|
| 1 | +# Incident lane — Codex sign-in lockout after an applied integration (#5261) |
| 2 | + |
| 3 | +Status: OPEN, dispatched out of band and ahead of the phase 2 bundles. This is a user-reported |
| 4 | +incident, not a roadmap item. |
| 5 | + |
| 6 | +## What was reported |
| 7 | + |
| 8 | +A Windows 11 user on 2.59.0 configured the Codex integration, repeatedly failed to add an account |
| 9 | +pool, then could not call any model from Codex. After restarting, Codex would no longer sign in at |
| 10 | +all: the client showed only "Unable to load sign-in requirements" and a Retry button. The reporter's |
| 11 | +own analysis is that every Codex request was being routed to the local proxy on port 10100 and hung |
| 12 | +there. |
| 13 | + |
| 14 | +## Why this outranks a model-routing bug |
| 15 | + |
| 16 | +The damaging part is not the failed inference. It is that the user is locked out of Codex sign-in |
| 17 | +while the integration stays applied. A reboot does not clear an applied integration, so the lockout |
| 18 | +survives it, and the failure surface offers no recovery the user can act on — the only visible |
| 19 | +control is Retry against the endpoint that is failing. A user in that state cannot reach the |
| 20 | +product that would let them undo the change. |
| 21 | + |
| 22 | +## What the lane must establish from source, not assume |
| 23 | + |
| 24 | +1. What the Codex integration apply path actually writes, and whether its scope covers only |
| 25 | + inference or also the authentication and sign-in bootstrap. |
| 26 | +2. What happens to those requests when the proxy is absent or hung: a bounded timeout, a fail-open |
| 27 | + path, or an indefinite wait. |
| 28 | +3. Whether Windows can reach a state where the integration stays applied while the service is not |
| 29 | + running, including after a reboot. |
| 30 | +4. Whether a recovery path exists that does not require the proxy to be running, and whether a user |
| 31 | + in the failure state can discover it. |
| 32 | +5. The account-pool add failure that triggered the sequence, recorded precisely rather than assumed |
| 33 | + to share a cause. |
| 34 | +6. Whether anything already on `dev` since 2.59.0 changes this path. |
| 35 | + |
| 36 | +## Fix standard |
| 37 | + |
| 38 | +Having the integration applied must not be able to lock a user out of Codex sign-in. Either the |
| 39 | +authentication and sign-in bootstrap stays off the proxy path, or an unavailable proxy produces a |
| 40 | +detectable failure and a discoverable recovery. Whichever holds, it is fixed by a regression that |
| 41 | +simulates the dead-proxy state. |
| 42 | + |
| 43 | +## Execution constraint specific to this lane |
| 44 | + |
| 45 | +The incident is a configuration change that locked a user out. The lane therefore may not run the |
| 46 | +proxy, start or restart the service, or modify any credential or configuration file on the working |
| 47 | +machine; doing so would reproduce the damage locally. Verification is static source review plus |
| 48 | +exact-head hosted CI. |
0 commit comments