Physix Frontier · News Briefing Card (Hacker News · Nov 1, 2026)

Codedict cuts Claude Code token use by 63% with local index

KEY FACTS

Codedict builds a code index locally with tree-sitter and never calls AI during the process.
The index records functions, classes, routes, SQL tables and call relationships, and stores them in a SQLite file.
In benchmarks, all 12 questions were answered correctly while token consumption dropped 63%.
The tool passes retrieval results to the AI alongside the question through hooks, cutting answer rounds from 5 to 20 down to 1 to 2.
The install script writes three hooks into the Claude Code configuration and keeps a backup of the original settings.

KEY DATA

63%token reduction

12/12correct answers

about 250test project file count

35,000 to 50,000single-round context tokens

PHYSIX OBSERVATION

The key to cutting tokens is not the model but the retrieval method: let the local index do the locating first, and the AI only handles answering. This suggests context engineering still has plenty of room for optimization, and it also means the code retrieval layer may become a standard component of AI coding tools rather than an accessory feature.

Source: Hacker News report