STABLE CHANNEL — recommended. MCP server for cross-Claude-Code-chat orchestration plus an optional control-plane daemon. Peer messaging and discovery, chat read/search across sessions, live context and rate-limit monitoring via a chained statusLine wrapper and a PostToolUse OAuth hook, context and rate-limit guards, idle-beep notification, bundled role skills. The opt-in daemon (`install --systemd`) supervises peer lifecycle: spawn/stop/restart with sanitized env and a fork-guard, peer_compact orchestration (audited inject path, auto-watchdog off by default), team_layout declarative reconcile, lifecycle events routed into offline peers' inboxes. Without the daemon the plugin behaves exactly as before — no regression. See docs/SETUP-DAEMON.md and ADR-008 in docs/architecture.md.
4 skills · 1 MCP server · hooks
DEVELOPMENT CHANNEL — ahead of stable, not yet promoted. Do NOT install alongside `claude-bridge`: both register an MCP server. Closes an eight-tool quality gate (peer_restart, peer_spawn's environment whitelist, model_info, session_stats, peer_chat_read, peer_chat_search, rate_limit_status, team_adopt) and adds three lifecycle tools — team_release (drop a peer from daemon state without touching its process), team_reconcile (report where state and reality disagree, read-only), team_restart (rolling restart that refuses unrestartable peers up front and stops at the first failure) — plus `claude-bridge-daemon send` for delivering a message into a peer's inbox from outside the fleet. Two rounds of live piloting since: adoption carries a peer's model and launch parameters, a restart no longer resumes an id that names no transcript nor reports success over a process that died, an adopted peer stays in its team's tmux session even if its window is already gone, peer_list reports the peer's pid rather than the bridge's, and team_adopt can be scoped to one host session. 554 tests. v0.10.11 closes the path none of the pilots had taken: the fleet launches `claude` as a bare name, and a relaunch composes its environment from nothing, so the name did not resolve — adoption now records the absolute path resolved through the peer's own PATH. A peer that was the only window of its session can come back (the session is recreated), and a failed restart keeps its record instead of erasing the peer from the control plane. v0.10.12: a record that outlives a failed restart no longer claims to be live — every post-spawn failure marks it unknown with no pid. v0.10.13/.14 fix an outage this project caused itself: a relaunched peer built its environment from the DAEMON's values, whose PATH under systemd has no nvm — so twenty-one peers lost node, their statusLine, their hooks and their own MCP server at once. A peer's own environment is now captured and reused (still filtered by the same wh
4 skills · 1 MCP server · hooks