Roadmap and limits

The full plan is in docs/PLAN.md.

Milestones

M0, PTY layer: done. Unix verified by tests; Windows compiles and needs a manual run.
M1, agent API: done. Create, list, kill, attach, scrollback replay, token auth.
M2, hub: mostly done. hosts, ls, new, attach, kill work. Still to come: an interactive TUI with a session list you can attach to and return from.
Web UI: done. clawsh web serves a browser dashboard with a full-page terminal (loopback only). Next: LAN access (#6), editing hosts from the page (#8).
Auto-agent plugin: proof of concept. The Claude Code plugin starts an agent if none runs and lists sessions started outside clawsh as external. Linux verified; Windows and macOS need a manual check.
M3, status and alerts: next. Claude Code hooks will report working, idle and needs_input; the hub will show badges and ring the bell when a session needs you.
M4, hardening: agent as a system service, hub auto-reconnect, Windows Job Objects, a resize policy for multiple viewers.

Later: split-pane views, mDNS discovery, TLS or Tailscale, sessions that survive agent restarts.

Only sessions started through clawsh are managed. A claude launched in a plain terminal shows up in ls as external when the plugin is installed, but can't be attached to or killed.
Sessions stop when the agent stops. claude --resume recovers the conversation.
Status is running until the first hook arrives; hooks are not installed automatically yet (clawsh hook exists, install-hooks is a separate issue).
On Windows, killing a session doesn't yet kill claude's child processes.
With several viewers, the last resize wins.
On Windows, typing into a session right after it exits can misbehave in the agent. Tracked in issue #12.