Fix/2109 nested lists - #394
Open
conradmugabe wants to merge 12 commits into
Open
Conversation
…gins LLMs indent sublists by 2 spaces, but CommonMark requires the child to reach the parent's content column (3+ under "1. "), so nested bullets rendered as detached sibling lists and 2-space ordered children were flattened into their parent with wrong numbering. Add normalizeListIndentation, a post-preprocessLaTeX pass that re-indents undershooting children to their parent's content column, and wire it into the chat <Markdown> pipeline. Nested ul/ol also inherited the top-level my-6 margin, visually detaching them from their parent item; collapse it to my-1 via the [ul_&]/[ol_&] variants so only nested lists tighten. Refs #2109 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01REe8vMW5v7NeDPpXEqn69Z
A round of browser-verified fixes to preprocessLaTeX, each from a real
mentor reply that rendered broken:
- Preserve \\ row separators in $$ environments (pmatrix/aligned rows
collapsed to one) by masking math before the line-break pass.
- Keep CommonMark escapes \# and \_ intact so they render literally
instead of becoming a heading or italics.
- Promote whole-line $$...$$/\[...\] spans to fenced display form so
remark-math classifies them as block math; mid-sentence spans stay
inline.
- Escape \$ inside math spans as \text{\textdollar} (remark-math honors
no escapes inside $...$, so the inner $ terminated the span and paired
orphans swallowed prose into math).
- Normalize \[/\(-wrapped lone styling commands to dollar form so
\[ \textbf{...} \] headings unwrap to **...** instead of KaTeX
rendering literal stars; mask math while the styling conversions run
so genuine math keeps its LaTeX commands.
- Convert \begin{verbatim} to a masked code fence.
- Convert self-nested itemize/enumerate innermost-first so the outer
begin no longer pairs with the inner end and leaks raw tokens.
- Convert aligned environments made entirely of &\verb|...| rows (code
blocks in math costume) to fenced code blocks, preserving list-item
indentation; mixed or genuine aligned math is untouched.
- Convert streaming tails of itemize/enumerate as items complete, and
guard unclosed fences mid-stream.
Update the journey 61 block-math comment to reflect the display
promotion.
Refs #2109
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REe8vMW5v7NeDPpXEqn69Z
The canvas markdownToHtml pipeline corrupted fenced code: the preprocessMarkdownForHtml fixes are whole-document regexes that injected blank lines and merged comment lines inside fence bodies, and the marked-highlight extension emitted per-token spans whose newline text nodes TipTap drops, merging code lines (@tool\ndef -> @tooldef). Mask fenced and inline code (including an unclosed streaming fence) before the markdown fixes run and restore it verbatim after, mirroring the chat pipeline. Drop marked-highlight/highlight.js entirely: no consumer renders the spans, and plain code blocks survive TipTap intact; adjust ProseMirror CSS so pre>code stops double-boxing. Wire normalizeListIndentation into markdownToHtml so canvas nesting matches chat. Refs #2109 Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01REe8vMW5v7NeDPpXEqn69Z
LLMs abuse display math as a block container: an entire
\begin{itemize}...\end{itemize} (nested sub-lists included) wrapped in
\[...\]. The wrapper survived list conversion as $$ around a Markdown
list, the math masking then hid the block from the \textbf/\textit
unwrap, and the message rendered as one flat prose paragraph with
literal \textbf{...} and inline "-" marks.
Strip \[...\]/$$...$$ delimiters whose body is exactly an itemize or
enumerate environment, before any delimiter conversion, so the list
converts normally. The lazy body only terminates at an \end{env}
directly before the closing delimiter, so a self-nested environment
cannot end the match early; genuine display math never matches.
Refs #2109
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REe8vMW5v7NeDPpXEqn69Z
The blanket brace-expansion >=5.0.8 override (f92a8bd, added for GHSA-mh99-v99m-4gvg) forces v5 under minimatch@9, whose CJS build calls braceExpansion.default(...) -- v5 only exports { expand }. Every istanbul coverage run (vitest --coverage, and with it scripts/check-test-coverage.sh in the pre-push hook) crashed with "brace_expansion_1.default is not a function", blocking all pushes. Plain `pnpm test` never hits the path, which is why the override looked green. Scope-override test-exclude's glob to >=11.1.0: glob 11 resolves minimatch@10, which consumes the renamed @isaacs/brace-expansion instead, so the vulnerable package leaves the coverage chain entirely. Coverage runs work again, the pre-push gate passes (all changed files >= 95% lines), and `pnpm audit --audit-level=high` stays clean with no ignore rules. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01REe8vMW5v7NeDPpXEqn69Z
The default prompt told models to "Always use LaTeX formatting for
presenting your responses" -- and they complied, emitting headings as
\[\textbf{...}\], lists as \begin{itemize}, and code as
\begin{verbatim}: document-mode LaTeX that KaTeX cannot render and that
produced every #2109 breakage class seen in production.
Replace it with a structured prompt that scopes LaTeX to mathematical
expressions inside $...$/$$...$$ delimiters and plain prose for
everything else, alongside tightened response-style, retrieval-context,
greeting, and scope rules. The constants test now pins the invariants:
the math-delimiter instruction must be present and the old "LaTeX for
presenting" phrasing must never return.
Only newly created agents pick this up -- existing mentors keep their
stored prompt, so the preprocessLaTeX repair layer remains the safety
net for them and for already-stored messages.
Refs #2109
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REe8vMW5v7NeDPpXEqn69Z
The last three #2109 fix rounds (math-wrapped headings, aligned-\verb
code blocks, display-math-wrapped lists) were pinned only by
preprocessLaTeX string tests, but their failure modes live in the
remark-math/KaTeX interaction that only a rendered <Markdown> exercises.
Add render tests for each shape: \[\textbf{...}\] headings become real
<strong> (no literal stars, no KaTeX), \(\textit{...}\) becomes <em>,
aligned \verb rows become a <pre><code> block with one line per row,
\begin{verbatim} keeps a $ in its body unescaped, and the wrapped-
itemize shape renders as a real <ul> with bold/italic labels and inline
\(\alpha\) still rendering as math inside a converted list item. A
conservative guard pins that wrappers with trailing prose are not
unwrapped and genuine display math stays math.
Refs #2109
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01REe8vMW5v7NeDPpXEqn69Z
# Conflicts: # pnpm-lock.yaml # pnpm-workspace.yaml
Contributor
PR E2E - FAILEDFull report, traces and screenshots
|
bnsoni
added
run-tests
bnsoni
added a commit
that referenced
this pull request
Jul 30, 2026
GitHub keeps only ONE pending run per concurrency group. A third dispatch CANCELS the previously-pending one before it starts a single job — silently, with no diagnosis on the PR. os#390 was dropped exactly this way: #394 took the group, #390 queued, #367 arrived and evicted #390. Its check went red having tested nothing. This was masked while stg2 shared the load across two groups; pinning everything to stg1 made a three-deep queue routine. Two changes, and both are needed: Wait for a free dispatch slot before firing. The caller polls until no run for this env is queued/pending, so the pending slot is never contested. Costs nothing new — the caller already polls for up to 2h; this just moves the wait earlier, and prints 'N run(s) queued ahead' instead of failing silently. Retry on eviction. Two callers can still check simultaneously and both dispatch. A central run that ends 'cancelled' having started ZERO jobs is an eviction, not a result — re-dispatch (up to 3 attempts) rather than reddening the PR. Neither alone suffices: the wait has a residual race, and retry alone would let PRs thrash. Dispatch and wait merge into one step so the retry can span both. The gate now reads steps.dispatch.outputs.conclusion. The liveness guard (busy-vs-offline) and phase streaming are carried over unchanged.
conradmugabe
added
run-tests
michael-on-code
approved these changes
Jul 31, 2026
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Checklist
Changes
Screenshots
Screencast.from.2026-07-30.02-15-33.mp4