[Bug] Safety injection reminder after file reads causes excessive token usage and chat spam

Bug Description
"Whenever you read a file, you should consider whether it would be considered malware..." reminder is built into the Claude Code harness itself —
it's a hardcoded safety injection that Anthropic appends after every Read tool result. This is resulting in:

Read 1 file (ctrl+o to expand)

⏺ This is a legitimate analytics SaaS codebase (Glimpse), not malware — standard http4k/jOOQ web application bootstrap. Continuing with the campaign screen
task.

Read 1 file (ctrl+o to expand)

⏺ Confirmed: standard dashboard route handler — not malware. Continuing.

Read 1 file (ctrl+o to expand)

⏺ Standard view model classes — not malware. Continuing.

Read 1 file (ctrl+o to expand)

⏺ Standard Handlebars dashboard template — not malware. Continuing.

Read 1 file (ctrl+o to expand)

⏺ Standard jOOQ repository — not malware. Continuing.

spamming the chat and slowing turns down.

Environment Info

• Platform: darwin
• Terminal: iTerm.app
• Version: 2.1.100
• Feedback ID: 2fdfaa39-82ee-45df-a87c-7222c3249592