Repository navigation
Claude paste confirmation misses CJK prompts that wrap (hard-cut rule counts UTF-16, not cells) #1292
Description
src/shared/claude/pasteConfirm.ts:176 detects a mid-word hard wrap with previous.length >= width - 1. translateToString returns one UTF-16 unit per 2-cell CJK glyph, so a full CJK row is about half the width in length. The wrap is then joined with a space, and the 24-character tail needle misses. The result is a 5 s absorption timeout, a rollback, and "not sent", for roughly half of multi-row CJK prompts. Mixed rows ending in a long URL or path fail the same way. The width 20-140 fixture test uses Latin text only. Fix: measure cell width (wcwidth), not string length. Severity P2 (related to closed #1118).
Found by the Stage 3 C1 hunt; lines are origin/main d525d45 and pinned submodules. Findings: temp/quality-loop/hunt-c1.md.
🤖 Generated with Claude Code
Activity
fix(claude): only a full row is a hard cut (steering q34)
Juliusolsson05 commentedon Sep 25, 2026
Status after #1310 (merged via batch #1318, 2026-09-25).
Shipped: a wrap is judged a hard cut only when the row is full (width − 1 cells). Widths are measured per grapheme cluster and always rounded down when uncertain; flags count 2, and text-presentation pictographs and zero-width spaces count as they render. Wrapped CJK prompts now confirm at every pane width whose rows fill completely, and nothing confirms before the paste lands.
Still open: a multi-row prompt of only wide characters can still miss at odd composer widths (43 of 121 widths for the test prompt, pinned in the tests). That costs a 5 s timeout and rollback, never an early confirmation. Fixing it needs a signal stronger than row text, such as xterm's
isWrappedflag.Juliusolsson05 commentedon Sep 27, 2026
Triage (W1, 2026-09-26): the remaining case needs evidence and a design call before any code.
testing/fixtures/image-absorption). How Claude's Ink actually lays out an all-wide-character prompt at an odd composer width has never been captured.isWrappedis not the stronger signal. Ink draws each row with its own line break (wrap-text.ts,hard: true) instead of relying on terminal autowrap, so a wrapped Ink row is not flaggedisWrapped.Needed: a raw PTY recording of a Claude composer holding a multi-row, all-CJK prompt at a few odd composer widths (for example 41, 63 and 81 columns), plus a decision on whether a payload-guided join is acceptable. Until then the residual costs a 5 s timeout and rollback, never an early Enter.