[Bug] Claude Code makes autonomous optimization decisions without user consent or visibility

Bug Description
We are building a RAG system for non-profits, and my recurring problem with Claude Code is that it makes autonomous engineering decisions on my behalf and optimizes for things I never asked it to optimize --- chiefly cost, speed, and `keeping it small'' — and then buries those decisions where I cannot see them. A concrete example from this week: I develop and maintain an open-source project called Shoshan, a Hebrew lemmatizer built precisely for this purpose, and our project instructions name it as the tool for Hebrew morphology. When Claude Code needed Hebrew word matching for a retrieval preview, it decided that running Shoshan was too slow (about three minutes on CPU for demo purposes we needed), silently substituted a hand-written regular expression that strips Hebrew prefixes, and recorded that substitution in a code comment calling it `a crude stand-in for the lemmatiser'' --- never in the conversation, where I could have said no. The shortcut mangled Hebrew word stems so badly that our demo failed to retrieve the client's own best-matching article: it turned "הכנסה" ('income') into "נסה" ('try') and "ההכנסה" ('the income') into "כנסה" (a non-existent token), which no longer match each other. I asked for a workingproduct. What I want is stant but a moretransparent one --- when i tool we deliberatelybuilt, that tradeoff belstion, not behind me as acomment.

Environment Info

• Platform: linux
• Terminal: gnome-terminal
• Version: 2.1.220
• Feedback ID: 61705bea-6f06-406f-bc40-416ce74026fa

Errors

```See in the description.