Sites Git credentials: secure authentication from Codex and token revocation behavior

ChatGPT
, ,
1

We’re trying to deploy an existing ChatGPT Site from Codex Desktop on macOS. The source and build are tested, and the Sites connector successfully issues a short-lived repository write credential through create_source_repository_write_credential.

Authentication worked for a read-only Git operation. However, Codex’s automatic approval review rejected a push because passing the Bearer token through git -c http.extraHeader=... exposes it in the execution-tool transcript.

Git supports --config-env=http.extraHeader=SITES_AUTH_HEADER, but we haven’t found a supported way to transfer the connector-issued token into that environment variable without including it in another logged tool argument.

Could OpenAI clarify:

  1. What is the supported way to securely authenticate a Sites source push from Codex Desktop?

  2. Does issuing a new source-repository credential immediately revoke the previous token, or can both remain valid until their respective expiration times?

  3. Is there an explicit revocation operation for these credentials?

  4. Is there a supported upload or publishing flow that handles source authentication internally?

We’re asking specifically about Sites source-repository write tokens, not OpenAI API keys, ChatGPT login tokens, or Site visitor authentication.

2

By using the CLI in codex it worked (AI steps):

Here’s how authentication worked:

  1. Requested a short-lived, repository-scoped token from the Sites connector.
  2. Held the response in tool-session memory without printing it.
  3. Passed the token to Git through a command-scoped environment variable, using --config-env=http.extraHeader=SITES_AUTH_HEADER.
  4. Disabled shell tracing with set +x and redacted the token from returned command output.
  5. Cleared the stored credential afterward.