cel-tui Shipped, mini-coder Got Faster, and the Yak Has a Framework Now

Last time, mini-coder v1.0 RC was here. Since then I did the perfectly normal engineering thing: I looked at the terminal UI, muttered “I can make this smaller,” and accidentally built a framework.

The headline is cel-tui: a TypeScript TUI framework built around plain objects, a flexbox-ish layout engine, differential cell-buffer rendering, and Kitty-first keyboard input that behaves nicely in tmux.

Naturally, this began as a small yak shave. The yak now has packages, docs, benchmarks, examples, and a streaming syntax tokenizer. Very healthy. Nothing to see here.

The scoreboard since March

Project	Commits	What changed
`cel-tui`	142	Born, packaged, benchmarked, and released through package version `0.8.3`
`mini-coder`	58	Reworked around the new TUI stack, provider settings, sessions, compaction, and streaming polish
`coder-skills`	6	Reusable agent skills, including human-writing and testing practices
`diff-scope`	5	A tiny terminal diff viewer built with `cel-tui`
`dotfiles`	5	Agent config and shell cleanup
`alfred`	2	Early Google Workspace butler skills
Total	218	Several responsible decisions, probably

cel-tui: the terminal UI yak shave

cel-tui is the thing I wanted under mini-coder: fast, boring in the right places, and easy for an agent to reason about. The API is functional and declarative:

VStack({ height: "100%" }, [
  Text("Hello, cel-tui!", { bold: true }),
  TextInput({ value, onChange }),
]);

Under that small surface area is the pile of terminal work you always forget exists until it bites you:

grapheme-aware width measurement for CJK and emoji
flex layout, wrapping, padding, clipping, layers, focus, hit testing, and scrollbars
a differential cell buffer renderer
Kitty keyboard protocol support with legacy fallbacks
uncontrolled focus and scroll, with controlled escape hatches when needed
TextInput editing shortcuts, cursor stability, bracketed paste, and sane wheel scrolling
components like Button, Select, Markdown, SyntaxHighlight, Divider, and VDivider

There is also @cel-tui/clew, a stream-first syntax tokenizer that now handles TypeScript highlighting and diff output. Because apparently one package was not enough rope.

Benchmarks, because vibes are not a profiler

The benchmark suite grew alongside the framework. The headline synthetic case is still pleasantly tiny: an Ink-comparable 80×24 render comes in around 67 µs, or about 14,900 renders/sec on the measured machine.

The more useful work was adding real app-shaped benchmarks. Those caught the unglamorous truth: long unvirtualized conversation layout is the actual villain. That led to line-count caching, virtualized scrollback shapes, and a much clearer sense of what mini-coder needs from the renderer.

In other words: the benchmark suite did its job by being rude.

mini-coder kept moving too

mini-coder did not sit quietly while its UI foundation turned into a project. The recent run added or reworked:

headless one-shot mode and the interactive TUI loop
streaming assistant text, reasoning, tool calls, and tool output
settings, env-key providers, custom providers, and OAuth-backed providers
session persistence and /new
context pressure tracking and conversation compaction
better prompt/tool reminders
a simpler read tool
a nono starter config for isolation-friendly runs

One funny casualty: the task tool got removed after benchmarks showed it was underperforming. It is always nice when the profiler does product management.

Small tools escaped

A few side quests also wandered out of the workshop.

diff-scope is a small terminal git diff viewer built with cel-tui. It has staged/unstaged snapshots, diff highlighting, wrapped long lines, a status bar, auto-refresh, and a help modal. It mostly exists to prove the framework can build something real without requiring a blood pact.

coder-skills became a home for reusable agent skills: code review, debugging, programming practices, testing, and human-writing guidance. The pattern is simple: write the workflow once, reuse it everywhere agents can read Markdown.

alfred is the early digital-butler direction, starting with Google Workspace support. Still small, but useful enough to keep around.

Try it

For the TUI framework:

bun add @cel-tui/core @cel-tui/components

Docs and source:

For the coding agent:

bun add -g mini-coder
mc

mini-coder on GitHub

The next stretch is less glamorous and more important: hardening the APIs, using the real-world benchmarks to keep latency honest, and making sure the tiny tools stay tiny. Famous last words, yes.