Hacking Claude Code for Fun and Profit (sibylline.dev)

A developer demo/repo shows how to “hack” agent workflows by putting a proxy (the Bifrost LLM router) between agents and providers so you can transparently rewrite prompts, inject guardrails, route to multiple backends, and emit OpenTelemetry traces for debugging and observability. The demo wires Claude Code CLI through Bifrost and an observability stack (OpenTelemetry Collector → ClickHouse + Grafana Tempo → Grafana), so every run produces spans (claude.webhook and claude.cli) you can inspect with TraceQL or SQL. The pattern generalizes: any agent that accepts OpenAI-/Anthropic-style base URLs can be routed through Bifrost (e.g., /anthropic → OpenRouter or /v1 → OpenAI-compatible), letting you maintain existing clients while adding middleware logic. Technically, the repo is Docker Compose–ready (docker compose up --build) and includes a webhook harness, pre-instrumented Claude Code, Bifrost gateway, OTLP ingest (4317/4318), and exposed endpoints (e.g., http://localhost:8787 for the webhook, 6080 for Bifrost, Grafana on 6300). Auto-instrumentation requires a privileged container (or disable with BIFROST_ENABLE_OTEL_AUTOINSTRUMENTATION=0). You can query traces in Tempo or run SQL against ClickHouse to pull span attributes like final responses and request payloads. The core takeaway: proxy-first agent engineering gives powerful, production-ready control over context, routing, and observability—critical for debugging, optimization, and safely operating agent fleets.