Rework context snapshots into opt-in, LLM-summarized handoffs + UI polish

[DevMando]

Summary

Reworks the context snapshot feature from auto-salvaged deterministic dumps into opt-in, LLM-summarized handoffs, and folds in a round of UI polish.

On a model switch (or Take snapshot), the outgoing conversation is buffered and offered via a native card. The user picks a summarizer model, optionally names it, and the snapshot is born with a real LLM recap. Ignoring the offer discards it — pure opt-in, no silent auto-save.

Snapshots

• Removed the deterministic "Light" recap and the two-tier Light/AI model. ContextSnapshot is now { OriginModel, SummarizerModel, Recap, Name, MessageCount }.
• New SnapshotEnhancer — a bare-kernel Ollama summarizer (no tools/plugins, can't touch the live chat) that map-reduces the entire conversation so it handles long histories. Prompts are framed as a handoff briefing to the next model: substance over filler, verbatim specifics (paths, versions, URLs), explicit done-vs-pending, plain prose.
• Native create-snapshot card in the chat: current model selected instantly (background model-list load, no lag), optional name, local · free / cloud · uses tokens tags, [Name] [model] [Create] on one row.
• Import accumulates multiple snapshots (was a silent overwrite that kept only the last) and dedupes by id so the same snapshot won't stack twice. The import chip and card title show the snapshot's name when set.
• Quieter "context cleared" chip — only shown when there was actually a conversation to clear.

UI polish

• Slide animations for the snapshots panel and the Settings/MCP pages.
• Settings/MCP nav items toggle back to the last active agent instead of reloading in place.
• Settings: Connection-first tab order (the setup that matters), and "Make Default for New Agents" hidden on the app-wide Appearance tab (it can't save it).
• Removed the redundant export button from the main chat header (still in the tab's ⋯ menu).

Notes

• All changes are in the Desktop tree; the pinned MandoCode harness submodule is untouched.
• Builds clean (dotnet build, 0 warnings / 0 errors). Prompt quality was validated live against local Ollama models during development.
• Not yet exercised end-to-end in a full manual UI pass.

🤖 Generated with Claude Code