Trajectory segmenter for coding agent logs

Slice long Claude Code sessions into readable phases so you can see where it went off the rails

Problem

A developer spent weeks trying to write a simple paper with Claude and, despite heavy context engineering and material prep, the output degraded into an unintelligible word salad. Meanwhile the community keeps building tools to make AI output understandable because nobody wants to read the entire long tail of decisions and garbage. The problem is not model strength — it is that once a session grows long, humans lose any sense of when it started drifting.

Solution

Parse local Claude Code session logs, automatically split them into phases based on goal switches, tool-call bursts, retry storms and context trimming, and show duration, token burn and success rate per phase. A timeline highlights likely drift points, and any phase can be exported for review or a bug report.

Users

Indie developers and small teams who lean heavily on Claude Code, Cursor and similar coding agents

MVP

Read local session logs; rule-based phase segmentation; timeline view with duration and tokens; mark suspected drift points; export a single phase as markdown

Revenue

Free base version; subscription unlocks multi-project comparison, team sharing and report export; also an entry point for enterprise agent observability

Why now

Coding agent sessions are getting longer and more expensive while observability is still a log dump; local log formats are stable, so no server or big data is needed