AI coding tools
Ripwire 0.4 Uses a Code Graph to Narrow the Search Scope for Agents, but Saving Tokens May Not Reduce Task Costs
Ripwire uses tree-sitter, a symbol relationship graph, and Personalized PageRank to generate reproducible file rankings and change-impact scopes for coding agents. First-party tests show better localization rates than five comparison tools, but an independent review found that compressing context may still degrade answer quality and increase overall agent costs.
Ripwire 0.4.0 was released on September 7, aiming to reduce the cost of coding agents repeatedly searching and reading large amounts of source code to understand a repository. It is a C++23 CLI with zero runtime dependencies and can also operate as an MCP server. It uses tree-sitter locally to parse code, extracts symbols, calls, and import relationships, and then ranks them with Personalized PageRank based on the task description. Its output can include not only candidate files but also function signatures, callers, complexity, Git change frequency, change-amplification scope, and recommended tests—all without embeddings, API keys, or a hosted vector index.
[The project’s evaluation](https://github.com/redhat-et/ripwire) on 60 primarily Python-based LocBench samples achieved a strict file@10 score of 58.3%, meaning that every ground-truth file appeared in the top 10; the best comparison tool scored 40.0%. But this measures only whether the files were found, not whether a correct patch was produced. In its Django experiment, Ripwire reduced context tokens to about 5% of those used by a direct grep-and-read approach, yet only 5 of 12 answers passed the strict criteria, compared with 11 of 12 for grep-and-read.
Even more notable are the project’s six published Codex agent experiments: Ripwire ranked the ground-truth patch file first in all three runs that used it, yet adding it to the workflow increased median output tokens by approximately 80.2% and wall-clock time by approximately 40.7%. An [independent technical review](https://wavect.io/blog/ripwire-ai-repo-context-review-2026/) attributed the problem mainly to the agent loading too many skill instructions and making routine tool calls, rather than to the ranker itself. Engineering teams should therefore treat Ripwire as a localization tool to activate when the scope is unclear, while retaining direct-search and testing paths. The evaluation metric should also be the cost of producing a patch that passes acceptance tests, rather than the number of tokens saved by a single query.