14days-build-claude-code-cli

Follow the 14-day progression starting with a Typer CLI entry, REPL, and mock provider, then add a real Anthropic provider with tool_use handling on Day 2. Each day adds one capability such as file tools, safe edits, bash permissions, or sessions.

Register tools with a name, description, and JSON input_schema, then pass them to create_message. Loop on the response: when stop_reason is not end_turn, execute each tool_use block and append tool_result messages with matching tool_use_id values.

Yes, the project defaults to DeepSeek's Anthropic-compatible endpoint by setting ANTHROPIC_BASE_URL to https://api.deepseek.com/anthropic with your auth token. It also works with Claude or any Anthropic Messages API compatible service.

Use a read-before-edit pattern: read current content, verify the model's old_content matches the actual file to detect concurrent changes, show a unified diff preview, and write only after user confirmation.

This is an intentional security boundary from Day 3 that resolves paths against the current working directory and rejects anything outside it. You can expand allowed roots or replace the hard block with a permission prompt.

No, it is explicitly a teaching project focused on understanding agent harness architecture, tool calling, and permission systems. Production robustness, full Claude Code feature parity, and enterprise-scale performance are stated non-goals.