penguin-harness/changelog/0.1.1/2026-07-22-skills.md

Skills: local serving and fine-tuning join AI App Development, plus deck authoring

Three new built-in skills extend the AI App Development group so agents can stand up and tune the models they build on:

• vllm — install and serve models with vLLM's OpenAI-compatible server, including the tool-calling flags agent harnesses need (--enable-auto-tool-choice --tool-call-parser …), quantization and memory options, and troubleshooting.
• ollama — pull and serve local models with Ollama, its OpenAI-compatible endpoint, and context-length tuning via OLLAMA_CONTEXT_LENGTH or a Modelfile.
• llamafactory — fine-tune with LlamaFactory (dataset registration, llamafactory-cli train/export/chat/api driven by the upstream example YAMLs), then serve the result with vLLM or with Ollama via a Modelfile import.

Both serving skills share a guided workflow: ask the user which model to serve (recommending the small Qwen3.5-0.8B default when they have no preference) and which engine they prefer — Ollama as the approachable default, vLLM when throughput matters — then serve, verify, and register the endpoint with the penguin CLI (models appear in penguin config model list only after model add). Each skill's icon follows the project's official mark.

The hard root-separation guardrail lives where configuration is taught — the penguin-cli skill (now v5) and the penguin-sdk skill: configuring Penguin's own model uses the default root, but models configured for an AI app under development must use the app's own project directory (--root ./penguin_data) unless the user chose otherwise — never the global ~/.penguin/data. The serving skills point at that rule rather than restating it.

bento-slides joins the Office Productivity group, so an agent asked for a presentation produces a real deck rather than a wall of bullets. A Bento deck is one self-contained .bento.html file whose document is plain JSON in a #bento-doc script block; the skill teaches the agent to edit that block in place, to fetch the app itself from bento.page when starting from nothing, and — the part that matters — to map source material onto the right feature (charts for numbers, tables for grids, morph transitions for a subject that changes across slides, state slides for drill-downs, ken-burns and count-ups for motion) instead of defaulting to text slides. It carries the gotchas that otherwise cost a round trip: chart series data must be plain numbers, morph needs stable shared element ids, assets are embedded as data URIs, and docId is never regenerated on an existing deck. Adapted from the Bento project's own skill (MIT, © 2026 The Bento/Suite authors) with attribution in the skill body; the authoritative schema stays at https://bento.page/agents.md.

The agenthub-models skill tracks the AgentHub 0.4.1 upgrade: it documents the new supported-model registry, the config parameters a client may now reject outright, and the Gemini 3.6 / Kimi K3 / GLM-5.2 families with their reasoning-effort knobs, alongside a routing table rewritten from the release's actual client-matching order.