v1.3.0 — Claude Code provider

Added

• provider input: cursor (default) or claude. provider: claude runs Claude Code's built-in /code-review, /security-review and /simplify skills through the Claude Agent SDK and posts findings as inline review comments, like Cursor.
• claude-code-oauth-token (from claude setup-token, Claude Pro/Max) and anthropic-api-key for Claude auth.
• review-level (low | medium | high | xhigh | max), passed to Claude's /code-review.
• auto-describe works on both providers.

- uses: Vilancer/bugbit@v1
  with:
    provider: claude
    claude-code-oauth-token: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
    review-modes: code-review, security-review, simplify
    review-level: medium

Changed

• cursor-api-key is only required when provider is cursor. model defaults to composer-2.5 on Cursor and to Claude Code's default model on Claude.

Security

• Claude sessions are read-only: one plain read-only git command at a time, no file edits or web access, and no reads of .git/config, .env*, keys or /proc.
• Neither provider loads settings, rules, hooks or MCP servers from the PR branch.
• bugbit removes its credential inputs and the Actions runtime token from the environment before any agent starts.
• On both providers, comments, descriptions and labels are stripped of the GitHub token, the provider key and token-shaped strings before they are posted.
• Cursor's shell can't be restricted per command (its SDK has no hooks, and its sandbox doesn't start on GitHub-hosted runners); the scrubbing above is the backstop.

@v1 is updated to this release.