CodeDNA: An In-Source Annotation Standard for Deterministic Navigation of Multi-File Codebases by LLM Agents

CodeDNA is an experimental inter-agent communication protocol implemented as in-source annotations. The writing agent encodes architectural context directly into source files; the reading agent decodes it. The file is the channel — no RAG, no vector DB, no external rules files. The protocol addresses a specific problem: AI coding agents spend a significant fraction of their context window exploring irrelevant files. CodeDNA embeds navigation metadata (reverse dependencies, domain constraints, public contracts) directly in source files using 4 layers: a project manifest (.codedna), module headers (exports, used_by, rules), function-level rules, and semantic naming. Benchmark results — SWE-bench, 5 Django tasks, 3 LLMs, ≥3 runs/task at T=0.1: - Gemini 2.5 Flash: +13pp F1 (60% → 72%, Wilcoxon p=0.040 ✅) - DeepSeek Chat: +9pp F1 (50% → 60%, p=0.11) - Gemini 2.5 Pro: +9pp F1 (60% → 69%, p=0.11) Positive Δ on 3–4/5 tasks per model. Results measure file localization F1 (navigation efficiency), not end-to-end patch correctness. All 3 models complete. Fully open source (MIT), vendor-neutral, zero dependencies. GitHub: https://github.com/Larens94/codedna Website: https://codedna.silicoreautomation.space