fix(ai-release-notes): migrate off retired GitHub Models API

[simon (susishopware)]

Why

GitHub Models was fully retired on July 30, 2026. Since then, every release workflow using ai-release-notes fails with a 401 from the legacy inference endpoint (models.inference.ai.azure.com) and no release gets created — first observed on databus-conductor v1.10.0 (failed run). The last successful run was on July 29, one day before the shutdown. This affects every repo consuming this action.

What changed

• Replaced the hardcoded GitHub Models endpoint with a configurable OpenAI-compatible api-endpoint input (default: https://api.openai.com/v1/chat/completions) and a new api-key input.
• Graceful degradation instead of failing the release: when no api-key is configured, or the AI call errors, the action logs a warning and uses the raw GitHub-generated changelog as release notes. A broken or missing AI provider can never block a release again.
• The model input is passed through unchanged (default stays gpt-4o), so it works with any provider speaking the OpenAI chat completions format — OpenAI directly, the Anthropic compatibility endpoint, Azure OpenAI, etc.
• Removed the now-obsolete models: read permission requirement from the docs; only contents: write is needed.

Migration for consumers

Consuming workflows keep working as-is (they fall back to raw notes with a warning). To restore the AI rewrite, pass a key:

- uses: shopware/github-actions/ai-release-notes@main
  with:
    product-name: "My Product"
    api-key: ${{ secrets.OPENAI_API_KEY }}

Open question for the maintainers

We deliberately kept this provider-agnostic. On our side (Pipe Fiction team) we'd request an OpenAI key from IT and wire it up as an org/repo secret — but if there is already an established company-wide pattern for AI API access in Actions (central proxy/gateway, an existing shared secret, or a preferred provider), we're happy to adapt this PR to that instead. Input welcome.

🤖 Generated with Claude Code