Evals

Surfaces: web app (a session’s Evaluation tab; Evals) · CLI (kcap eval) · your agent (reads results, runs kcap eval for you)

The web app is at https://<your-workspace>.kcap.ai for a cloud workspace; on the local edition the same pages are served from your machine. The desktop app opens it from Help and support → Open dashboard, and kcap local open opens the local one from a terminal

An evaluation scores a finished session for safety, plan adherence, quality and efficiency, with evidence for each finding, and turns what it found into lessons. It can run on a configured server provider, on Kurrent-managed AI, or on your own machine through the daemon. The questions come from your workspace’s catalog, which admins can edit

What you can do

Evaluate a session. In the web app, open a finished session from Sessions, choose its Evaluation tab, then Evaluate session. Pick the engine if more than one is offered, pick questions or whole categories, and Start. Queued runs show their position; during a run the checklist fills in; the retrospective follows. You can leave the page. Re-evaluate runs a fresh one. From the terminal:

kcap eval <session>
kcap eval --chain --model opus <session>
kcap eval --questions safety <session>
kcap eval --list-questions

From your agent, which runs kcap eval for you and reads the result back:

Evaluate my last session and give me the results

What did the evaluation of session 3f9c… find, and what should I fix first?

Compare the last two evaluation runs of this session

Repositories and projects can also evaluate every session automatically when it ends; admins turn that on under Turn automatic evaluations on

Read a result and act on it. The Evaluation tab puts actions beside the score:

• For you: fixes and recommendations with their evidence. Copy the list as a checklist, or one fix into your working session
• For the agent: guidance for future sessions. Promote in Knowledge sends a finding to the repository’s curation queue
• Kept working: what went well and is worth preserving
• Patterns the judge applied: the retained lessons it consulted

Expand a passing question to see why it passed. A comparison chip shows whether the overall score rose or fell since the last run. Copy as Markdown takes the whole evaluation into an agent conversation

Check how much the judge saw. An evaluation can finish with limits on its evidence: a turn-detail budget reached, a tail-trimmed transcript, fewer questions judged than asked. Read those notes before trusting a score, and use the transcript to settle a disputed finding

Compare across repositories and work items. Evals in the sidebar, with a scope selected, has two views: Scopes ranks the repositories, work items or sessions below the current level by score and coverage; Evaluation shows the current level’s aggregate findings and retrospective. Drill down with a row, back up with the breadcrumb. Coverage is the share of a scope’s sessions that have an evaluation. A score built from three evaluated sessions out of three hundred says little about the repository, so read the coverage column before comparing scores

Help me get unstuck

Evaluate is disabled. The session is still active, or has resumed. Wait for it to end

The run failed with a vendor error. A breaker may have tripped for that vendor: bad credentials or exhausted quota. Admins see and reset it under Administration → Evals → Eval Queue

My daemon run never started. The daemon needs a judge harness installed, normally Claude Code, and a sign-in that is current. Use the server engine instead if one is configured

The score dropped after I changed nothing. The judge is a language model, and a language model does not make the same call twice: a repeat run can read the same evidence differently and score a question up or down. Compare the questions judged and the evidence notes before treating a change in the number as a change in the work

The findings quote a credential. Please file a support issue with Kurrent through the Feedback and Support feature; see Support. Evaluation suggestions should not include secrets; the capture redaction may have missed a pattern