A terminal agent that ranks truth above user intent

What it does CodeWhale is a pair of Rust binaries — a dispatcher CLI and a TUI runtime — that turn DeepSeek V4 into a terminal-native coding agent. It edits local workspaces, streams reasoning blocks, and gates destructive actions behind user approval. The whole system is built around the idea that an agent needs a “harness” of rules and feedback loops, not just a model call.

The interesting bit The harness is literally a Constitution (prompts/base.md) with seven articles and a nine-tier hierarchy of authority. When user intent conflicts with stale project rules or tool output, the model knows exactly which source wins — live verification outranks assumptions, and the duty of truth is non-negotiable even against user requests. DeepSeek V4 helped write this constitution, and because V4’s prefix caching makes the long document cheap to re-read each turn, the model treats it like an open-book reference rather than a memorized prompt.

Key highlights

• A cheap deepseek-v4-flash “Fin” call auto-routes each turn to the right model and thinking level
• Sub-agents run concurrently (up to 20) with non-blocking launch and bounded transcript retrieval
• Every turn snapshots the workspace to a side-git repo for /restore or revert_turn
• Post-edit LSP diagnostics from rust-analyzer, pyright, gopls, and others feed back into context
• macOS Seatbelt sandboxing is active; Linux Landlock is detected but not yet enforced

Caveats

• Sandbox coverage is uneven: macOS Seatbelt is active, Linux Landlock is detected but not enforced, and Windows sandboxing is not yet advertised
• The Homebrew formula still uses the legacy deepseek-tui name
• The feature set — auto mode, Constitution, and cost routing — is built specifically around DeepSeek V4; other providers are not discussed

Verdict Developers who want a terminal-native DeepSeek V4 agent with built-in guardrails, automatic cost routing, and per-turn rollback will find this a structured alternative to raw API chat. If you need broad model provider support or fully cross-platform sandboxing, look elsewhere.