v0.9.8

Added

• Tool and subagent calls are now exposed as artifact-update A2A events so any SSE consumer can observe invocations, arguments, and results; opt-in per request via userMessage.metadata["tool-status-update"]; also added to preview
• Preview now shows context size and token usage
• OpenAI provider support via llm kind openai
• Model and model parameters are shown in evaluation runs in MLflow

Changed

• Consider static assertions in tests now in "output_correctness" and "success_rate" metrics of evaluation runs
• The generic call action tool is now the default, matching @cap-js/mcp; per-action tools are opt-in via cds.env.mcp.per_action_tool: true. This replaces the former cds.env.agents.per_action_tool flag (which defaulted to per-action) and resolves the conflicting defaults between @cap-js/agents and @cap-js/mcp

Fixed

• @agent.hitl / @Common.IsActionCritical are now honored when actions are exposed through the generic call action tool (previously human-in-the-loop gating only took effect with per-action tools)
• Human-in-the-loop edit decisions can no longer repoint the generic call action tool to a different action than the one approved
• Correctly handle non @cap-js/mcp tools during PII masking
• Correctly report A2A messages as failed when AI Core is configured but no AI Core credentials can be found
• Correctly report input tokens for Anthropic models on AI Core. OpenTelemetry specifies that input_tokens includes cached tokens, but Anthropic did not include the cached tokens in input_tokens
• Summarize on timeout prompts are now correctly shown in MLflow
• Evaluation run metrics are now correctly aggregated in MLflow
• Thinking steps are now correctly separated also inside a turn
• Tasks are now moved into failed state on server crash to avoid quota issue on unstable servers
• @agent.llm is now correctly considered for feature toggled services
• SAP HANA Cloud based masking is now retried in case of HANA Cloud NLP connectivity issues
• HITL tool rejections are now surfaced to the model as a deliberate user decision instead of a tool failure, so it no longer invents technical reasons (e.g. insufficient stock) for an action the user chose to reject