Skip to content

Agent handbook

Use this guide to change or evaluate the managed stack. For day-to-day operation, start with the AI workbench.

AgentChoose it forManaged configuration
Codexprofile-led implementation, focused child roles, and managed MCP profileshome/dot_codex/
Claude Codeplugin/hook workflows and source-grounded or adversarial local roleshome/dot_claude/
CursorIDE/Agents Window, Composer 2.5 subagents, and self-hosted computer usehome/dot_cursor/
OpenCodegenerated MCP namespaces and a constrained review agenthome/dot_config/opencode/
Piits rendered local-MLX or declared cloud-provider setuphome/dot_pi/agent/

Choose the harness whose configuration and interaction model suit the task, then inspect the active session before relying on its tools or account state.

Edit the declared input, then render or reconcile it. Repairs inside ~/.codex or an installed skill directory do not survive sync.

Start from the source of truth

home templates and skills

chezmoi

package manifests

installers

deployed configuration

new session

observed inventory and evidence

Wide diagrams scroll horizontally. Focus the figure to use arrow keys.

NeedStart hereIt does not establish
Change guidanceInstruction architectureTool, account, or permission availability
Add a capabilityInventories and ownershipA live or authorized service
Select Codex routingProfiles and rolesA sandbox or guaranteed savings
Assess a resultValidationCorrectness beyond the performed check

Instructions govern reasoning; skills provide procedures; MCP servers provide configured tool routes; plugins provide marketplace extensions. None independently proves a login, a live service, or permission for an external action.

  1. Read project instructions and inspect the worktree.
  2. Pick the narrowest profile and skill that fit.
  3. Use a direct command for deterministic work; delegate only an independent bounded unit.
  4. Keep the command, artifact, and assertion together.
  5. Exercise the result in its consumer and state what that observation does and does not prove.

For persistent state, use browser workspaces or durable tasks. For an interface-specific route, see browser and desktop evidence or domain workflows.