A co-authored-by search for Cursor's name returned 10,033,042 public commits

The guard is one line. It covers exactly one of the two failures. `const unusable = (n) => n === 0 || n >= total;` ^[7] Co-authored commits are a strict subset of all commits, so a count equal to the total means the qualifier filtered nothing ^[7]. The post argues that if GitHub stops parsing `co-authored-by:`, the query degrades to free text, matches everything, and the card renders 100% AI co-authored ^[6].

Searching by display name walks straight through that check. The name form returned 10,262 Claude commits and the address form 10,255 ^[11]. Against a 26,251-commit total, 10,262 is neither zero nor the ceiling, so `unusable` returns false ^[25]. The gap is about 0.07% of the count ^[26], and each of those commits carried a second trailer naming a human collaborator, counted as an AI ^[12].

That post asserts free-text degradation without measuring it. The control it reports is that an invented qualifier returns zero, and that changing the value changes the count ^[5]. A search that fell back to free text would also return zero for an invented qualifier, because no commit message contains that string. The zero fits a parser that rejects unknown keys just as well as one that quietly stops parsing a known one.

The Cursor query ran without an author filter, so there was nothing to compare it against: `co-authored-by:cursor+is:public` returned 10,033,042 ^[9]. "Ten million felt too round," the author wrote ^[28]. Printing five of the matching commit messages showed the real trailer, `Co-authored-by: Cursor <[email protected]>`, and a name search matches any co-author who happens to share it ^[10].

Keying on the address is forced by the trailers themselves. Aider writes whichever model it used into the display name, so the same tool emits `Co-authored-by: aider (anthropic/claude-sonnet-5) <[email protected]>` on one commit and `Co-authored-by: aider (ollama/gemma4:e4b-mlx) <[email protected]>` on another ^[14]. The nine trailers in the post were read off real public commits instead of vendor documentation ^[15], and one tool the author went looking for skips the dedicated trailer entirely, so the search came back with only human contributors ^[16].

One repo in that history ran 960 of 1,107 commits, about 87% ^[1][29], so the aggregate rate describes one person's mix of work. It transfers to your team only if your private-to-public ratio resembles his and your agents write trailers at all ^[16]. Restricting the same tool to public repositories moves the figure by a factor of about 19 ^[27]. "A public-only number isn't a smaller version of the truth, it's a different number," wrote the developer, who publishes on dev.to as yosh1 ^[18][23]. The counts also move as he commits, so every figure here is one snapshot ^[22].

The first card grew bars from zero with `<rect height="0"><animate attributeName="height" to="28"/></rect>`, and any renderer that ignores SMIL sees height 0, which is what static renderers, social preview generators and thumbnail pipelines do ^[19]. The fix declares the resting geometry and lets the animation supply only a starting value ^[20]. There is a test asserting the animated output contains no `height="0"` ^[21].