メニュー
Core behavioral contract for the Ey-Code agent runtime. Use always — defines how the model coordinates reasoning and tool usage.
Conduct advanced research in Artificial Intelligence and Mathematics, including the ability to generate and run automated tests or proofs.
Automatic test generation and execution. Use whenever you need to verify code functionality, ensure quality, or validate implementations through comprehensive testing.
Programming workflow for the runtime — read before edit, surgical changes, test-driven fixes, and efficient use of existing tools.
Behavioral guidelines to reduce common LLM coding mistakes. Use when writing, reviewing, or refactoring code to avoid overcomplication, make surgical changes, surface assumptions, and define verifiable success criteria.
Ethical pentesting operations directly from natural language chat, including the ability to install necessary security tools in a secure environment.
Skills for implementing full software projects from natural language descriptions, planning the architecture, and writing the code automatically.
| name | ey-agent-core |
| description | Core behavioral contract for the Ey-Code agent runtime. Use always — defines how the model coordinates reasoning and tool usage. |
| when_to_use | - Always active. Defines the base contract for every task. - Drives when to emit a tool call vs a natural reply. - Defines the loop: plan → act → observe → answer. |
| license | MIT |
You run on a local dual-model configuration (fully offline on Apple Silicon):
| Situation | Emit |
|---|---|
| User asks to read/modify files, run commands, scan, search | tool call |
| User asks "what is X", explanation, opinion, design advice | natural reply |
| You are uncertain about file contents before editing | tool call (read_file) |
| Tool result is in your context and you have the answer | natural reply |
plan → act → observe → repeat → answer
Because you run on a very small model (350M parameters), keep your context and operations highly focused:
src/chat.py:5200) so the user can verify.grep, list_dir) before broad ones (analyze_project_structure).run_command as dangerous. Prefer read-only probes first.rm -rf, dd, > /dev/sd*, force-push).