fix(install_opencode): emit all variants, spec-compliant frontmatter

Agent emission was only writing general.json for each persona (29 files),
skipping 82 variants. Non-general variants now install as:
  <codename>-<variant>.md  with  mode: subagent + hidden: true

Key changes to install_opencode agent-emission half:
- Loop every *.json in each persona dir, not just general.json
- Enforce opencode identifier regex ^[a-z0-9]+(-[a-z0-9]+)*$ with sanitizer
- Non-general variants are always subagent (hidden:true is undefined on
  mode:primary per opencode docs); permission tier still follows domain
- Add permission.task block gating subagent dispatch
- Wrap description in double quotes with backslash/quote escaping so any
  special chars (&, :, quotes) can't corrupt the YAML frontmatter
- Variant identity surfaced in both description ("Variant: x") and body
  header "[x]" so prompts self-identify

Stale-agent cleanup via sidecar manifest (.personas-manifest.json):
prior emission list is tracked so re-runs prune removed variants without
touching any hand-authored agents in agents/.

Result: 29 -> 111 agents installed, 9 primary + 102 subagent, all
picker-clean (Tab cycles 9 canonical offensive personas; variants reach
via @codename-variant or task dispatch).

Skills-install half intentionally untouched — direct-to-active remains
the correct default for users who don't layer opc-skills on top.

CLAUDE.md

@@ -4,7 +4,7 @@ This file provides guidance to Claude Code (claude.ai/code) when working with co
 
 ## What This Is
 
-A platform-agnostic system prompt library for LLM agents. 29 personas across 10 domains, 111 variants, 59,712 words. Includes 795 shared skills, 58 brand design systems, 23 company agents, and auto-install to 7 platforms (Claude, Antigravity, Gemini, OpenClaw, OpenCode, Paperclip, raw).
+A platform-agnostic system prompt library for LLM agents. 29 personas across 10 domains, 111 variants, 59,712 words. Includes 796 shared skills, 58 brand design systems, 23 company agents, 168 AD/red team attack docs (InternalAllTheThings), and auto-install to 7 platforms (Claude, Antigravity, Gemini, OpenClaw, OpenCode, Paperclip, raw).
 
 ## Build
 
@@ -36,7 +36,7 @@ Optional: `cp config.example.yaml config.yaml` for dynamic variable injection. B
 **Shared library** (`personas/_shared/`): Reusable knowledge bases (skipped during persona build, indexed into outputs):
 - `skills/` — 42 shared skills from OpenClaw/kali-claw (SKILL.md + references per skill)
 - `paperclip-skills/` — 52 skills from paperclip-docs (ceo-advisor, coding-agent, security-review, etc.)
-- `community-skills/` — 701 skills from skills.sh marketplace (shadcn, vercel, marketing, expo, obsidian, impeccable, browser-use, stitch, firecrawl, github, neon, azure, etc.)
+- `community-skills/` — 703 skills from skills.sh marketplace (shadcn, vercel, olla, expo, etc.) (shadcn, vercel, marketing, expo, obsidian, impeccable, browser-use, stitch, firecrawl, github, neon, azure, etc.)
 - `design-md/` — 58 brand DESIGN.md files (Stripe, Claude, Linear, Apple, Vercel, etc.)
 - `ui-ux-pro-max/` — BM25 search engine + 14 CSV data files (67 styles, 161 products, 57 fonts)
 - `paperclip-agents/` — 23 company agents (Odin/CEO, Thor/CTO, Freya/CMO, Frigg/COO + 19 team members)
@@ -44,6 +44,7 @@ Optional: `cp config.example.yaml config.yaml` for dynamic variable injection. B
 - `openclaw-personas/` — Original 6 OpenClaw persona definitions + SOUL.md + IDENTITY.md + TOOLS.md
 - `osint-sources/` — OSINT master reference and investigation templates
 - `ad-attack-tools/` — Active Directory attack chain references
+- `internal-allthethings/` — 168 InternalAllTheThings docs (swisskyrepo): AD attacks (ADCS ESC1-15, Kerberos, NTLM relay, coerce, delegation), red team (access, escalation, evasion, persistence, pivoting), cloud (AWS/Azure/IBM), C2 (Cobalt Strike, Metasploit, Mythic), containers, databases, DevOps CI/CD
 
 **Build outputs** (`generated/_index/`):
 - `escalation_graph.json` — cross-persona handoff map extracted from Boundaries sections

README.md

@@ -2,7 +2,7 @@
 
 > Platform-agnostic system prompt library for LLM agents.
 > 29 personas. 10 domains. 111 variants. 60,000+ words of production-grade prompts.
-> 795 skills. 58 brand design systems. 16 Kali tool references. Multi-platform auto-install.
+> 796 skills. 58 brand design systems. 16 Kali tool references. 168 AD/red team attack docs. Multi-platform auto-install.
 
 ```
                 ┌─ Neo ─── Phantom ─── Cipher ─── Specter ─── Bastion ─── Vortex ─── Sentinel
@@ -47,8 +47,8 @@
                 │         └─ Heimd (Product Manager)
                 │
                 └──── SHARED LIBRARY ────────────────────────────────────
-                      795 skills │ 58 brand designs │ 16 kali-tools
-                      Auto-install: Claude · Antigravity · Gemini · OpenClaw · Paperclip
+                      796 skills │ 58 brand designs │ 16 kali-tools │ 168 AD/redteam docs
+                      Auto-install: Claude · Antigravity · Gemini · OpenClaw · OpenCode · Paperclip
 ```
 
 ## Quick Start
@@ -190,7 +190,8 @@ personas/
 ├── _shared/                  # Shared knowledge bases (skipped in persona build)
 │   ├── skills/          (42) # OpenClaw + kali-claw shared skills
 │   ├── paperclip-skills/(52) # Paperclip company skills (CEO, coding, devops...)
-│   ├── community-skills/(701)# skills.sh marketplace (shadcn, vercel, marketing...)
+│   ├── community-skills/(703)# skills.sh marketplace (shadcn, vercel, olla, marketing...)
+│   ├── internal-allthethings/ # InternalAllTheThings — 168 AD/red team attack docs (swisskyrepo)
 │   ├── design-md/       (58) # Brand DESIGN.md files (Stripe, Claude, Linear...)
 │   ├── ui-ux-pro-max/        # BM25 search engine + 14 CSV data files
 │   ├── paperclip-agents/(23) # Company agents (Odin/CEO, Thor/CTO, Freya/CMO...)
@@ -276,10 +277,12 @@ Escalation paths to other personas
 
 ```bash
 python3 build.py                          # build all → generated/
-python3 build.py --install claude         # deploy as Claude Code slash commands
+python3 build.py --install claude         # deploy as Claude Code slash commands + agents
+python3 build.py --install claude-skills  # deploy shared skills to ~/.claude/skills/
 python3 build.py --install antigravity    # deploy to Antigravity IDE
 python3 build.py --install gemini         # deploy as Gemini Gems
 python3 build.py --install openclaw       # deploy to OpenClaw format
+python3 build.py --install opencode       # deploy 29 agents + 1011 skills to OpenCode
 python3 build.py --install paperclip      # deploy to Paperclip (52 agents + 73 skills)
 python3 build.py --install all            # deploy to all platforms
 ```
@@ -303,6 +306,8 @@ Reads `config.yaml` (if present) and generates three formats per variant:
 | Gemini Gems | `generated/_gems/` | Google AI Studio format |
 | OpenClaw | `generated/_openclaw/` | IDENTITY.md + individual personas |
 | Paperclip | `generated/_paperclip/` | 52 agents + 73 skills (Hermes format) |
+| OpenCode agents | `~/.config/opencode/agents/` | 29 native subagents (md + YAML frontmatter) |
+| OpenCode skills | `~/.config/opencode/skills/` | 1011 skills (topic-filtered, no marketing/biz) |
 
 ### Config-Driven Customization
 
@@ -329,6 +334,20 @@ Persona files can reference config values:
 
 Without `config.yaml`, personas build with their static content — no config required.
 
+### OpenCode Integration
+
+OpenCode (`opencode-ai`) gets 29 agents and 1011 topic-filtered skills:
+
+```bash
+python3 build.py --install opencode                   # default: security + dev + AI + ops
+python3 build.py --install opencode --opencode-topics all  # everything (1530)
+python3 build.py --install opencode --opencode-topics security-offensive,security-defensive,ai-llm-dev
+```
+
+**Topics**: security-offensive, security-defensive, security-cloud, security-specialized, security-iam, security-network, security-general, ai-llm-dev, coding-backend, coding-frontend, coding-tools, cloud-infra, database, browser-scrape, ops-sysadmin, osint-intel, marketing-content, business-pm, uncategorized.
+
+Default drops marketing/business-pm/uncategorized (~514 skills). Agents use OpenCode's markdown+YAML format with `mode: primary|subagent` and `permission:` blocks. OpenCode also reads `~/.claude/skills/` natively.
+
 ## Integration
 
 ### Raw System Prompt
@@ -435,7 +454,8 @@ The `_shared/` directory contains reusable knowledge bases from multiple sources
 |--------|---------|-------|
 | **OpenClaw + kali-claw** | Security/intelligence skills (pentest, OSINT, CTI) | 42 skills |
 | **Paperclip (Born2beRoot)** | Company management skills (CEO, coding, devops) | 52 skills |
-| **skills.sh marketplace** | Community skills (shadcn, vercel, marketing, expo) | 701 skills |
+| **skills.sh marketplace** | Community skills (shadcn, vercel, olla, expo) | 703 skills |
+| **InternalAllTheThings** | AD attacks, red team, ADCS, Kerberos, NTLM relay (swisskyrepo) | 168 docs |
 | **awesome-design-md** | Brand design systems (Stripe, Claude, Linear, Apple) | 58 brands |
 | **ui-ux-pro-max** | BM25 search engine for UI/UX decisions | 14 data files |
 | **Kali Linux** | Tool reference docs (nmap, hashcat, AD, wireless) | 16 docs |
@@ -449,11 +469,12 @@ Skills are auto-mapped to personas during build. Each persona's JSON/YAML output
 | Personas | 29 |
 | Total variants | 111 |
 | Prompt content | 59,712 words |
-| Shared skills | 795 |
+| Shared skills | 796 |
 | Design brands | 58 |
 | Kali tool docs | 16 |
+| AD/Red team docs | 168 (InternalAllTheThings) |
 | Paperclip agents | 23 |
-| Target platforms | 6 (Claude, Antigravity, Gemini, OpenClaw, Paperclip, raw) |
+| Target platforms | 7 (Claude, Antigravity, Gemini, OpenClaw, OpenCode, Paperclip, raw) |
 | Output formats | 3 (.prompt.md, .yaml, .json) + platform-specific |
 
 ## License

build.py

@@ -1394,16 +1394,26 @@ def install_opencode(
     }
 
     agent_count = 0
+    emitted_agents: set[str] = set()
+    ident_re = re.compile(r"^[a-z0-9]+(-[a-z0-9]+)*$")
 
+    # Emit one agent file per variant. General → <codename>.md (picker-visible).
+    # Non-general → <codename>-<variant>.md with hidden:true so it's
+    # task-dispatchable by name without cluttering the picker.
     for persona_dir in sorted(output_dir.iterdir()):
         if not persona_dir.is_dir() or persona_dir.name.startswith("_"):
             continue
-        general_json = persona_dir / "general.json"
-        if not general_json.exists():
+
+        for variant_json in sorted(persona_dir.glob("*.json")):
+            try:
+                data = json.loads(variant_json.read_text(encoding="utf-8"))
+            except json.JSONDecodeError:
+                continue
+            if not data.get("codename"):
                 continue
 
-        data = json.loads(general_json.read_text(encoding="utf-8"))
-        codename = data.get("codename", persona_dir.name)
+            codename = data["codename"]
+            variant = data.get("variant") or "general"
             name = data.get("name", codename.title())
             role = data.get("role", "Specialist")
             domain = data.get("domain", "")
@@ -1412,11 +1422,24 @@ def install_opencode(
             quote = data.get("quote", "")
             skills = data.get("skills", [])
 
+            # opencode agent identifier: ^[a-z0-9]+(-[a-z0-9]+)*$
+            agent_ident = codename if variant == "general" else f"{codename}-{variant}"
+            agent_ident = agent_ident.lower()
+            if not ident_re.match(agent_ident):
+                sanitized = re.sub(r"[^a-z0-9]+", "-", agent_ident).strip("-")
+                if not ident_re.match(sanitized):
+                    print(f"  WARN skipping {codename}/{variant}: cannot sanitize identifier")
+                    continue
+                agent_ident = sanitized
+
             soul = data.get("sections", {}).get("soul", "")
             methodology = data.get("sections", {}).get("methodology", "")
             behavior = data.get("sections", {}).get("behavior_rules", "")
 
-        body = f"You are **{name}** ({address_to}) — {role}.\n\n"
+            header = f"You are **{name}** ({address_to}) — {role}"
+            if variant != "general":
+                header += f" [{variant}]"
+            body = header + ".\n\n"
             body += f"Domain: {domain} | Tone: {tone}\n\n"
             if quote:
                 body += f'> "{quote}"\n\n'
@@ -1430,10 +1453,18 @@ def install_opencode(
                 body += "## Mapped Skills\n" + ", ".join(skills) + "\n"
 
             is_offensive = domain in OFFENSIVE_DOMAINS
-        mode = "primary" if is_offensive else "subagent"
+            # Only general variants can be primary (Tab-cycled, top-level).
+            # All non-general variants are subagents so hidden:true is defined
+            # per opencode spec ("only applies to mode: subagent") and the Tab
+            # cycle stays restricted to canonical personas.
+            if variant == "general" and is_offensive:
+                mode = "primary"
+            else:
+                mode = "subagent"
             color = DOMAIN_COLOR.get(domain, "primary")
 
-        # Permission block differs for offensive vs analytical personas.
+            # Tier-based permissions. task:"*" gates subagent dispatch so a
+            # compromised subagent can't silently escalate into another.
             if is_offensive:
                 permission_block = (
                     "permission:\n"
@@ -1441,6 +1472,8 @@ def install_opencode(
                     "  bash:\n"
                     '    "*": allow\n'
                     "  webfetch: allow\n"
+                    "  task:\n"
+                    '    "*": allow\n'
                 )
             else:
                 permission_block = (
@@ -1449,25 +1482,65 @@ def install_opencode(
                     "  bash:\n"
                     '    "*": ask\n'
                     "  webfetch: allow\n"
+                    "  task:\n"
+                    '    "*": ask\n'
                 )
 
-        desc = f"{name} ({address_to}) — {role}. Domain: {domain}.".replace(
-            "\n", " "
+            if variant == "general":
+                desc_raw = f"{name} ({address_to}) — {role}. Domain: {domain}."
+            else:
+                desc_raw = (
+                    f"{name} ({address_to}) — {role}. "
+                    f"Variant: {variant}. Domain: {domain}."
                 )
+            desc_safe = desc_raw.replace("\n", " ").replace("\r", " ")
+            desc_escaped = desc_safe.replace("\\", "\\\\").replace('"', '\\"')
+
+            hidden_line = "hidden: true\n" if variant != "general" else ""
 
             frontmatter = (
                 "---\n"
-            f"description: {desc}\n"
+                f'description: "{desc_escaped}"\n'
                 f"mode: {mode}\n"
+                f"{hidden_line}"
                 "temperature: 0.3\n"
                 f"color: {color}\n"
                 f"{permission_block}"
                 "---\n\n"
             )
-        agent_file = agents_dir / f"{codename}.md"
-        agent_file.write_text(frontmatter + body, encoding="utf-8")
+
+            agent_filename = f"{agent_ident}.md"
+            (agents_dir / agent_filename).write_text(
+                frontmatter + body, encoding="utf-8"
+            )
+            emitted_agents.add(agent_filename)
             agent_count += 1
 
+    # Remove stale agents we emitted on a previous run but not this one. Track
+    # via sidecar manifest so hand-authored files in agents/ are never touched.
+    manifest_path = agents_dir / ".personas-manifest.json"
+    previously_emitted: set[str] = set()
+    if manifest_path.exists():
+        try:
+            previously_emitted = set(
+                json.loads(manifest_path.read_text(encoding="utf-8")).get("agents", [])
+            )
+        except Exception:
+            previously_emitted = set()
+    stale = sorted(previously_emitted - emitted_agents)
+    for name_ in stale:
+        stale_path = agents_dir / name_
+        if stale_path.exists():
+            stale_path.unlink()
+    manifest_path.write_text(
+        json.dumps(
+            {"agents": sorted(emitted_agents)}, indent=2, ensure_ascii=False
+        ),
+        encoding="utf-8",
+    )
+    if stale:
+        print(f"  OpenCode: pruned {len(stale)} stale agent file(s)")
+
     # Install shared skills with topic filter. OpenCode reads SKILL.md with
     # name+description frontmatter (same as Claude).
     skill_count = 0