fix(web): preserve single tildes in sent user messages#23204

Problem

The rich composer shows Install the skill as ~/folder-a/skills/x, then follow the guide (~/folder-a/guide.md). literally, but sending it removes the two single tildes and strikes through the intervening user-message text. The stored prompt remains intact.

Change

Configure the existing GFM parser to require double tildes for strikethrough in sent user messages, and apply the same configuration to user-message Thread Find extraction. Single tildes, including intentional ~solo~, become literal; ~~double~~ and assistant markdown keep their existing behavior.

Scope and approval

Fixes #16105 GitCafe: https://git.cafe/pingdotgg/t3code/issues/16105 Triaged bug via the automated issue flow: #16105. This addresses the reported mismatch between rich composition and sent web user messages. Regular and automation user rows share the fixed renderer. No stored text or provider protocol changes.

Native Electron is unverified. Mobile uses a separate native parser and remains affected; it needs a follow-up. Other user-authored markdown surfaces are outside this fix.

Verification

• Real app reproduction on a12d6ed942: sent the exact prompt through the web dev app, then repeated with Codex and Claude selections and a 390px viewport. The sent user bubble struck the paired paths while SQLite retained the original text. Double-tilde and code controls behaved as expected.
• Browser regression in MessagesTimeline.browser.test.tsx: failed on main because the rendered user message lost its tildes; now 10/10 browser tests pass. Controls cover double-tilde user text and unchanged assistant parsing.
• Focused shared tests in markdownPipeline.test.ts and threadFindText.test.ts: 12/12 pass. Search extraction preserves user tildes and still removes double-tilde formatting.
• Targeted lint and web/shared TypeScript checks pass. The 22 focused tests also passed after merging main.
• Real read-RPC comparison: both searchThread and searchThreadStream return zero matches before and one correct user-message match after for the raw home path, literal single-tilde text, hidden tail, automation body, and legacy-context prose. Hidden context output and structured-context labels remain excluded.

Evidence

Real web dev apps at b5a226e49c (before) and 8ac56474a3 (after), Chrome 154 on Linux, DPR 2, scale 1, frozen UTC clock. CSS viewports 1440×1000 and 390×1000; light and dark themes. Narrow captures are web, not native mobile.

1440px, light

1440px, dark

390px, light

390px, dark

Interaction recordings contain narrated, media-timed chapters for expand/collapse, find navigation, hidden-tail reveal, automation, and context exclusions. Text captions occupy added black padding; app pixels are retained.

Supplemental after recordings: automation reveal, legacy/context exclusions. Actual read-RPC transcripts: before, after.

Visible changes: user single tildes and their former strike decorations change; restored characters change bubble width and narrow wrapping, shifting later rows by 23 CSS pixels. Translucent-composer blending and bottom-row cropping follow that reflow. Desktop header, assistant, and scroll-button pixels match exactly; narrow assistant content matches after its layout shift. No CSS changed. Full-frame diffs are unaligned and unmasked.

Independent reconstruction from schema only plus the committed fixture reproduced the complete after screenshot with zero changed pixels, without masking or alignment: metrics.

Unverified: native Electron; native mobile, whose separate renderer remains affected and needs a follow-up; a real UI-generated structured terminal-context chip. The empty-record fixture proves search exclusion only. Projection fixtures prove rendering/read RPC behavior, not provider orchestration or event replay.

Full audited evidence index, reproduction bundle, and checksums. Independent audit verified all 89 checksums and 39 public artifact hashes. Both final reviewers returned MERGE.

Models and harnesses: Claude Opus 5.5 through Claude Code implemented the fix and tests; GPT-6.1 Sol through Codex coordinated triage, fixtures, and proof. Both reviewed the diff.

Takeover

Original fix by @juliusmarminge. No code or test changes; final-diff runtime evidence and independent verification added below. The original description is preserved. This update supersedes its historical mobile-unverified wording: plain mobile is verified by the maintainer's EAS iOS static fixture; no mobile fix is claimed here. No desktop-specific files changed, so Electron is outside this diff's required surface set.

Evidence

Pinned before b5a226e49c, after 8ac56474a3. Fixture matrix, reused capture index, new runtime index, independent proof audit. The table below applies the corrected file-path surface scope to the planning matrix.

Current-VM pinned-font reconstruction reference (dark1440): Arial interface/composer resolves Liberation Sans; code uses Liberation Mono; eight font assets hashed. Same fixture, frozen clock, DPR2, scale1 and layout on both commits. Recipe/font index. Historical fallback identity remains unknown; this additional pair is the reproducible reference, not an assertion of identical historical pixels.

Pinned-pair visible differences: single tildes restored, accidental user strikes removed, and resulting user-bubble width/content movement; header, assistant, code, composer and scroll control match exactly. Runtime supplements: Chrome 154/Linux, CSS1280×800, DPR2, scale1, light/system, no clock freeze. Recordings retain application pixels and add timestamped captions below. Canonical static pairs retain their original frozen-clock environment. Supplemental full-frame differences include timestamps, provider output, sidebar/layout and a corrected baseline context attempt; the index lists them explicitly. Diff0 and DiffT >16 are supplied, with a separately labeled prompt crop. No claim that every supplemental pixel difference comes from the parser.

Fixes pingdotgg/t3code#16105