Reduce the initial tool surface with reliable discovery

[ericleepi314]

Summary

• align clawcodex initial tools with Claude Code’s proven default set
• defer 21 clawcodex-specific tools while keeping them registered
• repair ToolSearch tool_reference emission and next-turn schema loading
• announce deferred tool names and send the required Anthropic beta header
• translate beta options through the Anthropic SDK and subscription mcp_ naming
• make headless init logs report the actual initial request tool set
• enable supported experimental betas in Harbor subscription runs only

Evidence

• enabled schema payload: 50,626 -> 30,290 chars (-40%)
• first-call prompt: 27,371 -> 19,816 tokens (-27.6%)
• pypi-server Harbor reward: 1.0, no exceptions
• 159 focused regression tests pass

Benchmark

eval/harbor/jobs/2026-07-23__09-06-29/pypi-server__Gq7Sezw