Loading Accelerated Logic…
GA Release

The Multi-Agent WebGPU Engine & Interactions API are now generally available.

Part 20: Settings, Appearance & The Troubleshooting Bible

Master system settings, theme customization, LLM parameter tuning, and reference the ultimate error troubleshooting bible.

Category: Reference • Read Time: 32 min read • Updated: 2026-11-10

1. System Settings Navigation

Access the Settings Modal via the header gear icon. The multi-panel modal features a real-time instant search bar (Find a setting...) spanning 13 configuration tabs:

Engine & API AI Configuration, Gemini API Key, Model Fallbacks, Image Generation.
Parameters & Agents LLM Parameters, Multi-Agent Systems, MCP Tools, Custom Skills.
Cloud & Data Appearance & Themes, Fact Memory, Google Drive Cloud, Local Backup & Export.

2. Appearance & Palette Themes

Customize the editor UI workspace with 3 instant themes saved in browser storage:

🌙 Dark Mode (Default)

Deep midnight contrast (#0f172a) optimized for late-night coding sessions and OLED power savings.

☀️ Light Mode

Crisp daytime readability with bright cards, clear contrast, and slate typography for office environments.

📜 Sepia Warm Mode

Eye-safe warm reading palette (#fbf0d9) that minimizes blue light exposure during long reading sessions.

3. LLM Parameter Engine Tuning

Fine-tune model generation characteristics under Settings → LLM Parameter Engine:

Sampling Temperature (0.0 – 2.0): Set to 0.0 - 0.3 for deterministic code and factual tasks. Use 0.7 - 1.0 for general chat, and 1.2 - 2.0 for creative brainstorming.

Global System Instructions: Inject custom system prompts ahead of every chat execution (e.g. "Always respond in TypeScript with Tailwind utility classes").

4. The Comprehensive Troubleshooting Bible

Bookmark this matrix for instant fixes to common runtime issues across providers, contexts, sandboxes, and cloud storage:

A. Provider & API Connection Errors

  • "Please select an AI model before starting the chat": Open the Model Organizer (click model pill) → Choose Provider Puter.js → Equip at least one model card.
  • "Gemini API key rejected (401)": Obtain a new key from aistudio.google.com/app/apikey. Ensure the Generative Language API is enabled in Google Cloud Console.
  • "Gemini quota or rate limit reached (429)": Check the provider quota and retry later. Model fallback chains are saved in settings but are not currently applied by request execution.
  • "Local Ollama Disconnected / CORS": Launch Ollama with CORS origins allowed: OLLAMA_ORIGINS="*" ollama serve.

B. Context, Image & File Errors

  • "Context Used: 95% / 100%": Unpin unused files from the Project Context bar, or start a New Chat inside the active folder to reset context.
  • "Model cannot see images": Switch to a model bearing the Visual badge (e.g. Gemini 3.5 Flash, GPT-4o, Claude 3.5 Sonnet).
  • "Image Generation Safety Refusal": Rephrase prompts to avoid restricted trademarks or brand names.

C. Web Sandbox & Preview Errors

  • Blank White Preview Screen: Open Web Sandbox → inspect the Console drawer for missing imports (e.g. useState) or missing default exports.
  • Unstyled Tailwind Components: Ensure src/index.css contains @import "tailwindcss"; or index.html includes the CDN script.
  • GitHub Push 401 Bad Credentials: Generate a GitHub PAT with the repo scope enabled.

5. Building Your Own Troubleshooting Playbook

The Bible above covers the common errors. Your personal playbook covers your recurring ones. Ten minutes once saves an hour every month:

  1. Capture the exact error string. "429" is noise; "quota_exceeded: tokens per minute (limit 6000) on model X" is actionable. Screenshot or copy it into a note named after the signature.
  2. First suspect ordering: (1) provider/quota, (2) local runtime state (Ollama/WebLLM loaded?), (3) context size, (4) browser/site data corruption, (5) app bug — in that order, because it is also the order of frequency.
  3. Bisect by substitution: same prompt + different model. If it passes, the issue is provider-side; if it fails identically, it is your prompt/context/data.
  4. Bisect by environment: same prompt in a clean browser profile. Fixed there → your site data is corrupt (export what you need, then clear). Still broken → app-side, report with the repro steps.
  5. Write it down. A personal notes.md of signature → fix pairs accumulates value indefinitely; the playbook is yours, not the app's.

6. Settings Search — Finding Any Toggle in Seconds

The Settings modal's instant search matches tab names, section titles, and toggle labels. Query with the noun you remember, not the menu path:

fallback

Jumps to the Model Fallbacks / fallback chain area — remember, saved order is not applied at request time (Part 8).

memory

Lands on Fact Memory settings — AI-create/delete permissions and the reset controls.

drive

Opens Google Drive Cloud sync — selectors, encryption vault behavior, disconnect.

mcp

Straight to MCP Tools & Servers — search toggle, server entries, built-in tool switches.

temperature

Lands on LLM Parameter Engine — sampling controls and global system instructions.

robot

Display section — Robot Companion visibility and density options.

7. Three Tuning Profiles — Beginner, Daily Driver, Privacy-First

Knob Beginner Daily Driver Privacy-First
Model routingAL Router defaultsPuter/OpenRouter mix by taskOllama / WebLLM / Gemini Nano only
TemperatureDefault0.2–0.3 code, 0.7 chat0.0–0.3 (deterministic, fewer surprises)
Multi-AgentAuto RouterMulti for builds, Single for Q&ASingle (fewest outbound handoffs)
MemoryDefault onOn, audited quarterlyOff or minimal, never personal facts
Drive backupOff until first export drillOn, chats+skills+memoriesOff; snapshot export to encrypted local disk
AI create/delete permissionsDefault (all off)Create on, review each Keep card; delete offAll off
Fallback orderIgnore itSaved for reference; switch manually on failureIgnore it (nothing auto-fails-over)

8. The Full Reset Path — Escalation When Nothing Helps

Work down this ladder; stop as soon as the symptom disappears. Export a snapshot before step 4 — it is the only destructive rung:

  1. Regenerate. Same model or a different one — separates flaky generation from systematic breakage.
  2. New chat + same prompt. Eliminates poisoned context/history as the cause.
  3. New browser tab / hard refresh. Drops wedged workers (the Python worker survives 20s, but a sick tab lies).
  4. Export snapshot, then clear site data (Settings → Purge / browser site settings). Wipes corrupt IndexedDB/localStorage mirrors — the classic fix for "app loads blank but a fresh profile works".
  5. Import the snapshot back on a clean state. If the symptom returns, the corruption was in the data — identify the poisoned slice by importing granular exports one at a time (models, then skills, then chats...).
  6. Still broken in a clean profile? The problem is environmental (proxy/VPN/firewall vs. the app or your provider) or an app regression — capture repro steps and report. Do not keep purging data; you are no longer in a data problem.