One feature of claude code that people use in practice (https://code.claude.com/docs/en/hooks#prompt-based-hooks), allowing injecting context into a agent's prompt when it makes tool calls. We could support this functionality using handlers