42 of my 79 repos looked empty. The code was there the whole time.

I have a lot of small tools on GitHub — offline CLI things for audio, release metadata, studio admin. Each one has a description. Each one built fine locally. I had never once looked at them the way a stranger does.

When I finally did, 42 of my 79 public repositories rendered as completely empty pages. No README, no files, nothing. Not broken — blank.

The code had been pushed. It just wasn’t on the default branch

The agent that built these repos worked the way agents usually do: branch, commit, push. Every project ended up on codex/initial-implementation. What never happened was the last step a human does without thinking — getting that work onto main.

So each repo had a main that was the default branch and had no commits at all, plus a feature branch holding the entire project. Git was happy. The push succeeded. The work was genuinely public — you just had to know to switch branches to see any of it.

Detecting it across a whole account

The cheap test: ask for the git tree of the default branch. If the branch has no commits, the tree 404s. That is one API call per repo and no cloning.

for r in $(gh api user/repos --paginate -q '.[].name'); do
  d=$(gh api repos/OWNER/$r --jq .default_branch)
  gh api repos/OWNER/$r/git/trees/$d --jq '.tree|length' >/dev/null 2>&1 \
    || echo "EMPTY DEFAULT BRANCH: $r ($d)"
done

The fix is one call, and it is not a merge

Because main has no commits, there is nothing to merge. You are just creating the ref at the tip of the branch that holds the work:

sha=$(gh api repos/OWNER/REPO/commits/codex/initial-implementation --jq .sha)
gh api -X POST repos/OWNER/REPO/git/refs -f ref=refs/heads/main -f sha="$sha"

The feature branch is untouched, so it is reversible. And it exposes nothing new — that branch was already public. All it changes is which branch a visitor lands on.

Two things I found while I was in there

Once the repos rendered, the next question was whether they led anywhere. Only 8 of 79 READMEs linked to anything else I make. Someone finds a loudness checker, likes it, and hits a dead end. And only 12 of 80 repos had topics, which is most of how small tools get found at all.

Automating the topics taught me something worth repeating: I generated them from each repo’s own description with keyword rules, and a dry run caught it tagging a tool called exportheadroom with acoustics — because “headroom” matched room. It also got live-performance from “delivered”. Word boundaries fixed both.

What it actually cost

None of this is clever. It is the unglamorous half of shipping: the work existed, was finished, was public, and was reaching nobody. If you have let an agent create repositories for you, the check above takes about a minute.