[BUG] ENABLE_TOOL_SEARCH auto mode not triggering despite MCP tools exceeding 10% threshold

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?

What's Wrong?

MCP Tool Search auto mode is not triggering despite MCP tools consuming 24.8% of the context window — well over the documented 10% threshold.

Environment

• Claude Code Version: 2.1.14
• Model: Opus 4.5 (Claude Max)
• OS: macOS
• MCP Servers: Sentry, Atlassian, IDE (multiple servers with 40+ tools)

Workaround

Explicitly setting ENABLE_TOOL_SEARCH=true works correctly:

ENABLE_TOOL_SEARCH=true claude

Result with explicit true:

• Total context: 21k/200k tokens (11%)
• MCP tools: shows "loaded on-demand" with no token counts
• Free space: 134k (66.8%)

Additional Context

The old experimental flag ENABLE_EXPERIMENTAL_MCP_CLI=true (which I had been using since December 2025) produced the same on-demand behavior. After removing it and expecting auto mode to take over per the 2.1.7 release notes, the tools reverted to loading upfront.

Setting ENABLE_TOOL_SEARCH=true explicitly restores the expected behavior, but auto mode (the documented default) does not trigger despite being well over the 10% threshold.

What Should Happen?

What should happen?

Per the docs and v2.1.7 release notes: "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."

MCP tools should show as "loaded on-demand" with minimal token usage.

Actual Behavior

All MCP tools load upfront, consuming 49.5k tokens (24.8% of context):

• Total context: 71k/200k tokens (36%)
• MCP tools: 49.5k tokens (24.8%)
• Free space: 84k (42.0%)

Each tool shows individual token counts (e.g., mcp__sentry__get_issue_details: 923 tokens).

Error Messages/Logs

Steps to Reproduce

Steps to Reproduce

1. Configure multiple MCP servers (in my case: Sentry, Atlassian, IDE)
2. Start a fresh Claude Code session with no ENABLE_TOOL_SEARCH env var set
3. Run /context

Claude Model

Opus

Is this a regression?

I don't know

Last Working Version

_No response_

Claude Code Version

2.1.14

Platform

Anthropic API

Operating System

macOS

Terminal/Shell

iTerm2

Additional Information

<img width="678" height="381" alt="Image" src="https://github.com/user-attachments/assets/5a9ea68f-dcdf-49f2-b653-f5cf707fac54" />

<img width="662" height="746" alt="Image" src="https://github.com/user-attachments/assets/65b7635f-cdd4-4b44-9fa4-d81ce58ba601" />