## Technical Decisions (from research) Based on collaboration, here are the chosen approaches: ### 1. Technology Stack: **Go** - Compiled, single binary - Libraries: `github.com/creack/pty`, `nhooyr.io/websocket`, `golang.org/x/term` - Integrated into existing Dockerfile.sandbox build process ### 2. WebSocket Client - Library: `nhooyr.io/websocket` - Connects to Rails ActionCable via `host.docker.internal:3000/cable` - Subscribes to `AgentChannel` with agent type identifier - Auto-retry on disconnect (5s backoff) ### 3. Claude Code Integration: **PTY (Pseudo-Terminal)** - Uses `github.com/creack/pty` for natural terminal emulation - Handles window resize signals (SIGWINCH) - Raw terminal mode for proper Tab/Arrow key handling - Claude runs with `--dangerously-skip-permissions` flag ### 4. Message Flow ``` Rails → WebSocket → Go Bridge → PTY → Claude Code ↓ Visual indicator: "[RAILS COMMAND] ..." ``` ### 5. Console/Observability - Direct PTY output to stdout (full Claude UI visible) - Rails commands prefixed with blue `[RAILS COMMAND]` indicator - stderr for bridge status messages ### 6. Process Management - Bridge spawns claude process on startup - Automatic reconnection to Rails on WS disconnect - Claude crash handling through process exit ### 7. Configuration: **Environment Variables** ```bash AGENT_TYPE=worker|reviewer|orchestrator RAILS_WS_URL=ws://host.docker.internal:3000/cable ``` ### 8. Deployment: **Integrated into Dockerfile.sandbox** - Multi-stage build: Go builder stage → Node stage - Binary copied to `/usr/local/bin/agent-bridge` - Existing `run-claude-agent.rb` modified to exec `agent-bridge` instead of `claude` ## Files to Create/Modify **New:** - `agent-bridge.go` - Main bridge application **Modify:** - `Dockerfile.sandbox` - Add Go builder stage - `run-claude-agent.rb` - Change `exec claude` to `exec agent-bridge`