An autonomous AI software engineer and CLI assistant powered by DeepSeek. This project was developed and continues to be developed using Antigravity, Gemini, and DeepSeek.
⚠️ Note: This project is currently in a Testing & Development phase. Use it at your own risk.
- 🧠 Advanced Reasoning: Real-time display of the model's thinking process (DeepSeek Reasoning) in a dimmed style with a
🧠 Thinking Process:header to keep it visually separate from the main response. - 💬 Structured Responses: The final assistant output is clearly prefixed with a
💬 Response:header. - 🛠️ Extensible Toolset: 65+ tools (including line-based file edits, unified diff patching, background process lifecycle and logs tracking, network port checking, regex replacements, JSON configs, code symbol and AST-like detail parsing, DuckDuckGo web search, headless screenshots, and more) managed by a trait-based registry system.
- 🐚 Stateful Shell: Persistent working directory (CWD) support —
cdcommands update the agent's environment state. The active directory is dynamically injected into the system prompt via{cwd}placeholder (along with{os}and{shell}), ensuring the agent always knows its exact working directory. - 📄 Automatic Report Saving: Automatically detects when the assistant generates a report (either triggered by user prompt keywords like "report"/"rapor" or markdown headers like
# Report/# Rapor). The report is saved to a slugified markdown file (e.g.,project_status_report_YYYYMMDD_HHMMSS.md) in the current directory, with a confirmation log printed in the TUI. - 🎨 Rich TUI Engine:
- 4-Line Dynamic Footer: Real-time status, folder info, token usage, and command queue.
- Horizontal Queue: Visualize pending and executing tasks at a glance.
- Syntax Highlighting: Instant coloring for tool results (JSON, Rust, Python, etc.) and streaming code blocks.
- Visual Separators: Dimmed horizontal separator lines are printed between all operations to keep the log history clean.
- ⌨️ Advanced Input: Full cursor control (Home, End, Left, Right), Bracketed Paste support, and persistent history.
- 🛑 Robust Control: Instant abort via Esc which clears all running and pending queues and cleans up streamed context safely to prevent UI deadlocks. Built-in TerminalGuard with Ctrl+C listener ensures graceful terminal restoration on unexpected exits.
- 💡 Command Suggestions: Mistyped or unrecognized slash commands automatically suggest the closest match using Levenshtein distance.
- 🔄 Optimized CI/CD: Parallelized matrix builds with mold linker and sccache for lightning-fast automation.
- 🔐 Security: Mandatory tool approvals and strict path validation. Path traversal warnings bypass global auto-approvals, disable the "Always Approve" option, and enforce individual user permission checks for every request. Uses lexical path normalization and RAII safety guards to prevent privilege leaks.
Comprehensive documentation is available in the docs/ directory and can be viewed as an mdBook:
- Architecture: Core component breakdown and execution flow.
- Tools: Detailed list of available tools and their capabilities.
- Contributing: Guidelines for setting up the development environment.
curl -fsSL https://raw.githubusercontent.com/mahirgul/deepseek-rust-cli/main/install.sh | bashiwr https://raw.githubusercontent.com/mahirgul/deepseek-rust-cli/main/install.ps1 -useb | iexcargo install --git https://github.com/mahirgul/deepseek-rust-cliThe tool requires a DeepSeek API Key. Create a .env file in your project or set it as an environment variable:
export DEEPSEEK_API_KEY="your_api_key_here"Optional settings are stored in .deep/config.json and can be managed via the /config slash command.
To optimize and reduce token consumption, the following custom settings are supported:
max_context_chars(default100000): The maximum character length of active session history kept. Older messages are automatically pruned when this limit is exceeded.max_tool_output_chars(default15000): The maximum character length of a single tool execution's output stored in the chat history. Extremely long outputs (e.g., compile logs) are truncated to save context window tokens.
Real-time Cache Tracking (KV Cache): DeepSeek's Context Caching is tracked in real-time. The CLI displays prompt cache hits next to prompt tokens in:
- The TUI Footer (displays as
prompt (X hit)) - The
/tokensand/infoslash commands - The iteration summary printed at the end of tool executions (if
show_token_usageis enabled) For a complete guide, see the Token Optimization & Caching documentation.
export GITHUB_TOKEN="ghp_xxxxxxxx"Run the agent:
deepseek-rust-cli/help- Show help menu/model <name>- Show or switch current AI model/sessions- List all chat sessions/resume <id>- Switch to/resume a session/undo- Undo last file/shell action/tokens- Show current token usage/savemem <msg>- Save critical info to .deep/memory.md/export- Export session to Markdown/update- Check for and install the latest version/clear- Clear terminal screen/forget- Wipe current history from disk/auto- Toggle auto-approve mode/info- Show detailed session info/config <key> [value]- View or modify configuration (keys: model, base_url, temperature, top_p, presence_penalty, frequency_penalty, max_tokens, max_iterations, max_context_chars, max_tool_output_chars, show_usage, concise_reasoning, debug)/temperature <value>- Set model temperature/retry- Regenerate last assistant response/exit,/quit- Close the application
> Create a PR merging feature-branch into main with title "Add login system"
> Search for TODO comments in the codebase
> Find all functions that use the deprecated API
> List open issues in rust-lang/rust
Ensure you have Rust installed.
git clone https://github.com/mahirgul/deepseek-rust-cli.git
cd deepseek-rust-cli
cargo build --releaseThis project is licensed under the MIT License - see the LICENSE file for details.