boxlite-ai

Show what this AgentLite instance can do — installed skills, available tools, and system info. Read-only. Use when the user asks what the bot can do, what's installed, or runs /capabilities.

Quick read-only health check — session context, workspace mounts, tool availability, and task snapshot. Use when the user asks for system status or runs /status.

Delegate sub-tasks to remote ACP (Agent Client Protocol) peers such as Claude Code or Codex via the built-in acp_* host actions. Use when the user asks to hand work to another agent, wants a second opinion, or needs long / sandboxed execution that shouldn't block this chat.

Add /compact command for manual context compaction. Solves context rot in long sessions by forwarding the SDK's built-in /compact slash command. Main-group or trusted sender only.

Add Discord bot channel integration to AgentLite.

Add Gmail integration to AgentLite. Can be configured as a tool (agent reads/sends emails when triggered from WhatsApp) or as a full channel (emails can trigger the agent, schedule tasks, and receive replies). Guides through GCP OAuth setup and implements the integration.

Add image vision to AgentLite agents. Resizes and processes WhatsApp image attachments, then sends them to Claude as multimodal content blocks.

Add Ollama MCP server so the container agent can call local models for cheaper/faster tasks like summarization, translation, or general queries.

Use when iterating on a fix after an adversarial code review (Codex, human reviewer) surfaces design-level findings — not typos, lint, or formatter work. Frames the review/fix loop as implement -> [review -> REFLECT -> reproducer -> fix -> verify -> handoff] where each round runs in a fresh agent to prevent context accumulation. Provides a four-item cheap-checks list (timeline draw, upstream contract read, full lifecycle trace, variant enumeration) that catches recurring failure classes. No git commits during the loop — all changes commit together after zero-finding termination. Trigger when a review pass surfaces 2+ findings on a single change, when sibling bugs of an "already-fixed" issue keep appearing across rounds, or when the user asks "why did I miss this?". Do NOT use for single-line fixes where the cause is obvious from the diff, or for initial implementation work (use CLAUDE.md rules).