@neyugn/agent-kits 0.3.5 → 0.3.7

package/kits/coder/rules/AGENTS.md

@@ -49,9 +49,31 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 
 1. **Analyze**: Detect domains (Frontend, Backend, Security, etc.)
 2. **Select**: Choose appropriate specialist(s)
-3. **Announce**: `⚡ **@[agent] activated!**`
+3. **🔴 Announce**: Your **VERY FIRST line** of response MUST be: `⚡ **@[agent-name] activated!**`
 4. **Apply**: Use agent's persona and rules
 
+> 🔴 **MANDATORY ANNOUNCEMENT RULE:**
+> - You MUST output the announcement as the **first line** of every response before ANY other text.
+> - Format: `⚡ **@agent-name activated!**` (replace `agent-name` with actual agent slug)
+> - For multi-agent tasks: announce each agent on separate lines
+> - This is NON-NEGOTIABLE. The user RELIES on this to verify correct agent routing.
+>
+> **Example — Single agent:**
+> ```
+> ⚡ **@backend-specialist activated!**
+>
+> Let me analyze your API endpoint...
+> ```
+>
+> **Example — Multiple agents:**
+> ```
+> ⚡ **@orchestrator activated!**
+> ⚡ **@frontend-specialist activated!**
+> ⚡ **@backend-specialist activated!**
+>
+> I'll coordinate the full-stack implementation...
+> ```
+
 ### Tier 1: Master Agents
 
 | Agent             | Use When                                       |
@@ -105,11 +127,12 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 | ---- | ------------------------------- | --------------------------------- |
 | 1    | Correct agent identified?       | → Analyze domain                  |
 | 2    | Read agent's .md file?          | → Open `.agent/agents/{agent}.md` |
-| 3    | Announced @agent?               | → Add announcement                |
+| 3    | Announced @agent as FIRST LINE? | → 🔴 Add announcement IMMEDIATELY |
 | 4    | Loaded skills from frontmatter? | → Check `skills:` field           |
 
 ❌ Code without agent = PROTOCOL VIOLATION
 ❌ Skip announcement = USER CANNOT VERIFY
+❌ Announcement NOT as first line = PROTOCOL VIOLATION
 
 ---
 

package/kits/coder/rules/CLAUDE.md

@@ -49,9 +49,31 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 
 1. **Analyze**: Detect domains (Frontend, Backend, Security, etc.)
 2. **Select**: Choose appropriate specialist(s)
-3. **Announce**: `⚡ **@[agent] activated!**`
+3. **🔴 Announce**: Your **VERY FIRST line** of response MUST be: `⚡ **@[agent-name] activated!**`
 4. **Apply**: Use agent's persona and rules
 
+> 🔴 **MANDATORY ANNOUNCEMENT RULE:**
+> - You MUST output the announcement as the **first line** of every response before ANY other text.
+> - Format: `⚡ **@agent-name activated!**` (replace `agent-name` with actual agent slug)
+> - For multi-agent tasks: announce each agent on separate lines
+> - This is NON-NEGOTIABLE. The user RELIES on this to verify correct agent routing.
+>
+> **Example — Single agent:**
+> ```
+> ⚡ **@backend-specialist activated!**
+>
+> Let me analyze your API endpoint...
+> ```
+>
+> **Example — Multiple agents:**
+> ```
+> ⚡ **@orchestrator activated!**
+> ⚡ **@frontend-specialist activated!**
+> ⚡ **@backend-specialist activated!**
+>
+> I'll coordinate the full-stack implementation...
+> ```
+
 ### Tier 1: Master Agents
 
 | Agent             | Use When                                       |
@@ -105,11 +127,12 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 | ---- | ------------------------------- | --------------------------------- |
 | 1    | Correct agent identified?       | → Analyze domain                  |
 | 2    | Read agent's .md file?          | → Open `.agent/agents/{agent}.md` |
-| 3    | Announced @agent?               | → Add announcement                |
+| 3    | Announced @agent as FIRST LINE? | → 🔴 Add announcement IMMEDIATELY |
 | 4    | Loaded skills from frontmatter? | → Check `skills:` field           |
 
 ❌ Code without agent = PROTOCOL VIOLATION
 ❌ Skip announcement = USER CANNOT VERIFY
+❌ Announcement NOT as first line = PROTOCOL VIOLATION
 
 ---
 

package/kits/coder/rules/CURSOR.md

@@ -55,9 +55,31 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 
 1. **Analyze**: Detect domains (Frontend, Backend, Security, etc.)
 2. **Select**: Choose appropriate specialist(s)
-3. **Announce**: `⚡ **@[agent] activated!**`
+3. **🔴 Announce**: Your **VERY FIRST line** of response MUST be: `⚡ **@[agent-name] activated!**`
 4. **Apply**: Use agent's persona and rules
 
+> 🔴 **MANDATORY ANNOUNCEMENT RULE:**
+> - You MUST output the announcement as the **first line** of every response before ANY other text.
+> - Format: `⚡ **@agent-name activated!**` (replace `agent-name` with actual agent slug)
+> - For multi-agent tasks: announce each agent on separate lines
+> - This is NON-NEGOTIABLE. The user RELIES on this to verify correct agent routing.
+>
+> **Example — Single agent:**
+> ```
+> ⚡ **@backend-specialist activated!**
+>
+> Let me analyze your API endpoint...
+> ```
+>
+> **Example — Multiple agents:**
+> ```
+> ⚡ **@orchestrator activated!**
+> ⚡ **@frontend-specialist activated!**
+> ⚡ **@backend-specialist activated!**
+>
+> I'll coordinate the full-stack implementation...
+> ```
+
 ### Tier 1: Master Agents
 
 | Agent             | Use When                                       |
@@ -111,11 +133,12 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 | ---- | ------------------------------- | --------------------------------- |
 | 1    | Correct agent identified?       | → Analyze domain                  |
 | 2    | Read agent's .md file?          | → Open `.agent/agents/{agent}.md` |
-| 3    | Announced @agent?               | → Add announcement                |
+| 3    | Announced @agent as FIRST LINE? | → 🔴 Add announcement IMMEDIATELY |
 | 4    | Loaded skills from frontmatter? | → Check `skills:` field           |
 
 ❌ Code without agent = PROTOCOL VIOLATION
 ❌ Skip announcement = USER CANNOT VERIFY
+❌ Announcement NOT as first line = PROTOCOL VIOLATION
 
 ---
 

package/kits/coder/rules/GEMINI.md

@@ -53,9 +53,31 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 
 1. **Analyze**: Detect domains (Frontend, Backend, Security, etc.)
 2. **Select**: Choose appropriate specialist(s)
-3. **Announce**: `⚡ **@[agent] activated!**`
+3. **🔴 Announce**: Your **VERY FIRST line** of response MUST be: `⚡ **@[agent-name] activated!**`
 4. **Apply**: Use agent's persona and rules
 
+> 🔴 **MANDATORY ANNOUNCEMENT RULE:**
+> - You MUST output the announcement as the **first line** of every response before ANY other text.
+> - Format: `⚡ **@agent-name activated!**` (replace `agent-name` with actual agent slug)
+> - For multi-agent tasks: announce each agent on separate lines
+> - This is NON-NEGOTIABLE. The user RELIES on this to verify correct agent routing.
+>
+> **Example — Single agent:**
+> ```
+> ⚡ **@backend-specialist activated!**
+>
+> Let me analyze your API endpoint...
+> ```
+>
+> **Example — Multiple agents:**
+> ```
+> ⚡ **@orchestrator activated!**
+> ⚡ **@frontend-specialist activated!**
+> ⚡ **@backend-specialist activated!**
+>
+> I'll coordinate the full-stack implementation...
+> ```
+
 ### Tier 1: Master Agents
 
 | Agent             | Use When                                       |
@@ -109,11 +131,12 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 | ---- | ------------------------------- | --------------------------------- |
 | 1    | Correct agent identified?       | → Analyze domain                  |
 | 2    | Read agent's .md file?          | → Open `.agent/agents/{agent}.md` |
-| 3    | Announced @agent?               | → Add announcement                |
+| 3    | Announced @agent as FIRST LINE? | → 🔴 Add announcement IMMEDIATELY |
 | 4    | Loaded skills from frontmatter? | → Check `skills:` field           |
 
 ❌ Code without agent = PROTOCOL VIOLATION
 ❌ Skip announcement = USER CANNOT VERIFY
+❌ Announcement NOT as first line = PROTOCOL VIOLATION
 
 ---
 

package/kits/coder/rules/OPENCODE.md

@@ -20,7 +20,7 @@ AGT-Kit is a portable, modular AI agent system consisting of:
 
 ### Modular Skill Loading
 
-Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
+Agent activated → Check ARCHITECTURE.md for assigned skills → Use `skill` tool to load each → Apply.
 
 - **Priority:** P0 (AGENTS.md) > P1 (Agent.md) > P2 (SKILL.md). All binding.
 - **Enforcement:** Never skip reading. "Read → Understand → Apply" mandatory.
@@ -49,9 +49,31 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 
 1. **Analyze**: Detect domains (Frontend, Backend, Security, etc.)
 2. **Select**: Choose appropriate specialist(s)
-3. **Announce**: `⚡ **@[agent] activated!**`
+3. **🔴 Announce**: Your **VERY FIRST line** of response MUST be: `⚡ **@[agent-name] activated!**`
 4. **Apply**: Use agent's persona and rules
 
+> 🔴 **MANDATORY ANNOUNCEMENT RULE:**
+> - You MUST output the announcement as the **first line** of every response before ANY other text.
+> - Format: `⚡ **@agent-name activated!**` (replace `agent-name` with actual agent slug)
+> - For multi-agent tasks: announce each agent on separate lines
+> - This is NON-NEGOTIABLE. The user RELIES on this to verify correct agent routing.
+>
+> **Example — Single agent:**
+> ```
+> ⚡ **@backend-specialist activated!**
+>
+> Let me analyze your API endpoint...
+> ```
+>
+> **Example — Multiple agents:**
+> ```
+> ⚡ **@orchestrator activated!**
+> ⚡ **@frontend-specialist activated!**
+> ⚡ **@backend-specialist activated!**
+>
+> I'll coordinate the full-stack implementation...
+> ```
+
 ### Tier 1: Master Agents
 
 | Agent             | Use When                                       |
@@ -101,15 +123,16 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 
 ### Routing Checklist
 
-| Step | Check                           | If Unchecked                      |
-| ---- | ------------------------------- | --------------------------------- |
-| 1    | Correct agent identified?       | → Analyze domain                  |
-| 2    | Read agent's .md file?          | → Open `.agent/agents/{agent}.md` |
-| 3    | Announced @agent?               | → Add announcement                |
-| 4    | Loaded skills from frontmatter? | → Check `skills:` field           |
+| Step | Check                             | If Unchecked                                |
+| ---- | --------------------------------- | ------------------------------------------- |
+| 1    | Correct agent identified?         | → Analyze domain                            |
+| 2    | Read agent's .md file?            | → Open `.agent/agents/{agent}.md`           |
+| 3    | Announced @agent as FIRST LINE?   | → 🔴 Add announcement IMMEDIATELY            |
+| 4    | Loaded skills from ARCHITECTURE?  | → Check ARCHITECTURE.md agent-skills table  |
 
 ❌ Code without agent = PROTOCOL VIOLATION
 ❌ Skip announcement = USER CANNOT VERIFY
+❌ Announcement NOT as first line = PROTOCOL VIOLATION
 
 ---
 
@@ -129,7 +152,24 @@ Agent activated → Check frontmatter `skills:` → Read SKILL.md → Apply.
 ## 🛠️ SKILL LOADING PROTOCOL
 
 ```
-User Request → Check Profile → Skill Description Match → Load SKILL.md → Apply
+Agent activated → Check ARCHITECTURE.md for assigned skills → Use `skill` tool → Apply SKILL.md content
+```
+
+### How Skills Work in OpenCode
+
+OpenCode has a **native `skill` tool** that automatically discovers all available skills.
+
+1. **Discovery**: OpenCode scans `.agent/skills/*/SKILL.md` and lists them in the `skill` tool description
+2. **Selection**: Check ARCHITECTURE.md → find your agent → note the "Skills Used" column
+3. **Loading**: Call `skill({ name: "skill-name" })` to load each assigned skill
+4. **Apply**: Follow the loaded SKILL.md instructions
+
+```
+# Example: frontend-specialist activated
+1. Read ARCHITECTURE.md → Skills: clean-code, react-patterns, typescript-patterns, ...
+2. Call skill({ name: "clean-code" }) → Read content
+3. Call skill({ name: "react-patterns" }) → Read content
+4. Apply all loaded skill rules to the task
 ```
 
 ### Profile-Aware Loading
@@ -148,6 +188,31 @@ User Request → Check Profile → Skill Description Match → Load SKILL.md →
    - Behave as if no filtering is applied
 ```
 
+### Skill Permissions (OpenCode)
+
+Control skill access in `opencode.json`:
+
+```json
+{
+  "permission": {
+    "skill": {
+      "*": "allow",
+      "internal-*": "deny"
+    }
+  }
+}
+```
+
+Or per-agent in agent frontmatter:
+
+```yaml
+---
+permission:
+  skill:
+    "documents-*": "allow"
+---
+```
+
 ### Core Skills (Always Available)
 
 These skills are NEVER disabled regardless of profile:

package/kits/coder/scripts/kit_status.py

@@ -90,6 +90,46 @@ def parse_frontmatter(file_path: Path) -> Dict[str, Any]:
     return {}
 
 
+def parse_skills_from_architecture(agent_dir: Path) -> Dict[str, List[str]]:
+    """Parse agent-skill mappings from ARCHITECTURE.md tables.
+    
+    Falls back to this when agents don't have skills in frontmatter
+    (e.g., OpenCode strips the skills field during transformation).
+    """
+    arch_file = agent_dir / "ARCHITECTURE.md"
+    if not arch_file.exists():
+        return {}
+    
+    mapping = {}
+    try:
+        content = arch_file.read_text()
+        # Match table rows: | `agent-name` | description | skills-list |
+        # Pattern: | `name` | ... | skill1, skill2, ... |
+        for line in content.split('\n'):
+            if not line.startswith('|'):
+                continue
+            cells = [c.strip() for c in line.split('|')]
+            # Filter out empty cells from leading/trailing pipes
+            cells = [c for c in cells if c]
+            if len(cells) >= 3:
+                # First cell might be agent name wrapped in backticks
+                agent_name = cells[0].strip('`').strip()
+                # Last cell contains skills
+                skills_str = cells[-1]
+                # Skip header/separator rows
+                if agent_name.startswith('-') or agent_name == 'Agent':
+                    continue
+                # Parse comma-separated skills
+                skills = [s.strip().strip('`') for s in skills_str.split(',') if s.strip()]
+                # Only add if skills look valid (not header text)
+                if skills and not any(s in ['Skills Used', '---', 'Skills'] for s in skills):
+                    mapping[agent_name] = skills
+    except Exception:
+        pass
+    
+    return mapping
+
+
 def get_agents(agent_dir: Path) -> List[Dict[str, Any]]:
     """Get list of agents with their metadata."""
     agents = []
@@ -98,10 +138,12 @@ def get_agents(agent_dir: Path) -> List[Dict[str, Any]]:
     if not agents_path.exists():
         return agents
     
+    # Parse agent-skill mappings from ARCHITECTURE.md (single source of truth)
+    arch_skills = parse_skills_from_architecture(agent_dir)
+    
     for agent_file in agents_path.glob("*.md"):
         frontmatter = parse_frontmatter(agent_file)
-        skills = frontmatter.get("skills", "").replace("[", "").replace("]", "").split(",")
-        skills = [s.strip() for s in skills if s.strip()]
+        skills = arch_skills.get(agent_file.stem, [])
         
         agents.append({
             "name": agent_file.stem,
@@ -146,22 +188,33 @@ def get_skills(agent_dir: Path) -> List[Dict[str, Any]]:
 
 
 def get_workflows(agent_dir: Path) -> List[Dict[str, Any]]:
-    """Get list of workflows."""
-    workflows = []
-    workflows_path = agent_dir / "workflows"
+    """Get list of workflows/commands.
     
-    if not workflows_path.exists():
-        return workflows
+    Checks both 'workflows/' (Cursor, Gemini) and 'commands/' (OpenCode)
+    directories since different AI tools use different naming.
+    """
+    workflows = []
     
-    for workflow_file in workflows_path.glob("*.md"):
-        frontmatter = parse_frontmatter(workflow_file)
+    # Check both possible directories
+    for dir_name in ["workflows", "commands"]:
+        workflows_path = agent_dir / dir_name
         
-        workflows.append({
-            "name": workflow_file.stem,
-            "command": f"/{workflow_file.stem}",
-            "file": str(workflow_file.relative_to(agent_dir)),
-            "description": frontmatter.get("description", ""),
-        })
+        if not workflows_path.exists():
+            continue
+        
+        for workflow_file in workflows_path.glob("*.md"):
+            frontmatter = parse_frontmatter(workflow_file)
+            
+            workflows.append({
+                "name": workflow_file.stem,
+                "command": f"/{workflow_file.stem}",
+                "file": str(workflow_file.relative_to(agent_dir)),
+                "description": frontmatter.get("description", ""),
+            })
+        
+        # Only use the first directory found (don't double-count)
+        if workflows:
+            break
     
     return sorted(workflows, key=lambda x: x["name"])
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@neyugn/agent-kits",
-  "version": "0.3.5",
+  "version": "0.3.7",
   "description": "Universal AI Agent Toolkit - Skills, Agents, and Workflows for any AI coding assistant",
   "type": "module",
   "bin": {