tree-trace/treetrace

When you work with an AI coding agent, you spend time steering it: correcting a wrong assumption, pulling it back from a risky approach, repeating a constraint it ignored, or redirecting it after it went down a dead end. That steering information disappears at the end of the session. TreeTrace captures it locally and converts it into regression tests, failure records, and a memory pack the next agent can read before it starts. You run npx treetrace inside your project directory. It finds Claude Code session files automatically, or you can point it at any conversation file with a flag. It builds a tree of the session: which paths were accepted, which were abandoned, where corrections happened, and what the correction said. From that tree it produces several output files. The failure log records moments where the agent did something problematic and what fixed it. The eval file contains model-agnostic test cases that a CI system or eval harness can check against. The lessons file summarizes what went wrong. The agent memory file is a compact document the next agent can read so it starts the session already aware of past constraints and decisions. Security tracking is a specific focus. The tool watches for moments when an agent touched authentication code, printed a secret, loosened access control, skipped a test, or introduced a potential injection point. When it finds one of those patterns and a human correction follows, it records the pair as a regression eval so the same mistake can be caught in future sessions. All analysis is done locally using rule-based heuristics, not an AI model. There is no judge, no API call, and no upload. Secrets found in the output are flagged and must be resolved before any files are written. The tool has zero runtime dependencies and works on Node.js 18 or newer. An MCP server is included so AI coding tools can query the stored memory and eval data during a session without leaving the tool.