Claude/Codex Hooks and Skills

Maxx ships a maxx-agent helper and one bundled skill for Claude Code and Codex. The skill teaches agents how to call Maxx’s local control surface, while the helper translates explicit agent lifecycle events into sidebar status.

What Gets Installed

• maxx-agent: open, list, rename, prompt, and close visible Maxx tabs, then use the returned durable Control API session_id for follow-up work.
• Advanced maxx +control guidance inside the same skill for coordinating groups of child tabs from a parent session.
• Hook configuration for supported CLIs so Maxx can display agent-declared activity state.

Claude Code discovers personal skills under ~/.claude/skills or $CLAUDE_CONFIG_DIR/skills. Codex discovers user skills under ~/.agents/skills. Maxx writes managed skill files with an ownership marker so uninstall can remove only the files it owns.

Install from Maxx

Open Maxx Settings and use the Claude Code or Codex install controls. Settings also expose default permission and sandbox modes for agent-spawned sessions; explicit CLI flags still take precedence.

Verify the Boundary

Supported agent status indicators come from explicit hook events, control API declarations, or terminal-native mechanical facts such as bell/progress escape sequences. They do not come from scraping terminal output.

maxx-agent new-tab returns a session_id handle for the tab it creates, and the child can expose an explicit answer through sessions set-result. Parents should retrieve that answer with sessions get <session_id> and read result, result_source, and result_at. Workflow state still needs explicit declarations such as set-state, set-summary, metadata, or event payloads.

See the no-inference rule for the full product constraint.