LLM tool calling · Agents
Exploring how language models call tools and carry out real work. What began as function-calling model training grew into inference parsers, SDK middleware, and a coding-agent runtime.
Model training · Archived
Early post-training experiments for function-calling models: LoRA configurations, tool-calling notebooks, and Hermes / Llama chat templates. Development has ended.
Python · Upstreamed to vLLM
A parser for Hermes-style tool calls from Llama models on vLLM. It handles tool-call tags split across streaming tokens; the implementation was contributed upstream to vLLM.
TypeScript · Vercel AI SDK
Middleware that turns text output into structured tool calls for models without native function calling. It provides model-specific parsing, streaming reconstruction, and schema-aware argument coercion.
TypeScript · Archived
A coding agent built on the Vercel AI SDK, with file, editing, and shell tools for interactive and headless use. Kept for reference; continued development lives in pss-runtime.
TypeScript · Successor to Plugsuits
A reusable agent runtime for threads, the model loop, hooks, and storage. A separate coding-agent application adds workspace tools, a TUI, and a headless runner.