feat(tui): highlight past user inputs with the original userMessageBackground band

[ericleepi314]

Problem

After a few turns, previous user inputs are hard to find in the transcript — they rendered as bare ❯ text, visually identical in weight to assistant prose. The original Claude Code draws every past user input on a full-row background band, which is what makes prompts scannable when scrolling (user-supplied reference screenshots).

Change (faithful port from the original source)

• theme.ts — new userMessageBackground token: dark rgb(55,55,55) / light rgb(240,240,240), the exact original utils/theme.ts values. Skinnable via user_message_bg, DEFAULT_THEME fallback. Also ports the original selection blues (dark rgb(38,79,120) / light rgb(180,213,255)): the previous dark selectionBg #373737 was byte-identical to the new band, so selecting text on a past user row would have been invisible.
• messageLine.tsx — transcriptRowBand(msg, t) paints the inner row Box for user rows and slash echoes (UserPromptMessage.tsx:76 / UserCommandMessage.tsx:62 parity). Multi-line inputs band every row as one solid block (continuation indent and empty lines included — verified per-cell). Slash echoes borrow the user pointer + text color (previously muted under the system · dot). The user gutter glyph drops its non-original bold — the band, not bold text, carries the emphasis.
• appLayout.tsx / useMainApp.ts / lib/virtualHeights.ts / domain/blockLayout.ts — the ─── inter-turn dash becomes a monochrome-only fallback: it renders (with its 2-row height accounting) only when color is disabled, via the shared showsInterTurnSeparator(...) predicate fed from TRANSCRIPT_COLOR (process.stdout.hasColors). With color available the band replaces it — the original has no dash. Render gate and height estimator share one predicate, so virtualization stays in lockstep.
• lib/forceTruecolor.ts — the bundled chalk predates NO_COLOR (verified: level 2–3 on a PTY with NO_COLOR=1), so the pre-chalk bootstrap now translates NO_COLOR → FORCE_COLOR=0 (the channel chalk honors) unless the user set FORCE_COLOR explicitly. The app is now genuinely monochrome under NO_COLOR, and the fallback signal agrees with what the renderer emits.
• diffView.tsx — structuredDiffSupported() now gates the raw-ANSI ColorDiff path on TRANSCRIPT_COLOR too (ColorDiff builds SGRs without chalk, so FORCE_COLOR=0/TERM=dumb sessions previously kept colored diffs inside an otherwise monochrome transcript; they now fall back to the chalk-suppressed markdown path).
• Tests — band routing, slash pointer swap, theme value pins + skin fallback/override, selection≠band distinctness guard, separator gate matrix, band-mode same-height + monochrome +2 heights, NO_COLOR→FORCE_COLOR=0 translation + explicit-FORCE_COLOR-wins, structured-diff no-color gating; refreshed one stale pre-existing expectation.

Adversarial review loop (all catches fixed in this PR)

1. Codex /codex:adversarial-review Add initial project source, docs, tests, and config #1 — removing the separator unconditionally left no-color terminals with no turn boundary → restored as the monochrome fallback.
2. Critic — TRANSCRIPT_COLOR diverged from the renderer's chalk under NO_COLOR (band + dash double emphasis) → root-caused to chalk predating NO_COLOR; fixed via the bootstrap translation; verified on the full env matrix incl. NO_COLOR+COLORTERM=truecolor.
3. Critic — dark selectionBg == band → original selection blues ported + regression test.
4. Codex verification pass — structured diffs still colored under FORCE_COLOR=0/TERM=dumb → structuredDiffSupported() aligned to the same signals + tests.

Final verdicts: Critic APPROVE; Codex separator no-ship lifted, its remaining recommendation implemented verbatim.

Evidence

📸 Visual captures (before / after / multiline+slash / NO_COLOR): https://claude.ai/code/artifact/09bee85d-b365-4258-a84d-13a9f3bb5378

pyte PTY harness (fake NDJSON agent-server, per-cell bg audit):

surface	before	after
past user row	bg default×100	bg 373737 across the row
multi-line input (incl. empty line)	no band	every row banded, one solid block
slash echo	muted · system style	❯ + text color on band
assistant/system/composer	no band	no band (unchanged)
color mode separators	─── above turn 2+	none (band replaces)
NO_COLOR (± COLORTERM=truecolor)	full color anyway (chalk bug)	genuinely monochrome + ─── fallback

• ui-tui vitest: 8 failed = main baseline exactly (fixed one stale test, zero new failures). tsc --noEmit clean; eslint clean on touched files.
• Known pre-existing artifact (reproduced on the unmodified build): inline growth repaint can leave stale right-edge cells from shifted footer/rule rows at some geometries. Renderer follow-up, out of scope.
• Future parity candidates: bash ! input band (bashMessageBackgroundColor), memory # input highlight, queued-prompt band reuse.

🤖 Generated with Claude Code