GitLab Orbit Local halves Claude Code tokens for impact analysis

The test targeted a common pain point: when an AI agent traces a change's impact across a large codebase, it typically greps one call level at a time, reading search results and files along the way. The author's premise was that a prebuilt graph of call relationships would let the agent answer in one query instead.

In the trial, setting up Orbit for Claude Code alone was not enough; Claude often reverted to grep after an initial Orbit call. The author then added a ready-made recipe to the project instructions, which prompted Claude to use a recursive query for multi-level callers. The beta caveats matter: package-crossing calls and TypeScript path-alias imports may not appear in the graph, so the author recommends treating Orbit Local as one additional tool rather than a sole replacement.