## Why
GitHub Actions workflows can use short-lived OpenAI credentials instead of storing a long-lived OpenAI API key in GitHub secrets.
## What changed
- Add `--github-oidc` with required audience, identity provider, service account, and GitHub token-request URL settings. Read the GitHub request credential from stdin and connect it to the existing federated authentication flow.
- Require the default OpenAI Responses endpoint in OIDC mode before reading credentials, and complete the initial token exchange before publishing server info.
- Close the npm launcher's copy of stdin after the native child starts.
- Document OIDC setup, credential handling, renewal behavior, and caller responsibilities. Keep API-key authentication as the default.
## Testing
Add tests for required OIDC flags, upstream restrictions before credential reads, rejection of unsafe or empty settings without exposing sensitive input, and failure when a configured custom CA cannot be loaded in the CLI blocking context.
GitOrigin-RevId: 7de6fd4c3491e7ccd6a2c2ccdee0bc2eedd0e642
## Why
The Responses API proxy could reuse a federated credential after an upstream `401`, and concurrent requests could each retry a failed token exchange.
## What changed
- Invalidate the rejected credential so the next request obtains a fresh token. Forward the `401` without replaying the request or waiting for an ongoing renewal; leave credentials intact on `403`.
- Track rejection per token issuance so a late `401` cannot invalidate a newer credential.
- Back off failed exchanges exponentially from 1 to 30 seconds, starting when the exchange fails. Reject requests during backoff without reusing the old token, and reset the delay after a successful exchange.
## Testing
Add tests for response forwarding without replay, late and concurrent rejections, shared renewal failures, backoff timing and reset, and credential-safe renewal errors without following redirects.
GitOrigin-RevId: eb1ed6848d74fdba1534919cace5126f4a0595cf
## What changed
Add a federated `Auth` variant that exchanges JWT subject tokens for OpenAI access tokens using identity provider and service account IDs. Serialize concurrent exchanges, cache credentials, and renew proactively with a fresh subject token.
Validate token types, bearer credentials, payload sizes, and lifetimes. Bound caching by both relative and absolute expiry, discard cached credentials before renewal, and disable redirects for federated requests. Keep credential buffers zeroizing and diagnostics free of secrets.
## Testing
Add HTTP fixture tests for concurrent cache reuse, renewal with fresh subjects, expiry across host suspension, failed renewal, oversized payloads, invalid subjects, and unsupported or missing issued token types.
GitOrigin-RevId: e7e6dacd4a834fb73166caf146d4cdc50a895f56
## What changed
Add a `GithubOidc` implementation of `SubjectTokenSource` that requests a fresh token using a bearer credential read from stdin. Validate HTTPS request URLs and replace the audience query parameter while preserving other parameters. CLI integration is not included.
Add a token HTTP client that disables redirects and request logging, applies timeouts, and bounds response bodies. Keep token response buffers in zeroizing storage and exclude credentials, request URLs, and response bodies from errors.
Generalize the locked stdin credential reader for bounded OIDC bearer tokens while preserving API key validation.
## Testing
Add tests for fresh token acquisition, short reads, credential size limits, injection rejection, redirect refusal, and credential-safe errors and logs. Extend API key tests to retain the stricter character restrictions.
GitOrigin-RevId: fd7767178f475b8161734f29d29108c795839185
## What changed
Introduce a shared `Auth` abstraction that supplies sensitive authorization headers and the client's redirect policy. Preserve the existing static API key allocation and default redirect behavior, and return an empty `502` response if header acquisition fails.
Unify the stdin authentication entry point around platform-specific `read_from_stdin` helpers.
## Testing
Add a test verifying that both compact and pretty `Debug` output redact API keys.
GitOrigin-RevId: 6a924534f00c1b209b24e8b0af117091cfa3ec07
Stabilizes the Responses API proxy header test by splitting the coverage
at the right boundary:
- Core integration test now verifies parent/subagent identity headers
directly from captured `/responses` requests.
- Proxy dump unit test now verifies those identity headers are preserved
in dumped request JSON.
- Removes the flaky real proxy process + temp-file dump polling path
from the core test.
This makes Responses API proxy request/response dumping first-class by
adding an optional `--dump-dir` flag that emits paired JSON files with
shared sequence/timestamp prefixes, captures full request and response
headers and records parsed JSON bodies.
`codex-responses-api-proxy` is designed so that there should be exactly
one copy of the API key in memory (that is `mlock`'d on UNIX), but in
practice, I was seeing two when I dumped the process data from
`/proc/$PID/mem`.
It appears that `std::io::stdin()` maintains an internal `BufReader`
that we cannot zero out, so this PR changes the implementation on UNIX
so that we use a low-level `read(2)` instead.
Even though it seems like it would be incredibly unlikely, we also make
this logic tolerant of short reads. Either `\n` or `EOF` must be sent to
signal the end of the key written to stdin.
This PR expands `.github/workflows/rust-release.yml` so that it also
builds and publishes the `npm` module for
`@openai/codex-responses-api-proxy` in addition to `@openai/codex`. Note
both `npm` modules are similar, in that they each contain a single `.js`
file that is a thin launcher around the appropriate native executable.
(Since we have a minimal dependency on Node.js, I also lowered the
minimum version from 20 to 16 and verified that works on my machine.)
As part of this change, we tighten up some of the docs around
`codex-responses-api-proxy` and ensure the details regarding protecting
the `OPENAI_API_KEY` in memory match the implementation.
To test the `npm` build process, I ran:
```
./codex-cli/scripts/build_npm_package.py --package codex-responses-api-proxy --version 0.43.0-alpha.3
```
which stages the `npm` module for `@openai/codex-responses-api-proxy` in
a temp directory, using the binary artifacts from
https://github.com/openai/codex/releases/tag/rust-v0.43.0-alpha.3.
This removes the `codex responses-api-proxy` subcommand in favor of
running it as a standalone CLI.
As part of this change, we:
- remove the dependency on `tokio`/`async/await` as well as `codex_arg0`
- introduce the use of `pre_main_hardening()` so `CODEX_SECURE_MODE=1`
is not required
---
[//]: # (BEGIN SAPLING FOOTER)
Stack created with [Sapling](https://sapling-scm.com). Best reviewed
with [ReviewStack](https://reviewstack.dev/openai/codex/pull/4404).
* #4406
* __->__ #4404
* #4403
This is likely the reason that I saw some conversations "freeze up" when
using the proxy.
Note the client in `core` does not specify a timeout when making
requests to the Responses API, so the proxy should not, either.
Details are in `responses-api-proxy/README.md`, but the key contribution
of this PR is a new subcommand, `codex responses-api-proxy`, which reads
the auth token for use with the OpenAI Responses API from `stdin` at
startup and then proxies `POST` requests to `/v1/responses` over to
`https://api.openai.com/v1/responses`, injecting the auth token as part
of the `Authorization` header.
The expectation is that `codex responses-api-proxy` is launched by a
privileged user who has access to the auth token so that it can be used
by unprivileged users of the Codex CLI on the same host.
If the client only has one user account with `sudo`, one option is to:
- run `sudo codex responses-api-proxy --http-shutdown --server-info
/tmp/server-info.json` to start the server
- record the port written to `/tmp/server-info.json`
- relinquish their `sudo` privileges (which is irreversible!) like so:
```
sudo deluser $USER sudo || sudo gpasswd -d $USER sudo || true
```
- use `codex` with the proxy (see `README.md`)
- when done, make a `GET` request to the server using the `PORT` from
`server-info.json` to shut it down:
```shell
curl --fail --silent --show-error "http://127.0.0.1:$PORT/shutdown"
```
To protect the auth token, we:
- allocate a 1024 byte buffer on the stack and write `"Bearer "` into it
to start
- we then read from `stdin`, copying to the contents into the buffer
after the prefix
- after verifying the input looks good, we create a `String` from that
buffer (so the data is now on the heap)
- we zero out the stack-allocated buffer using
https://crates.io/crates/zeroize so it is not optimized away by the
compiler
- we invoke `.leak()` on the `String` so we can treat its contents as a
`&'static str`, as it will live for the rest of the processs
- on UNIX, we `mlock(2)` the memory backing the `&'static str`
- when using the `&'static str` when building an HTTP request, we use
`HeaderValue::from_static()` to avoid copying the `&str`
- we also invoke `.set_sensitive(true)` on the `HeaderValue`, which in
theory indicates to other parts of the HTTP stack that the header should
be treated with "special care" to avoid leakage:
https://github.com/hyperium/http/blob/439d1c50d71e3be3204b6c4a1bf2255ed78e1f93/src/header/value.rs#L346-L376
ee