perf(mcp): schema caching for tools/list to reduce context bloat #29986

Problem

MCP injects full tool schemas (~37 tools from 5 MCPs) into every prompt. Context bloat up to 86% (Issue #17482).

From packages/opencode/src/mcp/index.ts:

• listTools() in defs() calls client.listTools() on every connection
• Cached only in-memory state (s.defs) — lost on restart
• tools() in the Service regenerates all tools from cached defs on each call
• No configurable TTL or persistent caching mechanism

Proposed solution

Cache tools/list response with configurable TTL:

interface MCPServerConfig {
  schemaCacheTTL?: number;  // seconds, default 3600
}

Impact

• 37 tools × ~486 tokens avg = ~18k tokens saved per prompt
• Affects: LEX_SUISSE (8 tools) + fedlex (7) + swiss-courts (7) + bag-epl (6) + mcp-biblioteca (9)

Workaround

Currently using subagent_type instead of raw MCP for INVESTIGADOR. But bridge skill (rag-mcp-bridge) still needs MCP tools.

References

• Issue [FEATURE]: Implement Dynamic/Lazy Loading for MCP Tool Schemas to Prevent Context Bloat #17482 (context bloat)
• _PIPELINE_RAG_MCP_BRIDGE.md §Fase 2 (tool retrieval)
• skill/rag-mcp-bridge/SKILL.md (routing table)