Concepts

Mushi is more than a feedback widget. It’s a four-layer system:

1. Capture — SDKs across web, mobile, and AI agents collect structured reports with screenshots, breadcrumbs, and device context.
2. Classify — A two-stage LLM pipeline (fast-filter → classify-report) tags severity, category, and component, with vision analysis on screenshots when relevant.
3. Connect — Every report is embedded into a knowledge graph (pgvector + Apache AGE) so similar bugs deduplicate and component hot-spots emerge.
4. Fix — Approved triage decisions hand off to an agentic fix orchestrator that runs in a sandboxed environment (E2B) and opens a PR via a scoped GitHub App.

Every layer is observable (Langfuse + OpenTelemetry), auditable (immutable audit log), and self-improving (judge loop with A/B-tested prompts).