Model ignored an explicit instruction to investigate through an MCP code graph and read source via shell instead

**What happened**
The user asked for an investigation of six indexed repositories through a GitNexus MCP server. The model made two graph queries per repository, then read the source with about forty-four Bash `cat`/`sed` calls across ten batches. The project's CLAUDE.md also said "Graph first; text search only for empty/UNKNOWN/literals" and "Never substitute grep for graph analysis".

**Expected**
The graph tools (query, context with include_content, cypher, trace, cluster resources) for the whole investigation.

**What the user said**
"Use gitnexus mcp and investigate how 6 harnesses I indexed are working with their agents." Mid-turn: "stop the shell reads, go back to the graph, and use clusters, tool_map, context with content, trace and cypher per harness before any file is opened by hand."

**Repro**
Configure an MCP code-intelligence server, instruct "use the MCP to investigate these repos", observe the model fall back to shell file reads after the first graph calls.

**Environment:** Claude Code CLI 2.1.270 (Claude Code), macOS, auto mode, model claude-fable-5-1, session of 2026-09-14.