[Edurique]

What

Follows Claude Code sessions from the Windows island, and lets the two things Claude waits on — a question, a permission — be answered where they are seen. The island stays a companion: everything done from it goes through Claude Code's own hooks and shows in Claude Code, whether the session runs in the Claude desktop app, VS Code or a terminal.

This is stacked on #57 (the GitHub panel): it reuses its panel layout, its diff and its syntax colours. Until #57 is merged, GitHub shows its commits here too; the ones that are this pull request are the last eleven. Without the panel it is about 4 200 lines in 28 files.

Windows only. Nothing under NotchBuddy/ is touched. As with the panel, the aim is to port it to Swift if you like where it goes.

Issues this closes

• Closes Multi-session support: track multiple Claude Code terminals simultaneously #18 — Multi-session support. Every Claude Code session is followed on its own (keyed by its session_id), a card names the conversation its request comes from, each session's request is answered on its own, and requests made at the same time wait their turn. The list of sessions shows where each one is at.

• Closes Overlapping Text on Windows #39 — Overlapping text on Windows. The screenshot there is two rows of the overview's ticker on one line. The ticker is replaced by the session's card, so there is nothing left to overlap.

• Closes Answer AskUserQuestion from the island (not only Allow / Deny) #94 — Answer AskUserQuestion from the island. The card shows the question and its options (one choice or several, several questions in a row) and the chosen answer goes back to Claude Code, so nothing is asked again in the terminal.
  All three on Windows. They are linked so that merging this closes them; if you would rather close them yourself, say so and I will turn these into plain mentions.

What it looks like

The session's card

As the prototype draws it: the step under way by its icon and its name, with the file or the command it is at, and a look at what it did — the lines of a file read, a command with the end of what it printed.

Once the turn is over: Done, and the first lines of what Claude replied.

The session panel

A click on the card. Mochi's column with the last steps, and the session's journal, read like a conversation in the order things happened: what was asked, each file read, each edit with its diff typed as it lands, each command with the end of what it printed.

Questions and permissions are in the journal too — the options with the ones picked, a tool that had to ask with what it was told — and it ends on Claude's reply, as it wrote it.

Behind the journal, every file the session changed, each with its whole diff.

Answering a question

A question from Claude's question tool, on the island: each option with what it means. One click answers.

Several to pick, then Send; several questions in a row, with a way back; an answer typed with Other…; Skip, or handing it back to Claude Code.

Permissions

A permission request for an edit shows the diff it would make before Allow. A command shows the command.

A turn's end

The finished card says what Claude replied, with the way to the whole reply and to what was changed. An error says so, with the way to where the session runs.

Several sessions at once

Each is followed on its own, and one is in front. With more than one, the session's card says how many there are; that chip opens the list of them, and a click puts one in front. The island's bar is untouched.

A request made while another is on the card waits its turn: the card says how many are waiting, and the next comes forward once this one is answered.

All of these are the real island playing a made-up session through its own hook handler (dev/claude-preview.html), not mock-ups.

How to read it

Eleven commits on top of #57 and a merge of main, each of which builds:

Commit	What	Where
1	The relay carries what a session does, and takes an answer back	hook/src/main.rs
2	Follow sessions from the island, and answer them	src/island/hooks.ts, src/views/session.ts, step.ts, markdown.ts, views.ts, core/state.ts, style.css, and pipe.rs / lib.rs for the answer
3	A preview page	dev/claude-preview.*
4	README and screenshots	windows/README.md, windows/screenshots/
5–11	Follow-ups from using it: the screenshots redone, a least air above and below the lines of every card, and the other sessions as a list behind a chip on the session's card	views.ts, session.ts, style.css

How it works

• Everything comes from the hooks Claude Code already sends. Nothing is asked of Anthropic and nothing leaves the machine.
• The relay forwards a few lines of a tool's response instead of dropping it whole: an edit's diff (from the structuredPatch Claude Code puts in the response), the first lines of a file read, the last lines a command printed, what a search found, the answers to a question. Forty lines at most.
• A question is answered through the PermissionRequest hook: allowing the question tool with updatedInput set to its own input plus answers. The session goes on as if the answer had been picked in Claude Code. Skip is a deny with a message saying so.
• Two things are read from disk by the relay: the file an edit asks permission for, to show its diff (never written), and the end of the session's transcript at a turn's ends, for the conversation's title and Claude's last message.
• Your third-party agents (coucou_agent) are untouched: an event tagged for an agent goes to that agent's pill, in a handler of its own with your logic; everything else is a Claude Code session.

How to try it

cd windows
npm install
npm run dev

then open http://127.0.0.1:1420/dev/claude-preview.html. A few addresses:

• ?view=session&live — the journal, an edit typed every few seconds
• ?view=session&asked — a question and a permission in the journal
• ?view=question — the question card (&many for three questions, &multi to pick several)
• ?view=approval&diff — a permission request with its diff
• ?view=overview&sessions — three sessions at once (the "3 sessions" chip on the card)

With Claude Code: npm run tauri dev, and ask Claude something that reads a file, edits one, runs a command and asks a question.

The rules of the house

• No new dependency: Cargo.toml, Cargo.lock, package.json and package-lock.json are unchanged.
• Never block Claude Code: still only a permission request waits for a human, with the same ack and timeout as before. A paused island, a session it does not follow, or a second request from the same session is declined at once, and Claude Code asks in its own window.
• Never approve without an explicit click: an answer, a skip, an allow and a deny are each a click on the island.
• ~/.claude/settings.json is not touched: the hooks installed today are enough.
• No telemetry, no network: nothing here makes a request.
• Nothing sensitive in the log: pipe.rs logs the event's name and the first word of a reply, never a diff, an output or an answer.
• 0 % CPU when hidden: the journal is drawn only while the panel is on screen; its animations run once and stop, and are paused with the rest when the island is folded (Windows: nothing in the views moves while the island is folded or hidden #56).
• Memory: a session keeps its last 80 journal lines and a file its last 12 edits; four sessions are followed at most; all of it goes with the session.
• Build: tsc --noEmit, vite build, cargo check pass; cargo test 70 passed (12 in the relay, 10 of them new); cargo clippy shows nothing new.
• Docs: a Claude Code section in windows/README.md. I left CHANGELOG.md to you.

Things you should know before merging

• The overview's ticker is gone on Windows. A session's card is the prototype's step-and-preview instead, and ticker.ts is removed. It is the part most likely to be a matter of taste.
• The Claude pill mirrors the session in front. State.sessions holds every session; the integration_claude task takes the front one's project, state and steps, so Mochi, the mini grid and the cards work as before.
• The transcript's format is Claude Code's own and undocumented. If it changes, the title and the last message simply stop showing; nothing else depends on it.
• Sessions started by the Agent SDK are ignored (CLAUDE_CODE_ENTRYPOINT starting with sdk): a plugin that runs a review on each commit sends the same hooks from the same folder and was taking the island away from the conversation in hand.
• The fallback step labels are in English ("Runs · cargo test") where they were the Mac's French ones, like the rest of the Windows island.

What I did not check

• Only the Claude desktop app was used for real. Nothing in the code is specific to it, but I have not run a session from a terminal or from the VS Code extension against this build.
• Answering a question from the island was done for real many times (one choice, several, two questions in a row). Skip, and the diff on a permission request, were only exercised on the preview's data.
• Several sessions at once, and a request waiting its turn, were exercised on the preview and in daily use with two conversations, not with four.
• Display scales other than 100 % and 125 %.