permission.skill / tools.skill in standalone agent .md frontmatter is parsed but silently ignored at config resolution

Title: permission.skill / tools.skill in standalone agent .md frontmatter is parsed but silently ignored at config resolution

Summary

Setting permission.skill or tools.skill in a standalone agent definition
file (e.g. ~/.config/opencode/agents/<name>.md) is accepted without error —
the value appears verbatim in the raw prompt field returned by
opencode debug agent <name> — but does NOT propagate to the resolved
tools/permission object used at request time. The only way to actually
restrict skill access for an agent is to define the override inside
opencode.json's top-level agent.<name>.permission.skill block instead.

This is confusing because both locations use identical YAML/JSON syntax and
neither documentation nor CLI output indicates that one is authoritative and
the other is not.

Steps to reproduce

1. Create ~/.config/opencode/agents/my-agent.md with frontmatter:

***
description: test agent
mode: primary
model: <any provider/model>
permission:
  edit: deny
  bash: ask
  skill:
    "*": deny
***

2. Run opencode debug agent my-agent
3. Observe: the prompt field shows the frontmatter exactly as written
   (including the skill: "*": deny block), but the resolved tools object
   at the bottom of the same output still shows "skill": true
4. Confirm the tool is still callable: run opencode run "<any prompt that might invoke a skill>" --agent my-agent — the model can still invoke the
   skill tool and load skill content

Expected behavior

Either:
(a) permission.skill / tools.skill in standalone agent .md frontmatter
should be honored at resolution time, matching the behavior documented
for opencode.json's agent.* block, OR
(b) if standalone agent files intentionally support a reduced frontmatter
schema that excludes tool/permission overrides, this should be stated
explicitly in the docs and ideally surfaced as a warning/error when an
unsupported key is present, rather than silently accepted and ignored

Actual behavior

Silent no-op. The only working fix is duplicating the agent's tool
restriction inside opencode.json:

{
  "agent": {
    "my-agent": {
      "permission": {
        "skill": { "*": "deny" }
      }
    }
  }
}

Impact

For agents with a large local + global skill catalog (105 skills in one tested
case, ~1.29MB of serialized skill metadata), this caused every request to
carry an unnecessarily large tool-calling prompt, costing significant prompt
evaluation time and obscuring the real working configuration path.

Environment

• OpenCode version: 1.18.15
• OS: macOS (Apple Silicon)
• Model provider: Ollama (local), qwen2.5-coder:14b-16k