/build

/ai-gateway:build

Take the user all the way from discovery to working code that uses an existing AI Gateway's models and/or MCP tools — in whatever they're building. This command runs the full workflow in the bundled skill skills/use-ai-gateway/SKILL.md end to end:

1. Discover & select the models and MCP tool servers in the gateway (Part 1 of the skill). Use $ARGUMENTS for the target gateway (ARM gatewayResourceId — required to list assets; a host alone is not enough, so resolve or ask for the id if only a host is given) and, if present, the desired use case. Ask the user to confirm the selected assets before continuing.
2. Retrieve a credential — list the gateway API keys and read the secret of the chosen key. The same gateway key authenticates both the model passthrough and the MCP tool servers, passed in the Api-Key header.
3. Integrate (Part 2 of the skill). Ask what the user is building and pick the path:
   • Call a model over the gateway's OpenAI-compatible passthrough (any language or raw HTTP).
   • Connect MCP tools from an MCP-capable client/app.
   • Scaffold a standalone agent with the GitHub Copilot SDK (github-copilot-sdk >= 1.0.0) if the user wants a ready-to-run agent project — default to Python unless they ask for TypeScript. Produce a complete, self-contained project: agent code, .env / .env.example, .gitignore (must ignore .env), requirements.txt (or package.json), and a README.md. Integrate into the user's existing project when they have one, respecting its language and conventions.
4. Run and verify the integration once and report the outcome, distinguishing genuine wiring bugs (unknown_model, 401/403, missing env var) from benign backend conditions (429/quota).

Guardrails

• Consumption only. This is strictly read-only against the gateway. Never issue ARM PUT/PATCH/DELETE calls and never create, provision, or delete gateways, models, tools, connections, or products. If the user asks to provision anything, tell them that is an administrator task done in the AI Gateway Portal.
• Never hardcode secrets. Read the gateway key from an environment variable and keep .env out of source control. Warn the user that the written .env holds a live key and to rotate it in the portal if it is ever exposed.
• Use properties.deployment.modelName (exact dots/casing) as the model identifier — prefer it over the ARM name or displayName (which some gateways reject with unknown_model). Runtime calls use the gateway's properties.gatewayUrl host and the /default/ workspace segment: <host>/default/models/openai/v1/... and <host>/default/toolservers/<name>/mcp (omitting /default/ returns 404).
• Authenticate models and MCP tool servers with the Api-Key header — not Authorization: Bearer (bearer-only auth is rejected, typically with a 401 or a misleading unknown_model).