v0.2.2

Added

• Support native GPT reasoning effort controls without synthesizing reasoning text.
• Route requests by model capability and available account capacity while retaining suitable session affinity.
• Add reversible Codex CLI and Claude Code setup, diagnosis, status, and restoration commands. Automatic setup and restoration are Linux-only; read-only commands remain cross-platform.

Fixed

• Keep catalog refresh and failed-account recovery off the request path, with single-flight initialization and bounded warm-up waits.
• Validate authentication and API regions before outbound requests, reject malformed credential imports, and preserve the last-good credential snapshot on source-read failures.
• Meter credits per physical generation on the originating account, handle repeated snapshots without double counting, and distinguish valid non-credit completion markers from malformed metering frames.
• Bind durable authentication and quota state to the current login, discard stale refresh results, and apply the same quota freshness rules to account eligibility and routing weight.
• Preserve concurrent client edits and recovery journals during setup failures, bound diagnostic responses and configuration files, clear competing Claude credentials and provider selectors, and reject unsafe plaintext proxy configurations.
• Avoid exposing account identifiers in capacity errors and prevent late requests from mutating replacement accounts or refreshing their affinity.

Other

• Use the short-lived GitHub Actions token for draft releases, validating all four binary artifacts and their checksums before publication.