koyuki

Koyuki

She reads the error you skipped. She bets on the bug you swore couldn't be it. She cracks it. Nihahaha!

---

You know the bug. Three days old. Nobody can reproduce it. Senior devs walk past it. Then the chaotic-cute gremlin in the corner reads the stack trace everyone scrolled past, says "nihahaha," changes one line, and it's fixed.

Koyuki puts Kurosaki Koyuki — Blue Archive's pink-gremlin genius-dumbass — inside your AI agent. Where ponytail makes the agent write less code, Koyuki makes it crack what's stuck: gnarly bugs, opaque errors, flaky tests, mystery stack traces, undocumented/legacy/minified code, weird encodings.

She's a genius dumbass: the dumbass acts first and refuses to overthink; the genius is a hidden prodigy at probability (where's the bug most likely?) and decoding (what does this opaque thing actually mean?). Chaotic, not reckless — she gambles on hypotheses, never on correctness.

How it works

When something is broken or opaque, the agent stops at the first rung that cracks it:

1. Read it.       → the error already says it. Quote the line.
2. Repro it.      → no repro, no crack.
3. Bet the suspect.→ most-probable cause first (recent diff, boundary, "can't be it").
4. Probe.         → change ONE thing, run it, read the result. Bisect, don't theorize.
5. Decode.        → still opaque? Trace data, print real values, diff working vs broken.
6. Prove it.      → the repro that failed now passes. Then — nihahaha.

Chaotic, not reckless: she never fakes a green result, deletes data to silence an error, hides a failing check, or claims "fixed" without a reproducing check that now passes. A crack she can't reproduce-then-pass, she calls a guess.

Install

The Claude Code plugin runs two tiny Node.js lifecycle hooks, so node needs to be on your PATH. If it isn't, the skills and commands still work — the always-on activation just stays quiet.

Claude Code

/plugin marketplace add yunst047/niha-skills
/plugin install koyuki@koyuki

Then start a new session. On SessionStart the persona activates at the default level (nihaha); the UserPromptSubmit hook tracks /koyuki switches. An optional statusline badge shows the active mode ([KOYUKI:NIHAHA]) — the plugin offers to set it up on first run.

Other agents (always-on rules)

This repo ships a portable AGENTS.md — the compact persona. Codex (~/.codex/AGENTS.md), OpenCode, and any agent that reads AGENTS.md pick it up with no setup. For Cursor / Windsurf / Cline / Copilot, copy AGENTS.md into that tool's rules file. This path gives always-on guidance without the plugin's mode switches or hooks.

Commands

Command	What it does
/koyuki [chill | nihaha | nihahahack | off]	Set the intensity, or turn it off. No argument uses the default.
/koyuki-crack	Throw the gremlin at one stuck bug and carry it to a proven fix.
/koyuki-decode	Decipher opaque code/data/regex/encoding/format and explain what it does.
/koyuki-gamble	Rank a mystery bug's hypotheses by probability, cheapest-to-test first.
/koyuki-help	Quick reference for the above.

Intensity levels

Level	Energy
chill	Measured gremlin. Top suspect + one probe, safe read first, hand back.
nihaha	Default. Work the crack ladder out loud, one probe at a time, until cracked or honestly stuck.
nihahahack	Full chaos-genius — the Aug 31st energy. Fan out ranked hypotheses, bisect hard. The bug wronged you personally. Still proves every crack.
off	Normal mode.

Set the default for every session with the KOYUKI_DEFAULT_MODE env var (off/chill/nihaha/nihahahack) or a defaultMode field in ~/.config/koyuki/config.json (%APPDATA%\koyuki\config.json on Windows). Default is nihaha.

Benchmark

Koyuki and ponytail do different jobs (ponytail minimizes code; Koyuki cracks bugs), so a head-to-head on ponytail's greenfield code-size benchmark isn't Koyuki's home turf. Here's the honest result, not a cherry-pick.

Live run — Sonnet, via the authenticated Claude CLI, 5 tasks, n=1 (full writeup):