Code knowledge indexes for agents

The idea

A code index at the level of signatures finds code in milliseconds. An Atlas profile, computed by code with no model, gives each repository's stack, commands, rules and module map.

The problem
Agents spent turns on grep and relearned how to build and test each repository.
Why it matters
On held-out queries, find puts the right file first 89% of the time (top 5: 100%; git grep: 42%).
Where it stands
In production as of

A step-by-step replay of this idea is not published yet. The library adds one when a recorded run shows it.

Where it lives in the product

  • Learned file ranker without an LLM

    Gradient-boosted trees rank a request's candidate files from about 140 counted signals and the Git history, with no LLM call.

  • Research from what is known, within budget

    Research opens with the area notes, earlier research and rules for the named paths, and reads only what they leave open. At its effort budget it gets one turn to ask for more.

  • Rulebook mined from history

    Commits, PRs, reviews, tickets, tests, CI and code are mined into rule candidates with evidence and scope. Accepted rules become checks, tested first against past changes. A rule that catches nothing in 90 days goes to its author to retire or keep.

Turn the numbers into shipped work.

Crewshal runs these choices for you: a persistent AI worker with memory and rules, on your Claude and Codex subscriptions.