Projects

Dependency-free Ralph-loop agent harness for small local models (VibeThinker-3B via Ollama) with grammar-constrained actions & .vibe/ reasoning traces.

Multi-agent orchestrator + Sonnet executor workflow: GitHub issues -> git worktree -> PR -> Stop-hook code-gated review -> merge.

Drop-in orchestration skills, custom subagents, deterministic PowerShell Stop hooks, and guard CI workflows for multi-agent repos.

One-pager flow-diagram and instruction generator with downloadable skill.md and MCP server for structured architecture handoff.

Empirical reproduction, metadata analyzer, and anonymized transcript audit of 516/1034/1552-token reasoning collapse clusters.

Five-shot benchmark evaluation and durable model_instructions_file remediation skill for shallow reasoning ceilings.

100+ foundational algorithmic, numerical, networking, text, and systems exercises implemented in Python.