feat: add Korean language compression modes

[levx-me]

Summary

• Add korean-lite/full/ultra intensity levels for Korean language compression
• Korean-specific rules: drop particles (은/는/이/가/을/를), use noun endings (함/됨/필요/완료), remove filler words (사실/기본적으로/정말), empty pleasantries (드리겠습니다), hedging (것 같습니다), connective endings → symbols (때문에→→), honorific stripping (시/으시)
• Hook support: /caveman korean command parsing, Korean per-turn reminder, alias resolution
• Benchmark infrastructure: --lang ko flag for run.py, Korean prompts, results

Benchmark Results

Korean caveman mode: 67% average token savings (range 57%–84%)

Task	Normal (tokens)	Caveman Korean (tokens)	Saved
React re-render bug	611	236	61%
Auth middleware fix	1176	191	84%
PostgreSQL pool setup	2363	539	77%
git rebase vs merge	1149	429	63%
Async refactor	739	316	57%
Microservices vs monolith	1225	508	59%
Security review	817	337	59%
Docker multi-stage	1210	442	63%
Race condition debug	1254	252	80%
Error boundary	2851	1011	65%
Average	1340	426	67%

Model: claude-sonnet-4-20250514, 1 trial per prompt. Comparable to English caveman mode (65% avg).

Why

Current caveman rules are English-optimized (drop articles a/an/the, English filler words). Korean has no articles — compression requires targeting different grammar elements: particles, verb endings, connective endings, honorifics, and Korean-specific filler patterns. Same gap that wenyan modes fill for Classical Chinese.

What changed

File	Change
skills/caveman/SKILL.md	Korean intensity table rows, rules, and examples
hooks/caveman-config.js	korean-lite, korean, korean-full, korean-ultra in VALID_MODES
hooks/caveman-mode-tracker.js	/caveman korean[-lite|-full|-ultra] parsing + Korean per-turn reminder
hooks/caveman-activate.js	korean → korean-full alias resolution
rules/caveman-activate.md	korean added to switch line
evals/prompts/ko.txt	10 Korean evaluation prompts
benchmarks/run.py	--lang ko support
benchmarks/prompts_ko.json	10 Korean benchmark prompts
benchmarks/results/	Korean benchmark results JSON

Before / After

Before (caveman full, responding in Korean):

"이 함수는 사실 기본적으로 불필요한 API 호출을 하고 있는 것 같습니다. 확인해 보시면 좋을 것 같습니다."

Inconsistent compression — English rules don't map to Korean grammar.

After (caveman korean-full):

"함수 불필요 API 호출 있음. 제거 필요."

Systematic compression targeting Korean-specific patterns.

Design decisions

• Rules section written in English (matching wenyan pattern) — LLMs follow English instructions more precisely
• Korean examples use natural compressed Korean, not literal translations of English caveman
• Three levels mirror English/wenyan pattern: lite (grammar intact), full (particles dropped), ultra (maximum compression)
• Korean-specific compression strategies: particle dropping, noun endings, connective ending → symbol conversion, honorific stripping

🤖 Generated with Claude Code

[qodo-ai-reviewer]

Hi, When skills/caveman/SKILL.md can’t be read, hooks/caveman-activate.js falls back to an English-only ruleset and switch list, so standalone installs won’t get Korean instructions even if the active mode is Korean.

Severity: remediation recommended | Category: correctness

How to fix: Add korean-aware fallback text

Agent prompt to fix - you can give this to your LLM of choice:

Issue description

Standalone path (no skills/caveman/SKILL.md) uses a hardcoded fallback ruleset that doesn’t mention Korean modes/rules, even though modeLabel can be korean-*.

Issue Context

Affects standalone hook installs; plugin installs likely have SKILL.md available.

Fix Focus Areas

• hooks/caveman-activate.js[47-111]

---

Qodo code review - free for open-source.