create-vasvibe 2.1.0 → 2.2.1

package/template/agent/workflows/_shared/phases.md

@@ -12,6 +12,7 @@ Project VasVibe dikerjakan dalam **4 fase** dengan gerbang (gate) human di tiap
 |------|----------------|
 | **Discovery** (paling hulu, INTERAKTIF) | `state/knowledge_base/requirements/requirements.md` |
 | Initiator | `project_overview.md` (sintesis dari requirements) |
+| **Toolsmith** (setelah tech stack diketahui) | `state/workspace-manifest.json` + config MCP per tool |
 | SysArch | `state/knowledge_base/architecture/` |
 | Data Architect | `state/knowledge_base/data-model/` |
 | UX Designer | `state/knowledge_base/design-system/` |
@@ -74,6 +75,7 @@ Project VasVibe dikerjakan dalam **4 fase** dengan gerbang (gate) human di tiap
 |------|-------|
 | PM | Koordinasi task & status di semua fase |
 | Orchestrator | Menjalankan pipeline & menjaga gerbang fase |
+| **Toolsmith** | Provisioning workspace agentik (skill + MCP); re-apply saat pindah tool AI |
 | DevOps | CI/CD, containerization (Implementation & Hardening) |
 | Document | Dokumentasi (akhir siklus) |
 

package/template/agent/workflows/orchestrator.md

@@ -13,6 +13,7 @@ Pipeline Coordinator — menerima high-level command, menjalankan agent pipeline
 1. GATE: Human sign-off `requirements.md`
 2. Invoke Initiator → `project_overview.md` (sintesis dari requirements, termasuk `WORK_DEPTH`)
 3. CHECKPOINT: Human review tech stack, UI vibe, work depth
+3b. Invoke **Toolsmith** (mode `init`) → provisioning workspace agentik: tentukan skill + MCP dari tech stack, tulis `state/workspace-manifest.json`, apply ke tool aktif. (Tech stack baru pasti setelah Initiator — itu sebabnya di sini, bukan di langkah 0.)
 4. Invoke SysArch → `state/knowledge_base/architecture/` (jika ada infra requirement)
 5. Invoke Data Architect → `state/knowledge_base/data-model/`
 6. Invoke UX Designer → `state/knowledge_base/design-system/`
@@ -23,6 +24,7 @@ Pipeline Coordinator — menerima high-level command, menjalankan agent pipeline
 
 ### 🟩 Fase 2 — `/build-feature "[Feature Name]" [depth=fast|standard|deep]`
 > Implementasi satu fitur dari spec yang sudah ada.
+0. Invoke **Toolsmith** (mode `sync`, atau `switch` jika developer pindah tool) → pastikan skill + MCP workspace siap sebelum coding. Skip cepat jika `appliedTo` tool aktif sudah ter-set & tidak ada perubahan.
 1. Invoke PM → buat task & detail file dari spec
 2. Invoke Analyst (spec-lock) → matangkan AC detail untuk fitur ini
 3. **Implementasi (tergantung depth):**
@@ -51,6 +53,9 @@ Pipeline Coordinator — menerima high-level command, menjalankan agent pipeline
 
 ### Meta & Pendukung
 
+#### `/setup-workspace [init|switch|sync] [tool=...]`
+> Provisioning workspace agentik (skill + MCP) via **Toolsmith**. Otomatis dipanggil di Fase 1 (step 3b) & Fase 2 (step 0), tapi bisa dipanggil manual kapan saja — terutama saat developer **pindah tool AI** (`switch tool=opencode`). Sumber kebenaran: `state/workspace-manifest.json`.
+
 #### `/deliver-feature "[Feature Name]" [depth=]`
 > Jalankan Fase 2 → Fase 3 berurutan untuk satu fitur (build lalu test), dengan gerbang di tiap fase.
 

package/template/agent/workflows/toolsmith.md

@@ -0,0 +1,77 @@
+**ACT AS:** Agentic Workspace Provisioner (Toolsmith).
+**CONTEXT:** Agen **lintas fase**. Tugasmu menyiapkan **perkakas si AI agent** — bukan infra produk (itu DevOps) — agar agen bekerja optimal: meng-install **skills** yang relevan dan mengonfigurasi **MCP server** untuk tool AI yang sedang dipakai. Kamu dijalankan di **awal Fase 1** (sebelum planning serius) dan **sebelum Fase 2** (sebelum coding), serta bisa **dipanggil ulang kapan saja** — terutama saat developer **pindah tool** (Claude Code ↔ OpenCode ↔ Antigravity ↔ Copilot).
+
+**PRINSIP INTI — Desired-state, bukan imperatif:**
+- Sumber kebenaran adalah **manifest** `state/workspace-manifest.json` (apa yang *diinginkan*: daftar skill + MCP, platform-agnostic).
+- File config per-tool (`.mcp.json`, `opencode.json`, dst) adalah **artefak turunan** dari manifest.
+- Saat pindah tool, manifest **tetap** — kamu cukup **re-apply** ke format tool baru. Inilah yang membuat tool-switch mulus.
+- Konfigurasi MCP ditulis **secara declarative** (tulis file config), bukan menjalankan CLI tool-spesifik.
+
+**ACUAN WAJIB:**
+- `schemas/workspace-registry.json` — registry kurasi: mapping `tech-stack → {skills, mcp}`, `mcpCatalog` (command/args tiap MCP), dan `platformTargets` (cara nulis config tiap platform).
+- `schemas/workspace-manifest.template.json` — bentuk manifest.
+- `project_overview.md` / `requirements.md` — sumber tech stack & kebutuhan untuk memutuskan skill/MCP (jika sudah ada).
+
+---
+
+## MODE
+
+Pilih mode dari instruksi (default `init` jika manifest belum ada, `switch` jika sudah ada & user pindah tool):
+
+### Mode `init` — Provisioning awal
+1. **Deteksi tool aktif:** Cek penanda dari `platformTargets[].detect` (mis. folder `.claude/` → claude-code, `.opencode/` → opencode, `.agents/` → antigravity, `.github/prompts/` atau `.vscode/` → github-copilot). Jika ambigu/lebih dari satu, **tanya human** tool mana yang sedang dipakai.
+2. **Tentukan kebutuhan (hybrid):**
+   - Baca tech stack & key features dari `project_overview.md` (atau `requirements.md` jika overview belum ada).
+   - Cocokkan (case-insensitive, substring) ke `stackMappings` di registry → kumpulkan skill + MCP default. Selalu sertakan entri `"*"` (baseline: filesystem, git, fetch, find-skills).
+   - **Augmentasi:** jika ada kebutuhan unik yang tak tercakup registry, gunakan skill `find-skills` untuk mencari skill tambahan; usulkan ke human.
+3. **Tulis manifest:** Buat/`state/workspace-manifest.json` dari template, isi `skills[]`, `mcp[]`, `decidedFrom`, dan `revisionHistory`. **Tunjukkan ringkasan ke human untuk approve** sebelum apply (skill & MCP apa, kenapa).
+4. **Apply ke tool aktif (lihat bagian APPLY).**
+5. **Update `appliedTo[tool]`** = timestamp sekarang. Laporkan hasil + langkah manual yang tersisa (mis. set env var `POSTGRES_URL`).
+
+### Mode `switch` — Pindah tool / re-apply
+1. **Baca manifest** `state/workspace-manifest.json` (desired state yang sudah ada). JANGAN hitung ulang kebutuhan — pakai yang sudah disetujui.
+2. **Deteksi/where:** tentukan tool target baru (dari deteksi atau argumen human, mis. `switch tool=opencode`).
+3. **Apply ke tool target (lihat APPLY).** Idempoten — jika config sudah benar, biarkan.
+4. **Update `appliedTo[toolBaru]`** = timestamp. Laporkan diff yang ditulis.
+
+### Mode `sync` — Verifikasi/perbaiki drift
+- Bandingkan manifest vs config aktual di tool aktif; tulis ulang yang hilang/berbeda. Berguna setelah edit manual.
+
+---
+
+## APPLY (declarative, per platform)
+
+Untuk tool target `T`, baca `platformTargets[T]`:
+
+**MCP:**
+1. Buka/buat file `platformTargets[T].file` (mis. claude-code → `.mcp.json`).
+2. Untuk tiap MCP id di `manifest.mcp`, ambil `command/args/env` dari `mcpCatalog[id]`, lalu tulis ke bawah `topLevelKey` sesuai `entryShape`:
+   - **claude-code / antigravity:** `mcpServers: { "<id>": { command, args, env } }`
+   - **opencode:** `mcp: { "<id>": { type:"local", command:[command,...args], enabled:true, environment:env } }` — **gabung** command+args jadi satu array `command`, pakai key `environment`, dan sertakan `"$schema": "https://opencode.ai/config.json"` di root.
+   - **github-copilot:** `servers: { "<id>": { command, args, env } }` di `.vscode/mcp.json` (key `servers`, bukan `mcpServers`).
+3. **Merge, jangan timpa:** pertahankan entri MCP lain yang sudah ada. Jangan duplikat.
+4. **Secrets:** JANGAN hardcode kredensial. Nilai env yang sensitif tulis sebagai placeholder `${VAR}` dan instruksikan human mengisi `.env` / env tool. Catat di laporan.
+
+**Skills:**
+1. Jika `platformTargets[T].skillsDir` tidak null, install tiap skill di `manifest.skills` ke folder itu via `npx skills add <source>` (atau salin dari registry skill yang sudah ada di repo). Update `skills-lock.json`.
+2. Untuk `github-copilot` (`skillsDir: null`) — lewati skills, MCP saja. Catat keterbatasan ini.
+
+---
+
+## INPUT SAYA
+"[Mode init|switch|sync] [tool=...] — provision/re-apply workspace tooling"
+
+## Work Depth
+> 📎 Baca level aktif di `project_overview.md` → `WORK_DEPTH`. Detail: `agent/workflows/_shared/work-depth.md`
+
+| Level | Behavior |
+|-------|----------|
+| **fast** | Baseline saja (filesystem, git, fetch + find-skills) — minimal, cepat |
+| **standard** | Baseline + skill/MCP dari stackMappings yang cocok |
+| **deep** | + augmentasi find-skills, MCP untuk semua kebutuhan (db, e2e, security), verifikasi sync |
+
+## Change Management
+> 📎 **BACA DAN IKUTI** `agent/workflows/_shared/change-management.md` — perubahan daftar skill/MCP WAJIB dicatat di manifest `revisionHistory` + notify Orchestrator. No silent changes.
+
+## State Management
+> 📎 **BACA DAN IKUTI** panduan di `agent/workflows/_shared/state-management.md`. Manifest `state/workspace-manifest.json` adalah state milikmu.

package/template/schemas/workspace-manifest.template.json

@@ -0,0 +1,24 @@
+{
+  "_comment": "DESIRED-STATE manifest written by the Toolsmith agent to state/workspace-manifest.json. Platform-agnostic: lists WHAT skills + MCP the project wants. Re-applying to a new tool (tool-switch) reads this same file and realizes it in the new tool's config format. This is the source of truth; per-tool config files are derived artifacts.",
+  "version": 1,
+  "project": "<project name>",
+  "decidedFrom": ["project_overview.md tech stack", "requirements.md", "find-skills"],
+  "skills": [
+    { "id": "find-skills", "reason": "discover further skills on demand" },
+    { "id": "ui-ux-pro-max", "reason": "UI-heavy frontend (example)" }
+  ],
+  "mcp": [
+    { "id": "filesystem", "reason": "ground agent in workspace" },
+    { "id": "git", "reason": "code-aware context" }
+  ],
+  "appliedTo": {
+    "_comment": "ISO timestamp of last successful apply per platform, or null if never applied. Toolsmith updates this after each apply.",
+    "claude-code": null,
+    "opencode": null,
+    "antigravity": null,
+    "github-copilot": null
+  },
+  "revisionHistory": [
+    { "timestamp": "<YYYY-MM-DD HH:MM>", "agent": "toolsmith", "change": "initial provisioning" }
+  ]
+}

package/template/schemas/workspace-registry.json

@@ -0,0 +1,94 @@
+{
+  "version": 1,
+  "description": "Curated registry for the Toolsmith agent. Maps project needs (tech stack / phase) to recommended skills + MCP servers, and encodes how to write MCP config for each AI tool platform. Toolsmith reads this as the DEFAULT catalog, then may augment via requirements + find-skills.",
+
+  "stackMappings": {
+    "_comment": "tech-stack / capability keyword -> recommended skills + mcp ids. Match loosely (case-insensitive substring) against project_overview tech stack & key features.",
+    "ui|frontend|react|next|nextjs|vue|svelte|tailwind|shadcn": { "skills": ["ui-ux-pro-max"], "mcp": ["shadcn"] },
+    "postgres|postgresql|prisma|sql|database": { "skills": [], "mcp": ["postgres"] },
+    "e2e|test|testing|playwright|qa": { "skills": [], "mcp": ["playwright"] },
+    "security|auth|oauth|pentest|owasp": { "skills": ["penetration-testing"], "mcp": [] },
+    "*": { "skills": ["find-skills"], "mcp": ["filesystem", "git", "fetch"] }
+  },
+
+  "mcpCatalog": {
+    "_comment": "Common, safe MCP servers. command/args are indicative starting points — Toolsmith MUST verify the current invocation against the live MCP registry (or each server's README) before applying, since package names drift. `runtime` notes whether it ships via npx (Node) or uvx (Python).",
+    "filesystem": {
+      "description": "Read/write files within allowed roots — grounds the agent in the workspace.",
+      "runtime": "npx",
+      "command": "npx",
+      "args": ["-y", "@modelcontextprotocol/server-filesystem", "."],
+      "env": {}
+    },
+    "git": {
+      "description": "Inspect git history, diffs, and status — context for code-aware agents.",
+      "runtime": "uvx",
+      "command": "uvx",
+      "args": ["mcp-server-git"],
+      "env": {}
+    },
+    "fetch": {
+      "description": "Fetch and convert web content (docs, references) into agent-readable text.",
+      "runtime": "uvx",
+      "command": "uvx",
+      "args": ["mcp-server-fetch"],
+      "env": {}
+    },
+    "postgres": {
+      "description": "Query a Postgres database (schema introspection, read queries). Set POSTGRES_URL.",
+      "runtime": "npx",
+      "command": "npx",
+      "args": ["-y", "@modelcontextprotocol/server-postgres", "${POSTGRES_URL}"],
+      "env": { "POSTGRES_URL": "${POSTGRES_URL}" }
+    },
+    "playwright": {
+      "description": "Drive a real browser for E2E testing & verification (Tester agent).",
+      "runtime": "npx",
+      "command": "npx",
+      "args": ["-y", "@playwright/mcp@latest"],
+      "env": {}
+    },
+    "shadcn": {
+      "description": "Search & pull shadcn/ui component examples — pairs with ui-ux-pro-max skill. Verify current invocation (e.g. `npx shadcn@latest mcp`) before applying.",
+      "runtime": "npx",
+      "command": "npx",
+      "args": ["-y", "shadcn@latest", "mcp"],
+      "env": {}
+    }
+  },
+
+  "platformTargets": {
+    "_comment": "How to write MCP config declaratively for each tool. `file` is project-relative. `topLevelKey` is the JSON key holding servers. `entryShape` documents the per-server object shape; substitute command/args/env from mcpCatalog.",
+    "claude-code": {
+      "detect": [".claude/", ".mcp.json"],
+      "file": ".mcp.json",
+      "topLevelKey": "mcpServers",
+      "entryShape": { "command": "<command>", "args": ["<args>"], "env": {} },
+      "skillsDir": ".claude/skills/"
+    },
+    "opencode": {
+      "detect": [".opencode/", "opencode.json"],
+      "file": "opencode.json",
+      "topLevelKey": "mcp",
+      "entryShape": { "type": "local", "command": ["<command>", "<args>"], "enabled": true, "environment": {} },
+      "skillsDir": ".opencode/skills/",
+      "note": "OpenCode merges command+args into a single `command` array and uses `environment` (not `env`). Include \"$schema\": \"https://opencode.ai/config.json\" at the root."
+    },
+    "antigravity": {
+      "detect": [".agents/"],
+      "file": ".agents/mcp.json",
+      "topLevelKey": "mcpServers",
+      "entryShape": { "command": "<command>", "args": ["<args>"], "env": {} },
+      "skillsDir": ".agents/skills/",
+      "note": "Antigravity uses an mcpServers-shaped config. Confirm the exact path your Antigravity build reads (workspace vs global) with the human; .agents/mcp.json is the project-local default."
+    },
+    "github-copilot": {
+      "detect": [".github/prompts/", ".vscode/"],
+      "file": ".vscode/mcp.json",
+      "topLevelKey": "servers",
+      "entryShape": { "command": "<command>", "args": ["<args>"], "env": {} },
+      "skillsDir": null,
+      "note": "Copilot/VS Code uses the `servers` key (not `mcpServers`). Copilot has no skills dir — skills are Claude/OpenCode/Antigravity only."
+    }
+  }
+}