Remote Control: slash commands (/clear, /compact, /context, /bug) route to model as plain text instead of executing
Summary
When driving a local Claude Code session via Remote Control (/rc) from the Claude mobile/desktop app, slash commands typed in the app intermittently route to the model as plain text messages instead of being executed as commands. The session never runs them, but the app sometimes shows a confirmation toast claiming they succeeded.
Environment
- Local CLI version: 2.1.173 (Claude Code), binary at
~/.local/bin/claude - Platform: macOS (Darwin 25.5.0), Apple Silicon Mac mini
- Mode: Remote Control session started with
/rc, controlled from the Claude mobile + desktop app - Auth: claude.ai (Pro/Max), full-scope login
Steps to reproduce
- Start a local interactive session on the machine and enable Remote Control (
/rc). - Connect from the Claude mobile/desktop app.
- In the app's input, type a slash command (e.g.
/clear,/compact,/context).
Expected
The command executes against the running local session (per docs, /clear, /compact, /context, /bug, etc. are supported over Remote Control as of 2.1.166).
Actual
/clear,/compact, and/contexteach arrived to the model as the literal text "/clear" / "/compact" / "/context" — they were never executed. Verified because the model still retained full prior conversation context after a supposed/clear./contextexecuted correctly once earlier in the same session (rendered the usage grid), then subsequent invocations stopped executing and routed as text — i.e. the failure is intermittent, not a hard block.- In the slash-command picker, typing
/clearautocompletes/maps to/auth, so submitting sends the wrong command. - The app sometimes displays "context cleared" / "compacting" confirmations even though the operation never reached the session (false-positive UI).
/bugfrom the app opened a feedback form, and on submit returned an error that it could not submit the report.
Impact
Context management (/clear, /compact) and bug reporting (/bug) are effectively unusable from the app on a Remote Control session. Workaround is to run commands from the local terminal directly, which defeats the purpose of remote control. The false-positive confirmation toasts are especially misleading — the user is told the operation succeeded when it did not.
This issue has 2 comments on GitHub. Read the full discussion on GitHub ↗