Caliber scores, auto-generates AI agent configs

This is the kind of unglamorous AI infra that actually compounds: not another prompt template, but a way to keep context from rotting. If Caliber stays accurate across real codebases, it could become the linting layer for agent instructions.

• –The 0-100 score is the wedge, because it turns stale-context anxiety into a deterministic audit.
• –Support for Claude Code, Cursor, Codex, OpenCode, and Copilot makes it useful across teams instead of trapping them in one assistant.
• –The review, diff, backup, and undo flow is what makes auto-regeneration safe enough for real repos.
• –The local-first angle is smart for model runners and private codebases, where keeping code on-machine matters.
• –Hitting 250 stars suggests the pain point is real, not just novelty.