feat(plugin): add opencode permission block to all agents

Each agent under plugins/claude-code/agents/ now carries both:

- Claude Code frontmatter: `tools:` and `disallowedTools:` (unchanged)
- opencode frontmatter:    `permission: { edit, bash, webfetch }`

Each host reads only its own fields and ignores the other's. Because
plugins/opencode/agents is symlinked to plugins/claude-code/agents, the
opencode plugin now exposes proper opencode permission semantics from
the same source files.

Permission mapping mirrors the Claude Code tool list:
- code-reviewer / context-scout: edit=deny, bash=deny, webfetch=deny
- coder-agent / task-manager:    edit=allow, bash=deny, webfetch=deny
- context-manager:                edit=allow, bash=allow, webfetch=deny
- test-engineer:                  edit=allow, bash=allow, webfetch=deny
- external-scout:                 edit=allow, bash=allow, webfetch=allow

Updates plugins/opencode/AGENTS.md to reflect the dual-frontmatter
approach.

Author: MomePP

plugins/claude-code/agents/code-reviewer.md

@@ -17,6 +17,10 @@ description: |
   </example>
 tools: Read, Glob, Grep
 disallowedTools: Write, Edit, Bash, Task
+permission:
+  edit: deny
+  bash: deny
+  webfetch: deny
 model: opus
 ---
 

plugins/claude-code/agents/coder-agent.md

@@ -16,6 +16,10 @@ description: |
   <commentary>A concrete implementation task with a specific file — coder-agent executes it.</commentary>
   </example>
 tools: Read, Write, Edit, Glob, Grep
+permission:
+  edit: allow
+  bash: deny
+  webfetch: deny
 model: opus
 ---
 

plugins/claude-code/agents/context-manager.md

@@ -2,6 +2,10 @@
 name: context-manager
 description: Manages context files, discovers context roots, validates structure, and organizes project context
 tools: Read, Write, Glob, Grep, Bash
+permission:
+  edit: allow
+  bash: allow
+  webfetch: deny
 model: sonnet
 ---
 

plugins/claude-code/agents/context-scout.md

@@ -17,6 +17,10 @@ description: |
   </example>
 tools: Read, Glob, Grep
 disallowedTools: Write, Edit, Bash, Task
+permission:
+  edit: deny
+  bash: deny
+  webfetch: deny
 model: haiku
 ---
 

plugins/claude-code/agents/external-scout.md

@@ -2,6 +2,10 @@
 name: external-scout
 description: Fetches external library and framework documentation from Context7 API and other sources, caching results for offline use
 tools: Read, Write, Bash, WebFetch
+permission:
+  edit: allow
+  bash: allow
+  webfetch: allow
 model: haiku
 ---
 

plugins/claude-code/agents/task-manager.md

@@ -2,6 +2,10 @@
 name: task-manager
 description: Break down complex features into atomic, verifiable subtasks with dependency tracking and JSON-based progress management
 tools: Read, Write, Glob, Grep
+permission:
+  edit: allow
+  bash: deny
+  webfetch: deny
 model: sonnet
 ---
 

plugins/claude-code/agents/test-engineer.md

@@ -2,6 +2,10 @@
 name: test-engineer
 description: Test authoring and TDD specialist - writes comprehensive tests following project testing standards
 tools: Read, Write, Edit, Bash
+permission:
+  edit: allow
+  bash: allow
+  webfetch: deny
 model: sonnet
 ---
 

plugins/opencode/AGENTS.md

@@ -20,10 +20,10 @@ truth — no copy drift.
 
 ## Known conversion gaps from Claude Code → opencode
 
-- Agent frontmatter still uses Claude Code's `tools:` field. opencode prefers
-  a `permission:` object (`edit`, `bash`, `read`, `task` → `allow|ask|deny`).
-  Unknown fields are ignored, so this is non-fatal — but tighten permissions
-  manually if you care.
+- Agent frontmatter carries both Claude Code's `tools:` / `disallowedTools:`
+  fields and an opencode `permission:` block. Each host reads its own; the
+  other ignores the unknown fields. If you adjust permissions, update both
+  so the agents stay in sync across hosts.
 - Commands embed `${CLAUDE_PLUGIN_ROOT}` paths. opencode does not export that
   variable; the `oac-hooks.ts` plugin sets it for child processes when it
   spawns the session-start script, but inline command bodies that read the