0.2.0-M12

The resident-desktop milestone: closing the window keeps the app serving from the tray and reopens straight onto the conversation you left, human approvals now show each tool's risk level and scale their friction to it, charts gain value sorting and a visual overhaul, Korea's national statistics service joins the MCP catalog, and Ollama models stay loaded for thirty minutes so a returning chat never waits out a reload.

Upgrading from 0.2.0-M11: the custom environment-variable aliases are gone. OBS_*, TOOL_STUDIO_FS_BASE, SPRING_AI_PLAYGROUND_MCP_*, SPRING_AI_PLAYGROUND_ACTUATOR_INCLUDE, SPRING_AI_PLAYGROUND_TRACE_SAMPLE, and the SPRING_AI_*_OBSERVE_* toggles no longer bind - use Spring Boot's relaxed-binding names instead (spring.ai.playground.observability.persist becomes SPRING_AI_PLAYGROUND_OBSERVABILITY_PERSIST, and so on; the configuration reference lists every one). SERVER_PORT and OPENAI_API_KEY keep working.

Highlights

• Desktop stays resident in the tray - closing the main window hides it instead of stopping the server, with a one-time "Still running in the tray" notification that reopens the window when clicked. A Recent Activity tray menu lists the pages you last had open and reopens the app directly on one of them, and Agentic Chat now writes the active conversation into the URL and the page title, so a reopened window returns to that conversation instead of a blank chat.
• Risk level shown in HITL approval dialogs, with friction that scales by level - every approval prompt carries the tool's risk chip, the rationale for that level, and a level-specific caution line. L4 renders the approve button in the error theme; L5 keeps approval disabled behind an "I reviewed the arguments and accept the risk" checkbox. Applies to both gates - the chat dialog and the MCP elicitation path used by external clients.
• Ollama models stay loaded for thirty minutes - spring.ai.ollama.chat.keep-alive and spring.ai.ollama.embedding.keep-alive default to 30m and ride on every request, so returning to a conversation after a pause no longer waits out a model reload. The request value wins over the Ollama server's own OLLAMA_KEEP_ALIVE default, and a single conversation can override it from the settings drawer's provider-options JSON.
• KOSIS National Statistics in the MCP catalog - a vendor-official pilot entry for Korea's national statistics service, operated by the National Data Office, pre-filled with its endpoint and separate /api/mcp path, marked PREVIEW, and shipped with per-tool risk levels for its ten read-only tools.
• Chart sorting and chart visual polish - renderChart and renderWindRose take an optional sort (asc / desc) so nominal categories can be ordered by value while time and ordinal categories keep their given order. A single-series rose defaults to one color per slice sorted ascending (nightingale style), and every chart type picks up a shared eight-color palette, rounded bar caps, gradient area fills, and themed axes, legends and tooltips.
• Safer defaults - beans is dropped from the actuator exposure, so a default install serves only health,info,metrics,prometheus; and the analytics page_view now sends a bare route path, so neither the ?conv= id nor a conversation title derived from your text can reach the wire.
• Daily MCP inventory and per-model usage telemetry - two aggregate events join the anonymous set: mcp_server_active (once a day per configured server) and model_usage (once a day per model used in the last 24 hours). Still content-free, still covered by the single SPRING_AI_PLAYGROUND_TELEMETRY_ENABLED=false switch.
• Cleaner chat transcript - user messages are right-aligned and message avatars and name labels are gone, so the transcript reads as a conversation rather than a labelled log.

Fixed

• The composer stays locked while a response streams, and relocks when you return to a conversation that is still streaming.
• Exposed built-in tool lists refresh when the popover opens; a combo with nothing to offer stays disabled instead of looking clickable.
• Chat no longer scrolls horizontally, and wide tables scroll inside their THINK / MCP TOOLS / RAG panels instead of stretching the conversation.
• The launcher validates the actuator shutdown response, reports a stop you asked for as a clean stop, drops App Args that duplicate a setting it already owns, and no longer opens a console window on Windows.

Docs

• A Tray and Menubar guide covering close-to-tray, Recent Activity, and the Updates submenu; per-level L3 / L4 / L5 HITL approval captures across the human-in-the-loop, architecture, and tutorial pages; refreshed chat, preset and visualization screenshots; and a redrawn integration diagram.