is important enough that one model's reasoning, or one critique loop, is not enough. This is an advisory council, not a vote and not an ambient committee. The driver or operator still owns the final decision.

The council pattern is for yes-and plus contest: first seed independent perspectives, then let them refine, challenge, and synthesize toward a recommendation while preserving dissent.

What this is

• A deliberative advisory process. The goal is a better recommendation, not a transcript full of opinions.
• The driver holds the pen. Council members advise. They do not decide, implement, or veto.
• Panel is a primitive, not the skill. A panel is isolated fan-out of independent views. A council may start with a panel, then escalate into deliberation, debate, sparring, red-team, or Delphi-style rounds.
• Dissent is first-class. Do not force consensus. A useful council can end with a recommendation plus a minority report.
• It is gated, not ambient. Open a council only when extra perspectives are likely to beat a single frontier model with enough reasoning time.

When to open a council

Open a council when the decision is both consequential and uncertain, especially:

• High-stakes trade-offs - architecture, product strategy, security posture, migrations, public commitments, compatibility boundaries, or workflow design.
• Multi-faceted questions - the answer depends on several valid lenses, such as user value, operational risk, maintainability, performance, and safety.
• Real forks - multiple plausible options exist and the choice changes future work.
• Anchoring risk - the main model may be locked into a frame, or a prior plan needs independent pressure.
• User asks for specialists - the operator wants a set of named models, personas, PAW SoT specialists, or perspectives.
• Contradictory evidence - existing findings conflict and need synthesis rather than another isolated critique.

When NOT to open one

Do not use a council for routine coding, factual lookups, mechanical refactors, simple bugs, ordinary test failures, or anything a single frontier model can answer and verify directly. Multi-agent ceremony is a cost; spend it only when diversity, adversarial testing, or synthesis is likely to change the outcome.

Modes and primitives

Use auto unless the caller specifies a mode. In auto, choose the lightest primitive that can resolve the uncertainty.

Primitive	Use when	Shape
panel	You need independent views, not interaction	Isolated fan-out -> synthesis
deliberate	You need collaboration and contest toward a recommendation	Panel -> focused exchange -> synthesis
debate	There are opposing positions or a binary/ternary fork	Assigned sides -> rebuttal -> judged synthesis
spar	One candidate recommendation needs a targeted pressure test	Prover -> verifier/challenger -> synthesis
red-team	A plan needs asymmetric attack/defense	Defender -> attacker -> mitigations
delphi	Authority effects or model dominance may distort the exchange	Anonymous or summary-mediated rounds

Do not treat votes as authority. If you report counts, report them as evidence about distribution of views, not as a decision rule.

Depth

Choose depth from the task, or accept explicit caller configuration.

Depth	Default shape	Use when
short	3 agents, one isolated panel round, one synthesis	Quick reversible trade-off
medium	3-4 agents, panel plus one focused interaction round	Architecture/product/workflow decision
long	4-5 agents, up to 3 rounds, required dissent report	High-stakes, irreversible, security, migration, or deep uncertainty

Cap ordinary councils at 3 rounds. Past that, drift, sycophancy, and transcript bloat usually cost more than they add. If a thread remains unresolved, preserve it in the minority report or reopen conditions instead of spinning.

How to run a council

1. Frame the decision. State the exact question, options under consideration, constraints, success criteria, non-goals, and what would change the recommendation.
2. Decide whether council is warranted. If a single-model answer with direct verification is enough, do not open a council. If the user explicitly asked for council, continue.
3. Configure the roster. Choose models, personas, lenses, and depth. Prefer heterogeneous frontier models for hard open-ended questions. If only one model family is available, make roles structurally different and keep the first round isolated.
4. Create the artifact set. Put it outside the repository unless the operator explicitly asks for a committed artifact. Use separate files for the brief, transcript, synthesis, and optional state metadata.
5. Launch one contained council-runner. Give it the brief, transcript, and synthesis paths. The runner spawns the members itself, runs the rounds in its own context, and returns only the synthesis path. The main session does not launch members or relay rounds (see Artifacts and containment).
6. Let the council deliberate contained. The runner runs an isolated panel first, then escalates to the lightest interaction primitive that targets live disagreement, flushing each turn to the transcript and writing the final synthesis.
7. Read the synthesis, including its dissent, by default. The main session integrates or briefs back from the synthesis artifact, and always reads the bounded minority_report and reopen_conditions fields so dissent is never silently dropped. It reads the raw transcript only on user request, failure, or an explicit audit trigger.
8. Close explicitly. State what the council concluded, what dissent remains, what would reopen the recommendation, and where the artifacts are.

Roster and model selection

• Persona diversity is the guaranteed floor; model diversity is requested, not guaranteed. Personas live in the role cards you control and are always delivered. A requested member model may silently fall back to a default model with no error, and the runtime exposes no way for a member to verify which model it is. Prefer model diversity for hard open-ended decisions, but never present it as a guarantee, and never rely on a member self-reporting its model as proof.
• Have each member self-report its provider_family (with unsure allowed) and record it in the synthesis provenance manifest. Treat a mismatch against the requested model -- especially a non-Claude/non-GPT request reporting Claude -- as a fallback signal to log and warn, not as verification either way.
• Keep panel members close enough in capability that one model does not dominate the others by prestige. If a model is much weaker, use it as a focused critic, fact-checker, or peripheral reviewer, not an equal voter.
• Use personas to create cognitive diversity: pragmatic implementer, operations skeptic, first-principles designer, user advocate, risk analyst, red team, maintainer, domain specialist, or orthogonal-methods designer.
• Reuse PAW Society-of-Thought specialists when relevant: security, performance, assumptions, edge-cases, maintainability, architecture, testing, correctness, release-manager. Reuse perspectives such as baseline, premortem, red-team, and retrospective as overlays.
• If a caller pins models or personas, respect those pins unless they create an invalid council, such as a same-model judge evaluating its own output without dissent preservation.

Council member contract

Every council member should output a structured turn:

agent_id: short-stable-id
model: requested-model-id
provider_family: self-reported provider/family/tier, or "unsure" (a weak probe, not proof)
persona: role or specialist name
epistemic_act: PROPOSE | SUPPORT | CHALLENGE | REFINE | CONCEDE | SYNTHESIZE
key_claim: one-sentence claim
confidence: HIGH | MEDIUM | LOW
grounds: evidence, assumptions, or reasoning that support the claim
warrant: why the grounds imply the claim
rebuttal_conditions: what would change or falsify this view
dissent_or_alignment: who this agrees or disagrees with, and why
relevance: CORE | ADJACENT | PARKED
smallest_change: the smallest direction-preserving fix, or "stop: <proof>"
what_gets_smaller: how this turn reduces uncertainty about the framed decision

Rules for members:

• Do not self-critique as a substitute for external critique.
• Do not concede because others disagree. Revise only when a specific flaw, missing evidence, or better frame is identified.
• If you agree, state why with evidence. Agreement without grounds is sycophancy.
• If you have no concern in your domain, say what you examined and why it passed. Do not fabricate a flaw to satisfy an adversarial prompt.
• Challenge specific claims, not personalities, model identities, or verbosity.
• Surface uncertainty explicitly.
• Classify each objection as CORE, ADJACENT, or PARKED against the framed decision. Do not promote an ADJACENT or PARKED concern into a pivot unless you prove the core decision fails without resolving it.
• A "this is broken, change direction" turn must carry the smallest direction-preserving alternative, or conclude stop with proof. Finding a defeater is not the goal; preserving or sharpening the decision is.
• Do not let stated confidence stand in for evidence. A confident assertion with weak grounds is weaker than a hedged one with strong grounds.

Keeping the council on course

Multi-agent deliberation has a characteristic failure mode: a member is rewarded for finding a "blocker," so a peripheral concern inflates into the center of the work, and the confidence with which members argue starts to drive how important a pivot seems. The council can then chase spikes and drift off the decision it was opened to make. Guard against it explicitly.

• Hold a decision vector. The brief states the core decision, what is being made smaller or settled, the non-goals and parked topics, and the stop condition. Every round carries it. Members may challenge the vector only by proving the decision cannot succeed without revising it.
• Conserve focus. Classify findings as CORE, ADJACENT, or PARKED. Adjacent and parked concerns are recorded, not allowed to steer the recommendation, unless a member proves the core decision fails without them.
• Reward sharper claims, not disagreement for its own sake. Distinguish three outcomes:
  • fake convergence - a vaguer, mushier shared claim ("it depends", mutual hedging). Forbidden; push back to the real crux.
  • genuine convergence - a sharper, more specific shared claim. A win.
  • named irreducible - agreement that the decision reduces to one either-way that argument cannot settle. The best outcome of a thread; name the open variable and move on. Do not stage disagreement to keep a thread alive, and do not treat agreement as failure. A robust answer re-derived from independent angles is a result, not fatigue.
• Make each round shrink the problem. Every round should state what got smaller, clearer, or closer to action. Two consecutive rounds that shrink nothing end the council.
• Separate confidence from grounding. Weight findings by evidence and rebuttal-resistance, not by how confidently they are asserted. Flag a recommended pivot as decision-blocking only with proof, never on conviction alone.

These rules apply across primitives, but the success condition differs by mode. In debate, sharpened disagreement or a named irreducible is the goal and easy consensus is suspect. In deliberate and ideation-style work, convergence toward a built recommendation is expected and fine; there the risk is vagueness and dissolved forks, not premature agreement. State which success condition applies before starting.

Artifacts and containment

The council's inner deliberation must stay out of the main session context by default. The main session launches one contained runner, waits for completion through the normal Copilot CLI subagent lifecycle, reads only the synthesis artifact, and briefs or integrates from it. The transcript stays auditable on disk but is not absorbed into the driver context unless there is a reason.

Use an artifact set:

council-<id>/
  brief.md        # curated context pack for all council members
  transcript.md   # full deliberation, flushed turn-by-turn, not read by the driver by default
  synthesis.md    # recommendation packet (with provenance manifest) read by the driver
  state.json      # optional status, timestamps, member ids, artifact paths

The contained council-runner

The driver launches a SINGLE council-runner subagent that owns the whole council in its own context and returns only the synthesis path. One in-context coordinator means there is no peer-coordination machinery at all -- no shared-file locks, turn tokens, or stale-lock recovery. The council requires a runtime where a subagent can spawn nested subagents (the members).

1. The driver writes brief.md (decision frame, context, options, non-goals, linked artifacts, roster, depth, mode, required output).
2. The driver launches ONE runner subagent, passing the brief, transcript, and synthesis paths. It does not launch members itself and does not relay rounds.
3. The runner spawns the members as nested subagents (each with its requested model and role card), runs an isolated panel round first, then escalates to the lightest interaction primitive only if material disagreement remains (max 2-3 rounds; stop at sharper convergence or a named irreducible).
4. The runner flushes each member turn to transcript.md as it happens (mandatory), so a hung, dropped, or silently-fallen-back member leaves recoverable state instead of a silent void. Containment means "the driver does not read the rounds," never "the rounds are unrecoverable."
5. The runner writes synthesis.md as rapporteur, including the provenance manifest, then obtains a faithfulness check from a member it did not author the synthesis as. It returns to the driver only the synthesis path plus a short meta report (rounds run, members + requested models, nesting outcome, any errors) -- no round contents, no synthesis body.
6. The driver reads synthesis.md, always surfaces its minority_report and reopen_conditions, reports the conclusion, and provides the transcript path for audit.

This optimizes for context hygiene: the driver is influenced by the council's conclusion and provenance, not by every intermediate argument, persona, tangent, or recency effect from the debate.

Check-in posture (default: light, contained)

By default the council is contained: the runner judges with what the brief gives it and the driver sees only the synthesis. The one exception is a bounded context-request -- when the panel surfaces a material fact the brief did not supply (the kind of gap that would otherwise force a low-confidence or assumption-laden finding), the runner may pause once at the panel-to-interaction boundary and ask the driver for it. The driver answers with the fact or says "proceed with what you have," and an unfilled gap is recorded as a stated assumption in the synthesis (open_questions / reopen_conditions). This is rare by design, batched into at most one request, and never pulls the deliberation into the driver's context -- only a targeted gap-fill. It does not make the driver the round orchestrator; the runner still owns every round.

Mechanism: launch the runner as a background agent so it can pause and resume; a context-request comes back as a tagged intermediate return (NEED-CONTEXT: ...), the driver replies once with write_agent, and the runner continues. A council that needs nothing just returns the synthesis.

This default is a posture, not a guardrail, and the operator sets it in either direction. Explicit operator direction always wins -- never treat containment as a reason to withhold a check-in the operator asked for:

• Quieter -- "closed council" (or "sealed"): the runner never asks; it records any gap as an assumption and judges on the brief alone.
• Default: light, material-gap context-requests only, as above.
• More involved: if the operator asks to be checked in with, to see emerging forks, or to be brought in on a high-spread call (e.g., "check in with me at the boundary," "surface any fork to me," "ask me freely"), do exactly that -- pause, return a bounded note (emerging CORE items + the fork), take the steer, resume.

Whatever the posture, keep what crosses back to the driver bounded -- a specific question, a fork, or a short note, never the rounds.

Context packs

Council members do not inherit the main session's full context. Give them enough context to deliberate well without forcing the main session to summarize the entire conversation into its own prompt.

Default brief.md contents:

• the decision question and why council is being opened;
• options under consideration;
• constraints, success criteria, non-goals, and known risks;
• relevant files, diffs, issues, PRs, research reports, transcripts, or session artifacts by path;
• the specific surfaces, sub-questions, or artifacts each member must cover -- a coverage checklist, so recall is structured rather than emergent;
• prior conclusions the council should treat as context, not as authority;
• roster, roles, models, depth, and mode;
• required output packet and audit triggers.

Optional deep context may include links to broader session artifacts or selected conversation excerpts. Full-session-history loading is an explicit opt-in, not the default: it can improve grounding, but it can also import irrelevant recency, private context, stale assumptions, and prompt noise.

Council agents should read brief.md first, follow linked artifacts as needed, and mark context gaps instead of guessing.

Transcript protocol

The runner is the single writer, so no locking is needed: it must flush each member turn to transcript.md as it completes (mandatory), with the member id, requested model, persona, round number, and a timestamp. Incremental flush is what keeps "contained" from meaning "opaque" -- it leaves recoverable state if the runner hangs or a member is dropped. A plain Markdown transcript is all that is needed -- the single writer means no structured event format or sequence bookkeeping is required.

Do not overbuild subagent lifecycle handling. If the runner hangs, use Copilot CLI's existing subagent status/notification behavior plus the transcript timestamps to decide whether to wait, inspect, or terminate. The skill defines deliberation and artifacts; it does not replace the host's subagent manager.

Audit model

The transcript is not hidden authority. It is deliberately out of the main context by default and available for audit on request. To avoid a circular audit model where every trigger depends on the synthesis confessing its own weakness, the main session always surfaces the bounded minority_report and reopen_conditions from the synthesis packet, regardless of triggers.

Read the raw transcript when:

• the user asks why the council concluded something;
• the synthesis reports low confidence or unresolved high-impact dissent;
• the synthesis is internally inconsistent or missing required fields;
• the run failed, timed out, or appears stale;
• a faithfulness check disputes the synthesis packet;
• a member's reported provider_family does not match its requested model (silent-fallback signal), or a non-Claude/non-GPT request reports Claude;
• the provenance manifest lists fewer members than requested, or a member turn is missing from the transcript;
• the runner returns a synthesis without a transcript flushed to the known path;
• the coverage manifest leaves a brief-listed surface not-examined (decide whether that gap matters before acting);
• the synthesis records a material context gap the brief did not supply (a signal the brief was thin, whether or not it was filled);
• the driver needs to verify a specific claim before acting.

Synthesis packet

The final answer should be concise enough for the main session, but structured enough to preserve the council's value:

decision_vector: the core decision this council was opened to make
recommendation: the best current recommendation
confidence: HIGH | MEDIUM | LOW
convergence: genuine-sharper | named-irreducible | unresolved
confidence_basis:
  panel_distribution: where agents agreed or split
  decisive_evidence: what actually changed the recommendation
  foreman_assessment: how strong the synthesis is
decisive_arguments:
  - claim: ...
    source_agents: [...]
    evidence: ...
minority_report:
  - agent: ...
    position: ...
    why_not_adopted: ...
parked:
  - topic: ...
    why_parked: recorded but not allowed to steer this decision
coverage_manifest:
  - surface: a brief-listed surface, sub-question, or artifact
    status: core-finding | examined-passed | not-examined
    by: member id(s) that covered it
open_questions:
  - ...
reopen_conditions:
  - If this condition changes, revisit the recommendation
audit_triggers:
  - when the raw transcript should be read before acting
faithfulness_check:
  by: agent id of a non-author member, or "independent-rapporteur"
  verdict: SUPPORT | DISPUTE
  note: one line asserting the packet matches the transcript
provenance_manifest:
  - member: agent id
    requested_model: the model id the runner asked for
    persona: role or specialist name
    provider_family: self-reported, or "unsure"
    model_identity: UNVERIFIED   # the runtime cannot confirm which model ran
    fallback_suspected: yes | no   # provider_family mismatches requested model
brief_path: path to the context pack
transcript_path: path to the artifact
synthesis_path: path to this packet

recommendation, confidence, decisive_arguments with their source_agents, minority_report, reopen_conditions, audit_triggers, the provenance_manifest, the coverage_manifest, and the artifact paths are required fields. The coverage_manifest maps each brief-listed surface to core-finding, examined-passed, or not-examined, so recall is checkable and blind spots are visible rather than silent. minority_report and reopen_conditions are never omitted: if there is genuinely no dissent, state that explicitly rather than dropping the field. The provenance_manifest records, per member, the requested model and self-reported provider_family with model_identity: UNVERIFIED, so the driver sees the requested roster and any suspected silent fallback without trusting model identity as fact. A synthesis missing any required field is itself an audit trigger.

The rapporteur may merge, deduplicate, classify, and recommend. It must not erase dissent, invent new findings that no member raised, or pretend that majority count is proof. It weights findings by grounding and rebuttal-resistance, not by stated confidence, and it does not elevate an ADJACENT or PARKED concern into the recommendation or a pivot without a member having proven the core decision fails without it.

Prefer an independent rapporteur that did not hold a decisive position in the deliberation. When the rapporteur is also a debater, a different member must add the faithfulness_check stub: a one-line SUPPORT or DISPUTE asserting that the synthesis packet faithfully reflects the transcript. A DISPUTE verdict is an audit trigger.

Modes

• integrate (default) - run the council, fold the synthesis into the work, and proceed. Use when the driver owns the decision.
• brief-back - run the council and report the recommendation, dissent, and risks to the operator without acting. Use when the operator owns the decision or explicitly asks agents to discuss and brief back.
• artifact-only - run or shape the council into an artifact without taking a decision. Use for planning, dogfood, or future implementation design.

Infer the mode from the caller. If unclear and the decision is consequential for the operator, prefer brief-back.

Caller customization

Callers may specify:

• models: gpt-5.5, claude-opus-4.8, claude-opus-4.7-high, etc.;
• personas or PAW SoT specialists;
• mode: panel, deliberate, debate, spar, red-team, delphi, or auto;
• depth: short, medium, long, or auto;
• transcript path or artifact policy;
• whether to return a concise brief or a full structured packet.

If the caller only asks for "a council", configure the smallest council likely to change the answer: usually 3 diverse perspectives, isolated first round, one focused interaction round only if needed, and a synthesis with dissent.