1.1.0 is the follow-through release for the web learning experience. The chat, artifact viewer, tutorial, blog, model chooser, quiz renderer, animation player, and benchmark plumbing all got touched so the app behaves more like a usable study environment and less like a collection of separate demos.
Image Generation Tooling
The browser generate_image tool now supports image models directly, including model, size, quality, mode, and diagram-kind controls. When a configured provider is unavailable, Keating falls back to richer browser-local SVG learning diagrams instead of empty placeholders. That gives the tutor a path to produce infographics, anatomy sketches, comparison cards, and structured study visuals inside the same artifact flow as plans, maps, and animations.
Generated image payloads also render in chat with copy affordances, so learners can keep the prompt, metadata, and generated artifact details instead of losing the useful parts behind a tool call.
Interactive Lesson Quizzes
Lesson-plan artifacts now expose Lesson and Quiz modes. The quiz mode turns the saved plan into an interactive quiz taker with multiple-choice, multi-select, true/false, short-answer, and transfer checks. Quiz attempts are scored in-place, missed topics are visible, and the learner can redo a quiz with a stronger focus on the sections they missed.
Adaptive quiz scoring was fixed at the same time: remedial fallback questions that are skipped because the learner already answered the primary question correctly no longer count against the raw score, partial-credit points, denominator, submitted payload, or persisted stats.
Cleaner Chat Output
Assistant output is now much easier to work with. Text blocks, code blocks, reasoning, tool arguments, tool output, generated images, and structured response blocks all have copy affordances where they are useful. Reasoning blocks parse <think> and <thinking> tags, deduplicate repeated reasoning, and auto-collapse after the response completes so older turns stop taking over the transcript.
Chat persistence also got a correctness pass. User messages are now preserved when a provider throws before streaming begins, which fixes the case where a first message appeared to disappear after an initial provider error. Tag-heavy messages such as question, goal, image, and reasoning outputs now render in the right place instead of showing raw tags, duplicated headers, or blank assistant bubbles.
Artifact and Animation Viewers
The artifact side panel was tightened up by removing the redundant header block and moving the close control into the search row. The animation player now renders playable visual motion instead of only displaying storyboard text, and artifact viewers gained clearer handling for generated lesson, quiz, image, and animation outputs.
Question cards now support collapsing, and the duplicate visible question header was removed from the expanded state. The compact summary only appears when the question is hidden, which keeps the transcript from repeating the same prompt twice.
Model and Provider Updates
The model chooser now supports provider filtering as well as search, making large provider lists easier to scan. MiniMax M3 compatibility entries were added to the MiniMax model lists, and Keating packages were updated to the Pi 0.78.0 package line where available.
Provider setup copy and tutorial navigation were also tuned so the browser, llama.cpp, Ollama, LiteLLM, and related setup buttons behave like in-page navigation controls without unexpectedly jumping the learner back to the top of the tutorial.
Benchmark and Runtime Fixes
Real-learner benchmark scoring, deterministic synthetic learner simulation, outcome thresholds, and real/synthetic blending were centralized in a shared browser-safe helper. Real and synthetic blend coefficients now stay non-negative and sum to one as learner feedback accumulates, so benchmark scores do not distort once a topic has more real outcomes.
Browser benchmark traces now report real learner outcome counts and synthetic fallback state correctly. Vite development mode also gained the same /api/agent-runtime/remote/** proxy path as the Nitro production server, so remote agent mode no longer works only in production builds.
Site Navigation Polish
The blog itself now has search, release-line filtering, update types, a version-oriented table of contents, featured docs, and resource links. The tutorial learned the same lesson: topic navigation is more obvious, hover states make interactive controls clearer, mobile button groups avoid cramped layouts, and selecting a tutorial section no longer force-scrolls the page to the top.
Redundant Keating and Features nav links were removed from the landing and blog contexts, while logo hover affordances were brought closer to the chat-page treatment for a more consistent site shell.