We moved from separate code-index processes to a shared warm graph. Retrieval got faster, but context sufficiency is still unclear.

Releases·Tool Release

Coding agents pay more to find code than to change it, and the interface is why

Ataraxy Labs introduced sem, an open-source tool addressing how coding agents spend most tokens finding code rather than editing it. Sem parses repositories into functions and call graphs, letting agents request a function with its callers, callees, and reaching tests in one response. It targets large, hard-to-navigate repos with slow test suites.

Ataraxy Labs introduced sem, an open-source tool addressing how coding agents spend most tokens finding code rather than editing it. Sem parses repositories into functions and call graphs, letting agents request a function with its callers, callees, and reaching tests in one response. It targets large, hard-to-navigate repos with slow test suites.

An open-source agent-harness maintainer published a benchmark showing code knowledge-graph lookups use 75-82% fewer tokens than grep-then-read for 'who calls this' queries across four repos. Commenters pointed out the benchmark measures only token cost, with no accuracy, recall, or precision metrics. The author added the limitation to the docs and is asking which of CodeRAG-Bench, CoIR-Retrieval, ContextBench, or SWE-Explore-Bench would test whether agents retrieve the correct callers.

Products·launched

Archex: local deterministic code retrieval and context assembly for coding agents

Archex is a local, deterministic retrieval and context-assembly engine for coding agents. It uses tree-sitter parsing, chunking, embedding, graph expansion, and ranked token-budgeted bundling to produce a context bundle with a receipt showing what was included, skipped, and whether it is safe to act on. It solves the problem of providing accurate, efficient code context to LLM-based coding agents, targeting developers building such agents.

Archex is a local, deterministic retrieval and context-assembly engine for coding agents. It uses tree-sitter parsing, chunking, embedding, graph expansion, and ranked token-budgeted bundling to produce a context bundle with a receipt showing what was included, skipped, and whether it is safe to act on. It solves the problem of providing accurate, efficient code context to LLM-based coding agents, targeting developers building such agents.

Products·building

Kivgraph — open-source semantic code graph MCP server for LLM coding agents

Kivgraph is a local, Apache-2.0 MCP server that indexes multiple repositories into a semantic code graph for LLM coding agents. It answers structural queries (callers, change impact, cross-repo consumers) directly instead of relying on long chains of grep calls and file reads. It targets developers using LLM agents on large codebases, reducing context cost and enabling provable cross-repository relationships.

Kivgraph is a local, Apache-2.0 MCP server that indexes multiple repositories into a semantic code graph for LLM coding agents. It answers structural queries (callers, change impact, cross-repo consumers) directly instead of relying on long chains of grep calls and file reads. It targets developers using LLM agents on large codebases, reducing context cost and enabling provable cross-repository relationships.

CodeNib is a new data system that provides coding agents with reusable lexical, dense, and structural views of repository context, maintained across commits and edits. It aims to reduce repeated discovery and obscure lifecycle costs by serving ranked search, symbol navigation, and bounded context through one runtime. The system was evaluated across 100 snapshots, mapping quality-cost frontiers.