[BunsDev]

Part of #277 (CF-4), tracker #273. This is the parser slice that the issue says can land before #274. Semantic turn rendering waits for #274's landed contract (PR #282) and is not in this PR: final answer by phase, outcomes, reasoning labels, and Jump to latest.

What changes

Markdown.tsx replaces its hand-written subset with micromark plus its GFM extension. The parsed tree renders straight to React elements, with no HTML string anywhere.

• Renders: tables (with column alignment), task lists, block quotes, nested lists, strikethrough, footnotes, reference links, horizontal rules, ####+ headings, and fences that keep their language (<pre class="code" data-lang="ts"><code>).
• Code blocks (components/transcript/CodeBlock.tsx): a language label (text when there is none) and an accessible Copy. Copy goes through main's clipboard:write and nothing else; a status line announces success or failure.
• Safe by construction:
  • Raw HTML and MDX-looking text render as text.
  • Only http(s) links become links, through the existing safeHref. javascript:, file:, data:, relative and autolinked non-web targets stay as the text that was written, as before.
  • An image becomes a link named Image: <alt>, so reading never fetches a remote image.
  • Task-list boxes are disabled.
• Streaming:
  • An unclosed fence runs to the end of the text, so a streaming fence shows its code so far.
  • While text keeps changing, it is split at blank lines outside fences (streamingChunks), and only the last piece is parsed again. Each piece is a memoized component.
  • Once the text rests for 300 ms it is parsed whole once, so the settled rendering is always the exact one.
• Settled renders sit in a 400-entry cache, so reopening a thread does not parse it again.
• Styles for the new elements use the existing border, surface and text tokens, so all three themes follow.

Dependency choice

Pinned exact, as renderer-only devDependencies bundled by Vite (like React and xterm): mdast-util-from-markdown@2.0.3 (Feb 2026), micromark-extension-gfm@3.0.0, mdast-util-gfm@3.1.0, @types/mdast@4.0.4.

micromark is the unified collective's CommonMark reference implementation, with spec-tested GFM extensions. Using it at the mdast level means Modex owns every element it outputs: there's no rehype, no dangerouslySetInnerHTML, and no component loading.

@create-markdown/core (already used by Notes) is a block-editor format, not a CommonMark/GFM parser, so it was not reused. react-markdown would pull in the whole remark/rehype pipeline for the same parse.

Measurements

Apple M3 Max, Node 24.18.1, macOS, renderer code loaded through Vite SSR. The machine was shared with other sessions (load average 21–50), so treat these as relative.

Case	Before (89c98f2)	After
7.9 KB reply streamed in 331 deltas, mean per delta	0.8 ms	0.15–0.18 ms (streaming pieces)
Same, worst delta	4.5–5.0 ms	0.7–1.0 ms
One full parse of that reply (what every delta would cost without the streaming pieces)	—	4.1–4.5 ms mean
First open of a 200-message, 239 KB thread	32–41 ms	135–178 ms
Reopening that thread	32–41 ms	7 ms
Renderer main chunk	436.6 kB (129.0 kB gzip)	518.0 kB (151.8 kB gzip)

The first-open cost is a real regression, and it is in this PR's evidence for review. A later CF-4 slice can defer parsing of folded work items if it matters in practice. The bundle now crosses Vite's 500 kB advisory warning; nothing gates on it.

Tests

• test/markdown-render.test.ts, 11 tests:
  • GFM elements, heading levels, fence language and indentation.
  • Mid-stream fences, every prefix of a streamed reply (no lost or doubled words), Unicode.
  • Inert raw HTML, link schemes, bare URL punctuation, reference links, images never fetched.
  • The streaming splitter (lossless, never inside a fence), and pieces rendering the same text as the whole.
• On 89c98f2 five of them fail for missing behavior. Four pass and stay as safety guards (prefixes, Unicode, raw HTML, link schemes). The two streaming tests failed with streamingChunks is not a function before it existed.
• test/markdown-render.test.ts also covers the bounded settled cache, and the code-block label and Copy markup. Both were seen failing before their code existed.
• e2e/conversation-rendering.spec.ts, 3 tests, through the real app and mock backend. The two Markdown tests fail with the 89c98f2 Markdown.tsx swapped in. The Copy test fails with the 505b488 one: no code-block header. All three pass at head. Copy puts exactly the block's code on the clipboard, and no tool, approval or new turn starts.
• Existing transcript specs sources, app, titles and agents passed locally: 28/28.

Checks on 68b89c8

Base 89c98f280292b13fd9b1c7af1bf2fe65cd2fa0d1, head 68b89c8dac5436012e498958bf6f4f6b89e88f61. Run locally on macOS (Apple Silicon), Node 24.18.1, pnpm 10.34.5, on a machine shared with other sessions:

Command	Result
pnpm install --frozen-lockfile	passed
pnpm build	passed
pnpm typecheck	passed
pnpm test	passed: worktree scripts 109/109, core 16/16, desktop 1,484 passed + 16 skipped of 1,500, desktop native/update 21/21, site 10/10, Store desktop 16/16
pnpm test:e2e	passed: desktop 320 passed + 3 skipped (21.5 min), site 7/7

The 3 e2e skips are the real-Open-Knowledge cases in knowledge.spec.ts, which skip without that local install. The 16 unit skips are the suite's existing platform and native skips. This PR adds none.

After the local run

• 135069f changes only a test helper. CodeQL flagged visible() in markdown-render.test.ts as "incomplete multi-character sanitization" (high): it stripped tags with a regex before decoding entities. It now skips tags character by character and decodes each entity once. On 135069f, markdown-render.test.ts passed 13/13 and the typecheck is clean, and CodeQL passes.
• No product code changed after 68b89c8, so the full local run above still describes what ships.
• CI on 68b89c8 (run 39108256492): unit + e2e (macOS) and iPhone companion passed.

After CF-1 (#282) merged

acba096 merges main (7780093) into this branch with no conflicts. A fresh full local run on acba096 (acba09642fe805a451714ace122e7198e6942f92) passed:

Command	Result
pnpm install --frozen-lockfile	passed
pnpm build	passed
pnpm typecheck	passed
pnpm test	desktop 1,511 passed and 16 skipped of 1,527; worktree scripts 109/109, core 16/16, desktop native/update 21/21, site 10/10, Store desktop 16/16
pnpm test:e2e	desktop 320 passed, 3 skipped (the real-Open-Knowledge cases), 20.0 min; site 7/7

CI on 135069f passed every check, CodeQL included.

Latest head: 046df0e

main kept moving (#284, then #285). ab8474a and 046df0e merge them in, both without conflicts.

• Full local run, 09ae3e5 (main at 8e26534 merged in): install, build, typecheck and pnpm test all passed. pnpm test:e2e ran 324 passed, 3 skipped, 1 failed. The failure, sidebar-details.spec.ts:40, was its setup's git init hitting No space left on device while the machine's disk was full. Run alone afterwards, that spec passed 3/3.
• 09ae3e5 also makes the Copy e2e click again until Copied shows. CI run 39112253838 had missed once there, likely because a late follow-up suggestion moved the transcript under the pointer. 9/9 local repeats passed.
• 046df0e (main at fb76aac merged in): typecheck, build, markdown-render 13/13, and conversation-rendering, typed-turns, sources and sidebar-details e2e 15/15. CI on 046df0e (run 39116117568) passed every check, including the required unit + e2e (macOS). Merge state: CLEAN.
• 74833f7 (main at d3711c6, Show the command that ran, its exit code, bounded output and honest source labels (CF-5, first part) #287 merged in, no conflicts): typecheck, build, markdown-render 13/13, and conversation-rendering, typed-turns, sources and conversation-evidence e2e 16/16. CI on 74833f7 passed every check. Merge state: CLEAN.

Reshaped at 29a74e8: the settle fix lives in #295

After review I found two more defects in this PR:

• a reply that finished streaming remounted its Markdown, dropping a selection or focus;
• tables collapsed at 900×600.

I fixed both here (083b8fb, 291b08b, verified on fc16df4). Then #295 appeared, another session's CF-4 completion, stacked on this branch at 74833f7, with its own fix for the same remount, its own continuity tests and its own theme captures. Keeping both fixes would conflict, so 29a74e8 hands that defect to #295:

• Markdown.tsx, markdown-render.test.ts and conversation-rendering.spec.ts return to 74833f7.
• transcript-continuity.spec.ts is dropped.
• The status clause is reverted.

Until #295 merges, this PR alone still has the remount-at-settle defect. #295 depends on this PR, so they land in order.

Kept here, because #295 does not touch styles.css:

• The narrow-table fix. At 900×600 with Review open, the transcript is 265 px wide. A table with an unbreakable path squeezed its Detail column to 58 px and its row to 1,009 px. Cells now keep at least 10em and wrap long tokens, so the table scrolls in its own box.
• Its theme check, in its own file: e2e/markdown-narrow.spec.ts, run in Graphite, Jev and OpenCoven at 900×600, checks:
  • no sideways page scroll;
  • the table and the code scroll in their boxes;
  • every cell is at least about 10 characters wide;
  • Copy stays visible.

On f1bbb28 (main at 04b7ec7 merged in): typecheck, build, markdown-render 13/13, and conversation-rendering, markdown-narrow, typed-turns and sources e2e 13/13. The full suites are left to CI. The last full local run was on fc16df4, all passed.

Ready: 44b21bc

44b21bc merges main at d3dfe76 (#294). Before pushing, typecheck, build and the conversation-fidelity, conversation-rendering and markdown-narrow e2e specs (8/8) passed. CI then passed every check. Merge state: CLEAN, current with main.

Not done here (left open on #277)

• Final answer by native phase, attempts, outcome and reasoning labels, and Jump to latest shipped separately in Lay out each turn from its typed items: final answer by phase, live items, attempts, Jump to latest (CF-4) #284.
• Reference-viewport screenshots, and keyboard and reduced-motion review of the Markdown elements.
• Human VoiceOver acceptance, recorded separately.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
modex	Ready	Preview	Oct 18, 2026 9:03pm UTC

[github-code-quality]

Code coverage overview

Languages: TypeScript

TypeScript / code-coverage-agent

main 7ec5f14	cf4-markdown-parser 5f96c1e	+/-
63%	63%	0%

View coverage of the 2 most impacted files.

File	main 7ec5f14	cf4-markdown-parser 5f96c1e	+/-
apps/desktop/sr...t/CodeBlock.tsx	0%	81%	+81%
apps/desktop/sr...ts/Markdown.tsx	0%	89%	+89%

ℹ️ Calculations represent line coverage. Results may change as calculations are updated.

---

_{Updated October 18, 2026 21:29 UTC}

[BunsDev]

Holding this PR (back to draft): I found a defect in it. While a reply streams, Markdown renders it as memoized pieces (Blocks). 300 ms after the text rests, it switches to plain elements from one full parse. The children's component type changes, so React unmounts every piece and mounts new DOM. A selection, keyboard focus, or a code block's Copied state inside a reply is lost the moment the reply finishes streaming.

#277 requires selection and focus to survive sustained updates. I'm adding a failing e2e test for it, then a fix: the settled parse keeps the same piece structure, so React updates in place instead of remounting. I'll mark this ready again with fresh CI.