kqiu10/sherlog

Sherlog is a tool that analyzes application failure logs, identifies why something broke, proposes a code fix, and then actually tests that fix to confirm it works. It uses multiple AI agents working together, built with a library called LangGraph and powered by Anthropic's Claude. One agent handles diagnosis, another proposes a fix, and a third verifies the result by running the project's existing test suite against a temporary copy of the code. The process starts when you give Sherlog a failure log from a broken build or crashed application. The tool can retrieve similar past failures from a database, read the actual source code files involved through read-only access (so it cannot change your code during investigation), and identify the root cause. It then proposes a specific change: which file, which line to remove, and what to replace it with. The verification step is what makes this different from asking an AI to look at a bug. Instead of an AI opinion about whether a fix seems right, Sherlog copies your project to a temporary folder, applies the proposed patch, and runs your actual test command. If the tests pass, the fix is accepted. If they fail, the loop feeds the failure back and tries again, up to a configurable limit. The original project is never modified. The README reports benchmark results where the test-based verifier caught all broken fixes in a 14-case set, while a purely AI-based opinion check missed one. The project requires Python 3.11 or later, Docker to run the PostgreSQL database that stores past incidents, and an Anthropic API key. Installation uses the "uv" package manager. A quickstart example shows Sherlog diagnosing a simple arithmetic bug: it reads the source, proposes a one-line patch, applies it in a sandbox, and reports a passing test. The project is in early development. The core pipeline from log ingestion through diagnosis, fix proposal, and self-correcting verification is described as functional. The license is MIT.