Code Saaya
The autonomous AI repository knowledge and documentation engine built by SaayaLabs. Converts any software repository into a pre-indexed, enterprise-grade knowledge base for human engineers and AI coding agents (Claude Code, Antigravity, Cursor, Aider, Devin). Delivers .saaya/repowiki/, CLAUDE.md, and AGENTS.md directly via GitHub Pull Requests, cutting agent token consumption by 80%.
AI coding agents waste tens of thousands of tokens per task repeatedly searching and reading source files to piece together database schemas, API contracts, and architecture rules. This causes slow execution, high API token costs, and frequent architectural hallucinations.
Stateful multi-pass LangGraph pipeline: planTaxonomy (domain partitioning) → generateCatalogs (50-100+ sub-articles) → generateModuleCards (6-file technology suites: overview, architecture, tech stack, conventions, commands, and YAML metadata) → writeArticles (deep Markdown with Mermaid sequence & ER diagrams) → buildMetadata. Incorporates an adaptive Token-Bucket Rate Limiter to prevent HTTP 429 throttling.
Engineered automated tech-stack auto-discovery scanning project imports and dependencies to construct 6-file module suites. Utilized Octokit base64 blob Git trees to deliver hundreds of documentation files in a single atomic GitHub Pull Request. Designed model-agnostic engine architecture supporting Claude 3.5 Sonnet, GPT-4o, DeepSeek R1, and local Ollama endpoints.
Saves 80% of token window usage for coding agents by providing pre-indexed context maps upon project load. Generates over 300 nested documentation files per scan, operates 100% free and open source, and eliminates agent context hallucination.
Pre-indexed architecture maps eliminate costly repetitive agent file-reading passes.
Generates root CLAUDE.md & AGENTS.md guidelines for instant agent workspace alignment.
Stateful agentic graph with checkpointing allows long-running scans to pause & resume.
Constructs base64 Git trees to commit complete documentation suites in one PR.
