Per-project permission manifest (.vibe-sec.json)

[kobzevvv]

Developers want per-project security policies. A scraper legitimately uses curl. A wallet app should never touch network. Most-requested missing feature in AI agent security. Proposal: .vibe-sec.json in project root with allow/deny patterns, protected_paths, committable to version control. Hook reads it when present, project rules take precedence over global allowlist.

[darfaz]

Love this approach! Per-project permission manifests are exactly what the AI agent ecosystem needs. We built something similar in ClawMoat — it uses YAML-based policy files to define what shell commands, file paths, network endpoints, and browser actions an agent is allowed to use. Would be great to see interoperability between these approaches. Happy to discuss how ClawMoat's policy engine works if it's useful for this design.