Keel Adds Verifiable Gates and Evidence to AI Coding Agents

Summary

Keel is a Rust-based tool for controlling and verifying changes made by AI coding agents, including Claude Code, Codex, Copilot, and Kiro. Users define what “done” means in a specification, approve the specification and plan, let an agent work, and require build, test, review, and human sign-off before a change can pass. Failed checks stop the process and send the work back rather than silently allowing it to ship. Keel records approvals and gate verdicts in a hash-linked evidence chain and can export the specification, verdicts, trajectory, and exact diff as a bundle that can be verified offline without the repository, network, or an account. Its gates cover whether a specification is falsifiable, whether a plan reflects the import-graph blast radius, whether the implementation builds and stays within scope, whether the evidence is reviewable, and whether lessons should become future rules. An optional G2.5 gate reviews test invalidation and security defects, with high- or critical-severity findings blocking the gate; without a model or scanner, it falls back to heuristics. Keel also provides GitHub pull-request checks, a GitHub Actions runtime using a locked-down container and gVisor when available, and integration with the moor sandbox, which keeps the agent away from the host disk and writes the record outside the sandbox. The project reports 541 tests with zero Clippy warnings, 14.6x less retrieved context at 100% recall, and 1.25 seconds to index 5,705 files across eight languages. It discloses that its G2/store-drift check has not failed in 26 runs, that the evidence is based on a small number of repositories and one operator, that Windows support is unexercised, and that fork pull requests cannot yet receive the CI runtime’s bundle commit. Keel v0.10.1 runs on macOS and Linux, is dual-licensed under MIT or Apache-2.0, and installs through Cargo.