thecode-agent 0.1.0

A secure, provider-independent coding agent for the terminal

TheCode

TheCode is a provider-independent AI coding agent with a premium Textual terminal interface. It can inspect code, search files, plan changes, edit files with approval, execute bounded commands, run tests, analyze Git, and retain local session history.

TheCode is an early product foundation. Review proposed actions and keep important work under version control.

Requirements and installation

• Python 3.12+
• Git
• Ripgrep is recommended

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -e ".[dev]"
cp .env.example .env
thecode config --setup
thecode doctor

Usage

thecode
thecode chat
thecode ask "how does authentication work?"
thecode plan "migrate JWT to OAuth2"
thecode run "fix the failing tests"
thecode review
thecode init
thecode config
thecode models
thecode sessions
thecode resume SESSION_ID
thecode undo SESSION_ID

thecode init creates .nexus/config.yaml and NEXUS.md without overwriting existing files. The former tablutech and nexus commands remain compatibility aliases.

Configuration

Global configuration is read from ~/.nexus/config.yaml; project configuration from .nexus/config.yaml takes precedence. Provider secrets belong in environment variables, never YAML.

llm:
  provider: openrouter
  model: anthropic/claude-sonnet-4
agent:
  max_steps: 30
permissions:
  mode: ask
context:
  max_characters: 120000

Supported providers are OpenRouter, OpenAI, Anthropic, and Ollama. Set the corresponding key from .env.example. Run thecode config --setup to choose the provider and model interactively. API keys are entered with hidden input and stored only in the project .env, which is ignored by Git.

Safety model

Every path is resolved against the project root, including symlinks. Sensitive files require explicit authorization. Commands use argument arrays, never shell=True; shell operators and destructive patterns are blocked. Write tools preserve previous content, and session rollback refuses to overwrite subsequent user edits.

Permission modes:

• safe: reads are automatic; writes and commands require confirmation.
• ask: reads are automatic; mutating actions request approval.
• auto: low-risk operations are automatic; high-risk actions still ask and critical actions remain blocked.

Architecture

The CLI composes an agent loop, a provider-neutral LLM interface, a typed tool registry, deterministic permission/security layers, repository services, and SQLite session storage. Providers normalize their native tool-call formats into LLMResponse; tools expose Pydantic JSON schemas and return ToolResult. This keeps provider, UI, and execution concerns independent.

Project instruction priority is NEXUS.md, AGENTS.md, CLAUDE.md, CONTRIBUTING.md, then README.md.

Development

ruff check .
mypy nexus
pytest

HTTP providers are designed for httpx.MockTransport; tests do not require live API access.