Specification
Look System
kanonak.org
agents@1.0.0
agents
Accept Edits
Auto-accept file edits; still confirm other destructive actions.
Agent
A named AI agent (called a subagent by some client products) that carries its own system prompt, description, tool access, and references to the skills and…
Agent ID
Short machine identifier for the agent. Lowercase letters, numbers, and hyphens only. Maps to the name field in Claude Code subagent frontmatter, the name…
Auto
Auto-accept anything the host considers safe.
Bypass Permissions
Skip all permission checks. Intended for sandboxed environments where there is nothing the agent can damage.
Color
Optional display color used by clients that show running agents in their UI. Recognized values follow the Claude Code palette: red, blue, green, yellow,…
Default
Confirm destructive actions. The standard mode for interactive sessions.
Description
Human-readable one-line description telling the host client when to delegate to this agent. The host uses this string to decide whether an incoming task…
Instructions
The agent's system prompt in markdown. Rendered verbatim into the body of the generated Claude Code or Gemini markdown file, and into the…
Max Turns
Optional. Maximum number of agentic turns the agent is allowed to take before it stops. Useful to cap runaway exploration for narrowly-scoped subagents.
Model
Optional. Model alias or full model id the agent should use (e.g., "sonnet", "opus", "haiku", "claude-opus-4-6", "gpt-5", "gemini-2.5-pro"). When omitted, the…
Permission Mode
A named policy governing how aggressively an agent is allowed to take actions without asking for confirmation. Modeled as a class so concrete modes (default,…
Permission Mode
Optional. The named permission policy the host should apply when running this agent. Clients that do not support a particular mode fall back to their own…
Plan
Read-only planning mode. The agent gathers context and proposes a plan but cannot execute destructive actions.
Speaks
Protocols the agent uses to talk to tools or other agents. Typical values point at kanonak.org/mcp for tool access and kanonak.org/a2a for agent-to-agent…
Target Client
Optional. Restricts this agent to a specific client or set of clients (e.g., only claude-code because the instructions rely on a Claude Code feature). When…
Uses Skill
Skills this agent knows how to use. Deploying the agent to a client SHOULD transitively deploy its referenced skills, so an agent that declares a skill via…
Uses Tool
Tools the agent is pre-approved to invoke. Matches the tools field in Claude Code subagent frontmatter. Leave empty to inherit every tool from the parent…