REST branches: GET /branches/master 301-redirects to the default branch, undocumented and contradicted by the 2023-02-07 changelog update #45642

What the docs say vs what the API does

The REST reference for branches lists 301 Moved permanently among the responses for GET /repos/{owner}/{repo}/branches/{branch}, but does not say when that redirect fires, or that Location may name a different branch than the one requested.

The changelog entry Links to deleted branches now redirect to the default branch (2020-07-17) announced this redirect, and an update appended to that same entry on 2023-02-07 withdraws it:

This behavior conflicted with other GitHub URL patterns, so we removed it. GitHub does redirect renamed branches, but not deleted branches.

That entry concerns web links. The REST API today still redirects, and it redirects to branches that no rename plausibly relates to master.

Measured 2026-08-28, redirect-following disabled

symfony/symfony        /branches/master   ->  301  Location=.../branches/8.2
scala/scala            /branches/master   ->  301  Location=.../branches/2.13.x
twitter/finagle        /branches/master   ->  301  Location=.../branches/develop
pytorch/pytorch        /branches/master   ->  301  Location=.../branches/main
kubernetes/kubernetes  /branches/master   ->  200  name=master   (no redirect)
symfony/symfony        /branches/nope-xyz ->  404

Nobody renamed master to 8.2, 2.13.x, or develop. Those are simply the current default branches.

Boundary

I am not claiming to know the cause — GitHub does not expose deleted-branch history, so "never had a master" is inference from a creation date, not a measurement.

Why this is worth documenting

Every HTTP client that follows redirects by default (urllib, requests, curl -L, most SDKs) turns the 301 into a transparent 200 describing a branch the caller never asked for. The failure mode is a successful read of the wrong resource — no error, no warning.

Any tool resolving a historical branch name is exposed. A pull request records its base ref as it was at the time, so old PRs name branches that are now gone:

b = get(f"/repos/{repo}/branches/{pr['base']['ref']}")   # 200. Looks fine.
cmp = get(f"/repos/{repo}/compare/{b['name']}...{head}") # answers about the wrong branch

In my own use — comparing a PR's head against the branch it targeted — this silently defeated a fix that existed specifically to stop comparing against the default branch. Across a sample of 379 pull requests in 108 repositories, 46 of 48 resolution failures were this redirect.

The guard callers need is name equality, not HTTP status:

if b["name"] != ref:  # the requested ref is gone; this is a different branch

Suggested doc change

Any one of these would remove the surprise:

1. Document on the branches endpoint when the 301 fires, and that Location may name a different branch.
2. Reconcile with the 2023-02-07 withdrawal, which currently reads as though this redirect no longer exists.
3. If the behaviour is intended to be renamed-branches-only, the observations above suggest it is broader than that.

Documentation alone would be enough. The behaviour is defensible; the contradiction with a published withdrawal is what makes it cost people time.