Kitaru post-import insights

Generate evidence-backed insight cards from imported Kitaru sessions. This distribution provides two independently selectable analyzers. Both profile every selected session and build deterministic candidates; the OpenAI analyzer then uses bounded analyst and editor calls to select candidates and write their descriptions.

Both entrypoints accept a list of session UUIDs for one agent and import, fetch each session with its nodes through KitaruAPIClient, and return list[InsightInput] for the task runner to persist. Each fetched trace is processed and released before the next fetch. Select either analyzer or both for an import. Each invocation persists its own cards; matching candidate names do not replace another analyzer's results.

The API client uses KITARU_API_URL and KITARU_API_TOKEN. Both analyzers accept agent_name for display context. The deterministic analyzer does not accept a model parameter or call OpenAI. The OpenAI analyzer uses gpt-5.6-luna with low reasoning effort when its model parameter is omitted; callers can override that parameter with a compatible model available to their OpenAI project. It fails when its credential is missing.

Install kitaru-post-import-insights for deterministic analysis, or kitaru-post-import-insights[openai] for OpenAI generation. The default catalog selects the corresponding requirement for each analyzer. The package does not depend on Langfuse or send insight-generation telemetry to it.

Release context

The exact kitaru==0.25.0+dev dependency requires the unreleased analyzer contract in PR #988: analyzer tasks pass session UUIDs, and the plugin fetches normalized sessions through the SDK. Replace the development pin with the compatible published core release floor during release preparation. This package must not be published while the development pin remains.

Links

• Kitaru documentation
• Source code
• Issue tracker

Licensed under Apache-2.0.

Investigation button labels

New cards persist cta_label beside investigation_prompt in metadata["kitaru.insights/v1"]. Both analyzers use a reviewed finding-type catalog, such as "Investigate tool retries" and "Investigate session duration", with "Investigate this finding" for unknown finding types. Labels describe the investigation topic, use Investigate [topic], and stay within 40 characters. No additional model requests are made.

The field is optional for historical records and accepts nonblank strings up to 40 characters. Consumers should render it as text and fall back to "Investigate this finding" when the label is missing or unusable. The existing investigation_prompt remains unchanged. A missing prompt should disable the action. Frontend wiring is separate from this backend contract. Existing stored labels are not rewritten. Older strict metadata readers reject the new field, so update any independently deployed readers before enabling new writers; retain reader compatibility on rollback.