A developer behind AxiomCode argues that coding agents waste tokens reconstructing deterministic code relationships (calls, inheritance, dependencies) via grep-and-follow exploration. They open-sourced AxiomCode's axiomcodegraph, which builds a code knowledge graph grounded in compiler/type information that agents can query before reading source. The open question posed: how much repository exploration should the LLM handle versus deterministic tooling?
A Reddit post by the tool's author describes a structural code-graph approach to giving coding agents blast-radius signal: tree-sitter parses repos into a nodes-and-edges graph stored in a single mmap'd file, served over MCP so agents can query callers of a function or the blast radius of a change. The author reports sub-100ms responses at 839k-node scale and invites better approaches.
A Reddit post by the tool's author describes a structural code-graph approach to giving coding agents blast-radius signal: tree-sitter parses repos into a nodes-and-edges graph stored in a single mmap'd file, served over MCP so agents can query callers of a function or the blast radius of a change. The author reports sub-100ms responses at 839k-node scale and invites better approaches.
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.
A concept for a service that indexes private GitHub repositories so developers can ask natural-language questions about code — where methods are called, what operations do, and how services depend on each other. The poster has built a small POC hosted on AWS ECS and plans to scale from one repo per tech stack to hundreds across an organization.
A concept for a service that indexes private GitHub repositories so developers can ask natural-language questions about code — where methods are called, what operations do, and how services depend on each other. The poster has built a small POC hosted on AWS ECS and plans to scale from one repo per tech stack to hundreds across an organization.
A Reddit post argues coding agents become unreliable when they flatten documents, structured data, relationships, procedures, and durable facts into one context window or vector index. It proposes routing requests to typed context surfaces — KB, DB, Graph, Skills, Memory — each with its own access pattern. The architecture is implemented in the open-source Codex MCP plugin OpenDCAI/DataMind.
A Reddit post argues coding agents become unreliable when they flatten documents, structured data, relationships, procedures, and durable facts into one context window or vector index. It proposes routing requests to typed context surfaces — KB, DB, Graph, Skills, Memory — each with its own access pattern. The architecture is implemented in the open-source Codex MCP plugin OpenDCAI/DataMind.
Researchers propose ACQUIRE, a QA-driven framework that improves LLM-based coding agents by identifying knowledge gaps before attempting repository fixes. Unlike prior fix-driven exploration, ACQUIRE generates targeted questions to gather precise context, reducing factual errors in automated issue resolution.
Researchers propose ACQUIRE, a QA-driven framework that improves LLM-based coding agents by identifying knowledge gaps before attempting repository fixes. Unlike prior fix-driven exploration, ACQUIRE generates targeted questions to gather precise context, reducing factual errors in automated issue resolution.