[Lakr233]

What was wrong

Every attach handed back the daemon's raw output buffer, trimmed from the front. A terminal fed it:

• began mid-sequence (a cut CSI printed its parameters, a cut character printed U+FFFD);
• lost every mode set before the cut — alternate screen, mouse reporting, bracketed paste, kitty keyboard flags — while the program still expected them;
• answered the old questions in it (DA1, CPR, XTVERSION, OSC 11, DECRQM…) into the program's input, and replayed clipboard writes and notifications;
• drew a ZMODEM transfer whose start had been trimmed as a screen of binary — the garbage after sz.

All four reproduced on a guest running 1.6.20.

What changed

• Rebuilt replay (ighostvtd-io): SessionOutputLog keeps the buffer with its stream offsets and the terminal state at its start, read once by TerminalStreamTracker (Ghostty's parser at the pinned commit). A replay writes that state back, starts at a sequence boundary, and leaves out questions, OSC 52, notifications, ENQ, answered kitty graphics and ZMODEM conversations.
• Continuation: an attach may name how much of the output the client holds (outputOffset); the reply is then the rest, unchanged (outputResumed). The tab keeps its position across transports, so a dropped link resumes with no gap and nothing printed twice. No reconnecting line is written into the terminal any more (the pill says it).
• Backpressure: the app pauses its transport while the terminal has more than 512 KiB unparsed, instead of letting the library drop the oldest bytes of a detached surface.

Both new fields are optional both ways, so every patch of the 1.6 line keeps talking to every other.

Tested

• make test: all suites, including the new Tests/ReplayStream (a fresh terminal and one that already showed the session both end in the whole stream's state, over random streams), and PTY harness checks on a real session and over the proxy wire.
• make deb, make ipa.
• jp-01 iOS 26.4 / 27.0.1 and jp-02 iPadOS 26.4 / 27.0.1: no old question answered, every mode back (DECRQM against the real surface), a trimmed transfer leaves a clean screen, a cut flooding tab reattaches through a rebuild.

🤖 Generated with Claude Code