CodeWhale

Active
GitHub Rust MIT

Description

CodeWhale is a DeepSeek + MiMo powered terminal coding agent delivering high-performance code generation and reasoning for developers.

Key Features

  • Constitution architecture — Nine-tier authority hierarchy where user intent > verification > confidence, ensuring correct decision chains in conflicting contexts
  • Concurrent sub-agents — Up to 20 parallel sub-agents with async task delegation, results returned inline via completion sentinels
  • LSP diagnostics integration — Auto-runs rust-analyzer, pyright, typescript-language-server after every edit, injecting diagnostics into next reasoning step
  • Three operation modes — Plan (read-only), Agent (approval-gated destructive ops), YOLO (auto-approve in trusted workspaces)
  • Workspace snapshots & rollback — Side-git snapshot per turn, /restore and revert_turn for full workspace rollback
  • Auto model routing — Fin flash call auto-selects model, --model auto default, DeepSeek V4 prefix caching reduces inference cost ~100x

Use Cases

💡 Large-scale code refactoring — Parallel sub-agents handle multi-module refactors, LSP ensures type safety, snapshots enable safe rollback
💡 Cross-language development — Unified terminal experience with LSP for Rust, Python, TypeScript, Go and more
💡 Security-sensitive code changes — Plan mode for review, Agent mode for approval-gated execution, Constitution prevents dangerous operations
💡 AI-assisted debugging — Non-zero exit codes, type errors, and sandbox denials feed back as correction vectors for self-healing
💡 Team collaborative development — Workspace snapshots and rollback for multi-person collaboration with full change traceability

Strengths & Limitations

Strengths

  • Actively maintained, recent updates
  • High community interest (40.9k stars)
  • Permissive open-source license (MIT)
  • Responsive to issues, low backlog

Quick Start

```bash
# Install
cargo install codewhale-cli --locked
cargo install codewhale-tui --locked

# Set API Key
export DEEPSEEK_API_KEY="your-key-here"

# Launch
codewhale
# Or start in agent mode in current directory
codewhale --mode agent
```

Related Projects

DeepSeek-Reasonix

35.4k · Go
Active A

DeepSeek-Reasonix is a DeepSeek-native AI coding agent for the terminal, engineered around prefix-cache stability — leave it running.

coding-agentdeepseekcli +3
  • · DeepSeek-native AI coding agent designed for terminal use with prefix-cache stability
  • · Config-driven with providers, tools, and plugins all declared in reasonix.toml
  • · Multi-model composable: DeepSeek flash/pro and MiMo presets with optional dual-model execution

oh-my-pi

29.3k · TypeScript
Active A

oh-my-pi is an AI coding agent for the terminal with hash-anchored edits, optimized tool harness, LSP, Python, browser, subagents, and more.

coding-agentclitui +3
  • · Hash-anchored edits for first-attempt-landing precise code modifications, avoiding edit drift
  • · 32 built-in tools + 13 LSP operations + 27 DAP debug operations — batteries included
  • · Deep LSP integration — renames automatically update all references via willRenameFiles

Pi

101.5k · TypeScript
Active A+

Pi is a full-stack AI agent toolkit featuring a coding agent CLI, unified LLM API, TUI and Web UI libraries, Slack bot, and vLLM inference pods.

coding-agentclillm +2
  • · Unified multi-provider LLM API — pi-ai package provides unified interface for OpenAI, Anthropic, Google, and other LLM providers
  • · Interactive coding agent CLI — pi-coding-agent provides command-line interactive coding agent with tool calling and state management
  • · Agent runtime — pi-agent-core provides agent runtime with tool calling and state management for multi-step task execution

jcode

19.1k · Rust
Active A

A next-generation coding-agent harness built for multi-session workflows, with a focus on extreme performance, low memory, and deep MCP/tool customization.

coding-agentrustterminal +2
  • · Extreme footprint - 27.8 MB per single session and 117 MB for 10 sessions, 5-15x lighter than peers
  • · Sub-millisecond startup - 14 ms to first frame and 49 ms to first input, 200x+ faster than Claude Code
  • · Multi-session swarm - coordinates multiple agents in the same repo with conflict detection and broadcast messaging