Configuration

ProbOS is configured with YAML validated by Pydantic models. Every setting has a default, so a short file is a complete configuration.

Which File Loads

probos and probos serve load the file given with --config PATH; without it, ~/.probos/config.yaml (written by probos setup); without that, the checkout's config/system.yaml. The file that loads replaces the others rather than layering over them: anything it does not list takes its built-in default.

• config/system.yaml is the full reference configuration — 181 sections, with the Ward Room, proactive crew cognition, the agentic loop and its tools, code execution, the approvals center and NATS turned on.
• ~/.probos/config.yaml from probos setup is minimal: the LLM tiers, self-modification, the knowledge repository, utility agents and a security profile. Subsystems whose built-in default is off stay off until you add them.

probos doctor prints which config file and data directory it checked. The Configuration Reference is generated from the models and lists every section and field with its default.

Key Sections

Section	Controls
system	Instance name, log level
pools, scaling	Default, minimum and maximum pool sizes, spawn cooldown, health checks, demand-based scaling
mesh	Gossip interval, Hebbian decay and reward rates
consensus, trust_dampening	Minimum votes, approval threshold, trust priors (Beta distribution), red-team pool size, cascade dampening
cognitive	LLM endpoints and models per tier, token budgets, timeouts
memory, dreaming, knowledge, records	Episodic memory, dream consolidation, the knowledge repository, Ship's Records
ward_room, proactive_cognitive, earned_agency	Crew communication, proactive crew thinking, trust-tiered autonomy
agentic_loop, dm_agentic, agentic_tools, execution, browser_tool, mcp	The agentic loop, its tools, code execution, the browser tool and MCP servers
approval_inbox, persistent_tasks, security, firewall	Approvals, durable tasks, permissions, and the communication contagion firewall
nats, federation	The event bus and multi-node federation
channels	Discord, Slack and webhook adapters (Telegram, Slack and Matrix also have probos channel <name> setup)

pools:
  default_pool_size: 3
  min_pool_size: 2
  max_pool_size: 7
  spawn_cooldown_ms: 500
  health_check_interval_seconds: 5.0

LLM Tiers

ProbOS routes LLM calls through named tiers, each with its own endpoint, model and sampling settings:

Tier	Use
fast	Classification, gist extraction, quick routing decisions
standard	General-purpose decomposition, agent instructions
deep	Complex multi-step reasoning, code generation, architecture design
vision, vision_fast	Image understanding
compute_use	Computer use
image_gen	Image generation

probos setup configures the fast, standard and deep text tiers; the others are optional. Switch the active text tier at runtime with /tier fast|standard|deep, and inspect the configuration with /models.

Standing Orders

Agent behavior is governed by a four-tier instruction hierarchy stored as Markdown in config/standing_orders/:

Tier	File	Scope
Federation Constitution	federation.md	Universal, immutable
Ship Standing Orders	ship.md	Per instance
Department Protocols	engineering.md, science.md, medical.md, …	Per department
Agent Standing Orders	architect.md, counselor.md, builder.md, …	Per agent, evolvable

compose_instructions() assembles them at call time and injects the result into every crew agent's LLM request. Crew personalities and seed callsigns live in config/standing_orders/crew_profiles/.

Vessel Ontology

config/ontology/ defines the ship's organization: departments, posts, the chain of command, and which agent type fills each post. See the Agent Inventory.

Environment Variables

A few settings can be overridden from the environment — for example PROBOS_NATS_ENABLED=false runs without NATS, and PROBOS_LLM_URL replaces the shared LLM base URL. ProbOS also loads a .env file at startup; see .env.example. Use --data-dir DIR to choose the data directory.