v0.6.3
AI Chat now runs on nr-llm's agent runtime
The backend AI Chat no longer drives its own MCP-only tool loop. Every chat turn is now handed to nr-llm's AgentRuntime, which runs the model over nr-llm's builtin tool registry — the ~45 backend tools nr-llm already ships (system log and last exceptions, system status, deprecations, record search, page content, and more). The assistant can look things up and act on the TYPO3 installation directly instead of being limited to whatever an MCP server exposed.
What changed
- Native nr-llm tools out of the box. Tool availability no longer depends on a configured MCP server; nr-llm's registry (with its RBAC and per-configuration tool gate) is used directly. Admin-only tools such as the log reader are offered when the acting backend user is an admin.
- A proper identity. A fixed identity/behaviour contract is always part of the system prompt: the assistant introduces itself as the TYPO3 Backend AI Chat by Netresearch, is steered to use its tools rather than ask you to paste data, never claims to be ChatGPT or made by OpenAI, and replies in your language. It no longer self-identifies as ChatGPT.
- Clearer failures. A missing or misconfigured nr-llm Task now fails the turn with a clear message instead of silently falling back to a tool-less chat. Non-completing agent outcomes (guardrail blocks, approval requests, provider errors) surface as a sanitized error on the conversation.
Upgrade notes
- Requires
netresearch/nr-llm ^0.23.1(the release that introduced the agent runtime). - The MCP client classes remain in the package but are no longer used by the chat turn.
- No configuration changes are required. File attachments (images/documents) continue to work.
This is an additive patch release.