Jeffy Loop: The Coding Agent That Won’t Let Itself Lie

A developer has built Jeffy Loop, a coding agent for Claude Code that enforces a disciplined audit-fix-verify workflow, preventing the model from claiming success without independent verification. The agent has converged on 52 open-source projects across 13 languages, with upstream fixes accepted in projects like bat, fasthttp, and jsoncpp, and has uncovered high-severity bugs behind green test suites.

Most autonomous coding agents are optimistic. They audit, fix, declare victory, and sometimes the victory is mostly vibes. Jeffy Loop takes the opposite stance. Built for Claude Code, it forces the model to act like a disciplined principal engineer: audit first, fix one verified task at a time, checkpoint everything, and refuse to stop until “done” survives independent checks. It starts from Geoffrey Huntley’s Ralph technique re-feeding one prompt in a loop and wraps it in real engineering method. Type /jeffy 10 and walk away. Map the public surface and run a breadth-first audit. Turn every finding into a backlog item with a runnable acceptance check. Execute one verified, checkpointed task per iteration. Revert any change that breaks the project’s own tests via a verify gate. Only stop when a fresh audit is clean, an adversarial evaluator countersigns, and a shell script re-checks the claim. “Done” is machine-enforced, not the model’s opinion. Real receipts 52 open-source projects converged across 13 languages 31 non-convergences fully documented Upstream fixes accepted in bat, fasthttp, jsoncpp, PapaParse, and chalk These include real High-severity bugs found behind green test suites in popular projects. Why it stands out It cannot claim it looked at code it never examined. Lessons become permanent rules. Progress means actual code movement. The engine itself is held to hundreds of behavioural checks. Prefer short, fresh-context runs over long sessions. Context pressure, stalls, and thrashing are all measured and cut short. If you’ve been burned by agents that rewrite working code or quietly delete failing tests, this is a different bet: process-heavy, transparent, and forced to prove its claims.