Findings

• [Major] Raw pipe splitting rewrites valid escaped-pipe tables — countSourceCells() treats every | byte as a column boundary, but GFM allows escaped pipes inside a cell. A valid two-column table like | A\|B | C | / |---|---| is detected as a three-column header, so the plugin pads the delimiter and re-parses it as a different table shape. Evidence: web/src/lib/remark-repair-tables.ts:36.

Summary

• Review mode: initial
• Found one correctness regression in the new table repair path.

Testing

• Not run (automation): bun is not installed in this runner environment.

Findings

• [Major] Escaped header pipes trigger false repairs — countSourceCells splits every |, so a valid two-column GFM table like | A \| B | C | with |---|---| is treated as three source columns. The plugin then pads the separator and re-parses, changing an already-valid table by adding a bogus empty column. Evidence web/src/lib/remark-repair-tables.ts:36
  Suggested fix:

  function countSourceCells(line: string): number {
      const trimmed = line.trim()
      const inner = trimmed.startsWith('|') ? trimmed.slice(1) : trimmed
      const stripped = inner.endsWith('|') ? inner.slice(0, -1) : inner
  
      let cells = 1
      let escaped = false
      for (const ch of stripped) {
          if (escaped) {
              escaped = false
              continue
          }
          if (ch === '\\') {
              escaped = true
              continue
          }
          if (ch === '|') cells++
      }
      return cells
  }

Summary

• Review mode: initial
• One issue found: valid tables with escaped literal pipes can be misclassified and rewritten. Residual risk remains around broader GFM table tokenization edge cases unless covered by tests.

Testing

• Not run (automation; Bun unavailable in this runner PATH)

HAPI Bot

Findings

• [Major] Repaired tables lose parser-extension nodes — parseTableBlock() builds a fresh parser with only remarkParse + remarkGfm, while the app markdown pipeline also has parser extensions such as remarkMath. When a broken table contains inline math, the original parse would produce math nodes, but the repaired replacement table is parsed as plain text, so rehypeKatex no longer renders it. Evidence web/src/lib/remark-repair-tables.ts:74
  Suggested fix:

  import type { Processor } from "unified"
  
  function parseTableBlock(processor: Processor, source: string): Table | null {
      const tree = processor.parse(source) as Root
      return tree.children.find((node): node is Table => node.type === "table") ?? null
  }
  
  export default function remarkRepairTables(this: Processor) {
      const processor = this
      return (tree: Root, file: VFile) => {
          const source = String(file.value)
          visitTables(tree as unknown as Parameters<typeof visitTables>[0], null, 0, source, processor)
      }
  }

Questions

• None.

Summary

• Review mode: follow-up after new commits
• The previous escaped-pipe finding was addressed. One remaining issue: the repair path reparses with a reduced markdown parser, which can regress rendering for tables that also use enabled syntax extensions. Residual risk is around parser-extension coverage for repaired table blocks.

Testing

• Not run (automation; bun is unavailable in this runner PATH)

HAPI Bot