pi-microsoft-decision-1

A Pi package and reusable pi-ai provider for Microsoft's Decision-1 classifier. The package name and provider ID are both pi-microsoft-decision-1. The model ID remains Microsoft's upstream Microsoft-Decision-1 identifier.

The package supports two integrations:

• Pi loads its extension entry point and registers the provider automatically.
• Other coding agents can register the exported provider with a pi-ai Models collection.

Install pi-microsoft-decision-1 from the Pi package gallery or add its npm package name through Pi's package manager. The package manifest declares ./index.ts as the Pi extension entry point and exports a reusable provider factory from the package root.

Verified with Pi 1.1.0 and Node 26.4.0.

Pi setup

Update models.json

Configuration change: this package registers the provider as pi-microsoft-decision-1. If your existing models.json entry is named azure-decision, rename that key and use the settings below:

NOTE: The baseUrl and deployment name needs to match your foundry endpoint and deployment name

{
  "providers": {
    "pi-microsoft-decision-1": {
      "baseUrl": "https://<resource-name>.services.ai.azure.com",
      "apiKey": "$AZURE_DECISION_API_KEY",
      "deploymentName": "<Your Foundry Deployment Name>"
    }
  }
}

The provider key is pi-microsoft-decision-1; the model ID remains Microsoft's Microsoft-Decision-1. baseUrl must be the HTTPS resource origin, without a project or API path. deploymentName is the Azure deployment name sent in the request. Do not add a custom models list for this provider.

The agent directory defaults to ~/.pi/agent. If PI_CODING_AGENT_DIR is set, the package reads models.json from that directory, not from its npm installation location. The provider reads only its own settings and does not change other providers or reuse their endpoint, deployment, or credentials.

Set the API key environment variable

Pi must inherit AZURE_DECISION_API_KEY when it starts. For example, set it in the shell before launching Pi:

export AZURE_DECISION_API_KEY="<your Azure Decision API key>"

To make it available in new terminal sessions, add the export to your shell profile, then restart Pi from that shell. Do not put the key directly in models.json, source code, or prompts. The $AZURE_DECISION_API_KEY value in models.json is an environment-variable reference, not the secret itself.

Use from codemode

To call the classifier from scripts, enable Pi's codemode tool by adding +codemode to defaultTools in settings.json. If defaultTools already contains tools, append this entry rather than replacing them.

Pi's public classifier type is bool; the adapter maps it to Microsoft's wire type noul and maps the response back. Choice and score probabilities, confidence, fractional scores, and raw Decision answers are preserved.

Use from a pi-ai coding agent

The package root exports createPiMicrosoftDecision1Provider. Register the returned provider with createModels() and configure the resource endpoint and deployment explicitly:

For this direct pi-ai integration, provide AZURE_DECISION_ENDPOINT, AZURE_DECISION_DEPLOYMENT, and AZURE_DECISION_API_KEY through the coding agent's environment (or provide the key through its pi-ai auth context). Set them before starting the agent:

export AZURE_DECISION_ENDPOINT="https://<resource-name>.services.ai.azure.com"
export AZURE_DECISION_DEPLOYMENT="Decision-1"
export AZURE_DECISION_API_KEY="<your Azure Decision API key>"

This integration does not require @earendil-works/pi-coding-agent; that package is used only by Pi's extension entry point.

Request behavior

The package sends one non-streaming request to:

POST <baseUrl>/providers/microsoft/v1/systemone
api-key: <AZURE_DECISION_API_KEY>
Content-Type: application/json

It validates configuration and response shapes, honors cancellation and request/response hooks, enforces a 60-second default timeout, rejects redirects, and does not retry requests. Errors expose the HTTP status and safe Azure error code without echoing provider payloads or credentials. The configuration checker reports settings and key presence but does not verify credentials or deployment existence.

Validation

Tests use mocked HTTP responses and a temporary Pi agent directory. They cover request construction, bool/choice/score answers, probabilities, usage, configuration precedence, malformed responses, API and network failures, cancellation, timeout, hooks, the Pi extension loader, and the pi-ai classifier bridge.

npm run check
npm test
npm pack --dry-run

The integration test requires a Pi installation. PI_CODING_AGENT_PACKAGE can point to its coding-agent package directory when it is not beside the current Node installation.

Security

Extensions run with the permissions of the Pi process. Review the source before installing. This package sends classifier requests to the configured Microsoft Foundry resource and makes no telemetry or package-time network calls.