Whole-file shell reads exceeded the output cap seven times; each persisted output was then re-read, so it was paid for twice

What happened
The model ran cat on entire 30 to 45 KB files in single Bash calls. Seven outputs exceeded the cap, were saved to disk with a 2 KB preview, and were then read back in full with Read. Each was paid for twice.

Expected
Targeted line ranges after an outline, or the MCP graph's context with include_content.

What the user said
"What went unused, against my task and exploting my budget"

Repro
cat a file over about 30 KB through Bash; the output is persisted; the model then Reads the persisted file.

Evidence
~/.claude/projects/<project>/<session>/tool-results/ held seven such files after the session.

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