v0.6.3

Thank you @onexoluxion for contributing PRs #29, #30, #31, #32, #33, #35, #36, and #37. Your Windows testing and detailed bug reports made this maintenance release possible.

Changed

• Claude Desktop launches now expose the selected model plus saved favorites — relay-ai claude-app puts the selected model first, appends available favorites in saved order, removes duplicates, and caps the catalog at 20 models total. Stale or unauthorized favorites are skipped with a warning, anonymous providers remain supported, and mixed regular/Cloud Code catalogs preserve routing order and OAuth provider identity. Claude Desktop still controls which discovered model it initially activates.
• Claude Desktop Favorites launches use the same catalog path everywhere — interactive Favorites selection and Admin UI Favorites launches now choose an available favorite as the starting model while exposing the remaining compatible favorites through the same ordered catalog.

Fixed

• Headless relay-ai claude launches fail clearly instead of crashing (#29) — when no provider/model is resolved and an interactive wizard would be required, non-TTY environments now receive an actionable error explaining how to pass --provider and --model.
• Windows builds use a cross-platform UI asset copy step (#30) — npm run build no longer depends on POSIX-only mkdir -p and cp commands. Docker builds also copy the required scripts/ directory before running the build.
• Claude Desktop is detected in the Windows AnthropicClaude install directory (#31) — Squirrel installations under %LOCALAPPDATA%\AnthropicClaude can now be launched normally.
• Concurrent transparent-proxy launches no longer delete a session that is still being created (#33) — cleanup gives incomplete session directories a grace period and no longer deletes them after transient read errors.
• Claude Desktop and Codex/ChatGPT app help works without a TTY (#35) — relay-ai claude-app --help and relay-ai codex-app --help print help instead of falling through to the interactive-terminal guard.
• Fresh Claude launches no longer inherit parent-session identity (#36) — Relay strips Claude Code's parent/child session markers before spawning a new session, preventing it from being misidentified as nested and preserving normal top-level behavior such as transcript saving.
• Claude Desktop cleanup no longer clobbers another live Relay session (#37) — cleanup and --restore now respect session ownership. Racing launches preserve the original _meta.json backup, lock writes are atomic, and config files remain in place while the active metadata still references them.
• Claude Desktop self-heals from a corrupt session lock instead of getting stuck — a .relay-ai.lock that fails to parse (torn write, disk-full, manual edit) used to be treated as a permanently live session: every future relay-ai claude-app launch refused to start, auto-recovery never kicked in, and --restore refused to fix it. The lock is written atomically (temp file + rename), so a corrupt lock at the real path can never belong to a running session — it's now treated like a stale lock and cleared automatically.
• Claude Desktop no longer duplicates every 1M-context Relay model — relay-ai claude-app now exposes one context-accurate picker entry per model: sub-1M models remain normal entries, while models with at least one million tokens appear once with a 1M label and retain their exact context limit.

Maintenance

• Windows-specific test assertions now normalize paths and platform behavior so the same suite passes on Windows and POSIX systems (#32).