config_fields.md

Config Fields

Field-level guidance for experiment configuration.

experiment

• id: unique experiment identifier
• seed: random seed for reproducibility
• max_steps: maximum controller steps (positive integer; used when controller.run omits max_steps)
• seeds: optional map of library-specific seeds

prompts

• optional prompt path overrides (system, persona, task, protocol, action_space, return_format, action, probe, persona_profiles)

agents

• id: agent identifier
• role: agent role label
• model: provider/name/decoding params
• persona_profile: optional persona string to control agent personality
• persona_profile_path: optional path to persona profile prompt

action_space

• enabled: list of enabled action primitives (message, decide, propose, respond, transfer)
• decide.reveal: default decision reveal strategy (sequential/aggregated/simultaneous)

controls

• communication: communication constraints (mode: broadcast/direct, max_messages_per_turn: positive integer)
• information_distribution: visibility controls (visibility: non-empty string)
• interdependence: task interdependence structure (structure: non-empty string)
• visibility_map: per-agent visible state fields by section (shape: {agent_id: {task_state: [fields], resources: [fields], turn_state: [fields], buffers: [fields]}})
• visibility_defaults: default visible state fields by section (shape: {task_state: [fields], resources: [fields], turn_state: [fields], buffers: [fields]})
• visibility_defaults_by_agent: per-agent default visible fields map (shape: {agent_id: {task_state: [fields], resources: [fields], turn_state: [fields], buffers: [fields]}})
• visibility_overrides: additional visible fields applied on top of defaults (shape: {task_state: [fields], resources: [fields], turn_state: [fields], buffers: [fields]})
• visibility_overrides_by_agent: per-agent visibility override map (shape: {agent_id: {task_state: [fields], resources: [fields], turn_state: [fields], buffers: [fields]}})
• visibility_excludes: excluded visible fields (shape: {task_state: [fields], resources: [fields], turn_state: [fields], buffers: [fields]})
• visibility_excludes_by_agent: per-agent excluded visible fields (shape: {agent_id: {task_state: [fields], resources: [fields], turn_state: [fields], buffers: [fields]}})

task

• type: registered task identifier
• task-specific fields documented in docs/task_schema.md
• hidden_profile fields: target_steps, shared_facts, private_facts

protocol

• turn_taking: turn-taking strategy (sequential or simultaneous)
• termination: termination settings (condition: max_steps or task_complete)
• step_mode: controller stepping (event or tick)
• proposal_expiry_steps: integer step threshold for proposal expiry (relative to creation)
• decision_timeout_steps: integer step threshold before decision reveal
• decision_quorum: integer number of decisions required for aggregated reveal
• memory_turn_limit: number of recent turns to keep in agent memory (default 7)
• visible_event_window: integer number of recent events to include per observation
• visible_event_types: list of event types included per observation
• visible_event_types_exclude: list of event types excluded per observation
• visible_event_types_by_agent: per-agent event type filter map
• visible_event_types_exclude_by_agent: per-agent excluded event type map (precedence: per-agent includes override global excludes unless per-agent excludes are set)
• visible_event_window_by_agent: per-agent event window map
• agent_memory_key_limit: max number of agent memory keys to persist per step
• observation_memory_key_limit: max number of memory keys included in observations
• agent_memory_key_limit_by_agent: per-agent memory key limit map
• observation_memory_key_limit_by_agent: per-agent observation memory limit map

probe

• cadence: probe cadence (per_action, per_turn, on_event)
• templates: list of probe template ids
• events: list of event types to trigger probes (when cadence=on_event)
• questions_path: path to JSON list of interview questions
• questions: inline list of interview questions

logging

• trace_schema_version: schema version tag
• output_dir: output directory for run logs
• observation_events: when true, emit observation/action proposal events