Models

The turn loop asks a ModelAdapter for one decision at a time and receives a stream of ModelEvents: text deltas, tool calls, usage. Everything above the adapter is vendor-blind, and swapping one changes no line of the harness.

The adapters

Package	Covers	Key
`@nativeharness/adapter-anthropic`	Claude — per-model facts for 11 models, adaptive thinking, effort levels	`ANTHROPIC_API_KEY`
`@nativeharness/adapter-openai`	OpenAI, xAI, DeepSeek, Moonshot, MiniMax, Zhipu — services whose API is chat-completions — plus Groq and Mistral through their compatibility endpoints	`OPENAI_API_KEY`, `XAI_API_KEY`, `DEEPSEEK_API_KEY`, `GROQ_API_KEY`, `MISTRAL_API_KEY`, …
`@nativeharness/adapter-google`	Gemini — `thought` parts, reasoning tokens split from output	`GEMINI_API_KEY` or `GOOGLE_API_KEY`
`@nativeharness/adapter-cohere`	Cohere v2 — the tool plan, streamed before tool calls	`COHERE_API_KEY` or `CO_API_KEY`
`@nativeharness/adapter-ollama`	local models — asks the model its real context window and whether it can use tools at all	none

new AnthropicAdapter({ model: "claude-opus-5", effort: "high" });
new OpenAIAdapter({ provider: "groq", model: "…" });
new OllamaAdapter({ model: "qwen3" });

Each adapter reports its model’s real context window, which is what compaction budgets against — not a guessed number.

Scripted, for tests

ScriptedAdapter from @nativeharness/agent plays a fixed sequence of decisions: text, tool calls, or a thrown error to exercise the failure path. It is how the harness’s own suite runs a whole turn — a real command, in a real sandbox — with no key, and how the CLI’s --model scripted:<file> works. It also declares itself deterministic, which is the one case the agent will run on a non-isolating sandbox provider: no model is deciding anything, so there is nothing to contain.

Choosing from the CLI

nhar chat takes --model, then $NH_MODEL, then the first key it finds in the environment, in this order: Anthropic, OpenAI, Google, Cohere, xAI, DeepSeek, Groq, Mistral. Adapters are loaded only when chosen, so an Anthropic user never loads the OpenAI SDK.

nhar chat <id> --model anthropic/claude-opus-5
nhar chat <id> --model ollama/qwen3
nhar chat <id> --model scripted:./turn.json

Writing one

A ModelAdapter is one method, decide, taking the system prompt, the messages, the tools and a token budget, and yielding ModelEvents — plus an info block naming the model, its context window and its capabilities. The interface is model-adapter.ts; the Ollama adapter is the smallest real one to read.

Getting started

How it works

Building your app

CLI

Operations

The adapters

Scripted, for tests

Choosing from the CLI

Writing one

​The adapters

​Scripted, for tests

​Choosing from the CLI

​Writing one

The adapters

Scripted, for tests

Choosing from the CLI

Writing one