diff --git a/build.py b/build.py
index 87a1eee..1d987f0 100755
--- a/build.py
+++ b/build.py
@@ -928,22 +928,63 @@ def print_summary(
 
 
 def install_claude(output_dir: Path):
-    """Install personas to Claude Code as slash commands + agents."""
+    """Install personas to Claude Code as slash commands + native subagents.
+
+    Native Claude Code agents are .md files with YAML frontmatter under
+    ~/.claude/agents/. Tool permissions are scoped per domain:
+      - offensive/dev/engineering: full toolset incl. Bash/Write
+      - intel/analysis/history/law: read + search + web only
+    """
     commands_dir = Path.home() / ".claude" / "commands"
     agents_dir = Path.home() / ".claude" / "agents"
     commands_dir.mkdir(parents=True, exist_ok=True)
     agents_dir.mkdir(parents=True, exist_ok=True)
 
+    # Domain → tool scope. "full" gets Bash + Write; "readonly" does not.
+    OFFENSIVE_DOMAINS = {
+        "cybersecurity",
+        "engineering",
+        "devops",
+        "software-development",
+        "ai-ml",
+    }
+
+    FULL_TOOLS = (
+        "Read, Edit, Write, Bash, Glob, Grep, WebFetch, WebSearch"
+    )
+    READONLY_TOOLS = "Read, Glob, Grep, WebFetch, WebSearch"
+
+    # Color palette per domain family for visual grouping in /agents picker.
+    DOMAIN_COLOR = {
+        "cybersecurity": "red",
+        "intelligence": "cyan",
+        "military": "orange",
+        "law": "yellow",
+        "economics": "green",
+        "politics": "purple",
+        "history": "blue",
+        "linguistics": "pink",
+        "media": "pink",
+        "engineering": "green",
+        "academia": "blue",
+    }
+
     cmd_count = 0
     agent_count = 0
+    orphans_removed = 0
+
+    # Purge legacy .yml agents from previous build (wrong format, unused by CC).
+    for stale in agents_dir.glob("*.yml"):
+        stale.unlink()
+        orphans_removed += 1
 
     for persona_dir in sorted(output_dir.iterdir()):
         if not persona_dir.is_dir() or persona_dir.name.startswith("_"):
             continue
 
-        # Install slash commands for all variants
+        # Install slash commands for all variants (general + specializations).
         for prompt_file in persona_dir.glob("*.prompt.md"):
-            variant = prompt_file.stem
+            variant = prompt_file.stem.replace(".prompt", "")
             codename = persona_dir.name
             cmd_name = (
                 f"persona-{codename}"
@@ -956,7 +997,7 @@ def install_claude(output_dir: Path):
             dest.write_text(command_content, encoding="utf-8")
             cmd_count += 1
 
-        # Install agent .yml for general variant (appears in /agents menu)
+        # Install native subagent for the general variant only.
         general_json = persona_dir / "general.json"
         if not general_json.exists():
             continue
@@ -975,43 +1016,45 @@ def install_claude(output_dir: Path):
         methodology = data.get("sections", {}).get("methodology", "")
         behavior = data.get("sections", {}).get("behavior_rules", "")
 
-        instructions = f"You are **{name}** ({address_to}) — {role}.\n\n"
-        instructions += f"Domain: {domain} | Tone: {tone}\n\n"
+        body = f"You are **{name}** ({address_to}) — {role}.\n\n"
+        body += f"Domain: {domain} | Tone: {tone}\n\n"
         if quote:
-            instructions += f'> "{quote}"\n\n'
-        instructions += "## Soul\n" + soul[:1500] + "\n\n"
+            body += f'> "{quote}"\n\n'
+        if soul:
+            body += "## Soul\n" + soul.strip() + "\n\n"
         if methodology:
-            instructions += "## Methodology\n" + methodology[:1500] + "\n\n"
+            body += "## Methodology\n" + methodology.strip() + "\n\n"
         if behavior:
-            instructions += "## Behavior\n" + behavior[:800] + "\n"
+            body += "## Behavior\n" + behavior.strip() + "\n\n"
         if skills:
-            instructions += "\n## Mapped Skills\n" + ", ".join(skills) + "\n"
+            body += "## Mapped Skills\n" + ", ".join(skills) + "\n"
 
-        agent = {
-            "name": codename,
-            "description": f"{name} ({address_to}) — {role}. {domain}.",
-            "instructions": instructions,
-            "allowedTools": [
-                "Read(*)",
-                "Edit(*)",
-                "Write(*)",
-                "Bash(*)",
-                "Glob(*)",
-                "Grep(*)",
-                "WebFetch(*)",
-                "WebSearch(*)",
-            ],
-        }
-        agent_file = agents_dir / f"{codename}.yml"
-        agent_file.write_text(
-            yaml.dump(
-                agent, allow_unicode=True, default_flow_style=False, sort_keys=False
-            ),
-            encoding="utf-8",
+        tools = (
+            FULL_TOOLS if domain in OFFENSIVE_DOMAINS else READONLY_TOOLS
         )
+        color = DOMAIN_COLOR.get(domain, "gray")
+
+        # Single-line description keeps YAML frontmatter safe.
+        desc = f"{name} ({address_to}) — {role}. Domain: {domain}.".replace(
+            "\n", " "
+        )
+
+        frontmatter = (
+            "---\n"
+            f"name: {codename}\n"
+            f"description: {desc}\n"
+            f"tools: {tools}\n"
+            f"color: {color}\n"
+            "---\n\n"
+        )
+        agent_file = agents_dir / f"{codename}.md"
+        agent_file.write_text(frontmatter + body, encoding="utf-8")
         agent_count += 1
 
-    print(f"  Claude: {cmd_count} commands + {agent_count} agents installed")
+    print(
+        f"  Claude: {cmd_count} commands + {agent_count} agents installed "
+        f"({orphans_removed} legacy .yml purged)"
+    )
     return cmd_count