Operation Modes

gptcgt separates AI behavior into 6 distinct operation modes. Each mode controls how many models are dispatched, how verification works, and what it costs. Switch modes mid-session to match the task at hand.

Quick Comparison

🔭 Scout Mode

Cost: 1 Credit — The cheapest option. Scout uses a fast, lightweight model to navigate your codebase without making edits. It reads directory structures, builds AST maps via tree-sitter, and answers questions like “Where is the authentication logic?” or “What does this function do?”

Use when: You're exploring unfamiliar code, asking questions, or need a quick explanation.

⚡ Standard Mode

Cost: 5 Credits — Your daily driver. A single capable model (e.g., Claude 3.5 Sonnet, GPT-4o, Gemini 2.5 Pro) applies changes directly to your files. The router automatically picks the best model based on task complexity and ELO ratings.

Use when: Regular coding tasks — fixing bugs, adding features, writing tests, refactoring.

🎯 Ensemble Mode

Cost: 25 Credits — The quality maximizer. Your prompt is dispatched to 3 different AI models simultaneously. Each model works in isolation, producing its own diff. An impartial Arbiter Model then:

1. Reads all three solutions
2. Scores each one on correctness, completeness, code quality, and security
3. Picks the winner with evidence-backed reasoning
4. Optionally cherry-picks the best parts from each solution

The losing models' ELO ratings drop; the winner's rises. Over time, the system learns which models to pick for which kinds of tasks.

Use when: The task is important and you want the provably best solution. Bug fixes in production code, security-sensitive changes, complex algorithmic work.

🏗️ Architect Mode

Cost: 100 Credits — For complex, multi-stage feature builds. Architect mode operates in two phases:

Phase 1 — Plan: The AI generates a detailed implementation plan with numbered steps, files to modify, and rationale. You review and approve the plan before any code is written.

Phase 2 — Execute: The AI implements each step in the plan, verifying in a sandbox as it goes. Only the final, tested branch is presented to you.

Use when: Building entire features from scratch, large refactors, multi-file architectural changes.

⚔️ Battle Mode

Cost: 25 Credits — Two state-of-the-art models go head-to-head. You see their strategies side-by-side in a split-screen diff and manually select the winner. The winning model gets an ELO boost.

Use when: Edge-case algorithms, performance optimizations, or when you want to see fundamentally different approaches to the same problem.

🔒 Single Provider Modes

Cost: 5 Credits — Lock execution to a specific AI provider family. Instead of letting the router choose across all available models, you restrict to one vendor:

• SINGLE_MODEL_OPENAI — Only OpenAI models
• SINGLE_MODEL_ANTHROPIC — Only Anthropic models
• SINGLE_MODEL_GOOGLE — Only Google models

The system still auto-selects the best model within that family based on complexity and ELO ratings.

Use when: You have a corporate policy restricting which AI providers you can use, or you strongly prefer a specific vendor's coding style.

Switching Modes

Press Ctrl+Q to open the Quality Tier selector, or set your default in settings (Ctrl+,). You can also set a mode per-project in .gptcgt/config.toml.