[Juliusolsson05]

Problem

When a Claude prompt is paste-like but not collapsed into a placeholder, Agent Code proves it landed by finding the prompt's last 24 characters in the composer (pasteTailNeedle, pasteAbsorbedVia and pollPasteAbsorbed in src/shared/claude/pasteConfirm.ts). Claude's Ink wraps with hard: true (vendor/claude-code-src/full/ink/wrap-text.ts), so a token longer than the line (an absolute path, a URL) is cut mid-token. Collapsing whitespace turned that cut into a space the needle does not contain. The send hit its 5 s budget, rolled back, and left the prompt in the composer. This is #1113's experience on the text path.

Change

activeClaudeComposerText reconstructs the text the composer holds, undoing the TUI's wrapping. Claude draws its divider across the full width, and a composer line is one column shorter (measured on the recordings). A line that fills the composer width and whose continuation doesn't start with a space was cut mid-token, so the two join directly; every other break is a soft wrap and joins with a space. The tail is then a plain substring of that text, for both the baseline and every poll. With no divider on screen, every break is treated as soft (the old behaviour). A misread errs toward a timeout, never an early Enter.

Two earlier approaches were withdrawn in review: stripping all whitespace (it accepted missing whitespace) and a pattern with optional whitespace everywhere (it matched earlier text). Both guessed where the wrap was.

wrap-ansi becomes an explicit devDependency. The lockfile already had it hoisted with dev: true; only the root entry changes. The test loads it through a typed require.

Tests (fail-first)

• Width sweep, 20–140 columns. The chrome comes from a recorded Claude Code 2.1.278 frame (testing/fixtures/image-absorption). The composer body is re-rendered at each width with Ink's own wrap options, for a prompt ending in …/promptDelivery.ts. The old code misses 20+ widths, including the issue's 46–50.
• A half-landed paste stays unconfirmed.
• foo\nbar does not confirm on a composer showing foobar, and does confirm on foo bar.
• The recorded unrelated opencode [Image #1] composer does not confirm a pending opencode[\nImage#1].
• A second paste over an already-present, hard-wrapped tail times out instead of confirming (baseline guard).

Mutation-checked, each individually: whitespace made optional, baseline guard removed, and baseline without wrap tolerance each turn a test red. Collapse instead of the pattern fails the sweep.

Review (2 Codex + 1 Pi)

Finding	Verdict	Disposition
A: stripping accepts missing whitespace. foo\nbar confirms on foobar, and a recorded unrelated image composer confirms a pending payload, so Enter is sent early	valid, major	fixed in 696578fa (asymmetric pattern), both are tests
B: the baseline transition is untested; dropping tailAlreadyPresent or reverting the baseline comparison survived	valid, major	fixed in 696578fa, test over a hard-wrapped pre-existing tail
tsc -b failed on an untyped wrap-ansi import	valid	fixed in 6ef099f5

| A round 2: optional whitespace between ANY characters matched an earlier part of the same paste (and read the recorded [Image / #1] as Image#1) | valid, major | fixed in 04db26dc: the composer text is reconstructed (a full line joins a mid-token cut directly, any other break is a soft wrap and joins with a space); both are tests |
| Pi F3: test geometry was pane − 2; the recordings are pane − 3 with ❯ + NBSP, so a hard-wrapped line is one column short of the divider, and the threshold would never have fired on a real screen | valid, major | fixed in 73df03a6: geometry calibrated on the recordings, threshold = divider − 1, and the test fails on the old threshold |
| Pi F1 (baseline guard untested), F2 (needle length unpinned) | valid, minor | F1 covered by the baseline test; F2 accepted (24 is a distinctiveness tuning, not a contract) |

Rounds: 2 (Codex A, Codex B, Pi). Round-2 findings were fixed by the author without a third round, per the at-most-two-rounds rule. Final state: B MERGE-READY, Pi MERGE-READY, A's round-2 finding fixed and tested.

Verification

npx tsc -b clean. src/shared/claude, TileLeaf and src/providers/claude pass.
Not verified here: a live Claude pane. Claude runs Bun.wrapAnsi, not the npm wrap-ansi; both implement hard.

Fixes #1118

🤖 Generated with Claude Code