Skip to content

Binding and security ​

Parent: Agents and delegation.

Every invocation goes through the same binder. Binding is runtime-specific:

  • runtime: rho: resolve model aliases and reasoning against the host config, apply optional provider/auth pins (keeping compatible parent auth, otherwise selecting an available target-provider login when auth is unset), settle fast mode from the agent's effective fast (an [internal_agents.<id>] entry turns it off; delegated and workflow runs never inherit the parent's /fast; fast: true on a model, provider, or auth without fast serving fails bind; host asides such as /side keep the session's /fast when the inherited model can serve it), render prompt policy, and intersect requested Rho tools with host-supplied capabilities. Host policy remains the upper authority boundary.
  • A per-call agent.model replaces only the Rho model/provider/auth selection, before binding reasoning, fast mode, and capabilities. It uses catalog/alias resolution and requires available credentials. It never rewrites the definition or its fingerprint, expands its tool allowlist, changes runtime, or relaxes host permissions. Omission keeps the existing definition policy, or the agent's [internal_agents.<id>] entry when it has one. See per-call model selection.
  • runtime: claude-cli: copy model byte-for-byte (or omit it when inherited), keep the Claude tool list, map optional reasoning: to Claude effort (low/medium/high/xhigh/max), and record inherit_claude_config. No Rho model-alias resolution and no mutation of the parent provider/model config. Rho-style @alias model values and reasoning: off / reasoning: minimal are rejected. runtime: claude-cli is delegated-only: interactive and automation roots cannot bind it. Auto and Allow edits later fail closed at spawn unless every Claude tool is a proven no-prompt built-in for that Rho approval class; unknown plugin and MCP names are not proven.