[BUG] MCPSearch not activating despite meeting 10% threshold (40k/200k tokens, 20%)

Preflight Checklist

• [x] I have searched existing issues and this hasn't been reported yet
• [x] This is a single bug report (please file separate reports for different bugs)
• [x] I am using the latest version of Claude Code

What's Wrong?

MCPSearch not activating despite meeting 10% threshold (40k/200k tokens, 20%)

Description

Expected Behavior:
According to the v2.1.7 changelog, when MCP tool descriptions exceed 10% of the context window, they should be automatically deferred and discovered via the MCPSearch tool instead of being loaded upfront.

Actual Behavior:
MCPSearch is not activating despite MCP tools consuming 40.1k tokens (20.0%) of a 200k context window—double the documented 10% threshold.

Environment

Version:
Claude Code 2.1.7 (native installation)

Model:
claude-sonnet-4-5-20250929

Platform:
macOS (Darwin 25.2.0)

Evidence

1. Context Usage (from /context command):

Context window: 200k tokens
MCP tools: 40.1k tokens (20.0%)
10% threshold: 20k tokens
Actual percentage: 40.1k / 200k = 20.0%

1. Feature Flag Status (from ~/.claude.json):

"cachedGrowthBookFeatures": {
"tengu_mcp_tool_search": true,
"tengu_tool_search_unsupported_models": null
}

1. MCPSearch Tool Absence:

When running /context, no MCPSearch tool appears in the MCP tools list (169 tools total, but no MCPSearch).

1. Settings Verification:

• MCPSearch is NOT in disallowedTools
• No MCP-related restrictions in settings
• Feature flag shows as true

MCP Servers Configured

Active servers:

• MCP_DOCKER (Docker gateway)
• kapture (browser automation)
• ai-pm (project management)
• plugin:context7
• plugin:episodic-memory
• graphiti
• n8n
• MCP_Docs
• claude-historian-mcp
• sequential-thinking
• claude-in-chrome

Disabled servers:

• home-assistant
• meraki
• unraid-mcp
• XcodeBuildMCP
• kali-tools
• paloalto
• obsidian

Changelog Reference

From ~/.claude/cache/changelog.md:
"Enabled MCP tool search auto mode by default for all users. When MCP tool descriptions exceed 10% of the context window, they are automatically deferred and discovered via the MCPSearch tool instead of being loaded upfront."

Configuration File Locations

Feature flags:
~/.claude.json (line 94: "tengu_mcp_tool_search": true)

Settings:
~/.claude/settings.json (no MCPSearch in disallowedTools)

MCP config:
~/.claude.json (lines 1766-1789: mcpServers configuration)

Steps I've Taken

1. ✅ Verified version 2.1.7 is installed
2. ✅ Confirmed feature flag is true
3. ✅ Checked MCP tools consume 40.1k tokens (20%)
4. ✅ Verified tengu_tool_search_unsupported_models: null
5. ✅ Confirmed MCPSearch not in disallowedTools
6. ✅ Tested with both Opus 4.5 and Sonnet 4.5 (same result)
7. ✅ Started fresh sessions to trigger MCP reload

Additional Context

• Total MCP tools: 169 tools across 10+ servers
• Context breakdown shows clear 20% MCP tool usage
• All documented conditions appear to be met
• Feature appears to be client-side flagged but not server-side activated

What Should Happen?

According to the v2.1.7 changelog, when MCP tool descriptions exceed 10% of the context window, they should be automatically deferred and discovered via the MCPSearch tool instead of being loaded upfront.

Error Messages/Logs

Steps to Reproduce

1. Install Claude Code version 2.1.7 (native installation)
2. Configure multiple MCP servers in ~/.claude.json to create >40k tokens of tool definitions. Example minimal configuration:

"mcpServers": {
"MCP_DOCKER": {
"command": "docker",
"args": ["mcp", "gateway", "run"]
},
"kapture": {
"command": "npx",
"args": ["-y", "kapture-mcp@latest", "bridge"]
},
"sequential-thinking": {
"command": "npx",
"args": ["-y", "@modelcontextprotocol/server-sequential-thinking"]
},
"ai-pm": {
"command": "node",
"args": ["/path/to/ai-pm/server.js"]
}
}

1. Verify feature flag is enabled (should be default in 2.1.7):

grep "tengu_mcp_tool_search" ~/.claude.json

1. Expected: "tengu_mcp_tool_search": true
2. Start Claude Code and create a new conversation
3. Run the /context command to view context usage
4. Observe the "MCP tools" section in the output

Expected Result:

• MCP tools should show significantly reduced token usage (~5-10k)
• MCPSearch tool should appear in the MCP tools list
• Most individual MCP tool definitions should be deferred

Actual Result:

• MCP tools show 40.1k tokens (20.0% of context)
• No MCPSearch tool in the tools list
• All 169 MCP tools loaded upfront

Note: The issue occurs regardless of which model is selected (Opus 4.5 or Sonnet 4.5) and persists across fresh sessions.

Claude Model

Opus

Is this a regression?

Yes, this worked in a previous version

Last Working Version

_No response_

Claude Code Version

Claude Code 2.1.7 (native installation)

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

_No response_