[Feature]: Add schema_deferred routing mode — keep tool names & descriptions, strip inputSchema #971

Problem

retrieve_tools mode has a blind-search problem: the agent sees only meta-tools and has no idea what upstream tools exist until it guesses a query. direct mode dumps all schemas upfront, eating 30K+ tokens for ~100 tools.

There's no middle ground: let the agent see all tool names + descriptions, but defer inputSchema until needed.

Proposal

Add routing_mode: "schema_deferred":

• tools/list returns all upstream tools with name + description, but inputSchema set to {}
• Agent calls the existing describe_tool(name) to get the full schema when it's ready to use a specific tool
• No new execution primitives needed — describe_tool and call_tool_* already exist

Config:

{
  "routing_mode": "schema_deferred"
}

That's it. One mode, one behavior change.

Why

• Agent gets full visibility into available tools (no blind search)
• Context cost drops from ~30K to ~3.5K tokens for 100 tools (~88% reduction)
• One extra describe_tool round-trip per tool call — acceptable trade-off

Proposed solution

Industry precedent

• Atlassian mcp-compressor "Low Compression" level: exactly this — strips schema, keeps name + description. 94 tools: 17,600 → 3,900 tokens. ([blog](https://www.atlassian.com/blog/developer/mcp-compression-preventing-tool-bloat-in-ai-agents))
• Alibaba Qoder: two-phase discovery — name + description first, LoadMcpTool(name) for schema on demand. ([blog](https://www.alibabacloud.com/blog/stop-feeding-your-model-every-tool-definition_603401))
• MCP official Best Practices: recommends a Catalog → Inspect → Execute progressive discovery pattern. ([docs](https://modelcontextprotocol.io/docs/develop/clients/client-best-practices))

Alternatives considered

No response

Which edition?

Personal

Pre-submission checklist

• I searched existing issues and discussions for a similar request