codex

mirror of https://github.com/pchuan98/codex.git synced 2026-07-01 00:31:56 +08:00

Files

T

cooper-oai d1aaf789ad [login] revoke existing auth before starting login (#27674 )

## Why

`codex login` previously persisted newly issued OAuth credentials and
only then attempted to revoke the superseded refresh token. The old
credential must be revoked before a replacement browser or device-code
flow starts, and successful login must not perform any post-login
revocation attempt.

## What changed

- Revoke and clear existing stored auth before browser or device-code
CLI login begins.
- Remove superseded-token detection and revocation from the shared token
persistence path; successful login now only saves the new credentials.
- Read the raw configured auth store during CLI cleanup so
environment-provided auth cannot mask the stored refresh token.
- Preserve `auto` storage fallback semantics when keyring deletion fails
by clearing the fallback auth file.
- Add a process-level CLI regression test that requires the revoke
request to precede every device-login request and occur exactly once.

If replacement login is canceled or fails, the previous local
credentials have already been cleared. Remote revocation remains best
effort, matching explicit logout behavior.

## Validation

### Process-level before/after reproduction

I compiled the real `codex` CLI from the pre-fix parent (`14df0e8833`)
and from the PR implementation (`25c002f23b`; the login behavior is
unchanged at the current head), then ran the same device-code flow
against a local HTTP mock OAuth authority.

Each run:

1. Used a fresh temporary `CODEX_HOME` configured with
`cli_auth_credentials_store = "file"`.
2. Seeded that temporary home with managed ChatGPT auth containing
`old-access` and `old-refresh` tokens.
3. Pointed `CODEX_REVOKE_TOKEN_URL_OVERRIDE` at the mock `/oauth/revoke`
endpoint.
4. Ran the compiled CLI as:

```shell
CODEX_HOME=<temporary-home> \
CODEX_REVOKE_TOKEN_URL_OVERRIDE=<mock-issuer>/oauth/revoke \
<compiled-codex> login --device-auth --experimental_issuer <mock-issuer>
```

5. Recorded every request received by the mock authority. The mock
marked `new-access` valid when `/oauth/token` issued it and invalidated
it if `/oauth/revoke` arrived afterward, reproducing the observed
session-invalidating failure mode. After login exited, the harness also
verified the persisted refresh token and probed a protected endpoint
with `new-access`.

| Build | Observed request order | CLI/persistence result | `new-access`
probe |
| --- | --- | --- | --- |
| Pre-fix | `usercode → device token → OAuth token →
revoke(old-refresh)` | Exit `0`; `new-refresh` persisted | `401` |
| PR | `revoke(old-refresh) → usercode → device token → OAuth token` |
Exit `0`; `new-refresh` persisted | `200` |

The PR run therefore issued exactly one revocation request, before any
request that initiated the replacement login, and issued no revocation
after token exchange.

### Regression coverage

`codex-rs/cli/tests/login.rs::device_login_revokes_existing_auth_before_requesting_new_tokens`
runs the real first-party `codex` binary against a `wiremock` OAuth
server with an isolated temporary `CODEX_HOME`. It asserts:

- the exact request sequence is `/oauth/revoke`,
`/api/accounts/deviceauth/usercode`, `/api/accounts/deviceauth/token`,
then `/oauth/token`;
- there is exactly one revoke request and its body contains
`old-refresh` with the `refresh_token` hint;
- the completed login persists `new-refresh`.

Local validation:

- `just test -p codex-login` — 130 passed
- `just test -p codex-cli` — 280 passed, including the new process-level
regression test
- `just bazel-lock-check`

d1aaf789ad · 2026-06-12 12:38:30 -07:00

History

app_server.rs

config: add strict config parsing (#20559 )

2026-05-13 16:08:05 +00:00

debug_clear_memories.rs

Move memory state to a dedicated SQLite DB (#24591 )

2026-05-26 20:07:25 +02:00

debug_models.rs

Add codex debug models to show model catalog (#18625 )

2026-04-20 05:42:22 +00:00

delete.rs