deepseek-code

A minimal Claude-Code-style CLI coding agent powered by DeepSeek.

This is a v0 closed loop: streaming chat, tool calls with per-call permission prompts, and six built-in tools (Read, Write, Edit, Bash, Glob, Grep). The architecture mirrors Claude Code at a high level — generator-based agent loop, Zod tool schemas, awaitable permission gate, Ink-based TUI — so growing the surface (slash commands, sessions, MCP, hooks) is mostly additive.

Ctrl+C: cancels the current turn (aborts the in-flight LLM stream and any running shell command). Press twice while idle to exit.

Quick start

npm install
cp .env.example .env
# edit .env and set DEEPSEEK_API_KEY
npm run dev

Or build and run:

npm run build
./bin/dsc.js

Flags

Flag	Effect
--accept-edits	Auto-approve Read/Write/Edit; still prompt for Bash
--dangerously-skip-permissions	Auto-approve everything
--cwd <dir>	Set working directory

Architecture

src/
├── index.ts              # CLI entry: arg parsing, render Ink root
├── tui/                  # Ink components
│   ├── App.tsx           # root, global keybindings
│   ├── Repl.tsx          # message list + input + permission prompt orchestrator
│   ├── MessageList.tsx
│   ├── PermissionPrompt.tsx
│   └── transcript.ts
├── agent/
│   ├── query.ts          # async function* — the agent loop
│   ├── runTools.ts       # tool dispatch with permission gate
│   └── systemPrompt.ts
├── llm/
│   ├── client.ts         # DeepSeek (OpenAI-compatible) streaming client
│   └── adapter.ts        # tool/message ↔ OpenAI format, stream accumulator
├── tools/
│   ├── Tool.ts           # interface
│   ├── registry.ts       # builtinTools list
│   ├── ReadTool.ts
│   ├── WriteTool.ts
│   ├── EditTool.ts
│   ├── BashTool.ts
│   ├── GlobTool.ts
│   └── GrepTool.ts
├── permissions/
│   └── canUseTool.ts     # awaitable permission gate
├── config/
│   └── settings.ts       # .env loader
└── types/
    └── message.ts        # Message + AgentEvent unions

The agent loop

src/agent/query.ts is an async function* that:

1. Calls the LLM with the current messages[] and tools[].
2. Yields text deltas as they stream; consumers (the TUI) update the assistant row.
3. When the stream ends with tool calls, dispatches them through runTools — which awaits the canUseTool permission gate before each execution.
4. Appends the assistant message + each tool result to messages[] and loops, until the LLM responds with no tool calls (or maxTurns is hit).

Every tool execution flows through canUseTool, which is the seam where the TUI surfaces a Y/N prompt and blocks the loop until the user decides.

Adding a tool

Implement Tool<Input> from src/tools/Tool.ts (name, description, inputSchema, isReadOnly, renderPreview, execute) and add it to src/tools/registry.ts.

Roadmap

• Slash commands (/clear, /compact, /help)
• Session persistence + resume
• WebFetch tool
• Hooks (PreToolUse / PostToolUse / Stop)
• MCP client
• CLAUDE.md-equivalent project-memory injection

License

MIT