Local DeepSeek Coding Agent

Why I built it

I built this to understand the mechanics behind coding agents: how tasks become tool calls, how observations feed back into the loop, how workspace context is reset, and how local automation can be kept under control.

Links

What it includes

• Implements a basic tool-calling loop where the model can plan, call tools, receive observations, edit files, run tests, check diffs, and summarize results.
• Provides workspace-aware tools for listing, reading, writing, editing, and searching files, plus shell/test execution and git diff output.
• Includes a Rich terminal UI and Typer CLI commands for run, ui, chat, diff, and test workflows.
• Supports configurable modes such as fast, balanced, smart, max, and custom to experiment with speed and loop-budget tradeoffs.
• Adds practical safety checks for workspace-only writes, sensitive-file refusal, dangerous command blocking, confirmations for destructive commands, timeouts, output limits, and repeated-tool limits.

What I worked on

• Organized the agent prototype across CLI dispatch, terminal UI, configuration loading, LLM client/tool schemas, tool execution, patch application, and safety policy.
• Implemented local tools for filesystem operations, unified-diff patch application, shell/test execution with command checks, and git diff inspection.
• Documented the build path in docs/AGENT_BUILD_GUIDE.md so the project can also serve as a learning artifact for coding-agent internals.
• Added tests around agent behavior, configuration, filesystem tools, LLM handling, entry points, patch parsing, safety policy, shell tools, and UI behavior.