refactor: replace file-based IPC with WebSocket communication

[echthesia]

Summary

Replace the file-based IPC system (stdin init, stdout sentinel markers, filesystem polling) with a single WebSocket server authenticated via per-container cryptographic tokens.

What changed:

• New src/ws-server.ts: host-side WS server with token auth, JSON-RPC 2.0 bidirectional protocol, heartbeat, buffered event queue, and per-group authorization
• New container/agent-runner/src/ws-client.ts: container-side WS client with automatic reconnection and reconnect guards
• src/container-runner.ts: token lifecycle (create before spawn, revoke on exit), orphaned container cleanup, graceful shutdown (signal via WS close → wait → force-stop)
• src/group-queue.ts: sendMessage/closeStdin use WS instead of files
• Deleted src/ipc.ts (filesystem watcher and polling)
• New src/ipc-handlers/ registry: skills register IPC handlers by message type without modifying core files (mirrors channel self-registration pattern)
• Container agent-runner and MCP server use WsClient instead of file reads/writes and stdout markers; stdin piped directly (no /tmp/input.json temp file)
• All skill templates updated for WS-based API

Why:

• Eliminates polling latency (was 1s host-side, 500ms container-side)
• Removes filesystem IPC volume mounts
• Gives delivery confirmation on messages (JSON-RPC requests, not fire-and-forget)
• Heartbeat + watchdog detect dead connections and trigger reconnect
• Handler registry makes IPC extensible without touching core

Scope: 45 files changed, ~3300 insertions, ~2400 deletions. stdin remains the secure bootstrap channel for secrets, prompt, and new wsUrl + wsToken fields.

Test plan

• npm test — all existing tests pass, plus 15+ new ws-server.test.ts tests covering auth, message routing, heartbeat, token revocation, orphan cleanup, and graceful shutdown
• Send a message, verify agent receives it over WS and responds
• Send a follow-up message mid-conversation, verify it's delivered to the running container
• Kill a container mid-conversation, verify the host detects the disconnect and cleans up
• Trigger a scheduled task, verify it spawns a container and completes
• Send messages to two different groups concurrently, verify isolation (each gets its own container and responses route back correctly)
• Stop the host process while containers are running, verify graceful shutdown (containers exit cleanly within timeout)

🤖 Generated with Claude Code

[TomGranot]

WebSocket IPC is an interesting direction but replacing the entire IPC layer is a big change. Keeping this as draft makes sense — let's discuss the design more before committing to this. @gavrielc thoughts on moving away from file-based IPC?

[Andy-NanoClaw-AI]

Hey @echthesia 👋 Wow — replacing file-based IPC with an authenticated WebSocket server is a genuinely ambitious and well-thought-out proposal. The JSON-RPC 2.0 bidirectional design is elegant!

That said, the codebase has evolved significantly since this was opened, and this draft now has severe merge conflicts that would make it very difficult to apply cleanly. The current IPC architecture is also deeply integrated with the container runner and scheduler.

We're planning to close this soon, but you are absolutely welcome to open a fresh PR against main if you'd like to revive it. This kind of architectural improvement deserves a clean shot! 🙌

[echthesia]

I ended up determining that WebSockets weren't the right way to do this anyway, so closing now.

[mvgrimes]

@echthesia, I'm curious why you don't think websockets are the best method for this anymore. I've been closely watching this PR and thinking about them for a similar project.

[echthesia]

@mvgrimes The connection lifecycle management introduced a lot of complexity and potential failure modes relative to using stdio, which is both buffered and available when and only when the container is running.