Vigil

Problem Debugging is slow, manual work: find the failing test, trace the bug, write a fix, re-test. AI coding tools can speed this up, but they can also change code on their own. Teams can't fully trust those changes: nobody reviewed them, there's no proof they fixed anything, and there's no record of who approved them. Wrong AI fixes cause errors and rework, so teams must choose between AI speed and control. Solution Vigil is a multi-agent bug-fixing tool where AI does the investigation and a human makes the decision. Nothing changes in the code until a developer approves it, and real tests prove the fix worked. Target users Development teams adopting AI coding tools who need oversight, engineering leads responsible for code quality, and regulated or high-stakes teams that need a record of every change. How users interact with it Scan: Vigil runs the project's tests and finds the failing one on its own. No bug report needed. Analyse: A Debug Agent and a Test Agent investigate in parallel. A Supervisor cross-checks their findings, chooses one fix and highlights the exact buggy line. Review: The developer sees the diff, reasoning and evidence, marked "Proposed: nothing has been applied." Approve or Reject: On approval, Vigil applies the fix on a separate git branch and runs the real tests. On rejection, nothing changes. Audit: Every decision is saved in an audit trail, exportable as Markdown or JSON. What makes it creative and unique Agents that check each other: a fix is chosen because the Test Agent's evidence corroborates the Debug Agent, not on confidence alone. Hash-locked approval: each proposal has a unique SHA-256 fingerprint. If the patch changes after review, the approval is refused, so you only approve what you saw. Proof, not promises: real test output is shown, failing before and passing after. Honest by design: broken or do-nothing fixes are rejected automatically, and the interface never claims something happened when it didn't. Built the way it works: IBM Bob 2.0 was our main builder. We used Agent mode and task lists to plan, edit and test step by step, approving every edit, the same human-in-the-loop idea Vigil brings to teams. How it addresses the issue effectively and efficiently Vigil takes over the manual steps of debugging: finding the failing test, locating the bug, writing the fix, branching, re-testing and documenting. The developer makes one decision instead. Every fix must pass real tests before anyone relies on it, so bad fixes are caught early, reducing errors and rework. Approvals are locked and logged, giving teams a verifiable record of what changed, why, and who approved it, without slowing down. Vigil is a working prototype, live at vigil-site.onrender.com and backed by 271 automated tests covering the approval gate, concurrent approvals and patching against real git repositories. Agents are read-only, secrets are redacted from logs, and Vigil only works inside allowed project folders.