blazeup-ai/pi-insights

Pi Insights is an extension for the Pi coding agent that analyzes your session history and generates a report about how you are actually using the tool. If you work with Pi across many sessions over weeks or months, this gives you a structured look at where your time and money are going and where your workflow could improve. The report covers several things. Basic statistics include token usage, cost, lines of code changed, commits made, tool errors, and parallel session counts. A model spend section identifies cases where you are using an expensive model on simple tasks or a cheaper model on complex ones that it handles poorly, and it estimates potential savings from adjusting. A suggestions section recommends Pi features, skills, or configuration changes based on your actual projects and tools, pulling from your existing setup so it does not suggest things you already have. A separate section flags patterns that are costing you time or money and suggests concrete alternatives. What distinguishes this from simpler analytics tools, according to the README, is that it tracks changes over time rather than producing a static snapshot. It compares the current week against previous weeks, weights recent sessions more heavily when computing satisfaction and friction metrics, detects whether costs and errors are trending better or worse, and only surfaces friction points that are still ongoing rather than problems you have already resolved. The pipeline works in five phases: scanning session log files, extracting deterministic stats from each session, using an LLM to classify goals and outcomes per session, aggregating everything with decay weighting and anomaly detection, and finally generating a self-contained HTML report (or Markdown if preferred) using a set of parallel LLM prompts. Results are cached so repeated runs are fast. Installation is one command within Pi. The extension requires Pi version 0.74.0 or later and an active AI model configured in Pi. It is released under the AGPL-3.0 license.