Expose Claude skills through available commands

[benjaminshoemaker]

Summary

Refs zed-industries/zed#55983.

This updates @anthropic-ai/claude-agent-sdk to 0.2.139, whose supportedCommands() output includes Claude skills from .claude/skills. Since claude-agent-acp already publishes supportedCommands() as ACP available_commands_update, this makes Claude skills discoverable in ACP clients such as Zed's Agent Panel using Claude's native /skill-name convention.

The newer SDK also adds a permission_denied system message variant, so this PR handles it with the existing ignored system-status events to keep the message handling exhaustive.

Testing

• npm run build
• npm run test:run
• npm run check
• git diff --check
• SDK behavior check with a temporary .claude/skills/test-skill/SKILL.md:
  • query(...).supportedCommands() returned test-skill with its description and argument hint.
• Manual Zed Agent Panel verification with a temporarily overlaid local claude-agent-acp package:
  • Opened ai_coding_project_base, which has project skills under .claude/skills.
  • Started an external Claude Agent thread.
  • Verified the composer placeholder says / for commands.
  • Typed / and saw project skills in the completion menu, including auto-verify, verify-task, ui-ux-pro-max, discover, and go.
  • Did not submit a prompt.
  • Restored Zed's cached package afterward.

Screenshots

I verified this visually in Zed's Agent Panel, but did not commit screenshots into the repo. The relevant visible state was the slash-command completion menu listing project skills such as auto-verify, verify-task, ui-ux-pro-max, discover, and go in a Claude Agent thread.

Notes

This intentionally uses Claude Agent SDK's existing command discovery instead of adding a parallel skill scanner in the ACP adapter.

Release Notes:

• Improved Claude skill discoverability in ACP clients that expose slash command completions.

[benjaminshoemaker]

Here's what it looks like in the agent view