Token usage in headless output; voice mode for API keys
v0.2.97
3 features4 enhancements7 fixesThis release3 featuresNew capabilities4 enhancementsImprovements to existing features7 fixesBug fixesAI-tallied from the release notes
From the original release noteView original ↗
- Headless JSON output now includes token usage and cost per prompt and session
- SDK turns now expose detailed token usage and cost information via Turn.usage
- Double-click or Enter on a previous user message now lets you edit and resubmit it directly from the transcript
- Text selection in scrollback now works better when starting on chrome, gaps, or while scrolling
- Shell commands using
rg(ripgrep) no longer require permission prompts by default - Voice mode is now available for API-key sessions
- New environment variables allow tuning scroll and draw cadence for high-refresh displays
- Background tasks started by the model in headless mode are now killed on exit instead of leaking
- Agent process leaks on failed spawns and missing-stdio teardown are now prevented
- Parked turn markers no longer appear after interjections and now count down as background tasks finish
- The
/contexttool definitions line no longer shows the cryptic disclaimer suffix - Terminal release waits boundedly for the process to exit and repeated waits share the drain grace
- Fixed a crash when the agent asked a question on narrow terminals
- Fixed misleading keyboard hints in the
/mcpspanel in minimal mode - Clipboard copy now reports success correctly when using iTerm2 over SSH
- Fixed scroll/input conflicts when plan approval appeared over an open edit block
- Fixed frequent MCP and skills reloads that could freeze sessions on devboxes
- MCP servers using HTTP (such as nebula for Slack) now automatically recover from disconnects
- Next reset time in
/usagenow shows in your local timezone instead of Pacific Time
Fetched July 12, 2026
