Hey HN, I think session transcripts written by coding agents like Claude Code and Codex are very interesting because they offer a detailed window into how work gets shipped. You can see the sequence of decisions that resulted in the final PR, what the agent got wrong, tools used etc. So I built a cli that analyzes these sessions and provides a local dashboard that shows what each session shipped (PRs, features), how much each PR cost, and recommendations for more effective usage. Concretely, it enriches each session with: - Outcome links: merged PRs, features shipped, files changed - Granular cost attribution to outcomes - Task complexity - Agent autonomy - Work type - Key decisions - Tool error categories and across sessions, identifies: - Agent rework / re-steer themes - Patterns of deviations from best practices Combined with the data already in the transcript like model, agent harness and repo, this data lets you answer questions like: - How much of my AI spend went into PR #2, or feature X? - Are my agents getting more autonomous over time on complex tasks? - What's my success rate on repo X vs. repo Y (or any other dimension you care about?) Works with Claude Code, Codex, OpenCode, and Pi. Everything runs and stays on your machine; enrichments that need an LLM can use your own provider key or a local model. The repo is at https://github.com/tuneloop/tuneloop, and you can try it by running `npx tuneloop@latest analyze`. Some things that are in the works next: - Skill invocation analysis – was re-work needed after a skill invocation? - harness/model comparisons on merged PRs – creates swe-bench style tasks from merged PRs on internal repos to help compare model/harness choices. - a version that offers this visibility at the team level. I’d love to hear from folks here if you find this interesting. Do you look at session transcripts much (perhaps via `/insights` on Claude Code or otherwise)? what do you find useful to track and actionable? |