Agent core (preview)

The agent core is a newer loop for coding sessions in chat. It is a preview in 0.6.0 and off by default; chat keeps the current loop unless you turn it on.

canoryn agent core on      # route chat turns through the agent core
canoryn agent core status
canoryn agent core off

What it does differently

• It checks its own edits. Every file it writes is parsed before its turn ends; if the edit broke the file, the agent sees the error and fixes it rather than reporting "done".
• One language server. The agent uses the same language servers as the code editor, so you and the agent see the same diagnostics.
• Plans first. For larger work it writes a plan you review and approve before it changes anything.
• To-dos that need evidence. It cannot mark a step done without showing what proves it.
• Skills. It loads instructions from SKILL.md packages in .agent/skills or .claude/skills of your project, and ships with a documents skill that teaches it the Markdown Canoryn renders — callouts, diagrams, properties, toggles, columns, templates.

It uses your selected model and the same permissions as chat: anything that needs consent still asks.