[itsrayeedscodespace]

Fresh cached credentials on AsyncAnthropic currently require a worker-pool round trip when the credential provider is synchronous. Unrelated blocking work can therefore stall a request before it reaches the HTTP transport, even though its token is already cached and needs no refresh.

Return fresh cached tokens under the existing cache lock before dispatching work. Cache misses, the advisory/mandatory windows, per-request (expires_at=0) providers and invalidation continue through the existing worker path. Provider calls remain off the event loop. Tokens with no expiry retain their existing caching behavior.

The regression uses AsyncAnthropic.models.list(), MockTransport and a real single-worker default executor. After priming the cache, an unrelated task occupies the worker until the request finishes. On upstream both fresh-token cases time out at the five-second test guard; with the change both requests complete before the worker is released. The test owns its loop/executor and releases the blocker in finally.

Validation on Apple M5/macOS arm64, Python 3.11.18:

• New cases on unchanged upstream: 2 failed, 7 passed.
• Full tests/lib/test_credentials.py: 456 passed, 2 skipped.
• Added boundary checks at 121/120/60/30/0/-1 seconds remaining, invalidation/force-refresh and per-request tokens; synchronous provider execution stays off-loop.
• A separate public-client Trio/AnyIO reproduction also blocks on upstream's occupied worker pool and completes with the change, with one provider call in both cases.
• Changed-file Ruff lint/format, Pyright (zero diagnostics), and git diff --check passed.

Seven warmed samples, alternating order, synthetic credentials and in-memory transport (no live API calls):

Case	Before	After
Cached lookup, future expiry	27.24 µs	0.18 µs
models.list(), future expiry, free worker	172.29 µs	128.52 µs
models.list(), future expiry, worker occupied for 100 ms	101.86 ms	0.73 ms

The last case demonstrates isolation from unrelated worker contention; it is an intentionally constrained executor, not a production throughput claim. With the normal network path, network latency can dominate the small scheduling savings. The fast path briefly acquires the existing mutex and reads the configured clock on the loop, just as the async-provider path already does.

Searched open/closed TokenCache, async_get_token, credential/cache/executor/worker reports and local commit history. #2058 changes fallback after advisory-refresh failure; closed #1841 concerned invalidation during refresh. Their diffs do not implement this fresh-cache path. Workload identity provider proposals #1903/#2010 address a different provider interface.

AI assistance was used for implementation, reproduction and verification.