clerk-js: SessionTokenCache serves server-expired session tokens when the device clock runs behind (regression of #672) #10185

Preliminary Checks

• I have reviewed the documentation: https://clerk.com/docs
• I have searched for existing issues: https://github.com/clerk/javascript/issues
• I have not already reached out to Clerk support via email or Discord
• This issue is not a question, general help request, or anything other than a bug report directly related to Clerk.

Reproduction

https://gist.github.com/AKhozya/06c6aab6ed3416e50fc8227823ac0a6c

Publishable key

Not included: the bug is in clerk-js's token cache and does not depend on an instance. Happy to share one if it helps.

Description

If the device clock runs behind the server, SessionTokenCache.get() keeps returning a session token after the token has expired by server time. Once the token is more than 5 s past exp, the backend's default allowance (DEFAULT_CLOCK_SKEW_IN_MS, packages/backend/src/jwt/verifyJwt.ts:20) rejects it.

Cause. When a token resolves, tokenCache.ts:411 sets live.createdAt = issuedAt, the JWT iat in server time. get() (line 194) computes elapsed = deviceNow - createdAt, so remainingTtl = (exp - iat) - (deviceNow - iat) = exp - deviceNow. That compares server time with device time directly. If the device runs S seconds behind, the cache serves the token for S seconds after it expired. The deletion timer (line 409, expiresAt - deviceNow) fires S seconds late as well.

History.

protectSession.ts:291 already describes this hazard for another store and avoids it by measuring elapsed time on one clock.

Steps to reproduce: apply the test from the gist to packages/clerk-js/src/core/__tests__/tokenCache.test.ts on main 44b09ba and run that file.

1. The server issues a token at T with a 60 s lifetime.
2. The device clock reads T − 40 when the token is cached.
3. Advance 50 s: get() returns the token (correct, 10 s remain).
4. Advance 20 s more: by server time the token expired 10 s ago.

Expected behavior: at step 4, get() returns undefined, as it does in the control test with no skew.

Actual behavior: at step 4, get() still returns the token. Result: Tests 1 failed | 69 passed (70); only the skew test fails.

Effect in an app. Our Expo app sends the session token to our own Express backend, which calls verifyToken and answers a failed verification with 401.

Our workaround, not yet deployed: clockSkewInMs: 60_000 on verifyToken, and one retry with getToken({ skipCache: true }) after a 401.

On a fix. If createdAt takes the device clock when the token resolves, the new test passes. #672 uses the device clock at caching time. If a token is already partway through its lifetime when it reaches the cache, get() counts its full lifetime from that moment. If the clocks agree, the deletion timer (exp minus device time) still evicts such a token at its real expiry, so get() serves it late only if that timer runs late. It also fails one existing test, broadcast resilience > a failing postMessage does not evict the freshly cached token, because that test's fixture has iat after exp. A fix that handles both a device clock running behind and a token that reaches the cache partway through its lifetime may need the offset between the server clock and the device clock, for example measured when a token is freshly minted. We have not tested such a fix.

Environment