insights

OMK Insights

Generate a usage report from KimiCLI session history. OMK collects bounded evidence and session packet files; the current agent writes the narrative report directly.

Route

• Default narrative report: run omk insights message <args>.
• Quick, local, statistical, or --no-llm request: run omk insights --no-llm <args> and summarize paths/counts only.
• If omk is unavailable, ask the user to run omk setup or reinstall oh-my-kimicli.

Pass through relevant flags such as --limit, --facet-limit, and --force. For token-saving requests, prefer --facet-limit 10.

Narrative Workflow

1. Run omk insights message <args>.
2. Read the generated insights-agent-task.md.
3. Read the generated manifest.json.
4. Read ranked session packet files from the manifest, starting with recommended_read_mode: "deep".
5. Extract session facets and write insights-facets.json.
6. Generate the final report.json directly. It must include metrics, facets, sections, and quality notes.
7. Generate the final self-contained report.html directly from report.json.
8. Self-review before the final response:
   • facets were written before the report
   • At a Glance was written after the other sections
   • suggestions are evidence-backed
   • weak sections are omitted or moved to Evidence Notes
   • headings avoid awkward machine translation
   • skill opportunities exist only when repeated usage, clear friction, or high leverage is visible

Treat the task file as the authority for paths, limits, language guidance, report shape, and the final user question.

Boundaries

• Do not run bare omk insights for the narrative workflow.
• Do not run omk insights render; that pipeline is intentionally removed.
• Do not run kimi --print.
• Do not manually scan ~/.kimi/sessions.
• Do not paste full raw transcripts into context.
• Do not create or update skills, hooks, or AGENTS.md during the report run. Ask first.
• No entries, 暂无信息, or an omitted section is valid when evidence is weak.
• Do not manufacture recommendations just to fill a section.
• If the task brief is insufficient, read the reported insights-input.json and session packet files. If they are too large, rerun with a smaller --limit or --facet-limit.

Final Response

Report only:

• HTML report path
• JSON report path
• sessions scanned/analyzed
• weak-evidence sections or data limits, if any
• one short question asking whether to create or update a skill, hook, or AGENTS.md instruction only when the report has a concrete skill opportunity with action create_skill, update_skill, add_hook, or add_agents_instruction

If there is no reliable candidate, say that the current evidence is not strong enough to create a targeted skill yet.