Mitos
An AI code-tracing workspace that turns GitHub repositories into interactive execution-flow maps with a durable Inngest pipeline, synced FlowBoard/Monaco navigation, and verifier-backed quality scoring.
Mitos addresses the hardest part of unfamiliar codebases: understanding how behavior actually travels through files, functions, branches, and runtime boundaries. Built with Next.js 16, React 19, ElysiaJS, Better Auth, Neon/PostgreSQL, Drizzle ORM, Inngest, and direct Gemini/DeepSeek provider integrations, the architecture is designed around durable background execution, trace-scoped security, evidence-grounded AI output, and a dense developer workspace that keeps diagrams, source code, timeline steps, validation logs, and prose explanations in sync.
- GitHub OAuth sign-in with access and refresh tokens encrypted at rest using AES-256-GCM.
- Repository and branch selection backed by the authenticated GitHub token, with exact commit SHA pinning for every trace.
- Custom trace parameters let users model runtime inputs such as roles, methods, payload shapes, or feature flags.
- Trace and run history persists generated traces, previous parameter sets, bookmarks, reruns, and deletion workflows.
- Public share links expose completed traces while keeping file and symbol access limited to the visible trace path.
- Mounted a modular ElysiaJS API inside Next.js App Router and shared route types with the browser through Eden type-only imports, avoiding backend code in the client bundle.
- Designed a durable Inngest trace pipeline with per-trace concurrency, cancellation, optional Redis worker locks, attempt-scoped CAS idempotency, and artifact offloading.
- Built a two-phase AI tracing system that separates bounded repository exploration, behavioral evidence extraction, canonical trace planning, and structured output drafting.
- Implemented universal directory relevance scoring so large repositories surface READMEs, manifests, docs, source files, entry points, MVC/API paths, and prompt-matching files before noise.
- Engineered a five-layer verifier that repairs coordinates from real source text, merges overlapping steps, audits import graph connections, scores quality, and locks verified coordinates after semantic review.
- Created repository intelligence infrastructure for tree syncing, import maps, path alias resolution, symbol parsing, proactive/lazy indexing, and relational BFS fallback traversal.
- Synchronized FlowBoard, Monaco, timeline, playback, keyboard shortcuts, hover states, and narrative content through a trace workspace optimized for desktop and mobile workflows.
- Hardened public trace sharing with trace-scoped file/symbol authorization, text archive fallbacks, GitHub token encryption, binary rejection, and Mermaid/SVG sanitization.
- Established a benchmark harness and diagnostics layer that tracks golden fixtures, anchor coverage, evidence grounding, tool behavior, stage timings, and generalization risk.
- GitHub OAuth, repository/branch picker, custom trace form, model configuration checks, and trace creation flow.
- Durable trace generation pipeline with repository resolution, symbol indexing, AI exploration, evidence packing, planning, drafting, verification, self-correction, and persistence.
- Interactive trace workspace combining Mermaid FlowBoard, Monaco code pane, timeline, playback controls, narrative drawer, validation drawer, command palette, and keyboard shortcuts.
- Trace and run history with bookmarks, parameter reuse, reruns, deletion, public share toggles, and copy-link actions.
- Responsive layout with mobile Diagram/Code/Steps switching, persisted panel visibility/order, and Motion-powered panel transitions.