- Calls queued behind a start that failed get its error instead of
starting the app again; a later call retries.
- The started process is reaped on a thread of its own.
- APPIMAGE counts only when the CLI runs from that image's mount.
- The guide names allow_agent_control as part of the listen condition.
- A watcher thread that can't start leaves agents off.
- The watcher test waits for the thread to end instead of a fixed sleep.
The ArtCraft Launcher's switch (AgentSwitch, mcp_flag, launcher_allows),
the agent settings, the connection file (agent-connection.json) and the
<id>-cli mcp side that finds the open window (Live: window while the app
accepts agents, otherwise headless or the closed-app choice; launcher
starts the app beside the CLI), behind the new feature agents. No UI
toolkit and no async runtime. Same formats as craft_automation::agents
on next, which can re-export these once the branches meet.
* feat: add craft-control (control channel protocol)
Port from solidcraft: the control channel any app may share, no UI toolkit.
- Per-launch token files (32 random bytes, user-only permissions)
- Loopback server with token-based authentication
- Client that talks to the loopback server
- Request/response types (ControlRequest, ControlResponse)
- Document watch for subscribed connections waiting on changes
- Effect type for requests still running when they finish
Desktop-only (wasm32 excluded): uses getrandom for token generation.
No external dependencies beyond serde and serde_json.
* fix: update test directory names from solidcraft-control to craft-control
* Record control crate dependencies in workspace lockfile
* re-trigger CI
* chore(lock): keep the wasm-bindgen family at the version main's runners install