Evals

Evaluations score recorded work for Safety, Plan adherence, Quality, and Efficiency, and turn findings into lessons for future sessions. Use a session’s Evaluation tab to review one run. Open Evals in the main navigation to compare results across your selected scope.

Run a session evaluation

1. Open a finished session and select Evaluation.
2. Choose Evaluate session. If multiple engines are available, select Use my daemon or a configured server engine.
3. Select individual questions or whole categories, then choose Start.

The question set comes from your server’s evaluation catalog, so its size and wording can change. Evaluations are available after a session ends; if it resumes, wait until it finishes before evaluating again.

A daemon evaluation uses a connected local daemon and its installed judge harness. A server evaluation uses a configured provider and does not require your machine to be online. Managed Kurrent AI evaluations are also available on Free; administrators control provider and automatic-evaluation settings. See Evals administration.

Queued runs show their position. During a run, the question checklist reports progress and results; the retrospective is prepared afterward. You can leave and reopen the page while evaluation continues.

Eligible sessions can also be evaluated automatically after they end. Repository settings and server policy determine which sessions qualify. You can still start an evaluation manually or use Re-evaluate to obtain a fresh result.

Read and use the result

The Evaluation tab puts actions alongside the score:

• For you collects fixes and recommendations, with evidence and recurring-pattern context. Copy the list as a Markdown checklist, or copy one fix into your working session.
• For the agent collects guidance for future sessions. Promote in Knowledge opens the repository’s curation queue.
• Kept working records strengths worth preserving.
• Patterns the judge applied shows retained lessons consulted during evaluation.

The result also shows the overall score, category scores, the judge and run details, and comparisons where prior results are available. Expand passing questions when you want to inspect what passed. Copy as Markdown copies the full evaluation for use in an agent conversation. When an earlier run is available, a comparison chip shows whether the overall score increased, decreased, or stayed the same.

Scores and verdicts also appear on session cards and in work-item retrospectives. They summarize the judge’s assessment of the recorded evidence.

Check evidence coverage

An evaluation can complete with limits on what a judge retrieved. Check the number of questions judged and any failure notices before interpreting its score. Evaluation records can also describe these evidence limits:

• Turn-detail budget reached means the judge had the compact session index, but full detail was delivered for only the reported number of turns.
• Tail-trimmed transcript means the supplied transcript was shortened to a character limit.
• Questions judged can be lower than questions requested if judges failed, timed out, or exhausted their run limits.

A complete run and complete evidence are different things. Use the session transcript to investigate a disputed finding, and re-evaluate after addressing a runner failure or changing the evaluation configuration. Older results may not include coverage information.

Compare a scope

Open Evals with All Repositories, a project, or a repository selected. The page has two views:

View	What it shows
Scopes	A sortable ranking of the repositories, work items, or sessions below the current level, with evaluation scores and coverage.
Evaluation	The current level’s aggregate scores, findings, and retrospective.

Open a row to drill down, and use the scope breadcrumb to move back up. Opening a session leads to its Evaluation tab. Load more expands rankings when the page has not loaded the full set.

Compare coverage as well as scores: a repository with few evaluated sessions represents a smaller sample. All results respect the sessions you can see.

Evaluate from the CLI

kcap eval --list-questions
kcap eval <sessionId>
kcap eval --model opus <sessionId>
kcap eval --chain <sessionId>
kcap eval --questions safety <sessionId>
kcap eval --skip efficiency <sessionId>

Use --list-questions to inspect the current catalog before selecting questions. CLI evaluations use the local evaluation runner; choosing a server engine is available in the dashboard.

Feed lessons into future work

Retained findings appear in Knowledge. Start with Docs to review grouped lessons and their member facts, then approve useful agent guidance as skills. This is the preferred way to curate accumulated findings without working through every fact individually. Docs and skills explains review, installation, and the caps on automatic injection.

Use Facts and curation for targeted corrections or an individual lesson’s delivery, and Errors and plan validation for narrower checks on a session.