Separate llm context for buddy?

[flivni]

I wonder if adding <!-- buddy: [reaction here] --> to context could have unintended consequences?

• Attention tax. Every turn Claude has to remember and execute an extra ritual. On a hard task, "append the comment in character referencing something specific from this turn" competes for the same planning budget as the actual work.
• Persona bleed. The instructions set a personality ("dragon", "peak stat = sass"). Strong persona framing could leak into the main response: tone shifts, more asides, occasionally in-character asides outside the comment.
• Instruction conflicts. "NEVER skip this. Every single response must end with..." is a hard override. If another system prompt says "end with X," now there are two absolute rules and the model picks one.

An alternate approach that uses a separate llm context might work:

• Stop hook calls claude -p with inexpensive model (like haiku) and disowns (so not blocking)
• Then statusline tick picks up the change
• Recursion guard so that claude -p doesn't also call buddy. E.g. via env guard, BUDDY_IN_PET_CALL.