VeriContext – Preventing Stale Documentation for LLM Agents

Directory-level READMEs / AGENTS.md (“deep init” style) are great early on, but they inevitably drift as the codebase grows. Agents still treat those docs as truth, which makes decisions worse over time.

VeriContext tries to make those docs verifiable.

When a doc references code, it embeds a SHA-256 hash of the exact snippet inside an HTML comment. On verification, either it matches or it fails (fail-closed).

What’s useful in practice is the workflow around it: • You can install it as an agent skill, so the agent runs verification before finishing a task / finalizing a plan / committing, and updates the citations/hashes when it edits docs. • You can also enforce it mechanically via pre-commit hooks or CI, e.g. npx vericontext verify workspace ..., so stale docs simply can’t be merged.

This way docs aren’t “trusted” — they’re constrained to stay aligned with code.

I’d love feedback on trade-offs (strict fail-closed vs. ergonomics), and whether AST-aware verification would be worth it vs. raw snippet hashing.