OpenClaw merged PR #96476, "fix(link-understanding): strip markdown links whose label contains brackets", a P2 correctness and safety fix for link-understanding.
The affected path is extractLinksFromMessage, which strips markdown links before scanning a message for bare URLs. That distinction is intentional. A display-only citation such as [docs](https://docs.example) should not automatically become a URL that OpenClaw fetches for link understanding.
The bug appeared when the markdown link label itself contained a bracket. Real notes often contain labels like [my notes [v2]]. The old stripping expression could fail to recognize that whole markdown link, leaving the URL behind for the bare-link scanner.
The Repro
The PR gives this example:
Check [my notes [v2]](https://internal.example/doc) for details
Before the fix, the markdown link was not stripped. The bare-link scanner could then extract the inner URL, including a trailing parenthesis, and treat it as a link to fetch.
After the fix, OpenClaw recognizes the URL as part of a markdown citation and suppresses it from the bare-link extraction result.
Why It Matters
This is not just tidy markdown parsing. The PR calls out a correctness and safety issue: a display-only citation can be promoted into an actually fetched link.
That matters for any OpenClaw surface that receives pasted notes, transcripts, issue comments, or markdown-heavy messages. A user may include a private or internal URL as context, expecting it to stay as visible text. Link-understanding should not silently reinterpret that citation as a fetch target because the label includes a bracket.
The fix narrows that gap by allowing ] inside a link label as long as it is not the closing ]( boundary. Existing normal markdown links and bare URLs keep their previous behavior.
Validation
The PR includes a red-green proof for the old and new regex behavior. The old expression reproduces the bug by extracting https://internal.example/doc). The new expression returns an empty link list for the bracketed-label markdown citation.
It also includes parity checks for unaffected cases: ordinary bare links, simple markdown links plus a bare URL, loopback/private-range test inputs, and adjacent markdown links.
A new unit test covers the bracketed-label case directly: ignores markdown links whose label contains brackets. According to the PR, the test fails on the old regex and passes after the fix.
For operators, the takeaway is straightforward. OpenClaw link-understanding is now less likely to fetch a URL that was merely part of a markdown citation, especially in the messy note formats that agents and humans actually exchange.