The Cat, from your side

“I want to talk to the Cat.”

Enable the installed Insider App and open Projects. Choose a saved Cat, or select New Cat, choose a workspace, and send your first typed message. That message creates the Cat using the active OpenHands Agent Profile.

Select Talk to start Voice. Canvas checks availability before requesting microphone access. If permission is denied, allow the microphone in the browser and retry; you can keep typing meanwhile. Installation never starts the microphone.

The Cat rests when idle, listens with ears up, and stays awake while speaking or working. Text beside it explains connection problems, approvals, and other states that need attention.

“Let’s pick up where we left off.”

Choose the saved Cat in Projects. Open chat opens the same conversation in regular Canvas, where you can read tool activity and respond to approvals. Cat badges distinguish it in chat and the conversation panel, including the phone layout.

Moving between Projects and regular views keeps an active call attached to that Cat. Opening a different work conversation does not retarget the call. Ending the call keeps submitted messages and results saved; accepted agent work can finish after you leave.

A new call loads a bounded snapshot of the Cat’s saved context. Live captions and exact audio are not a permanent transcript. Open chat for the complete saved requests and agent results.

“How many conversations are in this backend?”

Ask the Cat to check. With the normal tools, enabled OpenHands API skill, and correct backend connection information, it can query Agent Server through its terminal. The backend’s conversation-count endpoint gives the total for that query; detailed listings require pagination.

Your work is a browser of loaded conversations. Its filters, Refresh, and Load more help you explore, but counting visible cards does not establish the full backend total.

If the Cat says it only has reasoning and finish tools, inspect its saved configuration. Choosing another LLM or starting Voice does not add tools. A new Cat created with a configured Agent Profile gets that profile’s tools and enabled skills.

“Help me with this conversation.”

Find the conversation under Your work and choose Discuss with Cat. Its identity and metadata accompany your next typed Cat message. Your draft stays intact. Selecting the card alone does not send a message to that worker or start work there.

The Cat can inspect or act through its configured tools and the backend API, subject to normal permissions and approvals. The full conversation shows which actions actually happened.

For Codex Voice, the board selection is not automatically included in the call. Put the target into a saved typed message before asking about it by Voice. Merely looking at a conversation elsewhere in Canvas does not select it for the Cat.

“Keep my message while I look around.”

An unsent Cat draft survives navigation within the active App, including selecting work to discuss. Sending remains an explicit action. A reload or backend change does not preserve that unsent draft; saved conversation history and unsent text have different lifetimes.

“Use your tools to investigate.”

Voice hands the request to the saved OpenHands Cat. That agent uses its normal model, tools, skills, and approval policy; Voice then speaks the saved result or a shorter excerpt. You can open chat to inspect the full answer and tool activity.

If the agent needs approval, review the action in full chat. Voice does not approve it for you. A spoken acknowledgment alone is not proof that the request reached OpenHands or that work finished.

The Voice guide explains the direct Codex handoff and the separate OpenAI API bridge.

“Wait, I meant something else.”

Stop speaking pauses local audio; Resume audio restores it. Mute controls the microphone. End call closes Voice. These controls do not cancel accepted OpenHands work; on Codex, pausing playback also does not cancel the provider’s generation.

If a new spoken request arrives while the Cat is still running, it is not queued or used to redirect that task. Voice reports that the new request was not sent. Wait for the task to settle, then repeat the correction. Inspect uncertain results in full chat before retrying an action.

The bridge suppresses an older pending spoken answer after a newer utterance. Audio already delivered cannot be recalled, and the original accepted work remains in the saved conversation.

“This conversation is getting long.”

Use /condense in Projects or regular chat to shorten the active context and keep the same Cat. Running tasks must finish and pending approvals must be resolved first. The command ends Voice before compaction; start Talk again afterward to use the refreshed context.

Condensation retains the saved event history and does not start another tool turn. If you want a separate subject and history, choose New Cat or use /new. The first typed send creates the new Cat; the previous one remains available in the picker.

“I’ll check the result later.”

End the call and return to the same Cat later. Work already accepted by OpenHands can complete and save its result. The App does not install a background scheduler or promise a notification after you close Canvas.