# @kindgi/dev-echo-provider

> Deterministic dev-only ModelProvider — no LLM key, no network, no cost. Emits a two-message scripted response (tool-call → text) so an agent turn completes in milliseconds. For dev, tests and tutorials. NEVER for production — every real deployment plugs in a real model provider adapter (Anthropic, OpenAI-compat, etc.) instead.

`npm install @kindgi/dev-echo-provider` · [source](https://github.com/kindgi/kindgi-sdk/tree/@kindgi/sdk@0.1.9/packages/dev-echo-provider)

Deterministic `ModelProvider` for dev, tutorials, and tests. No LLM key, no network, no cost. Every real deployment plugs in `@kindgi/adapter-model-anthropic` (or an equivalent adapter) instead.

## Shape

Emits a two-message script (tool-call → text):

1. **First invocation** (no tool result yet) — asks the runtime to call the first tool in `input.tools` with `{ message: <last user message> }`, or `demo.echo` when the call declares no tools.
2. **Second invocation** (post-tool) — emits a final assistant message: `Tool responded: <tool result>`.

**It says it isn't a model.** Every text answer starts with one line, `DEV_ECHO_WARNING_LINE`:

```text excerpt
⚠ dev-echo isn't a real model: it only repeats what it's given. Add an LLM provider key (Anthropic, OpenAI, Gemini, Groq, OpenRouter…) to get real answers.
```

An answer that is JSON (what a typed-output agent parses; dev-echo hands back the JSON it was given) stays bare, as does one asked for with `structuredOutput`. Every result also carries `warnings: [DEV_ECHO_WARNING]` (code `dev-echo-not-a-model`, with the commands that add a model), which an agent turn collects into its result's `warnings`.

State lives in the message trail, not on the instance — safe to share across concurrent runs.

## Usage

```ts excerpt
import { createDevEchoProvider } from '@kindgi/dev-echo-provider';
import { createProviderRegistry } from '@kindgi/capabilities';

const provider = createDevEchoProvider();
// Registry is tenant-scoped — register the provider for each tenant
// that should have it available.
const { registry } = createProviderRegistry([
  { tenantId, provider },
]);
```

The provider id is `DEV_ECHO_PROVIDER_ID` (`'dev-echo'`) and its single model is `DEV_ECHO_MODEL_NAME` (`'dev-echo-v1'`); `DEV_ECHO_PROVIDER_METADATA` exposes the metadata without creating a provider.

## Not for production

The provider's metadata says so explicitly — `metadata.description` is `'Dev-only deterministic provider — no LLM, canned scripted responses. NEVER for production.'`

If you find yourself reaching for this in a real deployment, wire a real provider (`@kindgi/adapter-model-anthropic`, `@kindgi/adapter-model-openai-compat`, etc.) instead.