DEV Community

Sulynn AI
Sulynn AI

Posted on

DevWit v0.7.30: the AI IDE that shows every token — now with a real terminal and a command palette

#productivity

Two months ago I posted DevWit here — an open-source AI IDE where every LLM request shows its full context manifest (per-item token costs, toggleable) and every agent file-write/shell-command goes through an authorization gate.

Since then it grew quietly. v0.7.30 shipped this week, and two of the new things are boring in the best way: a real terminal and a command palette.

Terminal panel (yes, it took this long)

The sidebar now has a fifth tab: a real shell session.

  • Real PTY (node-pty, with a pipe fallback), not a fake output pane
  • ANSI streaming — colors, bold/underline, the usual SGR set, rendered as you'd expect
  • Full keyboard passthrough — arrows, Ctrl+C, Tab-completion, IME input for CJK
  • Process-tree kill — closing a session takes down npm run dev and its children, not just the shell process
  • Sessions restart in one click; output is capped at 5000 lines so a chatty build can't eat memory

The honest limitation: it's a streaming renderer, not a screen buffer. Full-screen TUI apps (vim, htop) aren't supported — command output and REPL interaction are the target.

Command palette

Ctrl+Shift+P for commands, Ctrl+P for files — VS Code conventions. Fuzzy matching (subsequence + consecutive/word-start scoring), full keyboard loop (↑↓, Enter, Escape). It reaches the panels, the form toggle, settings, search, save, open-folder — the ten things you actually do all day.

The unglamorous part: eight rounds of adversarial review

Between v0.5.0 and v0.7.30, the entire codebase went through 8 rounds of adversarial code review (each round a fresh reviewer attacking a specific area: renderer, editor kernel, agent runtime, LSP/DAP/workspace, chat UI, MCP, first-run UX, terminal).

  • 147 findings, all fixed — including 5 criticals (a renderer-controlled path re-rooting bug that would have defeated the workspace containment, a debug-stop bug that could kill user processes)
  • Every fix got a regression test where feasible
  • The interesting ones are in the CHANGELOG: CSV formula injection in cost exports, a case-sensitivity escape in path containment on Linux, an opt-out beacon that briefly went to the wrong endpoint

This is the part nobody demos, but it's the part that decides whether an AI IDE touching your files deserves trust.

Numbers

  • 975 unit tests / 37 E2E suites (E2E drives the real packaged Electron app — real shell, real LSP, real git)
  • CI gates on lint + types + tests + quality checks for every push
  • Windows / macOS / Linux builds from public GitHub Actions

Install

  • winget: winget install eeyzs1.DevWit (0.7.28 live in the community repo, 0.7.30 in review)
  • Homebrew: brew install --cask eeyzs1/tap/devwit (0.7.30)
  • Or grab binaries from GitHub Releases

Still free, MIT, no accounts, telemetry opt-in and off by default.

If you care about context hygiene or compliance-friendly agent workflows, feedback is genuinely welcome. And if the direction resonates, a ⭐ on GitHub helps more than you'd think — it's the public signal needed to get free code-signing for Windows builds.

Top comments (0)