fix(codex): strip alt-screen + scrollback-erase from the codex byte stream

[aakhter]

Problem

When running a Codex session, the conversation history disappears (and becomes unscrollable) on every tab switch / pane refresh.

Codex's TUI emits, during startup and on each repaint:

• alternate-screen toggles — DECSET/DECRST 47 / 1047 / 1049
• scrollback-erase — CSI 3 J
• mouse-tracking enables — ?1000–1007

xterm.js faithfully obeys these: it switches to the alternate buffer (which has no scrollback), wipes the saved lines, and forwards the scroll wheel to the codex process. The net effect is that the user's prior conversation/shell output both vanishes and — even when still present in the buffer — can't be scrolled back to.

Fix

Strip just those sequences from the codex byte stream in two places, leaving the visible-viewport erases (2J / J) intact so codex can still repaint its own rows:

• Session._handleTerminalOutput — filters the live SSE/WS stream and the persisted terminal buffer at the source, for mode === 'codex'.
• GET /api/sessions/:id/terminal — applies the same strip (ALT_SCREEN_TOGGLE_PATTERN / ERASE_SCROLLBACK_PATTERN) to the replayed buffer, so a tab-switch replay keeps full scrollback. Mirrors the live-stream strip.

Only mode === 'codex' is affected; Claude/OpenCode/shell paths are untouched.

Tests

Adds test/codex-terminal-output.test.ts:

• alt-screen toggles and 3J are removed; 2J / J (visible-viewport erases) are preserved
• a ?1049h … 2J … 3J … ?1049l redraw collapses to just the visible content
• Ctrl+L redraws and native codex prompt/status/resume-picker redraws pass through without row-repair mangling

npm run test:ci passes locally (2735 passed, 0 failures); tsc --noEmit, eslint, prettier --check, and the frontend-syntax check are all clean.

[Ark0N]

Reviewed. The approach is right — strip at the source so neither the persisted buffer nor the live stream carries the scrollback-killing sequences, with the replay strip as backstop — and the 13 captured-stream tests are genuinely good coverage. Two gaps, both fixed in 7eda39e pushed to this branch:

1. Sequences split across PTY chunk boundaries slipped past the live strip._handleTerminalOutput filters per chunk, but chunk boundaries are arbitrary: \x1b[?1049h can arrive as \x1b[?104 + 9h. xterm's parser is streaming, so it reassembles and obeys the leaked toggle — and since the matching ?1049l (arriving intact) does get stripped, the terminal is left stuck in the scrollback-less alt buffer until the next tab-switch replay: exactly the bug this PR fixes, recurring intermittently. At ~8 bytes per sequence against ~4KB chunk boundaries the per-occurrence odds are small, but codex re-emits these on every repaint, so over a long session a leak is near-certain. Fixed with a small carry: a trailing partial digit-only CSI intro (≤7 chars) is held back and prepended to the next chunk; complete sequences are never held, and the carry resets in _resetBuffers. Two tests added.

2. Replay didn't strip mouse-tracking enables. The live strip removes ?1000–?1007, but the GET /terminal replay only stripped alt-screen + 3J — so a codex buffer persisted before this PR still re-hijacks the scroll wheel on every tab switch. Added MOUSE_TRACKING_PATTERN to the replay strip for parity.

Verified in a clean worktree: tsc, eslint, prettier --check clean; full test:ci green (2737 passed, +2 new tests).