SKILL.md

$2a

/council validate this plan                                    # verdict mode (the default)

/council --mode=brainstorm caching approaches                  # brainstorm mode

/council --mode=debate should we adopt event sourcing?         # debate mode (named personas duel)

/council --depth=quick validate recent                         # fast inline check

/council --mode=brainstorm --focus=research k8s upgrade paths   # research = focused brainstorm

/council --roster=security-audit validate the auth system      # preset persona roster

/council --depth=deep --runtime=mixed --roster=leadership-quartet validate product thesis

/council --adversarial validate the auth system                # verdict over 2 adversarial rounds

/council                                                       # infers mode from context

Council works independently — no RPI workflow, no ratchet chain, no ao CLI required.

Modes — the deliberation taxonomy

--mode selects one of exactly three deliberation patterns; verdict is the default. The taxonomy is frozen as an executable spec — references/council-modes.feature.

--mode

Pattern

Synthesis

brainstorm

diverge — agents generate options independently before any cross-talk

ranked set of ideas, perspectives, risks (no PASS/WARN/FAIL)

debate

contend — independent positions → adversarial 0–1000 cross-scoring → reveal round

ranked decision with recorded dissent

verdict (default)

converge — agents judge the artifact against the bar independently

one PASS / WARN / FAIL with consolidated findings

verdict runs when --mode is omitted. validate is a verdict alias; research folds into brainstorm (--focus=research). When --mode is omitted council infers it from natural language — trigger words in references/task-type-rigor-gate.md.

Mode and focus are orthogonal. --mode is the deliberation pattern; --focus is the subject. --depth, --runtime, and --roster are knobs, never modes. Every mode runs the same lifecycle — convene → brief → deliberate → synthesize → record — and deliberate always isolates each agent before any cross-talk. Full taxonomy, knob aliases (--quick/--deep/--mixed), and the lifecycle contract: references/modes.md.

Spawn backend (MANDATORY)

Council requires a runtime that can spawn parallel subagents and (for debate and --adversarial) send messages between agents. If no multi-agent capability is detected, fall back to --depth=quick (inline single-agent). Skills describe WHAT to do, not WHICH tool — see skills/shared/SKILL.md for the capability contract. Backend-specific spawn/wait/message/cleanup examples:

Claude Native Teams → references/backend-claude-teams.md · Codex Sub-Agents / CLI → references/backend-codex-subagents.md

Background Tasks → references/backend-background-tasks.md · Inline → references/backend-inline.md

Shared Claude feature contract → skills/shared/references/claude-code-latest-features.md (local mirror: references/claude-code-latest-features.md)

See references/cli-spawning.md for the council-specific spawning flow (phases, timeouts, output collection).

Debate mode ( --mode=debate )

--mode=debate convenes 2–4 named domain-expert personas who duel: each writes an independent verdict in character, every persona adversarially cross-scores every rival 0–1000, then a mandatory reveal round forces concessions and surfaces blind spots. Synthesis is a score matrix → a ranked decision with dissent kept verbatim. dueling-idea-wizards maps to --mode=debate --focus=ideas; /expert-council routes here.

Constraints: personas decide and the orchestrator only counts; the briefing goes on disk, never through argv; the reveal is never skipped. Per-phase persona-slate / duel / reveal / score-matrix templates: references/dueling-route.md.

Adversarial round ( --adversarial ) — a verdict intensifier

--adversarial is a verdict-mode flag, not the debate mode. It runs verdict over two rounds — R1 independent verdicts, R2 steel-manning revision via backend messaging — for high-stakes verdicts where judges are likely to disagree (security audits, architecture decisions, migration plans). Skip it for routine validation where consensus is expected. Full protocol: references/adversarial-protocol.md.

Incompatibilities:

--depth=quick and --adversarial cannot be combined. If both are passed, exit with error: "Error: --quick and --adversarial are incompatible."

--adversarial only applies to verdict mode. Combined with --mode=brainstorm or --mode=debate, exit with error: "Error: --adversarial is only supported with verdict mode."

Architecture

See references/architecture-flow.md for the context-budget rule, full Phase 1→3 execution flow diagram, reviewer-config loading, graceful degradation table, effort levels, and pre-flight checks.

Packet Format (JSON)

See references/packet-format.md for the full JSON packet schema (fields, output_schema, judge-prompt boundary rules) and the Empirical Evidence Rule for feasibility reviews.

Perspectives

Perspectives & Presets: Use Read tool on skills/council/references/personas.md for persona definitions, preset configurations, and custom perspective details.

Auto-Escalation: When --preset or --perspectives specifies more perspectives than the current judge count, automatically escalate judge count to match. The --count flag overrides auto-escalation.

Mixed-mode perspective assignment: Under --mixed, the perspective list is built once and each perspective is assigned to one Claude judge and one Codex judge with identical prompt and packet. This produces head-to-head pairs (perspective × vendor) so verdict differences isolate the vendor variable. Without --preset or --perspectives, both vendors run 3 generic judges each (6 total). With a 4-perspective preset like security-audit, plan-review, or leadership-quartet, both vendors run those 4 perspectives (8 total). Do not split perspectives across vendors — symmetric pairing is the whole point of --mixed.

Named Perspectives & Consensus

See references/consensus-and-output.md for named-perspective usage (--perspectives, --perspectives-file, YAML format, flag priority), consensus verdict rules (PASS/WARN/FAIL combination table, DISAGREE resolution), and the finding-extraction flywheel protocol. See references/personas.md for built-in presets.

Explorer Sub-Agents

Explorer Details: Use Read tool on skills/council/references/explorers.md for explorer architecture, prompts, sub-question generation, and timeout configuration.

Summary: Judges can spawn explorer sub-agents (--explorers=N, max 5) for parallel deep-dive research. Total agents = judges * (1 + explorers), capped at MAX_AGENTS=12.

Agent Prompts

Agent Prompts: Use Read tool on skills/council/references/agent-prompts.md for judge prompts (default and perspective-based), consolidation prompt, and debate R2 message template.

Output Format & Consensus Rules

Consensus verdict combination rules, DISAGREE handling, and the finding-extraction flywheel protocol live in references/consensus-and-output.md. Full report templates (validate, brainstorm, research) and debate-report additions live in references/output-format.md. All reports write to .agents/council/YYYY-MM-DD-<type>-<target>.md. Findings extraction targets .agents/council/extraction-candidates.jsonl; see references/finding-extraction.md for schema and classification heuristics.

Core consensus rules: All PASS -> PASS; Any FAIL -> FAIL; Mixed PASS/WARN -> WARN; cross-vendor disagreement -> DISAGREE.

Configuration

Minimum quorum: 1 agent. Recommended: 80% of judges. On timeout, proceed with remaining judges and note in report. On user cancellation, shutdown all judges and generate partial report with INCOMPLETE marker.

Env var

Default

COUNCIL_CLAUDE_MODEL

sonnet

COUNCIL_EXPLORER_MODEL

sonnet

COUNCIL_CODEX_MODEL

gpt-5.3-codex

COUNCIL_TIMEOUT

120

COUNCIL_EXPLORER_TIMEOUT

COUNCIL_R2_TIMEOUT

Flag

Description

--technique=<name>

Brainstorm technique (reverse, scamper, six-hats). See references/brainstorm-techniques.md.

--profile=<name>

Model quality profile (balanced, budget, fast, inherit, quality, thorough). See references/model-profiles.md.

See references/flags-reference.md for the full flag and environment variable reference (COUNCIL_TIMEOUT, COUNCIL_CODEX_MODEL, --deep, --mixed, --adversarial, --evidence, --commit-ready, --preset, --profile, and all other flags).

CLI Spawning Commands

CLI Spawning: Use Read tool on skills/council/references/cli-spawning.md for team setup, Claude/Codex agent spawning, parallel execution, debate R2 commands, cleanup, and model selection.

Examples

See references/examples-extended.md for the full example catalog and walkthroughs (fast single-agent validation, adversarial debate, cross-vendor consensus with explorers).

Troubleshooting

See references/troubleshooting.md for common error messages, causes, and solutions, plus the judge→council migration table.

Multi-Agent Architecture

See references/multi-agent-architecture.md for the deliberation protocol, communication rules, Ralph Wiggum compliance, degradation behavior, and judge naming convention.