Deploy

Deploy Skill

Requires spec.md, plan.md, .azure/deployment-plan.md, verify.md. If spec.md is missing, auto-run specify with the user's prompt, then auto-run plan and implement. If plan.md is missing, auto-run plan and implement. If implementation.md is missing, auto-run implement first. Load workspace context per copilot-instructions.md.

Execute

Preflight: az --version, azd version, and bicep --version all succeed; azd and the microsoft.foundry extension is installed (install with azd ext install microsoft.foundry) and up to date (check with azd ext list); azd auth login --check-status succeeds; azd env list shows the target env; azure.yaml exists and infra/main.bicep is present (skip if AZD template = none). If az is logged in but azd is not, run azd config set auth.useAzCliAuth true instead of the interactive azd auth login.

Packaging: validate before deploy:

• Hosted-agent deps: ship uv.lock when the project is uv-managed; otherwise a focused requirements.txt (runtime deps only). Do not ship the agent-framework meta-package — it pulls broken extras; pin the specific sub-packages you import (e.g. agent-framework-core, agent-framework-foundry, agent-framework-foundry-hosting) plus undeclared transitive imports (e.g. mcp). Validate with a clean install before deploy.

• Hosted-agent code deploy: generate/validate .agentignore to scope the remote payload.

• When remoteBuild: true, generate/validate service-specific .dockerignore to keep build contexts clean.

• Normalize Linux entrypoint line endings (LF) via .gitattributes or dos2unix/sed in the Dockerfile to avoid CRLF runtime failures.

• Confirm AZD environment, subscription, resource group, and region from .azure/deployment-plan.md before any state-changing command. Stop if the deployment plan records none for the AZD template (no azure.yaml to deploy).

• Make sure that instrumentation is enabled, so that telemetry is available immediately after deployment.

• For IaC changes, dry-run with azd provision --preview before azd deploy.

• Never print or commit secrets; values must resolve via Key Vault / managed identity.

• Stop local servers before deploying to avoid file locks.

• Deploy:

  • If azd provision was succeeded, now perform the azd deploy and wait for completion:

    azd deploy -e <AZD environment>
    

  • If azd provision was not previously run, run azd up instead to provision and deploy in one step:

    azd up -e <AZD environment>
    

• After deployment succeeds, gather the deployed endpoint URLs from azd show -e <AZD environment> -o json. The frontend URL is the primary user-facing endpoint and should be set in the azd environment variables as FRONTEND_URI.

• Post-deploy health: confirm ACA revision is active and provisioned (az containerapp revision list), hosted-agents sessions respond with the command azd ai agent show <agent-name> to retrieve the agent endpoint and playground URL, then azd ai agent invoke --agent-endpoint <agent-endpoint> {"input": "<message>"} to confirm the agent responds as expected, and check the logs by running azd ai agent monitor <agent-name>. Report the playground URL for each agent to the user.

• Repeat the HTTP-based E2E validation with the deployed endpoints and report pass/fail.

Common failures

Report

Do not claim deployment succeeded without azd deploy exit 0 AND the local E2E suite re-run against the deployed URL. Hand back the deployed Azure endpoint and summarize the updated ./docs/deploy.md. Generate or refresh root README.md linking to ./docs/spec.md, ./docs/plan.md, ./docs/implementation.md, ./docs/verify.md, and ./docs/deploy.md.