The session view combines the user/agent conversation with tool activity, approvals, diffs, files, usage, and delivery controls. Agents run on the server, so browser disconnects do not by themselves stop active work.
Start and resume
Use New session to choose a project, provider, model, permission mode, and thinking controls. Every existing Claude Code and Codex session appears in the session inbox and can be opened or resumed with its history.
The inbox groups sessions by attention and activity so a waiting approval does not disappear among recent but idle work. Star or archive sessions to keep the working set useful.
Continue in another desktop app
Claude Desktop
Use an explicit resume link to import a local YA-created Claude Code session
into Claude Desktop. Its Import CLI sessions button and /resume picker
can omit sessions created through the Claude Agent SDK, including YA sessions.
-
Wait for the current response to finish. In the session menu, choose Copy session id to copy the full Claude session UUID.
-
If YA still owns the provider process, choose Terminate in that menu. Closing the browser tab or interrupting a turn does not release the process.
-
On the Mac where the session files live and Claude Desktop is installed, open this link, replacing the placeholder with that UUID:
open 'claude://resume?session=YOUR_CLAUDE_SESSION_UUID'
Desktop imports the existing conversation even when it was absent from its
picker. Keep the normal SDK launch settings; no CLAUDE_CODE_ENTRYPOINT
override or transcript editing is needed. Continue from only one app at a time:
do not send another YA turn while Desktop owns the session.
This route was verified on macOS with Claude Desktop 2.37337.1 and Claude Code
2.1.296. Recent Claude Code CLIs also expose
claude --desktop --resume <session-id>; see Anthropic’s
Desktop handoff documentation.
ChatGPT desktop
ChatGPT desktop can automatically import YA-created Claude Code sessions from local Claude history. With automatic updates configured, no per-session import action in YA is needed. If the sessions already appear, you can open them there directly.
If they do not appear, check Settings → Import and its automatic updates setting. OpenAI documents importing supported recent Claude Code history, currently chats from the last 30 days; see Import from another agent. This imports history into ChatGPT chats; it is separate from Claude Desktop continuing the original Claude session.
Choose the transcript detail
The full activity transcript preserves the conversation, every tool call, and thinking when the provider makes it available. Expand individual tool and thinking blocks when you need their detail.
Use Conversation view to condense routine activity into a calmer reading flow. Collapsing a block or switching views changes only the presentation; the underlying session detail remains available in the full activity transcript.
Approvals and permission modes
When a provider asks to use a guarded tool, the session shows the proposed action and available choices. Review paths, commands, and unsafe Unicode before approving. Rejection returns control to the agent without pretending the tool ran.
Permission modes are provider-dependent. More permissive modes reduce prompts and increase risk; use them only in a project and environment where the agent is allowed to make those changes.
Steer, queue, and interrupt
- Steer sends guidance into a compatible active turn.
- Queue preserves a follow-up for delivery after the current turn reaches a safe boundary.
- Interrupt asks the provider to stop the active turn.
The UI reflects provider differences. A missing steering control does not mean the app will silently emulate mid-turn delivery. Queued text is delivered verbatim by default.
Use Project Queue when work should wait for the entire project—not merely the current session—to become quiet.
Search, recaps, and activity
Session search helps locate older work. Recaps summarize bounded activity when you return to long-running sessions. The global activity view shows what active agents are doing across sessions without requiring every transcript to remain open.
Fork and clone
Conversation forking and cloning are experimental and provider-dependent. A clone copies through the latest completed response; a boundary fork copies a completed prefix. The source session remains available. Actions are disabled when a safe completed boundary does not exist rather than copying a partial provider turn.
Attachments
Attach screenshots, photos, PDFs, or code files from the composer. Uploads are staged with the draft and materialized for the target session when sent. Review the destination session before sending sensitive material.