feat(tui): render assistant transcript with Markdown via OpenTUI markdown component

[knqiufan]

Summary

Closes #35

The TUI transcript panel previously rendered all messages (including assistant responses) as plain text via <text> + <span>, losing all Markdown formatting — code blocks, tables, bold/italic, lists, and links were displayed as raw characters.

@opentui/react already ships a <markdown> JSX component backed by MarkdownRenderable (with marked parsing + tree-sitter syntax highlighting), but it was never used in the transcript.

Changes

• Role-based rendering: Assistant messages now render through the <markdown> component with full Markdown support (code blocks, tables, bold/italic, lists, links). User/tool/system messages retain existing plain-text rendering unchanged.
• Extract transcript data-building logic into src/tui/transcript-items.ts: TranscriptItem interface, buildTranscriptItems, buildWelcomeTranscriptItem, resolveTranscriptIdentity, resolveTranscriptBackground, wrapMultiline, sliceByDisplayWidth — all previously private to app.tsx, now independently importable and testable.
• Add src/tui/transcript-syntax-style.ts: buildSyntaxStyleFromTheme() maps TUI theme colors to SyntaxStyle for code highlighting, so code blocks follow the active theme.
• Extend TranscriptItem with content: string (raw Markdown text) and useMarkdown: boolean fields.
• Update TranscriptEntry component: when useMarkdown is true, renders badge line + <markdown> block with table/code support; when false, keeps the original plain-text path.
• Add 21 unit tests in src/tui/transcript-items.test.ts covering buildTranscriptItems (per-role behavior), resolveTranscriptIdentity, resolveTranscriptBackground, buildWelcomeTranscriptItem, wrapMultiline, and sliceByDisplayWidth.

Before / After

Markdown element	Before	After
Code blocks (```)	Plain text	Syntax-highlighted via tree-sitter
Tables	Plain text	Rendered with borders and column layout
Bold / Italic	Raw ** / _	Rendered as <strong> / <em>
Links	Raw URL text	Rendered as clickable link
Lists / Headings	No hierarchy	Proper block-level rendering
Inline code	No distinction	Styled inline code
User/tool/system	Plain text	Plain text (unchanged)

Test plan

• Run pnpm vitest run src/tui/transcript-items.test.ts — 21 new tests pass
• Run lint, dep-guard, deadcode, tsc, prettier — all pass on this branch
• Start TUI (step exec), send a prompt that returns Markdown with code blocks — verify syntax highlighting renders
• Send a prompt that returns a Markdown table — verify table renders with borders
• Send a prompt with bold/italic text — verify text modifiers render
• Verify user input still renders as plain text (no Markdown parsing)
• Verify tool output still renders with compact preview (truncation logic unchanged)
• Switch themes (/theme sage, /theme default) — verify code highlighting colors update
• Scroll through long Markdown content in transcript — verify scrollbox behavior

[ZouR-Ma]

Thanks for the contribution!

To get this ready to merge:

• Reference an issue in the description with Closes #<number> (our link-check CI requires it; open one first if needed).
• Make sure pnpm check passes locally (lint, type-check, dependency/dead-code guards, formatting, and tests — also enforced in CI).

Full conventions:
https://github.com/stepfun-ai/Step-Realtime-CLI/blob/main/CONTRIBUTING.md

[knqiufan]

Thanks @ZouR-Ma for the review guidance!

• PR description already references Closes TUI transcript 不支持 Markdown 渲染，agent 响应中的代码块、表格、粗体等丢失格式 #35.
• Rebased onto latest main (ee6862f), new head: ac379b4.
• Ran checks locally: lint, dep-guard, deadcode, tsc, format:check, and src/tui/transcript-items.test.ts (21 tests) all pass.
• Full pnpm check on Windows previously had 13 unrelated test failures (symlink EPERM + shell timeout). Tracked in [chore] align utils tests with TESTING.md for Windows local pnpm check #74 and addressed in test(utils): align path and shell tests with TESTING.md on Windows #75 — local pnpm check is now green on Windows after that fix.

Ready for review when you have a moment.

[ZouR-Ma]

Status update: #32 has landed on main and touched app.tsx, and since you moved buildTranscriptItems into a new module, this PR now conflicts and needs a rebase first. We still want this feature — the list before it lands is the same as before, plus a new item 0:

0. (Must) Rebase to resolve the conflict. #32 is on main now; when you rebase, carry its hidden-entry filter (.filter((entry) => !entry.hidden)) into buildTranscriptItems in transcript-items.ts and add a matching case to the new test file.
1. (Must) Verify and handle Markdown behavior under streaming. Transient assistant entries (local-opentui-bridge.ts:185-189) grow incrementally during streaming and now feed into <markdown streaming={false}> — a full re-parse per delta, with unterminated code fences rendering broken mid-stream. Pick one: wire the streaming prop; render in-flight entries as plain text and switch on completion; or test the long-code-block streaming scenario and post the results.
2. (Must) Execute the seven manual test-plan items, with at least one pass on the installed step binary rather than pnpm tui:dev. The unit tests cover data building, not rendering, and the tree-sitter/Markdown assets go through the tsdown bundle — the empty-chunk incident in #25 shows that pipeline has bitten before.
3. (Must) Reopen issue #35. It was closed ahead of time; it should stay open and be auto-closed by this PR on merge — we can flip it back on our side.

CI is green across all three platforms, so that's not a concern. We want this feature — ping us once the rebase and verification are in.

[knqiufan]

Thanks for the detailed follow-up. I have rebased #39 onto current main (87c3445) and resolved the app.tsx conflict.

Changes made:

• Preserved fix(tui): hide plugin-injected system reminders from transcript #32's hidden-entry filter in transcript-items.ts and added coverage for hidden entries.
• Preserved the current collapsible tool and reasoning transcript behavior.
• Wired Markdown streaming state from the local OpenTUI bridge: transient assistant entries use streaming mode while deltas arrive, then finalize after the assistant message (including the tool-call boundary). Added coverage for an incomplete fenced code block during streaming.

Validation:

• pnpm check passes: 90 test files, 1,532 tests passed (24 skipped).
• The local TUI starts successfully from this branch.

Supplementary instruction:
PR#35 The Reopen control for the current account is disabled and requires maintenance personnel to perform the operation.