Skill Search/Retrieval mechanism — re-raising #20143 (auto-closed by inactivity bot, never reviewed)

Re-raising #20143, which was closed by the stale-issue bot after 7 days of inactivity — no maintainer ever responded to it, positively or negatively.

The original request is still accurate: skill metadata (name+description) is eagerly injected into context for every installed skill at startup (confirmed in "Equipping agents for the real world with Agent Skills": "the agent pre-loads the name and description of every installed skill"), with no ceiling and no filtering mechanism — unlike MCP tools, which already have a shipped deferred/search pattern (ENABLE_TOOL_SEARCH, ToolSearch).

This is not hypothetical: #20143 itself cites an org going from 67 to 183 skills in under a month with "no way to know which are used." The same context-budget and skill-selection-precision problem that justified building Tool Search for MCP applies identically to skills — nothing in the architecture explains why one got the deferred-loading treatment and the other didn't.

Reopening because the underlying problem was never actually evaluated, just auto-archived.