By yule-studio
Yule 부서 에이전트의 슬래시 명령어 + 스킬 번들 (생성물). SSoT 는 agents/grants/slash-command-grants.json + 레지스트리 markdown.
> 진입점은 [`/AGENTS.md`](../../AGENTS.md), 전역 규칙은 [`/CLAUDE.md`](../../CLAUDE.md).
이 문서는 engineering-agent lifecycle 의 각 단계가 어느 모듈에서 책임지는지 빠르게 찾기 위한 지도입니다. 코드 변경 전에 "이 책임이 어디에 있어야 하지?" 를 먼저 확인하는 용도이고, 큰 rename / 분할은 별도 브랜치로 진행합니다.
이 문서는 사용자가 Discord 에서 engineering-agent 를 실제 회사처럼 다루기 위한 운영 가이드입니다. 무엇을 입력하면 어떻게 흐르는지, 그리고 각 단계에서 사용자가 던질 수 있는 문구를 묶었습니다.
부서 게이트웨이가 LLM/RAG/research collector/agent evaluation 관련 작업을 받았을 때 이 역할이 담당한다. 다른 멤버가 자료를 모으고 합의하는 흐름이 LLM 측면에서 안정적으로 돌아가도록 설계와 개선을 자문한다.
부서 게이트웨이에서 도메인/서비스/API/데이터 계층 작업이 들어오면 이 역할이 담당한다.
External network access
Connects to servers outside your machine
Uses power tools
Uses Bash, Write, or Edit tools
Own this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimOwn this plugin?
Verify ownership to unlock analytics, metadata editing, and a verified badge. GitHub access is read-only (username + org membership).
Sign in to claimBased on adoption, maintenance, documentation, and repository signals. Not a security audit or endorsement.
ForgeKit is a personal agent-operations forge — you wire up AI providers, equip agents with the right skills for a job, run them under operator-gated runtime policy, and keep every result in a durable memory. It is not another chatbot: it orchestrates multiple providers, agents and tools to carry a real task to a bounded finish.
forge (a smithy — heat metal, hammer it to shape) + kit (a set of tools).
A monorepo with two cooperating apps:
forgekit-console — the operator TUI (Claude-Code-style). You configure the brain,
switch runtime modes, resolve work, and read honest status.engineering-agent (yule) — an always-on, role-based engineering runtime (Discord
gateway + member bots, SQLite job queue, Obsidian vault mirror).Everything is provider-neutral (Claude / Codex / Gemini / Ollama and any openai-compatible or enterprise endpoint sit behind one contract) and operator-gated (approval / budget / safe-class boundaries are real, not decorative).
| Concept | One line | Code |
|---|---|---|
| ForgeKit | the whole platform / execution environment | this repo |
| Hephaistos | the skill-forging core — turns a request into an equip plan (agent + skills + loadout + weapons + work packet) | forgekit_console/hephaistos/ |
| Nexus | the external knowledge source (areas / patterns / snippets / troubleshooting) Hephaistos reads (never copies) | read path: hephaistos/nexus_read.py |
| Armory | the catalog of Skills / Loadouts / Weapons Hephaistos forges from | hephaistos/armory.py |
| Work Packet | a structured, executable unit (goal / scope / forbidden / commands / verify / approval / evidence) | hephaistos/models.py |
| Runtime Mode | the operator posture (Shift+Tab) that actually changes routing / budget / approval | policy/runtime_mode.py |
Mine → library → smith: Nexus is the mine/library, ForgeKit is the forge, Hephaistos is the smith.
Status is one of working / partial / planned / blocked — see the matrices
in docs/operator-surfaces.md and docs/evidence-map.md.
| Capability | Status | Surface | Evidence |
|---|---|---|---|
| Multi-provider config + routing (no implicit ollama) | working | submit path, /mode | examples/runtime-teeth/ |
| Vendor-native usage ledger (live vs estimate) | working | /usage | examples/usage/ |
| Hephaistos resolve (request → equip plan) | working | /resolve /hephaistos /skills /loadout | examples/hephaistos/ |
| Armory breadth (7 categories, 25 skills, 8 loadouts) | working | /resolve | test_armory_breadth |
| Nexus read path (bounded, restricted-aware) | partial | /resolve nexus line | examples/hephaistos/nexus-read-foundation/ |
| Always-on bounded daemon + safe-class autopilot | working | forgekit runtime serve | examples/runtime/, examples/autopilot/ |
Provider /provider console config (set primary / list / doctor) | working | /provider | test_provider_surface |
| Nexus live repo connection | planned (not_connected until FORGEKIT_NEXUS_ROOT set) | — | — |
| Live Figma / YouTube / Google / Instagram | planned seam (never live in this tree) | /sources shows planned | — |
No fake-live. Anything not wired is shown as planned / not_connected / blocked /
restricted / unsupported_in_console, never as if it works.
pip install -e '.[console]' # console extra (textual + image render)
forgekit # open the operator console TUI
yule --help # the engineering runtime CLI (runtime/harness/doctor/…)
First run with no provider configured → the console reports setup-required and free-text
submit is held. You choose the brain — ForgeKit does not silently use a reachable
local Ollama (implicit local fallback is OFF by default).
npx claudepluginhub yule-studio/forgekitUnified capability management center for Skills, Agents, and Commands.
Brand Voice transforms scattered brand materials into enforceable AI guardrails — automatically. It searches across Notion, Google Drive, Confluence, Gong, Slack, and meeting transcripts to distill your strongest brand signals into a single source of truth, then applies them to every piece of AI-generated content. The more your team creates with Claude, the more consistent your brand becomes.