fix(desktop): prevent code blocks with no tag, text, or yaml from bei…

[gfpyc]

What does this PR do?

Fixes a bug where Hermes Desktop misclassifies fenced code blocks as prose — losing their CodeCard, stripping fence markers, and collapsing whitespace. Affected: text/plain/plaintext, yaml with bullet lists, markdown/md fences (rich-rendered as prose instead of code), and any non-COMMON language tag (gdscript, zsh, ...) whose content looks prose-like.

Root cause: Two functions in markdown-code.ts:

• isLikelyProseFence — NON_CODE_FENCE_LANGUAGES included '', text, plain, plaintext, md, markdown, so explicit text-like tags were never treated as code; proseLines >= 3 && codeSignals === 0 then misclassified zero-code-signal content as prose.
• isLikelyProseCodeBlock — the bulletLines >= 1 prose check caught YAML list items; the fallthrough caught text/plain/plaintext and any non-COMMON explicit tag (gdscript, zsh, ...).

Fix (2 files):

1. isLikelyProseFence — any explicit, valid language tag (VALID_LANGUAGE_RE) is respected and rendered as code. Covers text/plain/plaintext/md/markdown (previously in NON_CODE_FENCE_LANGUAGES) and gdscript/zsh and other non-COMMON tags.
2. isLikelyProseCodeBlock —
   • exempt known code languages (COMMON_CODE_LANGUAGES, e.g. yaml/markdown) from the bullet-prose heuristic so their lists render as code;
   • respect any explicit, valid language tag, same guard as isLikelyProseFence.

Bare fences ('') keep the existing behavior: untagged paragraph fences still unwrap to prose, and config/key-value listings (SSH config, .env, INI) stay fenced via the existing isLikelyStructuredText guard. Unknown tags (e.g. Streamdown's heads) still classify as prose.

Related Issue

Fixes #77253

Type of Change

• 🐛 Bug fix

Changes Made

• apps/desktop/src/lib/markdown-code.ts: classification logic — explicit-tag respect in both prose-detection paths
• apps/desktop/src/lib/markdown-code.test.ts: extended repro matrix (22 assertions)

How to Test

1. text fence with 3+ lines of keyword-free text — CodeCard
2. yaml fence with a bullet list — CodeCard
3. gdscript (or any non-COMMON tag) fence with prose-like lines — CodeCard
4. markdown/md fence — CodeCard with raw (highlighted) content
5. bare (no tag) fence with 3+ lines of plain prose — still unwraps (existing behavior)
6. text fence with SSH config / .env content — CodeCard (structured guard preserved)
7. python``/bash` control — still renders normally

Verification

Trigger precision (verified): the misclassification fires at proseLines >= 3 && codeSignals === 0... content with zero code signals (no JS keywords like for/while, no operators => }{``,} ;, no HTML tags). English sentences frequently contain for/while` and are accidentally immune, which is why line-count alone looked unreliable. No sticky state: after a misclassified block, subsequent well-formed blocks render fine.

Behaviour table (tested on Windows 11, packaged app, after rebase onto latest main @ 1f8fdc7 desktop v0.17.0):

Fence type	Before	After
text with keyword-free lines	Prose, collapsed	CodeCard
yaml with bullet list	Prose	CodeCard
gdscript with prose-like lines	Prose, no card	CodeCard
markdown/md fence	Rich-rendered as prose	CodeCard (raw, highlighted)
Bare fence (no tag), plain prose	Unwrapped (existing behavior)	Unwrapped (unchanged)
text with SSH config / .env	CodeCard (structured guard)	CodeCard (unchanged)
python`` / bash`	CodeCard	CodeCard

Tests: 22/22 pass (markdown-code.test.ts): explicit-tag matrix (text/plain/plaintext/markdown/md/gdscript/yaml) across both functions + upstream regression cases (SSH config, .env, bullet-prose with unknown tags, structured-text veto, bare-prose unwrap).

Notes

Lebased onto latest main (includes upstream isLikelyStructuredText guard, #84664). The fix was narrowed to explicit-tag respect — bare fences intentionally keep the upstream unwrap philosophy.

[spfcraze]

This was generated by AI during triage.

Summary:
isLikelyProseCodeBlock still returns prose (no CodeCard) for a non-COMMON explicit language tag such as gdscript or zsh when the fenced body is a bullet list, because the new explicit-tag guard sits after a bullet-prose heuristic that only exempts COMMON_CODE_LANGUAGES.

Problems:

• In apps/desktop/src/lib/markdown-code.ts the new explicit-tag guard (line 422) runs after the bullet-prose heuristic (lines 403-407), whose exemption names only COMMON_CODE_LANGUAGES (line 406), so a non-COMMON tag with bulletLines >= 1 and proseLines >= 2 returns prose before the guard is reached.
• Verified against the PR head logic: isLikelyProseCodeBlock('gdscript', '- item one\n- item two') returns true (prose) while isLikelyProseFence on the same input returns false (code); isLikelyProseCodeBlock is the render-time decision point consulted for CodeCard rendering, so such a fence renders without a CodeCard.

Solution:
Place the explicit-tag respect ahead of the bullet-prose check in isLikelyProseCodeBlock (or fold non-COMMON explicit tags into the exemption) and add a test exercising a non-COMMON tag whose body is a bullet list.

Evidence

no deterministic fact backs this claim — model belief, not executed or read evidence

---

Checked against 6cda16a — the tip of fix/code-block-prose-misclassification when this was written — and c896c09, main at the same moment.

[gfpyc]

@spfcraze Thanks for the triage report — confirmed and fixed.

You were right that the explicit-tag guard ran after the bullet-prose heuristic, and the gap was wider than reported: it covered not just gdscript/zsh but the whole NON_CODE family (text/plain/plaintext) too, since none of them are in COMMON_CODE_LANGUAGES.

What changed in 040b277:

• Moved the explicit-tag guard ahead of the bullet heuristic in isLikelyProseCodeBlock.
• • Kept the Streamdown heads case (bullet + inline markdown emphasis) as prose — the upstream test still passes.
• • Added whole-class regression tests: gdscript/zsh/text/plain/plaintext with bullet-list bodies (fence and code-block layers), plus the numbered Chinese prose repro Stone441 reported on macOS (text/prompt tags).
    Verified: 30/30 vitest pass, plus packaged-app rendering on Windows 11 — bullet lists under non-COMMON tags now render as CodeCards; heads and bare-prose unwrap behavior unchanged.

[gfpyc]

Hi @danspicytaco — thanks for cross-referencing this. Since #58023 and this PR both touch the fence-classification heuristics in apps/desktop/src/lib/markdown-code.ts from complementary angles, let me lay out the current coverage so we can consolidate rather than compete:

This PR (#77258) — 30 vitest cases covering:

• Any explicit valid language tag (gdscript, zsh, text, plain, plaintext, md, markdown, ...) renders as code, never downgraded to prose
• Bullet-list bodies under non-COMMON tags (the spfcraze regression class)
• md/markdown fences stay raw code blocks (mainstream AI chat UI convention)
• Chinese numbered prose under text/prompt fences (Stone441's macOS repro)
• Bare fences still unwrap as prose when the content is prose-like (upstream philosophy preserved)

#58023 covers:

• Shell fence aliases (zsh/fish/powershell/cmd) with $HOME/$USER markdown/math side effects
• End-to-end preprocessMarkdown regression coverage

Neither is a strict superset of the other — they touch different guard branches in the same file, so both change sets can merge cleanly if the maintainers want to consolidate. I'm happy to rebase/coordinate so we land a single fix quickly; this bug has been open since early August and is reproducible on Windows 11 (XpycT) and macOS (Stone441).

@alt-glitch @teknium1 — could we get a consolidation decision (and the workflows approved)? The P3 label may undersell this: it breaks code-card rendering for any non-COMMON tagged fence on Desktop.

[djedi]

Merge this please! :)