Add pluggable agent architecture for multi-agent support

[james-in-a-box]

Context

The Entireio CLI supports multiple AI agent backends through an abstract Agent interface with optional capability interfaces:

• HookSupport — lifecycle hooks (Claude Code, Cursor)
• FileWatcher — activity detection via file watching (Aider)
• TranscriptAnalyzer — agent-agnostic work detection
• TranscriptChunker — splitting large transcripts for storage constraints

They currently support Claude Code, Gemini CLI, Cursor, and Aider through this pluggable architecture.

Current State

Our system is tightly coupled to Claude Code. The sandbox entrypoint, hook system, transcript parsing, and checkpoint handler all assume Claude Code's JSONL transcript format, session directory structure, and hook API.

Proposal

Introduce an agent abstraction layer that would allow supporting additional AI coding agents in the future:

• Define an abstract agent interface with core methods (detect, read session, install hooks)
• Make transcript parsing agent-specific via the interface
• Support optional capabilities (hooks, file watching, transcript analysis)
• Keep Claude Code as the primary/default implementation

This is lower priority than other items since we currently only use Claude Code, but the architectural pattern is worth adopting proactively if we anticipate supporting other agents (Codex, Gemini, etc.).

Reference

See entireio/cli — agent/ package with registry pattern and capability interfaces.

Authored-by: egg