v0.5.47 — Design Mode, Skills, and a rebuilt agent team
- Design Mode lets you click nodes or draw on your 2D and 3D viewport to tell the AI what to change, without leaving the scene
- Skills replace scattered playbooks with one library you can extend, trigger with slash commands, and reuse across every project
- The orchestrator now delegates to a team of in-process experts that run in parallel, stream live in a side panel, and accept mid-run steering
- Permission modes give you Claude Code-style control over what runs automatically and what waits for your approval
Added
- Design Mode. Select a node in your 2D or 3D scene and a floating chat bar appears right on the viewport. Describe the change and it goes into the same chat, same model, same history as the sidebar. Cmd+Shift+D (Ctrl+Shift+D on Windows) toggles it.
- Draw on your game to build it. Drag on the viewport to circle an area, add a tree here, move this platform, fix the lighting in this corner. Summer captures a screenshot of your drawing and sends it with your message so the AI sees exactly what you mean.
- Fullscreen Iteration Mode. Collapse the chat sidebar or press Cmd+Shift+M to work fullscreen with a persistent bar at the bottom. The viewport fills the window while you iterate without losing the conversation.
- Skills framework. Summer ships playbooks for building games, lighting scenes, debugging, performance, and more. Type / in the chat to trigger one. Open Settings → Skills to read, customize, or write your own. Project skills override user skills override Summer defaults.
- Memory in Settings. Project decisions (art style, genre, naming rules) persist across chats and show up in a dedicated Memory section next to Skills.
- Slash commands. /make-game, /debug, /optimize, and /assets load the right playbook into your message so Summer starts in the right mode immediately.
- In-process expert team. The main agent plans and delegates. Explore reads your project. Builder writes scenes, scripts, and wires main scene, input, and project settings. Studio Asset Agent generates art and models. Summer expert verifies and repairs engine work. General Advisor gives a slower second opinion when the plan is unclear.
- Subagent side panel. Running experts show up as live tiles with transcripts you can drill into. Continue a thread, steer mid-run, or stop a run without losing context.
- Permission modes in the composer. Manual approves every edit and generation. Accept edits auto-runs file changes but asks before spending credits. Auto approve (default) runs edits and generation but still blocks destructive actions. Bypass auto-runs ordinary tools while critical recovery and catastrophic shell actions still require approval.
- Per-layer context breakdown. See how much of the context window goes to system prompt, tools, rules, skills, subagents, and conversation.
- Dictation in the composer. Hold the mic button to speak your prompt, release to insert, Enter to finalize.
- Work visibility. The chat shows what Summer is doing right now, including when it is waiting on a subagent or planning the next step.
Improved
- Auto mode default is now GPT-5.6 Luna at high reasoning. The orchestrator plans first and delegates build work instead of doing engine glue inline.
- Expert delegation runs inside one server invocation instead of queuing to Railway. Faster handoffs, live streaming, and durable run rows that survive reloads.
- Long asset jobs still run on Railway workers. When they finish, a durable wake resumes the chat automatically.
- Subagent transcripts render through the real chat UI with tool cards and markdown, not raw tokens.
- OpenAI-family models get first-class adapter support with cleaner reasoning history and fewer tool-schema rejections across providers.
- AI SDK 7 upgrade. Typed data parts, needsApproval on tools, and a hardened tool loop across the chat harness.
- Update banner shows phase-aware progress during slow applies and curated release notes for 0.5.47.
- Chat sidebar history finds projects by fingerprint first so conversations do not vanish on newer projects.
Fixed
- Design Mode overlay stays transparent on Windows WebView2 so the viewport shows through correctly.
- Checkpoint saves no longer interfere with active chat streams.
- Expert continuation chains group as one tile per agent instead of fragmenting across the panel.
- Snapshot cards deduplicate so you see one card per capture.
- SetResourceProperty accepts common field aliases from different models without rejecting valid edits.
- Theme provider nesting removed so light and dark mode apply consistently.
- Development database pool reuses connections across hot reloads.
Removed
- Legacy queued expert delegation (BullMQ → Railway → Redis poll). Experts are in-process now. Railway remains for long-running asset generation only.