Prism-Shadow/penguin-harness

Agent Creation turns a user requirement into a working agent configuration by writing AGENTS.md, setting identity metadata, and installing only needed skills. Use it to create or configure an Agent State under the agents directory.

Agent Evaluation runs one specified test agent on one benchmark case exactly once, privately scores the execution, and returns one protocol result. Use it when a run_subagent caller needs a single isolated agent evaluation without managing loops or scoreboards.

Agent Optimization improves one Test Agent through an evidence → hypothesis → Candidate → evaluation → accept or rollback loop. It uses a frozen Benchmark, Scoreboard, and public Test Traces as black-box feedback while delegating all evaluation to an agent-evaluation subagent.

AgentHub Models calls model APIs through @prismshadow/agenthub for streaming text, image generation, speech synthesis, embeddings, and the supported-model registry with one client. Use it when building AI apps that need a single TypeScript client for multiple model providers.

Benchmark Design builds and calibrates a multi-Case capability benchmark for one Test Agent and establishes a traceable Formal Baseline. Use it when you need to design, pilot, and freeze a benchmark without changing the test agent.

Bento Slides creates and edits Bento presentations — self-contained .bento.html decks whose document is JSON. Use it when the user wants a slide deck or presentation, whether from scratch, from source material, or by improving an existing file.

Data Analysis completes data-analysis tasks with bounded inspection, correct data semantics, native artifact handling, complete delivery, and risk-based verification. Use it when an agent must deliver a requested dataset or report artifact at an exact path with correct grain, units, and format.

Firecrawl searches the web and scrapes pages into clean markdown using the Firecrawl API. Use it when a task needs current web information or the content of a specific page, including public PDFs.

Humanizer rewrites or edits prose in any language so it reads like edited human writing in the register of books, newspapers and encyclopedias rather than default AI output. Use it when you need to remove AI fingerprints from generated text.

Humanizer rewrites or edits prose in any language so it reads like edited human writing in the register of books, newspapers, and encyclopedias rather than default AI output. Use it to revise drafts, remove AI tells, and match per-language idiom and typography.

LlamaFactory fine-tunes open-weight LLMs (LoRA/QLoRA, full-parameter, SFT, DPO) via YAML configs and the llamafactory-cli command. Use it to register datasets, train models, merge LoRA adapters, and serve or chat with the result.

Ollama Serving pulls and runs local models with Ollama, exposing an OpenAI-compatible endpoint on localhost:11434 for apps and agents. Use it to add a self-hosted model to PenguinHarness or any OpenAI client.

Penguin CLI manages model API keys, default models, and per-agent vault secrets for the Penguin harness. Use it when configuring model credentials and agent secrets.

Penguin Harness Development guides changes to the PenguinHarness monorepo, covering the two-repo symlink layout, CI-parity verification, record-and-ship changelog contract, and intentional seams. Use it when editing packages, model catalog, installers, release workflow, or auditing changelog entries and prose.

Penguin SDK is a TypeScript SDK for building self-contained AI apps with agents, sessions, and streaming thinking/tool messages. It includes a complete RAG recipe that ingests documents into a knowledge base and answers with citations behind a web UI.

Remote Claude Code runs and drives Claude Code on a remote Linux host over SSH, supporting persistent sessions, headless claude -p calls, and interactive TUI sessions inside tmux. Use it when you need to keep an SSH connection open across turns or hold a continuous conversation with Claude Code on a server.

Skill Porting teaches an agent to import skills from Claude Code plugin marketplaces, Codex plugins, npx skills, GitHub repos, and local folders, then normalize their frontmatter and verify the install. Use it when you need to safely bring external skills into penguin-harness without blindly running installers.

Skill Porting installs skills from Claude Code plugin marketplaces, Codex plugins, skills.sh registry names, GitHub repos, or local folders into Penguin's agent_state/skills directory after reviewing all files and normalizing SKILL.md frontmatter. Use it to add a third-party skill safely.

Software Engineering guides an agent through investigating and reviewing code, implementing bug fixes, features, and refactors with minimal scope, validating changes, and reporting verified outcomes. Use it for general software-engineering tasks when a request requires understanding, modifying, or verifying code.

VLLM Serving deploys and serves open-weight LLMs on local GPUs with vLLM behind an OpenAI-compatible endpoint, ready for chat and agent workloads. Use it when you need high-throughput GPU inference with tool calling for agent harnesses like PenguinHarness.

Web Design provides a Penguin visual language for generated web pages and app UIs, with GitHub-style simplicity, a single blue accent, light and pure-black dark themes, design tokens, and component recipes. Use it as the default for generated frontends unless specified.