🐛 - Bugbot treats fenced Markdown UI example fragments as broken live document links #873

Is there an existing issue for this?

• I have searched the existing issues. The fenced/anchor search returns an unrelated extension specification, not this finding defect.

Which actions are affected?

Pull Request

Which platforms are affected?

Linux (native GitHub Actions review); macOS for the independent local parser/source-value probes.

Description

Bugbot publishes a 99%-confidence broken-navigation finding for fragment tokens inside a fenced Markdown UI mockup. The original document renders those tokens as literal code, not live links. The actual GitHub Markdown API confirms that no corresponding anchor elements exist; the same two link strings outside the fence produce two real anchors as a positive control. This is a completed, fully contextual review with a wrong markup-state premise, not missing context or an actual failing navigation.

Consumer: vypdev/domus#28, exact head2155e4b58c46a65bd31c5994eef0fb67c0a904fd.
Original source: specification lines205–230.
Finding: vypdev/domus#28 (comment) .
Native review: https://github.com/vypdev/domus/actions/runs/41665073143/job/126018092821 .
Finding ID: specs/household-administration-browser.md:223:broken-example-navigation-anchors.
Severity low; correctness; confidence99%.

The finding says readers following the two links remain at the example because the document lacks matching headings. A reader cannot follow these strings from the rendered document: they appear inside pre/code. Adding document headings or replacing these with technical-documentation destinations would misinterpret an application mockup. The consumer UI uses edit buttons.

Reproducing the issue

1. Parse/render this exact minimal source, including its surrounding fence:

```markdown
## Links

[Household settings](#household-settings) · [Your preferences](#your-preferences)
```

2. MarkdownIt identifies one markdown fence token for the original representative view and zero inline link tokens targeting those fragments. This is the parser already used by the consumer docs tooling; its result is independently checked below.
3. Submit the full original immutable specification to GitHub POST /markdown with mode=gfm and context=vypdev/domus. Parse returned HTML: zero a[href='#household-settings'] / a[href='#your-preferences']; both original bracketed link strings remain literal under pre.
4. Positive control: render the same two links without a fence through that same API. Exactly two target anchors exist. This confirms the absence assertion is sensitive to real active links.
5. The native review nevertheless publishes the finding against the fenced line. The original rendered HTML, exact request inputs, parser proof and review telemetry are retained locally. Source identities are immutable and independently available through the links above.

Expected: fence-aware review does not report inactive example strings as broken live document navigation. If reviewing the internal completeness of a mockup, state that distinct scope and inspect intended application destinations rather than claiming a reader clicked a rendered source link.

Actual: a false broken-link finding is published with confidence99% and active navigation impact.

copilot Version

vypdev/copilot@v3 resolved3e7ceea08e6131360e026937d1c0cc1b120472e0; codex-cli0.166.0; configured codex/gpt-6-luna/smart. Current native logs verify the Action SHA and CLI version. Domus uses MarkdownIt and an independent real GitHub GFM API render.

Relevant Log Output

Complete context;11/11analysis partitions;candidate1/published1/open1;outcome completed. Native job126018092821 succeeds as a workflow while publishing this finding. All other current findings are resolved. Actual GitHub rendering: original targets0, unfenced positive-control targets2, literal code strings retainedtrue.

The newly implemented consumer C57 acceptance passes all eight maximum-Unicode locale/deployment/theme journeys and all36built catalog cases. No product code, alias/budget/identity policy, timeout, retry or coverage threshold changes. Green CI is not used as the sole counterevidence for a markup issue.

Additional context and comments

Please add an upstream regression separating real links from source examples in fenced/indented code, including a positive control and nested mockup headings. Verify claimed navigation against the rendered markup state before publishing or proposing document anchors. These upstream tests are proposed, not executed or modified here.

The consumer will clarify the mockup with the actual Edit household / Edit preferences button labels. That removes ambiguous pseudo-navigation source without implying that the original document contained live broken links. No manual thread resolution; no Copilot source/resources/configuration/workflows/models/credentials/bounds changes. Product acceptance continues independently on the existing issue/ref.