feat(skills): add agentation-fixer skill for watch & auto-fix loop

[armatrix]

Summary

The two-session workflow currently requires the Fixer (Terminal 2) to be set up with a manual prompt pasted into Claude Code. This PR adds a proper skill that encapsulates the full watch-and-fix loop.

What's included:

• skills/agentation-fixer/SKILL.md — new skill with:

  • One-shot mode (/agentation-fixer check): drain pending annotations, fix each, report summary
  • Watch loop (/agentation-fixer watch): continuous agentation_watch_annotations loop that auto-fixes annotations as they arrive
  • Intent routing: handles fix/change/question/approve annotations differently
  • Severity ordering: processes blocking before important before suggestion
  • Clarification flow: uses agentation_reply when annotation is ambiguous instead of guessing
  • allowed-tools declaration for all required MCP tools
  • argument-hint for CLI discoverability

• Updated two-session-workflow.md to reference the new skill (with manual prompt kept as fallback)

Motivation

The Critic (/agentation-self-driving) has a full skill with detailed instructions for browser automation. The Fixer side is equally important but was left as a raw prompt — which means every user has to figure out the watch loop, intent routing, and error handling on their own. This skill fills that gap.

Test plan

• Install skill: ln -s "$(pwd)/skills/agentation-fixer" ~/.claude/skills/agentation-fixer
• Run /agentation-fixer check with pending annotations → should fix and resolve each
• Run /agentation-fixer watch → should block and process annotations as they arrive
• Test with agentation-self-driving in a two-session setup → Critic creates, Fixer fixes
• Verify agentation_reply is used for ambiguous annotations instead of bad fixes

[vercel]

@armatrix is attempting to deploy a commit to the Benji Taylor's Projects Team on Vercel.

A member of the Team first needs to authorize it.