OMP Core Extensions

This document covers the mechanics and usage of @oh-my-pi/omp-stats and @oh-my-pi/swarm-extension.

@oh-my-pi/omp-stats

omp-stats is a Bun-based TypeScript observability service designed to log LLM usage metrics.

Architecture and Mechanics

• Engine: Runs as a Bun-based TypeScript service.
• Observability Store: Uses a SQLite database tracking metrics in messages and user_messages tables. It tracks request tokens, costs, durations, cache rates, and user frustration metrics (e.g., Yelling, Negation, Repetition, Anguish, Blame, Profanity).
• Telemetry Hooks: Key metrics are keyed by generation-specific IDs rather than generic agent IDs. This ensures multiple LLM calls from a single agent do not clobber state.
• Non-blocking I/O: SQLite inserts are non-blocking so as not to stall the primary workflow.
• Dashboard: Features a Single Page React Dashboard utilizing Recharts, Lucide icons, and TailwindCSS, running on port 3847 by default.

@oh-my-pi/swarm-extension

swarm-extension is designed for complex DAG (Directed Acyclic Graph) orchestration across multiple OMP agents.

Architecture and Mechanics

• Engine: Analyzes dependencies using waits_for and reports_to directives defined in YAML configs.
• State Tracker: Tracks execution state wave-by-wave inside .swarm_<name>/state/pipeline.json.
• Trace Context Propagation: Explicitly propagates trace contexts (like traceId and parentObservationId) across asynchronous boundaries to ensure distributed traces render accurately (e.g., in Langfuse).
• Command Entry: Can be invoked natively in the OMP CLI via the /swarm run command.