Comparing GitHub HTTPS ref discovery with HTTP/1.1

I constructed a tool called gitclone-doctor to compare the Git repository's selected HTTPS transport protocol with HTTP/1.1. This tool performs two sets of requests: one against the target repository using the chosen settings and another against the git/git control repository. The control repository uses the same frozen proxy and CA snapshot throughout the testing, ensuring that any settings specific to the target repository remain unchanged during the comparison.

The demonstration utilizes a controlled local environment, not a real GitHub incident. This setup generates a temporary certificate, accepts both HTTP/2 and HTTP/1.1 connections, and returns a deliberate 403 error for the target URL over HTTP/2. In contrast, it delivers a valid smart HTTP ref advertisement over HTTP/1.1. The control path experiences success with both protocols, while the test program examines the real Git trace and declines to provide a comparison if it cannot observe the relevant requests.

The report presents response protocols and numeric statuses, rather than raw stderr, headers, configuration values, or trace text. The child Git processes run in a rebuilt environment, devoid of helpers and extra headers, and disable redirects while requiring verified TLS. The tool rejects execution when a .netrc or _netrc file is present, as Git's libcurl may consult those files.

The report differentiates Git requests from direct network checks. Git employs the selected proxy when configured, whereas DNS and TLS checks operate independently, with their own deadlines. A direct DNS failure does not invalidate passing Git requests. The output is a JSON file containing fixed result codes and presence flags for helpers, headers, and URL rewrites, not the configured values.

This ref discovery check does not complete the clone pack transfer process and does not test LFS, submodules, or authenticated access. Although a shared failure may remain unresolved, a successful HTTP/1.1 response does not necessarily indicate a protocol defect. To observe the local example, run the command 'npm run demo'. The command-line interface (CLI) can be executed from GitHub using 'npx --yes --package=github:Arthur031221/gitclone-doctor gitclone-doctor git/git '.

I would appreciate feedback on whether the comparison is clear and whether its limitations are stated plainly. You can find the repository at https://github.com/Arthur031221/gitclone-doctor.