Seyon

Rootcode is a VSCode extension that makes AI coding agents actually work inside your engineering team, not just alongside it. It solves three problems no existing tool addresses. First, context: instead of semantic search or keyword retrieval, Rootcode builds a live dependency graph of your codebase using the TypeScript compiler API and a Python parser, then uses deterministic graph traversal to surface exactly the right files for every agent turn. No embedding databases, no extra LLM calls, no guessing. Second, memory: Rootcode watches your behavior after every response, edits, test outcomes, git reverts, and computes an implicit reward signal without you labeling anything. Failure patterns accumulate across sessions and get injected as lessons before the next one starts. A reinforcement learning bandit runs in parallel, learning which strategy the agent should use on each turn and improving its approach based on what actually works in your codebase. Third, adaptation: Rootcode mines your git history into a training dataset and fine-tunes a LoRA adapter on top of Meta's Muse Spark Open via Fireworks, privately, inside your infrastructure. The result is a model that knows your conventions and reviews PRs the way your best senior engineer would. The agent layer runs on Google ADK TypeScript, with specialist agents for development, QA in Docker, version control, and PR review via GitHub MCP. Large tasks are split across parallel cluster workers that execute in dependency order with write confinement, so no agent steps on another's work. Everything is built and running today.