[FEATURE]: Model-agnostic MCP tool search / deferred schema loading (works on OpenAI-style backends)

Enhancement

Add a config option (e.g. mcp.tool_search) that keeps MCP tool schemas out of the system-prompt prefix and exposes them through a single lightweight opencode-native search tool instead. When the model calls the search tool, opencode returns matching tool names/descriptions; on first real call to a discovered tool, its full schema is registered for the rest of the session.

Prior art / related issues

This has been requested before but is not implemented:

• #34873 proposed mcp.tool_search (closed automatically for a missing title prefix, not because it shipped)
• #23298 proposed Anthropic defer_loading passthrough (closed as completed; the feature is absent on dev)
• #45527 documents that absence and proposes wiring Anthropic's defer_loading + tool_search_tool_*
• #8625, #9461, #10090 (adjacent)

I am opening this separately because the previously proposed mechanism is Anthropic-specific and does not reach a large class of deployments (details below). This request is for a model-agnostic version.

Why the Anthropic defer_loading / tool_search_tool path is not sufficient

#45527 proposes surfacing Anthropic's defer_loading flag and the tool_search_tool_bm25_20251119 provider-defined tool. That works only when the backend speaks the Anthropic tool protocol natively. Many setups route Anthropic-format requests through a gateway that translates to an OpenAI-style responses API (e.g. a Fireworks-backed model behind a multi-protocol LLM gateway).

I verified this against a live gateway that accepts Anthropic requests on /anthropic/v1/messages but translates to an OpenAI responses call for a non-Anthropic model:

• Declaring {"type": "tool_search_tool_bm25_20251119", ...} → 400: tools[0].type must be 'function', 'web_search_preview' or 'code_interpreter'. The provider-defined search tool is rejected; the downstream API has no equivalent primitive.
• A plain tool with defer_loading: true alongside a non-deferred tool → 200, works. So defer_loading is accepted/passed through, but there is no search tool for the model to discover deferred tools with.

Net: for any model served via an OpenAI-style backend, defer_loading is accepted but useless without a search tool, and the Anthropic search tool cannot be declared. A mechanism that lives entirely inside opencode avoids this dependency.

Proposed design (model-agnostic)

• New config: mcp.tool_search: "auto" | "always" | "off" (default off for back-compat). "auto" could enable it when the configured MCP surface exceeds a size threshold.
• When enabled, MCP tools are partitioned out of the visible tool set in ToolRegistry.tools and replaced by one ordinary function-typed tool, e.g. search_tools({ query }). Because it is a normal function tool, every provider/backend accepts it, no provider-defined tool required.
• The search index is already available in-process (McpCatalog, Permission.visibleTools), so lookup is cheap and needs no new infrastructure.
• On the first real call to a discovered MCP tool, opencode registers that tool's full schema into the active request path for the remainder of the session (and emits a plugin event like mcp.tool.resolved for observability, as #34873 suggested).
• The model is told the tool exists via the search tool's own description; newly-resolved tools can be surfaced back to the model through a tool result or a short system note.

This keeps the per-turn prefix small for sessions that touch only a handful of MCP tools, regardless of model or gateway, and preserves prompt-cache reuse.

Benefits

• Users with many/rarely-used MCP servers pay a large token tax on every request today; this removes it for the common case.
• Works uniformly across Anthropic, OpenAI-style, and gateway-translated backends, unlike the Anthropic-only defer_loading path.
• Backward compatible (opt-in, default off).