pluxx-autopilot

name	pluxx-autopilot
description	Run the one-shot Pluxx import, refinement, and verification path.

Pluxx Autopilot

Use this skill when the user wants the one-shot path instead of a manually staged import and refinement flow.

Workflow

1. Identify the source and desired scope:
   • remote MCP URL
   • local stdio command
   • current working project
   • desired targets
2. Make the mode explicit:
   • quick
   • standard
   • thorough
3. Run the one-shot path:
   • pluxx autopilot --from-mcp ... --yes
   • include runner, mode, and review flags when they matter
   • use pluxx autopilot --resume for an interrupted compatible run
   • use pluxx autopilot --rollback when the user wants to abandon an interrupted run; rollback is local and does not require source or runner access
4. Summarize:
   • files created or updated
   • which agent passes ran
   • which verification checks ran
   • what still needs human follow-up

Rules

• Do not present autopilot as magic. Explain what it actually ran.
• Keep auth flags explicit when the MCP requires them.
• Treat autopilot as a native-shaping pass, not just a file dump:
  • promote repeated entrypoints into commands when that is stronger than a flat skill list
  • promote isolated reviewer/research/delegated work into agents or subagents when the host supports that more naturally
  • keep argument-rich workflows argument-rich instead of flattening them into vague generic commands
• If the user is learning or debugging, suggest the manual lifecycle when autopilot hides too much.
• If autopilot fails mid-run, say which stage failed:
  • auth
  • introspection
  • baseline
  • runner
  • boundary
  • review
  • verification (with failurePhase: post-agent-verification for failures after refinement)
• Preserve Autopilot recovery metadata while resume or rollback is still needed, and do not publish its raw contents.
• Resume only a compatible saved run. If behavior-affecting inputs changed, stop and start a new run instead of weakening the fingerprint check.

Output

• Explain what autopilot actually did.
• Call out what still needs human review.
• Make the next step obvious:
  • refine more
  • build/install
  • sync later
  • or ship