An ID before a session

We opened co claude in a new workspace and immediately handed its terminal to O Chat. The browser looked ready: it had a Claude session ID from the SessionStart Hook, a paired owner, and a Take control button. Its first message failed without asking for the file edit it had been told to make. A minute later the composer said the Host had not confirmed the message.

The failure was reproducible only before the first terminal turn. After a harmless prompt in Claude's terminal, the same browser action reached the Write approval and completed. The difference was on disk: before that prompt, Claude had announced a session ID but had not written a resumable transcript. Our headless browser runner treated every ID as an existing session and passed it to --resume.

The useful boundary

Happy Coder checks the native JSONL before setting the Claude SDK's resume option. If the ID has no valid stored message, it starts a new query. We do not need to copy its SDK or permission model to use that distinction. Our Station already sees the exact native lifecycle through scoped Hooks. It now records whether the local session has completed a turn. Handover before that point starts a new native session for the first browser message. The new ID is then used for later browser turns and return to the terminal. A handover with no browser message returns to a fresh terminal instead of trying to resume the empty ID.

There was a second boundary in the OIP path. The Host allocates a worker when the browser sends a provider message, but that is not proof that Claude has started. We already waited for native init and the matching SessionStart Hook before sending a positive ACK. When native startup failed earlier, we sent no ACK at all. The browser kept its draft, correctly, but waited for the timeout. The worker now sends the matching negative ACK when no positive one was sent, and the React client says Claude could not start the turn.

What we tested

A local O Chat page in Chromium paired with a real Haiku Station in a fresh workspace. We took control before typing any terminal prompt, sent a request to Write a file, saw the approval card, allowed it once, checked the file's bytes and Claude's reply, checked the narrow mobile layout, and returned control to the terminal. The earlier failure and the fixed desktop and mobile states are preserved in the 1.8.9b7 release evidence. The focused Core suite passed 124 tests; the React suite passed 188.

This fixes the first-turn path, not every difference from Happy. Station's browser approval remains deliberately limited to verified workspace edits; Bash, MCP tools and interactive Claude questions still lack browser decisions. We will revisit those policies when we can state and test the authority they would grant. For this preview, the important claim is narrower: a browser can start a real native Claude turn from an empty terminal and knows promptly if that start fails.