0.2.0-M11

The agent-loop milestone: Agentic Chat's tool calling runs inside a bounded, interceptable agent loop, attached images become native multimodal input with EXIF-aware analysis presets, the desktop launcher gains a tray, auto-update, and Ollama auto-start, and Home reworks into a live launchpad.

Desktop users on 0.2.0-M10: the M10 installers crashed on launch (whisper addon packaging). M11 is the fixed release - install it directly, or let the new in-app updater take over from here.

Highlights

• Bounded agent loop - the chat tool-calling loop is governed by AgentLoopManager: soft/hard round caps (16/18), an identical-call repeat guard, a per-turn interactive-tool budget, and an interceptor chain (loop guard, HITL approval, file upload, image reference). Batch HITL approvals land in a single dialog, and a cancelled upload or declined approval ends the request cleanly. Tunable via spring.ai.playground.chat.agent-loop.*.
• Multimodal vision input - attach up to five images per message; they are resized and EXIF-tagged in the browser, stored content-addressed on your machine, and sent as native multimodal input. describeImage re-summons an earlier image any time, and the Image analyst preset turns a batch of geotagged photos into an EXIF table and a pin map (renderTable + plotPointsOnMap). The tool catalog grows to 115.
• Desktop tray, auto-update, and Ollama auto-start - tray menu with settings deep links and model managers, stage-aware update checks (in-place install on Windows/Linux, guided download on macOS), automatic Ollama start when the profile needs it, and a dedicated Whisper voice-model manager window.
• Home as a launchpad - adaptive hero CTA, provider readiness chips, and a live System & Observability panel where every number is computed from the running services.
• Default model qwen3.5:4b + wider menu - vision works out of the box; qwen3.5 / qwen3.6 / gemma4 families, gpt-oss:20b, and deepseek-r1:8b join the menu, with GGUF builds kept alongside the -mlx catalog on Apple Silicon.
• Filesystem search, edit, and a destructive L5 tier - the filesystem family grows from 11 to 18 tools; destructive operations rate L5 and always pause for human approval.
• Visualization action cards + in-chat file upload - fourteen visualization tools render charts, diffs, and multi-point maps as review-then-act cards; requestFileUpload opens an upload dialog mid-conversation and converts spreadsheets to CSV in the browser.
• Dual-level tool risk chips and anonymous, opt-out usage telemetry (content-free, authored names masked, one switch to disable).

Fixed

• Key-order independent tool fingerprints (no false rug-pull flags).
• Built-in MCP loopback propagates the conversation id to the per-conversation workspace.
• Unknown tool calls get a recovery hint instead of crashing the stream.
• Stage-aware update banner (M < RC < GA), suppressed on desktop.
• No macOS keychain prompt before any secret has been saved.

Docs

• New Agent Loop architecture page and an OWASP MCP Top 10 coverage map; Multimodal Vision Input feature page; tutorials 13-15 (file upload, image + EXIF map, document detective).

What's Changed

• fix(electron): bundle whisper launcher modules in packaged app by @JM-Lab in #16
• fix(electron): background STT warm-up and credential-gated server launch by @JM-Lab in #17
• fix(electron): load whisper STT addon in packaged macOS builds by @JM-Lab in #18
• fix(electron): set whisper addon rpath at build time for packaged stt by @JM-Lab in #19
• 0.2.0-M11 - bounded agent loop, multimodal vision, desktop tray & auto-update, home launchpad by @JM-Lab in #21

Full Changelog: v0.2.0-M10...v0.2.0-M11