🎨 Gemini Scribe 4.11.0 - Design System, Plan Mode & Smarter Agent Runs
- 🎨 Design system overhaul — a refreshed visual identity: a signature Gemini accent (bold user message bubbles, a gradient send button, and a brand mark), a unified elevation/shadow scale, a gradient progress bar across all states, motion polish, and consistent icon sizing — all built on a new theme-adaptive design-token layer so the plugin adapts cleanly to light, dark, and custom themes (#1090, #1104, #1107, #1109, #1110, #1112)
- 📋 Plan Mode (opt-in) — a new toggle in the agent view (and a command) that has the agent lay out its plan before it starts acting, shown as a "Plan" pill; leave it off for the usual direct execution (#1046)
- ⚙️ Background execution by default for long-running tools — deep research and image generation now run in the background automatically, so the agent view stays responsive while they work and you can track them in the Background Tasks panel (#1085)
- ⏹️ Stop halts a streaming response immediately — the Stop button now cancels a mid-stream follow-up right away instead of waiting for it to finish, and follow-up requests stream live as they arrive (#1053, #1097)
- 🖥️ Ollama improvements — a single Ollama model field (replacing the per-task pickers) that persists independently, so switching Gemini → Ollama → Gemini no longer changes your Gemini chat model; automatic vision-capability detection so image-capable local models are recognized; and accurate token counts read from the model's own reporting (#1125, #1058, #1076, #1152)
- 🖼️ SVG attachments — SVG and SVGZ files are now rasterized to PNG before being sent to the model, so vector images work as attachments (#1082)
- 📱 Cleaner startup, especially on mobile — the plugin no longer triggers "attempted to load NodeJS package" notices when it loads, fixing startup toast spam and a latent mobile-compatibility gap (#1154)
- 🧠 Smarter long agent runs — context is now compacted during a long tool chain (not just before a turn), so extended agent runs stay within budget without dropping the work already in flight (#1074)
- 🔎 RAG polish — "Reindex All" is renamed "Rescan Vault" to match what it actually does, and the "show all files" affordance is now keyboard-accessible (#1056, #1159)
- 🐛 Fixes — Deep Research works again on @google/genai 2.x (#1151); agent tool logs no longer fold into the preceding reasoning block in session history (#1084); and error messages no longer mislabel non-network failures as connectivity problems (#1153)
- 🔧 Under the hood — the plugin's entire softened-lint backlog is now cleared and enforced at zero warnings and all 150 circular imports were eliminated, tidying the codebase and moving the Obsidian community-directory review toward clean (#1032)
4.11.0 is a large feature release. The headline is a full design-system overhaul: a new theme-adaptive design-token layer underpins a signature Gemini visual identity — bold user message bubbles, a gradient send button and brand mark, a unified three-step elevation/shadow scale, a gradient progress bar across all states, motion polish, and a consistent icon-size scale — so the interface feels more cohesive and adapts cleanly across light, dark, and custom themes. The agent gains an opt-in Plan Mode that lays out a plan before acting, background execution by default for long-running tools (deep research and image generation) so the agent view stays responsive, immediate Stop cancellation of a mid-stream streaming response, and mid-loop context compaction that keeps long tool chains within budget without losing in-flight work. Ollama support matures with a single model field that persists independently of your Gemini model (fixing a provider switch that previously reset your chat model), automatic vision-capability detection, and accurate token counts from the model's own reporting. SVG and SVGZ files can now be attached (rasterized to PNG first). Startup is cleaner — the "attempted to load NodeJS package" notices are gone, which also closes a latent mobile-compatibility gap. Fixes include Deep Research on @google/genai 2.x, session-history reasoning/tool-log separation, clearer error messages, and RAG affordance and keyboard improvements. Under the hood, the entire softened-lint backlog is cleared and enforced at zero warnings and all 150 circular imports were eliminated via a leaf plugin interface, moving the Obsidian community-directory review toward clean. Dependencies are unchanged (@google/genai 2.10.0).