Skip to content

digitarald/decide

v0.1.1MIT

Typed decisions for bounded classification, routing, and scoring through a user-configured Microsoft Foundry deployment.

GitHub

Decide

Typed decisions for bounded classification, routing, and scoring through a user-configured Microsoft Foundry deployment.

An experimental skill and a Python CLI for asking a decision model bounded questions. It returns booleans, choices with probabilities, or scores as JSON. Python 3.10 or newer is required; no Python packages are needed.

No behavioral evals or accuracy claims. Probabilities are model outputs, not measured accuracy. Keep deterministic checks and permission decisions in the caller.

Install

copilot plugin marketplace add digitarald/agent-parts-bin
copilot plugin install decide@agent-parts-bin

Start a new session. Example request:

Use Decide to design a routing policy for these synthetic support tickets. Show the request with a dry run, without calling a model.

In VS Code, install from digitarald/agent-parts-bin. See the marketplace instructions.

Setup and data handling

The package includes no Azure resource endpoint, credentials, or personal configuration. Each user supplies a Microsoft Foundry resource endpoint and deployment. Read setup and authentication. Agents must use --non-interactive and must not collect secrets in chat.

Live decisions send the supplied state and questions to that configured resource and may incur charges. Only send data approved for that destination. There are no automatic retries.

Endpoint and deployment settings stay in the user's configuration directory, outside the plugin. On macOS, setup stores API keys in Keychain; other platforms use environment credentials. Installation does not run setup, call the model, or change credentials. No hooks or MCP servers run automatically.

For offline inspection, run these from this plugin's directory:

python3 skills/decide/scripts/decide.py schema
python3 skills/decide/scripts/decide.py batch --non-interactive \
  --input skills/decide/assets/triage.json --dry-run
python3 -B -m unittest discover -s skills/decide/tests -v

The fixtures use synthetic input and example resource names. The response fixtures show the output contract, not an accuracy benchmark.

Host checks

  • Copilot CLI: confirm the installed decide skill is discovered, then run schema and a synthetic --dry-run using the installed skill's path.
  • VS Code: start a new chat, confirm the skill appears in Chat: Configure Skills, and ask for an offline dry run.
  • Live model access, routing quality, and native host invocation have not been verified for this plugin release. Only offline CLI and packaging checks are automated.