compressionenum
Context compression: none (higher token usage), moderate (balanced), aggressive (lower token usage). Default: aggressive
contextstring
Additional context to frame the deliberation
context_filestring
Path to a file whose contents will be used as context. Use this for large payloads that exceed inline string limits. If both context and context_file are provided, they are concatenated.
model_groupenum
Model selection. Cost is relative to smart, the default: cheap and fast run small models at roughly 1/4 of smart; anthropic-fable is roughly 2x smart. smart (current flagships — Opus 5, GPT-5.6 Sol, Gemini 3.1 Pro Preview; the baseline), mixed (random assignment across all providers — cheaper on average than smart, but quality varies seat to seat), fast (small models, lowest latency), cheap (small models, lowest cost — same model tier as fast), informed (search-capable models incl. Perplexity Sonar; pick when the answer turns on current facts), large (largest context windows; pick when the context payload is big, not when you want a better answer). Or single provider: openai, anthropic, anthropic-fable, google, perplexity. anthropic-fable runs Claude Fable 5 on EVERY agent — Anthropic's most capable model, but roughly 2x the token cost of smart, slower, and with an older knowledge cutoff (Jan 2026 vs May 2026 for Opus 5). Do NOT select anthropic-fable unless the user has explicitly asked for maximum capability and accepted the higher cost. Default: smart (MCP use cases tend toward high-stakes deliberation where flagship rigor is worth the cost).
questionstringобязательный
The question or topic for the panel to deliberate
roundsnumber
Polyhedron traversals (1-5). IMPORTANT: each agent sits on an EDGE of the polyhedron connecting two vertices (conversation nodes). In a single round, every agent participates TWICE — once contributing to its "left" vertex and once to its "right" vertex. So rounds=1 already produces ~2 responses per agent (total ≈ num_agents × 2). Higher rounds add additional full polyhedron traversals where every agent re-participates at both vertex endpoints reacting to the prior round's outputs. rounds=1 is already substantial deliberation; use 2+ only when agents need to react to other agents' completed positions across the full polyhedron — e.g., binary strategic decisions with sharply opposing arguments that benefit from second-pass reaction. Default: 1.
short_responsesboolean
Request concise panelist responses. Default: false
show_and_audit_reasoningboolean
Append Chain-of-X scaffolding to every panelist: Part 1 (reasoning quality requirements — explicit attribution, assumption surfacing, counterfactual contingencies, robustness assessment) and Part 2 (cross-analysis audit — agents flag prior agents' attribution gaps, confabulated specifications, probability/evidence mismatches, etc.). Substantially raises rigor for high-stakes deliberation. Increases token usage and run cost (~1.45x scaffolding-only multiplier). ON BY DEFAULT — omit this parameter to accept it. Pass false only if the user has asked for a cheaper run and accepted lower rigor.
topologyenum
Panel size. Agents sit on polyhedron edges, so the agent count is the edge count: small = tetrahedron, 6 agents (~12 responses per round); medium = octahedron, 12 agents (~24); large = icosahedron, 30 agents (~60). Time and cost scale with agent count — large is 5x small. Prefer small for exploration and most questions; escalate only when the question genuinely needs more distinct perspectives, not merely a better answer. Default: small
web_search_enabledboolean
Enable native web search (Anthropic web_search_20250305, Google googleSearch grounding) for agents to verify specific factual claims — dates, statutes, case citations, fees, version numbers, jurisdiction-specific rules, etc. Requires the Smart model group (other groups have search disabled or unavailable). Increases run cost (~3.25x search-only multiplier; ~6.5x when combined with audit scaffolding). Default: false
