Feature: Support MCP Prompt content injection as system message (hidden from user) #23480

Problem

When an MCP server delivers content via the Prompts primitive (e.g., user invokes /my-server:my-prompt), OpenCode displays the prompt content directly in the conversation UI as a visible user message. The LLM then echoes/repeats the full content before acting on it.

This breaks the use case of proprietary/confidential prompt delivery, where MCP servers need to send instructions to the LLM without exposing the raw content to the end user.

Expected behavior

MCP Prompt content should be injected as system-level context (invisible to the user in the TUI), similar to how system messages work. The user should only see the LLM's response/actions — not the raw prompt instructions.

Current behavior

1. User invokes /my-server:my-prompt
2. OpenCode fetches prompt content from MCP server
3. Content is displayed verbatim in the conversation UI as a user message
4. LLM echoes/repeats the instructions before executing them
5. Proprietary content is fully visible to the user

Why this matters

MCP servers that deliver proprietary skills/instructions (e.g., enterprise toolkits, paid skill libraries) rely on the content being consumed by the LLM without being displayed to the user. The MCP spec distinguishes Prompts from Tools precisely because Prompts are meant to be templates that guide LLM behavior, not user-visible content.

Proposed solution

One or more of the following:

1. Inject MCP Prompt content as a system message — content reaches the LLM but is not rendered in the TUI
2. Support the MCP audience annotation — the MCP spec (2025-06-18) defines audience: ["assistant"] on content items, which signals that content is intended for the LLM only, not the user
3. Add a configuration option — e.g., prompt_content_visibility: "hidden" | "collapsed" | "visible" to let users/servers control visibility

Additional context

• MCP spec reference: Content Annotations
• The audience field in annotations is specifically designed for this: "user" = show to user, "assistant" = show only to LLM
• Claude Code handles this correctly — MCP Prompt content is injected as system context and never displayed to the user
• This is different from Tool results (which are expected to be visible) — Prompts have a different semantic purpose

Environment

• OpenCode version: latest
• OS: macOS