omp-uwu

Your coding agent, but it says "hewwo" (◕ᴗ◕✿)

Long coding sessions can feel a little grey. omp-uwu is an omp extension that adds playful uwu-speak, kaomoji and optional pastel colors to your agent's chat. A little warmth around the work, not inside your code. ♡

Quick start · Styles · Commands · Reported scores · Contributing

Quick start

omp plugin install omp-uwu

Restart omp after installation, linking a checkout, or changing extension code. On omp 18.6.1, /reload-plugins refreshes discovery and commands but does not re-evaluate already loaded extensions; exit and run omp --continue to load the changed code and resume your conversation. On a fresh install, the defaults are:

| uwu | Style | Level | Locale | Colors | |---|---|---|---|---| | on | rewrite | mid | auto | off |

Settings persist in ~/.omp/agent/omp-uwu.json (or the active agent directory when overridden). Open /uwu status to adjust them, or try /uwu colors on for the pastel palette. Existing saved settings take precedence over the defaults.

A real TUI, a softer voice

This GIF shows omp's real TUI components, captured without a model call: /uwu colors on over the dark-sunset theme. The pastel editorial frame is presentation artwork, not additional app UI. The prose is uwufied; inline code, numbers and the corrected loop remain exact in this example. How the demos are made.

Choose your style

| Style | Where styling happens | History and model context | |---|---|---| | rewrite · default | Deterministically, after the reply finishes streaming | Styled text enters history and context once the awaited hook is supported | | prompt | The model follows a system-prompt instruction while streaming | Model-written uwu text; results depend on the model | | display · experimental | Only while the ANSI TUI or Tern native Markdown draws prose | Original text stays in history; no uwu instruction goes to the model |

Rewrite starts with a prompt fallback. The first turn uses prompt style until omp demonstrates support for the awaited assistant_message hook. If that hook never arrives—for example on omp 18.4.3—prompt style remains the fallback. Once supported, rewriting needs no extra style-prompt tokens; code/tool blocks and metadata are untouched. The TUI refreshes after streaming finishes, but clients that render only streamed chunks may continue showing the original text.

Selecting any style also turns uwu on. Level and locale apply to all three:

• min: keeps prose unchanged and only adds occasional uwu, owo, kaomoji or emoji at sentence ends. Code and other protected spans stay exact. Prompt style requests this behavior from the model; rewrite is deterministic.
• low / mid / max: fewer changes and decorations / original strength / stronger styling.
• auto: protects both English and Turkish critical words. It does not detect the language.
• en: protects English critical words only.
• tr: adds Turkish forms on top of English protection. Locale never translates replies.

Negations, warnings and common Turkish inflections are protected conservatively, not by a full linguistic parser.

Command reference

Commands stay plain, even when the conversation gets fluffy.

| Command | What it does | |---|---| | /uwu | Open the TUI settings dashboard (plain summary outside the TUI); does not toggle | | /uwu on · /uwu off | Turn it on or off explicitly | | /uwu rewrite | Rewrite finished replies deterministically (default, with the fallback above) | | /uwu prompt | Ask the model to write in uwu while it streams | | /uwu display | Experimental: style prose only on screen, in the ANSI TUI or a discovered Tern native hook | | /uwu level min\|low\|mid\|max | Decorations only, light, standard or strong intensity | | /uwu locale auto\|en\|tr | Which critical words to protect (never translates) | | /uwu colors [on\|off] | Kawaii palette, sparkles and animated working cat (toggles without an argument) | | /uwu preview <text> | Show a sample with current settings, without changing anything | | /uwu status | Open the dashboard (plain one-line summary outside the TUI) |

In prompt style, /uwu preview is a deterministic approximation, not a prediction of the model's reply.

Your little control panel

• Controls: switches, style/level/locale chips, a level meter and a live sample.
• Compatibility: what omp has actually shown it supports in this session—rewrite hook, display hook and fallbacks. These are observations, not guarantees.
• ↑/↓ moves focus; ←/→ or Enter/Space changes a value; Tab switches tabs.
• Enter on Save applies the draft. Esc discards it. Moving around or editing the preview never saves settings or changes existing messages.
• Reset to defaults (Enter/Space) resets only the draft: on, rewrite, mid, auto, colors off. Choose Save to persist it, or Esc to keep the previous settings.
• Command completion lists actions and supported values. Invalid actions explain the rejection; confirmations point back to /uwu status.
• Action labels and keyboard guidance use the normal text color; focus and selected values also have cursor/bracket markers, not color alone.

Pastels & sparkles ♡

In the ANSI TUI, /uwu colors on applies a pastel palette to chat Markdown and the user-message bubble. Whenever uwu is on, a (◕ᴗ◕✿) uwu status badge identifies it independently of colors. ANSI chat prose gives uwu/owo per-letter rainbows; kaomoji such as (◕ᴗ◕✿) and (ﾉ◕ヮ◕)ﾉ, and glyphs like ♡ ☆ ✧ ✿, get a pastel tint. Tern native prose uses semantic Markdown marks: one stable color per distinct token, drawn with Tern's current palette.

While uwu mode and colors are on and omp is working, a larger full-body kitty crawls around the Tern agent pane, alternating leg/tail poses every 240 ms and briefly turning to face its direction at each corner. It climbs the sides and crosses the top without moving the working label or taking keyboard/pointer focus. Tern animates the transparent, non-modal layer locally; reduced-motion mode leaves one static kitty. This requires Tern's styles protocol feature. ANSI terminals and older Tern versions keep the seven-column spinner walk (ᓚᘏᗢ → ᗢᘏᓗ, with alternating legs and tail poses); compact running-tool spinners stay unchanged. /uwu colors off or /uwu off removes the kitty. Restart omp after updating extension code; /reload-plugins does not activate the change on omp 18.6.1.

Enable both switches, then send a message:

/uwu on
/uwu colors on

In Tern, look around the pane edges while omp is working. In ANSI terminals, look bottom-left, beside the elapsed turn timer. The GIF above shows the ANSI spinner's two walking cycles, not the native roaming layer; the surrounding frame is presentation artwork.

Sparkles are colored at render time, through ANSI prose transforms or Tern native Markdown marks:

• No ANSI codes enter message text or history. Code, code blocks and link targets are never painted.
• The palette is an in-memory TUI theme; other components sharing those colors may change too. It pauses omp's automatic theme detection until omp restarts.
• The pastel theme is not serialized into Tern's Surface Protocol. The native kitty uses an accent-colored text node in a hoisted, non-modal overlay and a surface-scoped stylesheet; no theme files are written. Tern owns its motion clock and geometry. Without stylesheet support, the kitty replaces the working spinner and repaints every 240ms.
• An existing host transform, such as live-voice transcript coloring, takes precedence.
• ACP/RPC clients and plain output are not colorized.

What changes—and what stays exact

Rewrite and display target natural-language chat prose. Their deterministic protection rules leave recognized fenced/inline code, URLs, paths, numbers, quoted spans, identifiers, config keys and critical words byte-for-byte unchanged. These are syntax-aware rules, not a guarantee that every possible prose token or safety-sensitive phrase is recognized; display also has run-boundary limits described below.

Tool-call arguments, file contents written or edited through tools, subagent prompts/messages and tool blocks are outside these transforms. Commit messages are not a styling target. In prompt style, including rewrite's first-turn or unavailable-hook fallback, keeping code, commands, quoted errors, files and commit messages exact is an instruction to the model—not a byte-preservation guarantee from the plugin.

The playful bits include r/l → w, occasional th → d, na/ne/no → nya/nye/nyo, occasional stutter, rotating kaomoji and the odd cute emoji (✨ 💖 🌸 🎀). Examples include ٩(◕‿◕｡)۶, (ฅ^•ﻌ•^ฅ) and (づ｡◕‿‿◕｡)づ, with examples drawn from kaomoji.you. The intent is readable meaning, numbers and warnings—not a safety or coding-quality guarantee.

Before & after, from real replies

These are real, unedited replies from anthropic/claude-opus-5-5, using the same prompt and model with the extension off and on, rendered as images. Both code blocks in this example are identical byte for byte. See the prompt, plain transcript and uwu transcript. This is an example, not a coding benchmark.

Reported kawaii scores

| Reported configuration | Kawaii/uwu-bench score | |---|---:| | Opus 5.5 | 0.1% | | Opus 5.5 + omp-uwu | 98.5% | | Derived difference: 98.5 − 0.1 | +98.4 percentage points |

A visualization of two existing reported scores, not a new benchmark run. The original README reported 0.1% and 98.5%; the difference above is derived from those values. No raw benchmark results, harness, sample size or scoring methodology are checked into this repository, so reproducibility and uncertainty cannot be assessed here.

These are reported kawaii scores, not evidence of coding accuracy, speed, cost or safety improvements. The legacy HTML chart is a visualization with decorative terminal/config panels, not raw benchmark data. The README now uses the static SVG instead.

Display mode: know the edges

Display mode uses omp's private per-message render hooks, not public extension API. The plugin finds AssistantMessageComponent through the shared Container base class and patches its ANSI transform plus its own describe() method when present. It also patches the shared Loader/TUI prototypes narrowly for Tern's native working-row kitty animation. These host hooks are experimental; if discovery fails, styling or kitty animation do nothing (the ANSI palette is independent). There is no silent fallback to prompt or rewrite. /uwu status separately reports ANSI discovery and whether a Tern native describe() hook was actually installed.

• ANSI: styling happens on individual Markdown prose runs before wrapping. Inline formatting, links, newlines and streaming edits split runs, so results can differ from rewrite or /uwu preview. Recognized identifiers, numbers and quoted spans within a run stay protected; quotes split across runs cannot be protected as a whole. Blockquotes, some tables, headings and math may skip this path.
• Tern native: on hosts with the discovered describe() hook (development API: omp 18.6.1), display transforms copies of top-level assistant Markdown text and sparkles add semantic md.marks, never ANSI in Markdown. Thinking, code, badges, usage, raw messages and history stay unchanged. Native-only source ranges protect recognized quote/list fences, indented code and exact-length multiline backtick spans before the prose rules run on the original lines. Unfinished code stays protected while streaming until closure or container exit; these are conservative rules, not a replacement for a complete Markdown parser. ANSI and finalized rewrite keep their existing protection behavior.
• Native marks have one color per literal token, not per-letter rainbows. Tern styles every occurrence of that literal in prose, without the ANSI regex's word-edge guards. Short ASCII tokens such as :3, :), :D, <3, x3 and ^^ are skipped to avoid overly broad matches.
• ACP/RPC, print and export retain original display-mode text. Host-set transforms take precedence on both render paths.
• Changing display, level, locale or on/off refreshes messages already on screen, with or without colors.

Install options

omp plugin install omp-uwu          # from npm (recommended)
omp plugin install [email protected]    # pin a version
omp plugin uninstall omp-uwu        # remove

If you previously copied a loose uwu.ts into ~/.omp/agent/extensions/, delete it after installing the plugin. Otherwise /uwu is registered twice.

Contributing

Small, readable changes are welcome. Keep the prose soft and the technical behavior precise.

bun install
bun run check          # tsc against the omp package types
bun test
omp plugin link .      # use this checkout instead of the installed copy
omp --no-extensions -e ./src/index.ts  # or load only this checkout for a single run

The full-module integration tests exercise real omp Markdown/Assistant components, narrow wrapping, inline/fenced code preservation, unchanged raw messages, cache invalidation without colors and host-transform precedence. Native cases exercise describe() text/marks, quote/list/indented code and multiline backticks with colors off/on, incomplete streaming code, untouched thinking, guarded node shapes, preserved host marks/metadata, repeated-description identity, streaming child reuse and reload/disposal. They restore patched prototypes after each case.

An additional installed-host test automatically uses ~/.bun/install/global/node_modules/@oh-my-pi/pi-tui if present. Set OMP_UWU_HOST_TUI to a different pi-tui package directory to test another installation; otherwise that case is skipped.

For an interactive smoke check, start omp --no-extensions -e ./src/index.ts with an isolated test agent directory and run /uwu colors off, /uwu display, then /uwu status. --no-extensions avoids loading the installed plugin alongside this checkout:

1. Edit the draft and switch tabs. The sample should update, but existing messages and saved settings should not change.
2. Press Esc; nothing should be saved. Reopen, focus Save and press Enter; settings should persist, and colors, the badge and assistant paragraphs on screen should refresh.
3. Ask for ordinary prose plus inline and fenced code. Prose should be styled and rewrapped (try a narrow terminal); code should stay exact.
4. With colors still off, change level, locale and mode. Existing assistant paragraphs should refresh after Save.
5. Reopen the saved transcript or export: raw text should be unchanged, and the next model prompt should have no uwu instruction in display mode.

For a no-model Tern smoke, resume a copied session containing assistant prose such as really lovely uwu (◕ᴗ◕✿) owo ☆, short :3, inline uwu code and a fenced uwu block. In the Tern pane, set PI_CODING_AGENT_DIR to an isolated directory with omp-uwu.json containing {"enabled":true,"colors":true,"style":"display","level":"mid","locale":"auto"}. Set OMP_TUI_DEBUG to a unique socket (Windows example: \\.\pipe\omp-uwu-probe), then run omp --no-extensions -e ./src/index.ts --resume <copied-session.jsonl>. The debug server accepts newline-delimited JSON: {"op":"tsp","n":5} must report native:true; {"op":"doc"} must show styled assistant md.p.text and semantic md.p.marks, with no ANSI. Send {"op":"paste","text":"/uwu colors off"} and {"op":"keys","keys":"enter"}, then inspect doc again: sparkle marks should be gone while display prose remains styled. Check visually that only prose is tinted, not inline/fenced code. Repeat in a fresh session with PI_TUI_NATIVE=0 for the unchanged ANSI path. Do not infer native mode from TERM_PROGRAM or TERN_LENSES; inspect the debug result.

For the native kitty, mount a real Loader in an isolated test widget and call setWorkingRow() with a working label (no model or API key needed). With the styles feature, the kitty's non-modal overlay must be hoisted into layer, leaving the original working row intact. Observe movement around the pane, both directions, and a stable working label; typing and the stop control must remain usable. Movement must not send host animation frames. /uwu colors off and /uwu off must remove the layer; turning them back on resumes the cat. Inject {"op":"bytes","data":"\u001b_tsp;e;{\"ev\":\"motion\",\"reduce\":true}\u001b\\"} to verify a still kitty; send reduce:false to resume. Check a narrow pane for clipping. Removing the widget must remove its layer; reloading the extension must not duplicate cats or timers. Without styles, verify the existing seven-column working-spinner animation instead.

Demo assets

The three static README panels—readme-hero.svg, uwu-bench.svg and style-map.svg—are generated locally:

bun demo/readme-assets.ts

They use pastel instrument-panel layouts. The score panel visualizes only the two existing reported values and their derived delta; generating assets does not execute a benchmark.

The before/after replies come from real transcripts in demo/. record.sh captures one reply with the extension loaded and one without it, in a throwaway agent directory so personal rules and extensions do not affect the replies. The side-by-side image is drawn directly from those transcripts.

The three GIFs need no model. capture.ts renders omp's real TUI components with the kawaii palette over dark-sunset and sparkles installed, then saves their ANSI output:

• chat.json: the prompt typed into a user bubble, then the uwu reply streaming into an assistant message.
• dashboard.json: the /uwu status card while a scripted key sequence edits it. Edit the script in capture.ts to change what the GIF shows.
• kitty.json: two walking cycles from the real status-line brand segment, sampled every 240ms.

render.py draws the before/after image and all three GIFs with matching plum-and-pastel frames, pixel motifs and hairline dividers. It preserves the captured TUI content and transcript text; rebuilding the artwork does not record a new session. On Windows, Gadugi supplies the kitty glyphs missing from Consolas; elsewhere use --font with a font covering Canadian syllabics:

mkdir -p /tmp/omp-demo && cp ~/.omp/agent/agent.db* /tmp/omp-demo/
demo/record.sh
COLORTERM=truecolor bun demo/capture.ts
uv run --with pillow python demo/render.py

Model output varies between runs, so re-run record.sh until you get a reply that reads well. capture.ts is deterministic; re-run it whenever transcripts, the palette, dashboard or kitty animation change.

The old uwu-bench.html and uwu-bench.png remain available as legacy artwork, not benchmark evidence or the README's current graph. To recreate that legacy screenshot, edit the HTML and take a 1536×960 screenshot with headless Chrome:

chrome --headless=new --hide-scrollbars --window-size=1536,960 --screenshot="$PWD/demo/uwu-bench.png" "file://$PWD/demo/uwu-bench.html"

CI (ci.yml) runs check and tests on pushes to main and on pull requests. Pushing a v* tag runs them again, checks that the tag matches version in package.json, and publishes to npm with provenance (publish.yml).

# bump "version" in package.json, commit, then:
git tag v0.9.0
git push origin main v0.9.0

The workflow uses npm trusted publishing, so the repository has no npm token. npm only allows trusted publishing on a package that already exists, so the first version is published by hand with npm publish --access public. After that, go to the package's Settings → Trusted publishing on npmjs.com and add GitHub Actions with user NaC-L, repository omp-uwu and workflow publish.yml.

License

MIT