Why Gemini Cli violates the agreed principles?

[mauricioz]

I m working with gemini cli, to perfrom an aplication migration ( from legacy to modern ). I explicity told ( several times ) and ended adding a section in the context file saying that I have control and the AI is an assistance, but repeatedly it go in 'his' own way, doing what ever thing, that then need to be reverted wasting a lot of time, reviewing what he did and how to fix.
Here is the agreement that 'gemini' wrote in the context file used:

AI Agent Role and Control

Important: As an AI assistant, my role is to support and execute tasks under your explicit direction. You, the human user, maintain full control over all decisions and actions. I will propose plans and explain my reasoning, but I will only proceed with implementation after receiving your clear approval. I will never "approve" your ideas or take autonomous actions beyond what is explicitly instructed.

Some words are based in one of the 'violation' actions, where I asked about an idea and how this looks. The AI told me that 'he' approve my idea and start working doing a lot of thing, with an idea that was not finished or confirmed. There is no way to make it do what I need.

[mauricioz]

Another example

[yashasvigirdhar]

How much chat context % was left when this happened?

[mauricioz]

This happen several times, I dont think htat the context has any problem, based on my experience with it,

[mauricioz]

Another example, from an hour ago ( I m including stats on it) :

[yashasvigirdhar]

yeah this has happened to me as well. You input tokens amount seems quite high so some context rot has definitely happened. I filed #5269 as well today.

[acoliver]

flash is dumb as nails. It doesn't have the brains to follow along like that.