Fixed
- Forced sleep (explicit sleep request or forced eviction) stranded deferred captures in the spool; the drain only ran on a natural drowsy transition. Deferred captures now drain on every entry into sleep. Thanks @Volevanius (#115).
- The bench token-cost suite errored with
ModuleNotFoundError: anthropicwhenANTHROPIC_API_KEYis set but the (optional since v7.5) SDK isn't installed — every shell inside a Claude Code session. The exact-counter path now gates on the SDK being importable and falls back to the tiktoken proxy. Thanks @gepenniman (#123).
Changed
- The daemon now binds its socket and starts serving before the embedder finishes building, under a single boot-anchored deadline, so a cold start no longer blocks the first response on model load.
Full changelog: CHANGELOG.md