fix(web): clearing a running terminal keeps its screen and input modes - #36380

No code change in this round; the head is unchanged at 95ae7c8. Watched pull request #36380 until its checks and bot reviews finished: all 36 checks passed on 95ae7c8 (mergeable state clean), Macroscope approved with no findings (Correctness Check and Approvability Check both success), and CodeRabbit reported 'No actionable comments were generated' with all four of its pre-merge checks passed and no line comments. No finding rated high or above is open, so the watch was stopped.

CodeRabbit's description check, though passed, noted two things about the pull request description that are for the pipeline, not the code: the Evidence section is still a placeholder, and the 'Not in this change' bullets contain a duplicate and a contradictory line about whether the 9-byte overage after Clear is bounded (it is: at most the 9-byte erase sequence over budget until the next output, snapshot or restart). The description also says only clearOutput changed, while the diff also adds textStartOffset, terminalOutputReplay and the mobile call-site changes.

The fix as it stands: a cleared event appends an erase of screen and scrollback (ESC[H ESC[2J ESC[3J) to the client's retained output instead of resetting it, so attached emulators keep the running program's alternate screen, Kitty flags, cursor-key mode and mouse tracking; displayed text for the mobile text fallback starts after the clear; older native binaries get the full replay as initialBuffer; and Clear never trims retained output.

Fixes #36006

Reproduction on main

On origin/main 9efd960: started T3CODE_DEV_SEED=0 vp run dev in a fresh worktree, paired a headless Chromium (Playwright) web client, and opened the terminal drawer ('Toggle terminal drawer') on a new thread. In the terminal I ran a probe script that prints ESC[?1049h ESC[>7u ESC[?1;1000;1006h ESC= (alternate screen, Kitty flags 7, application cursor keys, SGR mouse), then stty raw -echo; exec cat > log, so every byte the viewer sends is logged. Clicked in the terminal and pressed ArrowUp and 'a': the program received ESC[<0;5;3M/ESC[<0;5;3m, ESC[1;1:1A/ESC[1;1:3A, and a plus ESC[97;1:3u. Then sent terminal.clear (the mobile Clear toolbar's RPC) for that thread/term-1 over the app's own /ws URL from the page; the server replied Exit Success. Repeated the click, ArrowUp and 'a' with the same program still running: the click produced nothing, ArrowUp sent ESC[A, and 'a' sent a with no release event. Also added a unit test in apps/web/src/terminal/ghostty/core.test.ts that applies a cleared event through applyTerminalAttachStreamEvent, readTerminalOutputUpdate and writeTerminalOutputUpdate on the real Ghostty WASM core; on main it fails, with alternateScreen, applicationCursorKeys and mouseTracking going from true to false and the Kitty flags from ESC[?7u to ESC[?0u.

Evidence

Evidence for #36006: real Chromium web client, 1280×800, light theme, fresh synthetic data on each commit. Enable the reported terminal modes once; click, press ArrowUp and a; invoke terminal.clear; repeat with the same program still running. The probe displays actual received bytes. Both sides work before Clear; only base loses mouse and Kitty events afterward. scenarios:prove exited 0.

Capture	Before fix	After fix
Commit	9efd960925	95ae7c8316
Before Clear
After Clear
Interaction	Before recording	After recording

Scenario source and probe.

Not shown:

• Native iOS/Android rendering and viewer reattachment with a full retained buffer were not exercised; this proof covers the reported attached web viewer.

Tests

None run locally this round (no code change). On the pull request at 95ae7c8: all 36 CI checks passed, including Test, Test Web 1 and 2, Test Server 1-8, Lint, and Build and Typecheck; Macroscope Correctness Check and Approvability Check succeeded; CodeRabbit completed with no actionable comments.

Not in this change

• TerminalManager.clear on the server still wipes its retained history, so a viewer that attaches or resubscribes (socket reconnect) after a Clear gets an empty snapshot and starts without the running program's modes; fixing that means changing server history/persistence semantics and is a separate change.
• Clear no longer resets the emulator's SGR pen or other non-text state (scroll region, cursor style); that matches the still-running program's own state and was left as is.
• Clear no longer frees the client's retained pre-clear output; it stays bounded by the existing max buffer bytes and is dropped on the next snapshot or restart.
• Older mobile native binaries that take initialBuffer from the materialized text still behave as on main (buffer becomes empty on Clear).
• TerminalManager.clear on the server still wipes its retained history, so a viewer that attaches or resubscribes after a Clear starts from an empty snapshot without the running program's modes.
• On older native binaries a history rollover (the buffer head trimmed at the retention limit) still stops the new buffer extending the old one and recreates the emulator; that is unchanged from main and unrelated to Clear.
• In the text fallback both the replay string and the displayed text are now materialized per update; the fallback is the rare no-native-view path and both are bounded by the max buffer bytes.
• Clear does not reset the emulator's SGR pen or other non-text state, and does not free the client's retained pre-clear output until the next snapshot or restart.
• On older native binaries with a full buffer, the first ordinary output after Clear (like any output on a full buffer, as on main) trims the head, so the buffer stops extending the previous one and the emulator is recreated; ordinary rollover is unchanged by this fix.
• After Clear the client buffer can exceed its byte budget by the 9-byte erase sequence until the next output, snapshot or restart.
• Clear does not reset the emulator's SGR pen or other non-text state, and does not free retained pre-clear output until it is trimmed or replaced.
• The pull request description needs tidying (placeholder Evidence section, duplicated and contradictory 'Not in this change' bullets, and it understates what the diff changes); I do not write to the forge.
• TerminalManager.clear on the server still wipes its retained history, so a viewer that attaches or reconnects after a Clear starts from an empty snapshot without the running program's modes.
• On older native binaries with a full buffer, ordinary output rollover still recreates the emulator, as on main.

Who did the work

• Written by Claude Fable 5.2, in Claude Code
• Reviewed by GPT-6.1 Sol, in Codex, and by Claude Opus 5.5, each on its own
• Proof captured by GPT-6.1 Sol and checked by Claude Opus 5.5

Run by the t3code factory in T3 Code.