AgriciDaniel/sync ? reverse-engineered prompt

Build me a local first Python tool that acts like a control plane for Claude Code and Codex.

I want to be able to pick a task role, have it routed to the right model and effort level, then run through a shared workspace where the brief, plan, evidence, and final verdict are all saved on disk. The key thing is that decisions should be based on what actually ran and passed, not just model output. Include a command line experience for checking setup, printing the current policy, routing tasks, running sandboxed checks, recording outcomes, and measuring results over time.

Please make it feel safe and evidence gated, with redacted logs, a read only gate for checks, and clear handling for invalid or incomplete runs. If there are current docs or best practices for Claude Code or Codex that help, look them up online and use them. Keep it practical, reproducible, and focused on real local use.