Extract Knowledge

extract-knowledge mines reusable business facts from a finished analysis and saves them as knowledge/*.md files. It runs automatically after a successful query when the reasoning revealed a non-obvious rule worth remembering — you don't invoke it directly.

What it does

After you reach a working query, the path to it often teaches something that isn't visible in the schema: a field encoding, a filter you must always apply, a join that has to go through a mapping table. extract-knowledge captures that as an atomic, reusable fact and stores it in the project's knowledge, so the next similar question already knows the rule.

What it captures

• Field encodings / enums — e.g. status = 'A' means an active account.
• Mandatory filters — e.g. always exclude rtype = 'D'.
• Join traps — e.g. two tables must be joined through a mapping table.
• Business-term-to-field mappings — which column a business term actually refers to.

It skips trivial queries and facts that are already recorded.

When it runs

It is triggered as part of other flows — for example during /init, /build-kb, and Session Summarize — whenever a finished analysis exposes a durable rule. You normally don't run it by hand.

What you'll observe

New entries appear under knowledge/ (indexed in AGENTS.md), each a short, atomic fact tied to your data. The next time a similar question comes up, the agent retrieves these facts instead of rediscovering — or missing — the rule.

Example

You ask "total active subscriptions last month". Getting it right requires knowing that status = 'A' means active and that cancelled rows carry rtype = 'D' and must be excluded. Once the query validates, extract-knowledge records two facts:

• status = 'A' → active account
• always exclude rtype = 'D' for active counts

Both become knowledge entries, so a later question about active users reuses them automatically.