The Unified Smart Bar

Every chat message is different. A quick unit conversion doesn't need the same firepower as debugging a production crash. And you shouldn't have to dig through settings menus to tell Meggy which model to use.

The Unified Smart Bar puts four controls into a single pill-shaped widget, right below the chat input field:

  1. Model / Category — choose a specific model or an entire category (Thinking, Fast)
  2. Thinking — a 3-way control: Off, On, or Deep, dialing how much the model reasons before answering
  3. Engine — dial the AI intensity and engine stubbornness from Quick to Try Till Done
  4. Sandbox — optionally tighten or relax the per-chat execution sandbox

Each conversation remembers its Smart Bar state independently, so you can have one chat pinned to a fast model for quick lookups and another using Try Till Done with Deep thinking enabled for a heavy research project.

The Engine preset determines the engine-side strategy (single pass vs. multi-approach with retries). The Thinking control independently controls the model-side reasoning budget. Selecting a higher Engine preset automatically defaults to using thinking-capable models.

Model & Category Selection

Tap the left section of the Smart Bar to pick which model (or model category) handles your messages. The dropdown is split into two tiers:

The Smart Bar label updates to show your current selection — either the category name or the specific model name — so you always know what's active.

Thinking Levels

The Thinking segmented control in the Smart Bar enables extended reasoning — sometimes called chain-of-thought. It has three discrete settings:

Level When to use What happens under the hood
Off Casual conversation, simple lookups No reasoning budget is sent — the model uses its built-in default
On Most everyday tasks where a little extra thought helps Sends a balanced medium reasoning budget to the provider
Deep Heavy research, hard debugging, multi-constraint planning Sends a high reasoning budget — slower and more thorough

When Thinking is on:

This is a per-conversation setting. Set it to Deep for a hard research thread, leave it on Off for casual conversation. The level persists across sessions.

Engine Presets

The right section of the Smart Bar lets you choose how hard Meggy tries. By default it shows Off — standard conversation with no engine overhead. Tap it to choose one of four engine presets:

Mode Prefix Icon What It Does Model Category
Quick L2 zap Single pass, no retries — fastest possible answer Fast
Standard L4 target Balanced multi-approach with automatic fallbacks Thinking
Thorough L6 microscope Deep research with task decomposition Thinking
Try Till Done L8 flame Full stubbornness ladder — Meggy never gives up Thinking

"Engine" vs. "Thinking" — these are independent. Engine controls the engine-side strategy (how many approaches Meggy tries before giving up). Thinking controls the model-side reasoning budget per attempt. You can run Quick + Deep, Try Till Done + Off, or any other combination — pick what fits your task.

When you pick a preset, two things happen:

  1. Your message is routed through the engine instead of plain chat — meaning Meggy can break it into subtasks, retry failures, and escalate approaches
  2. The model automatically switches to the category that fits the preset (fast models for quick answers, thinking models for complex problems)

Model Categories

Each execution preset has a preferred model category — a logical tier that maps to the models you've configured in Settings:

When you select a preset, Meggy picks the first model from the matching chain in your settings. You can always override this by selecting a specific model or model category from the model section of the Smart Bar — the preset sets the default, but you stay in control.

Stubbornness Levels

Behind each preset is a stubbornness parameter that controls how hard Meggy tries before giving up:

Mode Max Approaches Stubbornness Cap
Quick 1 Level 2
Standard 3 Level 4
Thorough 3 Level 6
Try Till Done 3 Level 8

When you select a preset, the Smart Bar pill updates to include a stubbornness-level prefix (e.g., L2 Quick or L8 Try Till Done) indicating the maximum stubbornness level.

Higher stubbornness means more retries with different strategies. At Level 8, Meggy cycles through every model in your chain, decomposes the task, and escalates approaches until the job is done — or the budget is exhausted.

Try Till Done (TTD)

TTD is the most powerful mode. It's designed for tasks where failure isn't an option:

TTD uses thinking-class models by default and pushes stubbornness to maximum. If the first approach fails, Meggy reformulates and tries again with a different strategy. If that fails too, it escalates to a more powerful model.

When to Use Each Mode

Scenario Recommended Mode
Quick math, conversions, or lookups Quick
Writing emails, summaries, or explanations Standard
Multi-step analysis or detailed code review Thorough
Critical tasks where you need a guaranteed result Try Till Done
Casual conversation or brainstorming Chat

Getting Started

  1. Open any chat conversation
  2. Look at the Smart Bar just below the input field
  3. Model row — tap to choose a model or category
  4. Thinking row — pick Off, On, or Deep
  5. Engine row — tap to choose an engine preset (Off / Quick / Standard / Thorough / Try Till Done)
  6. Type your message and send — Meggy handles the rest
  7. The message bubble shows which model actually processed your request, and the Thinking disclosure shows the reasoning trace when applicable

Each chat remembers its Smart Bar settings independently, so you can fine-tune per conversation.

What's Next?