A reattach stops disarming the modes a live program still holds

[karngyan]

The bug

"Scroll wheel is sending arrow keys · use PgUp/PgDn to scroll" after leaving a session idle and coming back — and no way to scroll at all.

Every backlog replay ended with emulator.stopReporting(), which clears mouse tracking and focus reporting in the local emulator. That call was aimed at revived sessions (PR #67), whose dead shell never wrote its own reset. But it fired after every replay, including a reattach to a live session whose program still holds those modes armed.

Claude Code's renderer holds the alternate screen plus full mouse tracking (?1049h ?1000h ?1002h ?1003h ?1006h). Normally a wheel tick becomes an SGR mouse report and Claude scrolls its own transcript. After the reattach clear, the emulator sits on the alt buffer (no scrollback) with tracking off — the exact state xterm.js answers by converting every wheel tick into arrow keys at the program's stdin (CoreBrowserTerminal.ts, the !buffer.hasScrollback branch). Claude sees arrow bursts, shows the hint, and scrolling stays broken until its next full redraw happens to re-assert the modes.

Evidence

Dumped the live session's ring through its holder socket. The replayed tail held an unmatched ?1049h plus tracking arms, with Claude's re-assertions landing hundreds of kilobytes apart — minutes of broken scrolling per reattach. A fresh tab attaching to the same session reproduced the stuck state instantly, which cleared the client-side leak theory: the replay tail is the program's true present state, and the client must not overrule it.

The fix

Delete the clear. A live session's replay ends at the program's present state, so the right amount of client-side settling is none. The revived case never needed the client's help: the daemon ends every revive preload with settleModes (internal/session/snapshot.go), so a dead shell's reset is already inside the replayed bytes. stopReporting and reportsPointer leave the emulator seam with their only caller — reportsPointer existed only to test the reset — along with the replaying bookkeeping that existed only to fire it.

Regression test: a replay carrying ?1003h must drain with nothing appended after it.

Testing

• pnpm vitest run — 88 files, 1837 tests pass
• pnpm run lint — clean
• TDD: regression test watched failing against the old code before the fix

🤖 Generated with Claude Code