Caliber open-sources AI agent config stack

The pitch is solid because it attacks the boring but expensive layer everyone rebuilds, and the deterministic scoring angle makes it feel more like infrastructure than prompt fluff. The strongest wedge is repo-aware maintenance of `CLAUDE.md`, `.cursor/rules/`, `AGENTS.md`, and `copilot-instructions.md` across multiple assistants. Deterministic audits are more credible than another LLM-generated setup wizard; catching stale paths, missing files, and config drift is genuinely useful. The open question is scope: config management alone may be too narrow unless it also handles secrets, run-state, evals, and debugging context. Integrations with LangChain or AutoGPT make sense, but only if Caliber stays the source of truth instead of becoming one more wrapper layer. The community feedback ask is the right move here, because the product will live or die on whether it covers real production edge cases, not just clean demos.