Chat & Composer
The prompt interface is the primary surface for communicating with Orbicus. Depending on whether you need exploratory technical discussion or automated code generation, you can switch between Chat and Composer modes.
Interaction Modes
Orbicus separates conversational discussion from active code modifications through two distinct modes:
| Mode | Purpose | Primary Behavior |
|---|---|---|
| Chat Mode | Open-ended codebase exploration, architecture reviews, debugging assistance, and conceptual Q&A. | Answers with text, architectural explanations, and copyable snippets without writing files directly to disk. |
| Composer Mode | Implementation, refactoring, test fixes, and multi-file code editing. | Inspects your codebase, forms an execution plan, edits files, and presents visual diffs for approval. |
You can toggle between modes at any point in a session depending on the task at hand.
Context Referencing with @
To ground the model's answers in specific files without manually copying code into prompts, use @-mentions:
Inspect @src/auth/jwt.ts and explain why token verification fails on expired cookies.
Mention Types
- Specific Files: Tagging
@src/components/Sidebar.tsxinjects the contents of that file directly into prompt context. - Directories: Referencing
@src/api/provides the agent with an overview of the directory structure and file signatures in that module. - Terminal Buffers: Reference recent build or test outputs so the agent can diagnose compiler and runtime errors directly.
Real-Time Step Execution
While generating answers or executing actions in Composer mode, Orbicus displays a chronological timeline of each operation:
- Tool Invocations: Transparently see when the agent reads a file, searches symbols via ripgrep, or inspects directory listings.
- Interactive Approvals: When an action requires verification (such as modifying source files or running terminal commands), execution pauses and presents an approval prompt.
- Cancellation: Press
Escor click Cancel at any moment to immediately halt generation without losing existing output.
Dynamic Model Switching
You are never locked into a single AI provider for an entire session. Using the model selector in the composer toolbar, you can switch between models mid-conversation:
- Use fast, low-latency models for quick syntax questions and code lookups.
- Switch to high-reasoning frontier models (such as Claude 3.7 Sonnet or GPT-4o) when initiating complex multi-file architectural refactors.
- Switch to offline local models (via Ollama or LM Studio) when working in air-gapped or restricted privacy environments.