Claude Code TUI corrupts in narrow panes (Ink renderer, not Zellij)

S5 · Instance observed 2026-04-23

The pattern

Reattaching a Claude Code session from a phone (via the cross-device SSH + Zellij pattern) produces a ~36-column pane. At that width, the Ink TUI fragments: tool-call labels wrap at the wrong boundaries, table column headers break apart, and individual words split mid-letter across successive lines. Easy to mistake for “phone terminal weirdness” — it’s actually a specific, well-reported, unfixed upstream Claude Code bug. Worth capturing so re-investigation isn’t needed the next time anyone hits it from a narrow pane.

Observed corruption pattern

Environment: Claude Code v2.1.118, Zellij 0.43.1, WSL2 on Windows, TERM=xterm-256color. Narrow pane via phone-over-Tailscale-SSH reattach (see cross-device SSH pattern).
Symptoms: mid-word splits in pconv list session-id columns; Bash(pconv list --workspace-toml auto tool-call labels wrapping mid-argument; gh auth status output splitting long words (account names, usernames) across three lines each — one to three characters per line.
Reproduces in tmux. Confirmed via upstream issue evidence — not a Zellij-specific bug.

Diagnosis — it’s Ink + Claude Code, not the multiplexer

Multiple converging upstream issues on anthropics/claude-code:

Issue	Status	Relevance
#42010	open, v2.1.88	Deep RCA of Ink rendering corruption: DECSTBM scroll mutating prev-frame buffer, style-pool cache collisions, wrapped-text style-segment desync. Fingerprint: “resize fixes it.” Exact match.
#43113	open	Feature request confirming Claude Code emits hard newlines at word boundaries instead of letting the terminal wrap. Directly explains the `cybersad\ner` split.
#22222	closed	Context indicator wraps character-by-character in narrow terminals. Same per-character shattering pattern.
#42231	closed, v2.1.89	`!` shell escape reports `COLUMNS=80` regardless of actual PTY width. Proves Claude Code hardcodes widths in some code paths.
#21037	closed (dup of #20825)	Multiple-choice prompts don’t line-break correctly in narrow tmux panes.

Zellij tracker is clean — zero matching issues on zellij-org/zellij. The one unrelated hit (zellij#4390) is about Claude Code’s /ide detection, not rendering.

Probable regression window

v2.1.113 switched Claude Code’s CLI to a native per-platform binary (large render-path change). The class of Ink bugs existed before this (#21037 lands at v2.1.19) but the native-binary transition is the most recent disturbance to the render pipeline and the most likely point where phone-narrow-pane usage started failing visibly. Not traceable to a specific changelog entry — the changelog doesn’t call out Ink changes.

Current: v2.1.118. No fix yet.

Workarounds (by cost)

Workaround	Effect	Caveat
`Alt + f` (Zellij pane fullscreen zoom)	Forces a repaint at full width. Most-cited fix.	Needs un-zoom to see other panes.
Drag to resize the pane	Triggers SIGWINCH → full repaint	Same as above, works in tmux too
Keep pane ≥ 80 columns	Avoids the hardcoded-80 threshold in #42231	Not practical on a phone in portrait
`claude -p "..."` (print mode, non-interactive)	Bypasses the Ink renderer entirely	Single-turn only; no interactive tooling

What doesn’t help (don’t waste time)

CLAUDE_CODE_NO_FLICKER=1 — tried by others in #46898, didn’t fix.
export COLUMNS=120 before launching — Claude Code ignores the env in some paths (#42231 proves this).
Wider Zellij config / forced pane minimums — Claude Code is the one mis-detecting; forcing the layout doesn’t reach the hardcoded-80 code paths.

Why low priority

Pure UX annoyance; no data loss, no silent corruption of anything persistent.
Single-keystroke workaround in Zellij (Alt + f).
Multiple open upstream issues already carry thorough evidence; adding another report would be noise.
Narrow-pane access (phone-reattach, small side panes) is a small fraction of typical Claude Code usage.

If it ever escalates — a layout that can’t zoom, or a regression that breaks the Alt + f workaround — revisit, repro on the latest build, and file upstream with the fresh evidence.

Promotion candidates

Stack pattern (tier 2): belongs as a “gotchas” subsection of cross-device SSH rather than its own pattern — the bug is situational to that pattern’s narrow-pane workflow, not a standalone decision tree. Fold in when the next pass through that doc lands.
Challenge (research/zz-challenges/): only if upstream drops the class of bugs without fixing the mid-word wrap. Unlikely — track record suggests one of the open issues will land a patch eventually.