verify-fix forces JSON for interactive verification results even when JSON was not requested #1830

What happened

I asked Codex in a normal conversation to verify whether a reported security finding had been fixed. I did not request JSON, but the final answer repeatedly came back as a results object with id, status, and a long evidence string. Source links and explanations were embedded in that string rather than presented as readable Markdown.

Environment

• Codex desktop on macOS 26.7.1
• Codex Security plugin 0.1.32
• Skill: codex-security:verify-fix

Steps to reproduce

1. Enable the Codex Security plugin.
2. Supply a security finding and the relevant source checkout.
3. Explicitly ask Codex to verify whether the fix remediates the finding, without requesting a machine-readable response.
4. Inspect the final answer.

Expected behavior

In an interactive conversation, I expect a readable verdict, evidence, and any validation gaps. JSON should remain available when I request it or when an automated caller requires it. The verification method and evidence standards should stay unchanged.

Observed instruction

The installed skill's Result Contract says: "Return exactly one JSON object." The same requirement is present in current main, with no interactive-output alternative.

Related: #815 narrowed the trigger to explicit security-fix verification requests while retaining the JSON contract. This report concerns deliberate interactive verification, not accidental skill selection.

I worked around this by adding a global AGENTS.md preference for readable security-verification summaries, without editing the plugin cache. I'd prefer the skill to support this directly.