Configuration¶

OpenCastor is configured via config/arm.rcan.yaml (or the path set by CASTOR_CONFIG).

Minimal config¶

agent: model: anthropic/claude-sonnet-4-6 harness: max_iterations: 8 thinking_budget: 2048 context_budget: 16384 cost_gate_usd: 0.05 retry_on_error: true drift_detection: true p66_consent_threshold: physical # physical | digital | none

contribute: enabled: false idle_after_minutes: 15 projects: - harness_research ```

Key tunables¶

Key	Default	Description
`max_iterations`	`8`	Max agent loop iterations per task
`thinking_budget`	`2048`	Token budget for model thinking/CoT
`context_budget`	`16384`	Max context window tokens
`cost_gate_usd`	`0.05`	Abort task if estimated cost exceeds this
`retry_on_error`	`true`	Auto-retry on transient errors
`drift_detection`	`true`	Detect and abort drifting tasks
`p66_consent_threshold`	`physical`	Minimum consent level for physical actions

Environment variables¶

Variable	Description
`CASTOR_CONFIG`	Path to config file
`CASTOR_RRN`	Robot Resource Name override
`ANTHROPIC_API_KEY`	API key for Claude models
`OLLAMA_URL`	Ollama endpoint (default: `http://localhost:11434`)
`OHB_MODEL`	Model for OHB-1 benchmark (default: `gemma3:1b`)

Champion configs¶

When the research pipeline finds a better harness config, it's stored as a pending update in Firestore — never applied automatically. You choose when to apply it:

bash castor harness apply-champion # apply to this robot castor harness apply-champion --fleet # apply to all robots

Or use the client app to review and apply with one tap.