AIToday
Large Language ModelsAI Coding AssistantsZenn AI/MLPublished: Nov 17, 2026, 16:00 JST

Claude Code tab handoffs moved to inbox signals

Send on LINE
Claude Code tab handoffs moved to inbox signals

3 Key Points

  1. What happened

    The developer switched only his own placed notes (user_note) to the inbox path; completion reports still travel the old paste route while a shadow copy records how they would have arrived.

  2. Why it matters

    The inbox is a file, so the body survives even if the one-line signal never reaches the tab — but routing and marking now ride on separate switches, a gap his own bug exposed.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

Context & Analysis

The problem this record describes is not Claude Code itself but the delivery layer around it. Sending a request or a completion report by pasting it into another tab's input field is simple, yet on September 9, 2026 a 9.7KB order was reported sent even though the input field stayed empty — the send script said sent, the hub log held the body, and the API returned 200. The cause was a wait loop that broke silently when the PC's CPU was saturated. A follow-up on September 18 checked arrival by looking at the other tab's screen and later exited with code 8 when arrival could not be confirmed, but that only adds checking on top of the paste route.

The redesign borrows from Anthropic's "Building always-on agents" by William Hu (2026-11-13). The idea is that the inbox, not the agent, is written to first; the tab receives only a one-line notice such as "📬 n 件" and reads the body with a take command. Items live in state/inbox/<tab>.json, move to state/inbox_log.jsonl when processed, and duplicate keys within 24 hours return the earlier id.

The record is candid about the risks of this arrangement. A Stop hook that keeps a session going is capped at 5 items and 8,000 characters per run and 6 times per hour per session, and the hub errs toward letting work pass when it cannot connect or does not answer within 4 seconds. The bug he hit himself — a note placed at 08:46 delivered again 30 minutes later — came from tying the processed mark to the hook switch while the tab had only the signal switch on. That is why the take path and the mark path are now described as something that must always line up.

FAQ
Why not use Claude Code's official tab-to-tab messaging (SendMessage)?
He rejected it as the main line because the destination name is auto-assigned and changes when a tab is relaunched, and because delivery and read records do not remain in his hub or appear on his screen. He keeps it only as an escape route for when the hub is down.
How does the new inbox avoid processing the same item twice?
Each item carries a dedupe key, and if the same key arrives within 24 hours the inbox skips adding it and returns the previous id. Items taken but not marked processed return to new if another session takes them, if the owner is not the current session, or after 30 minutes; after 3 returns it stops and notifies him.
What did the shadow measurement show?
At 12:29 on the same day, shadow-stats covered 4.1 hours and 11 completion reports addressed to main; 10 arrived while main was idle, and the one that arrived mid-generation took 27.1 seconds to be re-sent and pasted on the old path. He notes one item is too few to conclude anything.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleApple narrows macOS full disk access to selected folders in macOS 27.3 update