claude-code-trace

## [0.18.0] — 2026-09-20

Claude Code Trace can now tell you how well an agent session went, not just what happened in it.
Pick a session and it sends a redacted summary of the trace to [Jev, TypeSafe AI's System One
model](https://typesafe.ai/blog/introducing-system-one-models-and-jev), which answers a fixed set
of questions about the run as probabilities rather than prose. Back comes a score out of 100 across
six dimensions, plus findings like repeated work or thrashing that each link straight to the
messages that caused them. Nothing leaves your machine until you have seen the exact payload and
said yes. The whole feature is optional, marked beta, and off until you configure a key.

### Added

- **Jev-powered efficiency analysis**
([`0757a16`](https://github.com/delexw/claude-code-trace/commit/0757a16)). Choosing **Analyse
efficiency** on a session extracts the behavioural shape of the run — turns, tool calls,
failures, repeated calls, subagent count, context growth — and asks Jev nine typed questions
about it. The answers arrive as probabilities and become six scores: progress, tool use, focus,
exploration, recovery and token use, weighted into a single number out of