Large Bash outputs are persisted to disk under ~/.claude and must be re-read, doubling cost and writing files the user did not want

**What happened**
Seven outputs of 29 to 45 KB were written to `~/.claude/projects///tool-results/*.txt` with a 2 KB preview in context. Consuming them required a full Read of each file, so the content was paid for twice. The user later instructed that nothing be saved to files. A later attempt to delete those files was refused by the auto-mode classifier as "Session Transcript Tampering", so the user has to remove them by hand.

**Expected**
An in-context cap with a hint to narrow the command, or a setting to disable persistence; and a way for the user to have the model remove them.

**What the user said**
"YOU will not 'save anything to files' or otherwise."

**Repro**
Run a command that prints more than about 30 KB, then ask the model to delete the persisted file.

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

Reproduce the issue with a Bash command producing more than 30 KB, then inspect how results are written under ~/.claude/projects///tool-results/*.txt and how auto mode classifies deletion requests. Done should mean large output is handled without unwanted persistence or duplicate reading, and users can remove persisted results when requested.

Written by the indexing model from the issue text.