v0.18.0

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, 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). 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 100. Because the model
  returns typed decisions instead of generated text, the same transcript scores the same way twice.
  The dashboard shows the exact question behind each metric, so a number is always traceable to
  what was asked. Token efficiency weighs total tokens, context growth, turns and tool activity,
  and deliberately says nothing about money.

• Findings that link back into the trace
  (0757a16). Six kinds of finding
  come back — repeated work, possible thrashing, repeated exploration, failed retries, effective
  recovery and useful subagent — each with a likelihood and a range of messages. Clicking one in
  the dashboard jumps the conversation to the messages behind it, and the message list grows an
  inline marker above the first message of every finding, so simply scrolling the trace shows where
  the agent got stuck or dug itself out. Overlapping findings of the same kind are merged, so one
  rough patch no longer produces a wall of near-identical entries.

• Scores and live progress in the session picker
  (0757a16). Every analysed session
  carries a Dashboard · NN badge, coloured by how good the score is and labelled stale once
  the transcript has moved on past the analysis; clicking it opens the dashboard without leaving
  the picker. A session mid-analysis shows a progress bar in that spot instead, one that failed
  offers Retry analysis, and one never analysed offers Analyse. A separate strip tracks
  every running analysis across sessions and lets you cancel one or jump to its result when it
  lands.

• A privacy review before anything is sent
  (0757a16). The payload is
  redacted locally first — bearer tokens, named secrets, private key blocks, secret-shaped
  environment assignments, credentials in URLs and your home directory path are all replaced — and
  then the confirmation dialog shows you the exact redacted JSON that would leave the machine,
  along with where it is going. The default minimized mode sends the extracted signals plus
  selected excerpts; full transcript mode is there for runs that need more context. There is no
  way to send without confirming.

• API tokens held in the operating system credential store
  (0757a16). The desktop app saves
  Jev and provider keys to the macOS Keychain, Windows Credential Manager or the Linux Secret
  Service, never to settings.json. Web mode reads JEV_API_KEY from the server environment
  instead, and the HTTP API has no route for saving or clearing a token at all, so a browser cannot
  write one. In Docker, ./script/redeploy.sh asks for the key without echoing it, writes it to a
  dedicated claude-code-trace-secrets volume and mounts that read-only into the container — the
  browser neither sends nor receives it.

• Cached results that know when they are out of date
  (0757a16). One latest result per
  session is stored under the config directory, stamped with a SHA-256 fingerprint of the
  transcript and with the analysis, decision-set and score-formula versions. Reopening a session
  shows the stored result straight away; when the transcript grows or the scoring changes
  underneath it, the result is marked stale rather than quietly presented as current. Re-analysing
  replaces it.

Changed

• Rust 1.88 is now the minimum for building from source
  (0757a16). The platform
  credential-store crates require it. Pre-built releases are unaffected — this only matters if you
  compile the app yourself.

• Recommendation providers are restricted when configured from a browser
  (0757a16). Web and Docker modes
  accept OpenAI-compatible endpoints only on loopback (localhost, 127.0.0.0/8 or ::1), with
  no credentials, query string or fragment in the URL, because a browser is not a safe place to
  hand over a remote provider's key — use the desktop app for that. Docker also hides the Codex and
  Claude Code subscription providers, since the image bundles and authenticates neither CLI.
  Desktop keeps the full set.

Credit: Jev is developed by TypeSafe AI. This integration was built
in Claude Code Trace, which is an independent open-source project and not an official TypeSafe AI
integration.