Change synthetic context injection to assistant role

[Tarquinen]

Summary

• Switch prunable-tools context injection from user message to assistant message role
• Remove redundant instruction block that was no longer needed with the new approach

[zackleman]

Did this yield better results for you?

[Tarquinen]

Yes the performance is definitely better, models are trained to address user messages so getting them to ignore the message is very hacky, but its harder to work with for 2 reasons:

1. Assistant messages can't be injected wherever you want the way user messages can, for example anthropic models require all assistant blocks to either have thinking or not have thinking depending on if its enabled or not.
2. Can get around number 1 by injecting into the last existing assistant message, but this leads to weird UI bugs in opencode and only sometimes with some providers, not sure why or if there's a way around it. Looks like this:

I haven't looked into this too much yet, if anyone wants to help this would be big!

[Tarquinen]

figured this out finally, merged to dev in #187