[Bug] web_fetch collapses consecutive pipes in HTML table cells

[GCSimba]

Summary

web_fetch silently collapses consecutive literal pipe characters while converting HTML tables to GitHub-flavored Markdown. This changes cell text, including code operators, even when the HTTP request succeeds and the result is not truncated.

Reproduced against master (package version 0.2.1-alpha.2) at d743267388641bc76f17c45ce8b4c231aed1d32c. Master was fetched again after verification and still points to that commit.

Reproduction

Serve this HTML with Content-Type: text/html; charset=utf-8, then fetch it through the registered web_fetch tool:

<table>
  <thead><tr><th>Expression</th></tr></thead>
  <tbody><tr><td><code>first||last</code></td></tr></tbody>
</table>

The body row should be:

| `first\|\|last` |

But the baseline returns:

| `first\|last` |

The response is HTTP 200 and the tool metadata reports truncated: false. Runs of three pipes also collapse to one. Decimal/hex HTML entities, table headers and nested inline formatting are affected. A single pipe is already handled correctly.

This is not just table layout: changing || to | can change an expression's meaning in programming-language examples. The source page and programmatic HTML body are not modified; the corruption is in the model-visible Markdown result.

Root cause and minimal patch

renderTableCell() replaces an entire run matched by /\|+/g with one escaped pipe. Each literal pipe must instead be escaped separately. GFM's table specification also requires escaping pipes inside inline spans.

The original pipe count cannot be recovered from the corrupted Markdown alone: cells containing |, ||, or ||| can produce the same rendered text. The structured result is different. For successful web_fetch calls, execute() retains the fetched HTML in value.body.content, and tools/post-execute receives that canonical value alongside the rendered content. A hook using the retained HTML could re-render it with a corrected converter; a text-only consumer cannot reconstruct the count. Fixing the built-in renderer avoids requiring that workaround.

- .replace(/\|+/g, '\\|')
+ .replace(/\|/g, '\\|')

The production change is limited to that quantifier and its local JSDoc. The patch also updates the bilingual README pair, adds formatter and real-HTTP tool-registry regressions, and adds an independently owned recorded-session scenario. The existing HTTP snapshot fixture serves both pages; the original menu page and historical recordings remain unchanged.

No public TypeScript API, package export, dependency, configuration, RPC schema, Session format or network-access policy changes.

Patch links

• Minimal bug-fix commit — independently cherry-pickable.
• Fork branch.
• Separate documentation-test portability commit — replaces a privileged Windows file symlink with a directory junction, retaining both real-path escape assertions and explicit unlink cleanup. It does not change product behavior.

Verification

Node 24.19.0 and repository-declared pnpm 11.28.5, with frozen-lockfile installs:

• pnpm exec vitest run packages/web/tool-web/tests --coverage --coverage.include='packages/web/tool-web/src/fetch.ts': Windows and Linux, 99 passed; the modified file has 100% statements, branches, functions and lines.
• pnpm exec vitest run packages/web/web-fetch-http/tests packages/web/web/tests: Windows, 89 passed.
• pnpm run test:snapshot snapshots/session/headless.snapshot.ts -t 'replays web-fetch': Linux, both the original and new scenario passed. The new scenario runs real HTTP, the tool pipeline and Session persistence through dsh --profile headless; only model responses are replayed.
• pnpm run test:snapshot scripts/session-snapshot-corpus.corpus.ts: Windows and Linux, three checks passed.
• pnpm run build:lib: passed on Windows, including Host and Client library builds.
• pnpm run build:native-system: passed on Windows and Linux.
• pnpm run lint: passed on Windows, with one existing unused-disable warning.
• pnpm run test:docs: 21 gates passed.
• pnpm run doc-sync: Windows, all 43 gates passed, zero failed and zero skipped.
• pnpm run doc-typecheck: Windows, 81 documentation code blocks compiled.
• pnpm exec vitest run scripts/project-doc-site.spec.ts scripts/verify-doc-site-fragments.spec.ts website/tests/mermaid-viewer.spec.ts website/tests/image-viewer.spec.ts website/tests/code-groups.spec.ts website/tests/page-markdown-actions.spec.ts website/tests/raw-markdown.spec.ts: Windows and Linux, all 151 passed with no skips after the separate junction-fixture fix. No system permissions were changed.
• Built ESM formatter smoke and git diff --check: passed.
• Normal pre-commit and pre-push hooks: passed; pre-push rebuilt Host libraries and ran the Client typecheck. The fetched fork ref matches local HEAD.

Before the fix, the focused added regressions produced eight failures at the intended pipe-preservation assertions; the single-pipe control passed. After refreshing the new Session fixture, reverting only the quantifier made the assembled replay fail solely on the four table-body rows. Restoring it made the replay pass.

Validation limits

The new recorded-session scenario is POSIX-scoped: Windows emits JSON-escaped runtime-context paths and a platform-specific bundled ACL skill catalog, which prevent sharing this golden unchanged. Windows formatter and real-HTTP integration coverage passed. Linux assembled replay used the source profile, a locally built Linux native addon, and platform-neutral Typert contributors from the current Host build.

The full repository coverage suite, full snapshot suite, live-model API tests and remote CI were not run. This patch preserves pipe counts in HTML table cells; it is not a general lossless HTML-conversion guarantee, and existing truncation, hidden-content removal and colspan handling remain in place.

No equivalent Harness report or patch was identified in the checked public Discussions, issue/PR searches, commit messages, indexed code and sampled forks. That does not establish absence from private or unindexed work.

AI-assisted audit and implementation with Codex; both commits include Co-authored-by: Codex <noreply@openai.com>. Every verification command listed above was executed locally.

[argszero]

Root cause confirmed on the commit you verified against.

packages/web/tool-web/src/fetch.ts:54 (read at d743267388):

const escaped = content.trim().replace(/\n\r/g, '<br>').replace(/\n/g, '<br>').replace(/\|+/g, '\\|').padEnd(3, ' ')

/\|+/g consumes the whole run and the replacement emits one \|, so a run of N literal pipes in a cell becomes one. Your /\|/g is the right shape, and it also explains your single-pipe control: a one-character run is replaced identically under both patterns, which is exactly why the defect stayed invisible in the easy case. I checked the rest of the package and renderTableCell is the only site that rewrites pipes.

One point I would add to the body explicitly, because it forecloses a suggestion the fix will otherwise attract: the collapse is not recoverable downstream. web_fetch returns the already-rendered Markdown as its result, so any later consumer — including a tools/post-execute hook, which sees only the finished text — receives `first\|last` and cannot distinguish it from a cell that genuinely contained one pipe. The information is destroyed at line 54, and no plugin can reconstruct it; the renderer is the only place this can be fixed.

Your severity argument is the one that carries: a reader is shown first|last where the page said first||last, which in most languages changes what the expression means and not merely how the table lays out.

[GCSimba]

Thanks @argszero for checking the baseline and explaining the single-pipe control. I agree that the corrupted Markdown alone cannot distinguish an original |, ||, or |||; this is content corruption, not merely a layout issue. I have added that distinction to the report.

One small pipeline clarification: on this baseline, web_fetch.execute() retains the fetched HTML in the canonical result.value, and tools/post-execute receives that value as well as the rendered content. A hook using the HTML could re-render it with a corrected converter; it cannot recover the original pipe count from the damaged text alone. The built-in renderer remains the appropriate place for the minimal fix, rather than requiring a custom workaround.

Would you or the maintainers consider adopting the minimal fix commit upstream? It can be cherry-picked independently of the separate Windows documentation-test portability commit. The formatter, real-HTTP regression and assembled Session replay are included in that commit; the patch and recorded verification results remain unchanged.

I understand that external PRs are not currently accepted, so I have not opened one. I would be happy to follow up with any focused checks or adjustments needed for upstream inclusion.