sylvanus4

Meta-orchestrator that decomposes user goals into agent execution plans with model routing and dependency ordering, optionally pursuing them via Goal Mode with pursuing/paused/achieved/unmet/budget-limited lifecycle states. Maintains a persistent TODO backlog under state/todos/. Pipeline Mode auto-triggers for registered pipelines (report, persona, deep-research, research-report) and executes immediately. Use when: "jarvis", "자비스", "plan how to do X", "decompose this goal", "goal mode", "목표 모드", "pursue this until done", "jarvis goal", "jarvis plan", "자비스 계획", "자비스 목표", "long-running objective", "multi-turn goal", "멀티턴 목표", "장기 목표", "계획 세워줘", "jarvis todo", "자비스 TODO", "보고서 작성해줘", "리포트 만들어줘", "보고서 생성", "분석 보고서", "슬랙에 리포트 보고", "슬랙에 보고서 공유", "리서치해서 슬랙에 보고", "페르소나 분석", "관점 분석해줘", "관점 분석", "딥 리서치", "리서치해서 보고서로". Do NOT use for a single known skill (invoke directly), simple edits, or one-shot tasks. Do NOT use for daily pipeline (use today). Do NOT use for "딥 리서치" without "보고서" → use deep-research-pipeline.

End-to-end playbook to fine-tune and evaluate an LLM on ThakiCloud Metis (kubeflow-llm-training + vLLM serving on compute-h200). Covers served-model data generation, dataset→S3, TrainJob submission (with every admission gotcha), LoRA adapter serving for inference, and gate-based A/B eval. Use when "모델 학습시켜줘", "파인튜닝 돌려줘", "TrainJob 띄워줘", "학습 전후 평가/A/B", "데모 GPU로 학습", "fine-tune on the cluster", "train and evaluate", "serve the adapter and compare". General across models. Do NOT use for local/Colab training (use colab-unsloth-finetune) or serving-only ops (use demo-llm-inference-test).

Sync the #ai-platform-internal-qa Slack channel (demo.thakicloud.net QA reports) to a local corpus with full threads, filter out completion-tagged (완료/확인_완료) items, and dedup-check a new QA finding before posting it in the team's expert template. Use when "QA 이슈 올리기 전 중복 체크", "내부 QA 슬랙 싱크", "demo qa 중복 확인", "이 버그 이미 올라왔나", "internal-qa 채널", "check if this QA issue is a duplicate", "post QA finding". Do NOT use for GitHub issue dedup (use github-issue-dedupe), monitor/alert spam throttling (use slack-alert-dedup-cooldown), or general Slack posting (use scripts/slack_post_message.py).

Generate THAKI Design System (@thaki/tds) desktop-console UI from text / sketch / screenshot via a two-stage IR (UX-IR -> TDS-IR) with a deterministic compiler + validation gate + repair loop. Use when "TDS로 UI 생성", "화면 생성기", "디자인시스템 컴포넌트 트리 생성", "UI generator", "screenshot to TDS". Owns the TDS Component Registry (61 primitives) and the synthetic data engine for training a UI-gen model. Do NOT use for freehand React coding (use frontend-design) or Figma sync (use figma-to-tds).

Turn a known cartoon/comic character into a KakaoTalk-quality emoticon set (360x360 transparent, animated). Extracts a clean single-character reference, then generates held-cel sprite sheets with gpt-image-2 (body held, one part moves), splits/keys/stabilizes them, and compiles to KakaoTalk spec + GIF + PNG. Use when "이모티콘 만들어줘", "카카오 이모티콘", "캐릭터 이모티콘화", "움직이는 이모티콘", "emoticon set", "kakaotalk sticker set", "animate my character", "sprite emoticon". Do NOT use for one-off single stickers/messengers other than Kakao (use sticker-forge `stickerforge make`), plain image generation (use gpt-image-2), a pose sheet meant for dance/video (use pose-sheet-to-animation), or video (use video-producer).

Resume/continue development of 눈치맵 (nunchi-map) — a real-time local crowd/wait/on-site-photo app (대국민 눈치게임): enter a destination, see how crowded it is now, expected wait, and live '지금 뜨는 곳' ranking. Routes to the repos, explains architecture + how to run/verify, and lists the exact next gaps. Use when the user says "눈치맵", "누치맵", "nunchi-map", "대국민 눈치게임", "혼잡도 앱", "지금 뜨는 곳", or asks to continue/extend/deploy that app or its landing page. Do NOT use for ThakiCloud product repos (see [[thaki-platform-repos]]) or generic Go/React scaffolding unrelated to this app.

Evidence-grounded, leak-free skill-body evolution. Ports "Self-Harness" (arXiv:2606.09498) onto SKILL.md content: mines REAL execution-failure traces into clustered failure signatures (Weakness Mining), turns each cluster into a mechanism-targeted hermes mutation (Harness Proposal), and promotes variants through a leak-free non-regression gate with a sealed observe-only test split (Proposal Validation). Reuses hermes-skill-evolver as the variant engine — it does NOT replace it. Use when the user asks to "self-harness evolve", "evidence-grounded skill evolution", "evolve skill from real failures", "셀프하네스 진화", "실패 트레이스 기반 스킬 개선", "non-regression skill gate", "selfharness-evolve". Do NOT use for population/GEPA evolution on synthetic cases alone (use hermes-skill-evolver). Do NOT use for single-mutation quick fixes (use skill-autoimprove). Do NOT use for creating skills from scratch (use skill-creator). Do NOT use for the skill-ROUTER harness loop (use jarvis/runtime/SELF-HARNESS.md).

Translates ThakiCloud blog posts from Korean to English and Arabic and audits ko/en/ar parity so all three languages ship together in every batch. Use when adding or refreshing blog content that must ship in all three languages, translating a ko post to en/ar, or auditing language parity gaps. Korean+English triggers: 블로그 번역, 다국어 블로그, ko en ar 같이 배포, 트릴링구얼, 언어 패리티, blog translate, trilingual parity, i18n sync. Do NOT use for publish-cap/curation decisions (use tech-blog-manager), for build/push (use tech-blog-deploy), or for nightly quality scoring/rewriting (use tech-blog-evolve).

Audit a claim, design, process, or codebase for cargo-cult reasoning — copying the form of something that works without its mechanism, mistaking a name for understanding, or following a ritual whose original reason is gone. For each finding it names the missing mechanism the ritual is standing in for. Use when the user says "is this cargo cult", "are we doing this for a real reason", "check this for ritual thinking", "naming isn't understanding", "왜 이렇게 하는지 근거 있어", "이거 그냥 흉내내는 거 아냐", "카고 컬트인지 봐줘". Do NOT use to quiz a person (use grill-me) or to teach a concept forward (use feynman-technique).

Run a belief, result, decision, or plan through a self-skeptic pass: what would I observe if this were wrong, what is the cheapest experiment that could falsify it, and what am I not counting because it's inconvenient. Turns a comfortable conclusion into a testable one. Use when the user says "sanity check this", "am I fooling myself", "what would prove me wrong", "stress-test my reasoning", "pre-mortem this", "내가 착각하는 거 아냐", "이 결론 검증해줘", "반증해봐", "이거 진짜 맞아". Do NOT use to quiz another person (use grill-me). Do NOT use to spot copied ritual in a design (use cargo-cult-detector).

Force real understanding by explaining a concept in plain language until the explanation has no gaps. Pick the idea, explain it as if to a child, mark every place the words go vague or borrow jargon, return to the source, and re-simplify until the chain is unbroken. Use when the user says "explain this simply", "feynman technique", "do I actually understand this", "teach it back", "쉽게 설명해줘", "파인만 기법", "이거 진짜 이해한 거 맞아", "가르치듯 설명". Do NOT use when the user wants to be quizzed instead of taught (use grill-me). Do NOT use to detect ritual/fake reasoning in an existing claim (use cargo-cult-detector).

Rebuild a claim or design from base truths instead of analogy or convention. Strip inherited assumptions, isolate the irreducible facts that must be true, reconstruct forward from only those, and mark every place convention diverges from necessity. Use when the user says "first principles", "reason from scratch", "why do we assume that", "strip the assumptions", "rebuild this from base truths", "제1원리", "기본부터 다시", "그 전제 진짜 필요해", "원점에서 생각". Do NOT use for a self-skeptic falsification pass on a finished belief (use dont-fool-yourself). Do NOT use to detect copied ritual (use cargo-cult-detector).

Maintain an explicit ledger of what you don't yet understand. Given a topic, paper, codebase, or conversation, extract the honest gaps — the things named but not grasped — into a tracked list, and turn each into a concrete next step that would close it. Ignorance made visible instead of glossed over. Use when the user says "track what I don't know", "list my knowledge gaps", "what am I missing here", "build my ignorance log", "open questions notebook", "모르는 것 정리해줘", "지식 공백 목록", "이해 안 되는 것 추적". Do NOT use to teach a single concept forward (use feynman-technique) or to hold a standing set of research problems (use twelve-favorite-problems).

Keep a standing set of about a dozen open problems you care about, and test every new idea, paper, tool, or trick against all of them — so scattered inputs compound into progress on what matters instead of evaporating. Helps define the set, then pattern-matches new material against it ("does this help any of my twelve?"). Use when the user says "twelve favorite problems", "my standing problems", "does this help anything I care about", "match this against my open problems", "열두 개 문제", "내 핵심 관심 문제", "이게 내 문제에 도움 되나". Do NOT use to log session-level knowledge gaps (use notebook-of-things-i-dont-know) or to reason one problem from scratch (use first-principles).

Interrogate the user about something they claim to understand by walking down the chain of "why" — definition, then mechanism, then edge case, then cause — one question at a time, until it reaches the exact rung they memorized instead of understood. Use when the user says "quiz me", "test my understanding", "keep asking why", "poke holes in what I know", "find where I'm faking it", "이해했는지 시험해줘", "계속 왜냐고 물어봐", "내가 진짜 아는지 캐물어봐", "허점 찾아줘". Do NOT use to teach the concept forward (use feynman-technique). Do NOT use to audit a written artifact or claim (use cargo-cult-detector or dont-fool-yourself).

Schema-gated hyperparameter search (HPO/AutoML) skill for any training backend. Owns the closed loop — baseline eval, a validated search space, N trials, metric scrape, next-config recommendation, and an independent final eval of the best checkpoint — while deferring the single training run to whatever trainer you already have (Kubeflow Trainer v2, Ray Tune, SLURM, a local subprocess). Use when "hyperparameter search", "HPO", "AutoML", "tune the fine-tune", "sweep". Do NOT use for a single training run (call your trainer directly).