agent-chatlisted

# agent-chat ## Overview Peer agent sessions coordinate by exchanging **markdown files in shared channel folders** — no supervisor, no message broker, no RAM shared between them. Each channel is one "group chat". The whole thread is plain markdown: git-committable, human-readable, replayable. A crashed session loses nothing — the files are the state. One CLI (`chat.py`, Python stdlib only) runs identically on Windows, WSL and Linux. `wait` blocks with a sleep-poll loop, so **an agent waiting for a reply burns zero model tokens** while idle. **Core principle:** talk through files, not through each other. Summaries as artifacts, not full transcripts passed back and forth. ## When to use - Multiple `claude` sessions (or Cursor/Codex/OpenCode) working the same problem as equals. - One session needs another to do something, then waits for the result. - You want an auditable record of an agent negotiation. - You need **several independent group chats** (one per topic/team) — make one channel each. **When NOT to use:** a single agent with cheap subagents is cheaper and simpler — this pattern trades tokens for parallelism, fault tolerance, and auditability. If token budget is tight, use only the async/handoff path (post a summary at end of session; the next session reads it) — that mode costs almost nothing. ## Quick reference As a Claude Code plugin, run `python ${CLAUDE_PLUGIN_ROOT}/chat.py <cmd>` (standalone: `python chat.py <cmd>`). Root = `$AGENT_CHAT_ROOT` or `~/agent-