BoxDo Terminal 0.1.7

One shared pipeline behind the chat: what you see while an agent works is now the same quality on every agent, not just on Claude Code.

Fixed

• A command's output no longer disappears from its row. The output and exit code now land in the command's chat row, instead of the row looking like the command printed nothing.
• Rows went missing on Grok Build sessions. Every MCP call now gets its own row — several different calls used to collapse into one — and a terminal command shows the actual command line instead of a generic tool name.
• The permission-mode chip could describe a session wrongly. It was inferred from the screen, so a session running without permission prompts could be shown as one that will ask for them. It now comes from the agent's own signal, and no longer flips between two names for the same mode.

Improved

• Gemini CLI sessions show the agent's thinking, with their own timestamps and including thought lines that previously printed nothing — plus a model badge, which Gemini sessions never had.
• Grok Build sessions show thinking too, and announce the model once instead of on every reply.
• A refused action reads as refused on Kimi Code, rather than as a generic failure.
• "Needs you" is more dependable. The attention state now follows an official signal from the agent, with screen reading kept only as a fallback for the few things no agent publishes.
• Under-the-hood reliability work on the app's internal message queue.