0.3.2 release

0.3.2

New Features

• Agent Observability (Configurable Tracing) - Turn on agent.observability.tracing to export run traces to Langfuse, LangSmith, Datadog, Braintrust, or a generic OTLP collector. Optionally capture run content and redact sensitive data, and get stable trace references plus runtime trace grouping that tie benchmark, bootstrap, CLI, and chat runs into a single trace for easy correlation and debugging. #833 #864 docs
• Visual Artifacts: Dashboard Support - The new gen_visual_dashboard generates dashboard-style HTML (multi-chart card layouts) right from Chat, just like gen_visual_report, with local interactive preview served by datus --web — filters re-run live against your database, no SaaS backend required. Report and dashboard generation is also noticeably better: cleaner layouts, more reliable chart rendering (concurrent multi-chart queries to DuckDB no longer race), consistent use of the runtime ChartCard primitive, more accurate underlying data queries, support for more complex reports, and multi-round refinement. The compiled HTML's absolute path is now surfaced in the CLI message stream, so you can reopen the artifact after closing the browser tab. #829 #835 #842 #847 #848 #849 #853 #855 #863 #866 #867 #869 #894 #895 #901 #905 #907 docs
• Mid-Run User Input - Send additional instructions while an agent run is still streaming — type in the CLI/TUI or POST to the API; the text is picked up on the model's next step, saved to the session, and shown live as it is inserted, without interrupting the run. #824

Enhancements

• Semantic SQL Metric Extraction - When mining metrics from historical SQL, Datus tells apart brand-new metrics, metrics derived from existing ones, and plain references to existing metrics, so it won't create duplicates; it also keeps time grain, filters, and literal values. Supported metric types include count, distinct count, sum, average, min/max, conditional (filtered), ratio, expression, cumulative, and derived; for multi-table cases (cross-table calculations, non-equi joins, unions) it first combines the data into a single source and then defines metrics on top. End-to-end verified on a real warehouse (StarRocks): the generated metrics return the same values as the original SQL. #811
• Smoother CLI Chat & Streaming - Fixes a duplicated last paragraph in the terminal when streaming through Claude's native API (plus a related session-resume parsing fix); chat history now renders through one unified path, so history, resume/rewind, and mid-run user inserts all display consistently, with user messages shown in a bordered panel for clearer separation. #837 #852

Bug Fixes

• Phased Semantic-Model Validation - A semantic model can be validated before any metrics exist: the expected "no metrics yet" issue no longer aborts the flow, while genuine model errors still fail validation. #827 #850
• Reliable Bootstrap Result Handling - Bootstrap flows now correctly recognize each step's generation result (including failures), so successful output is no longer dropped and failures are no longer silently swallowed. #831
• Reference SQL Summary Path Resolution - Generated reference-SQL summary paths now resolve correctly across the different path forms; out-of-sandbox paths are safely skipped instead of crashing. #840
• Print Mode No Longer Hangs on Permission Prompts - datus -p (the non-interactive print mode, used in CI and scripting) now runs under the workflow execution mode and dangerous profile — consistent with /bootstrap and other non-interactive flows — so ASK/EXTERNAL permission prompts that previously blocked waiting for a human responder short-circuit cleanly. #891