llamacpp-tuner

Overview

Inspect a running llama-server, identify the bottlenecks that matter for a coding agent workload (long context, heavy prompt reuse, latency-sensitive tool calls), propose a tuned configuration, measure before/after, and report concrete deltas. The result is an actionable optimization with numbers behind it, not generic advice.

When to use

Use this skill when the user asks to tune, optimize, or speed up llama.cpp / llama-server for a local coding agent — phrases like "make my local model faster", "optimize llamacpp for Claude Code / Codex / Aider", "best llama-server config", or invokes /llamacpp-tuner.

Do not use this skill for: vLLM, Ollama (other than as a memory-pressure suspect to free), text-generation-webui, or proprietary inference engines. The benchmark and the knobs are specific to llama-server.

Workflow

The skill runs as a single sequential workflow. Each step gates the next.