What is mindlas?

Mindlas watches your AI coding assistant during long sessions and warns you when its work starts going off track. The idea is simple: a coding agent rarely fails right away. It fails slowly, dozens of turns in, as the conversation fills with noise and the original task gets lost. Mindlas reads the session live, spots that drift building up, and hands the agent a specific correction. Everything runs locally on your machine, so none of your code or session data ever leaves. The tool watches four specific problems. Context Rot tracks when the working context deteriorates, not just when the window is full. Verification Debt catches code changes that were never actually checked. Change Blast Radius measures when edits spread too far across files. Tool Failure Loop detects when the same tool keeps failing repeatedly. Each problem has one gauge and one paired correction. The gauges run on a simple loop: measure, trigger, explain, correct, remeasure, and report. Scoring is deterministic. The same session data always produces the same score, with no AI model grading another model's work and no network calls involved. Corrections change state directly rather than just prompting the agent to fix itself, which the developers found did not work in practice. Every correction records a before and after measurement to prove it did something. Setup takes about a minute if you have Python 3.11 or newer, git, and Claude Code installed. You can clone the repository and run a demo command to see a simulated alert and watch a correction bring a score from 100 down to 4. For live use, you install hooks that record session events and a status line that shows the gauges under your prompt. There are two install paths, a plugin option and a CLI option, and you pick one. Mindlas is honest about its limits. It does not guarantee correctness, prevent hallucinations, replace tests, or know what the model is thinking. Its claim is narrow: detect known deterioration causes from local signals, apply corrections, record the effects, and produce a scorecard you can export. The full README is longer than what was shown.