Custom prompts should only be added to context when invoked, not at startup - they occupy context and confuse the model

What feature would you like to see?

Prompts stored in ~/.codex/prompts/*.md are currently fully loaded into context on startup (cfr https://github.com/openai/codex/blob/main/docs/prompts.md#where-prompts-live). This behavior limits the custom prompts' usefulness, as they eat into the context window and their instructions are sometimes followed by the model when I didn't command it to.

I should be able to have prompts that apply to some circumstances but not others. By always having all prompts in-context, their instructions are applied to ALL situations, without the user having the ability to command otherwise, particularly in non-interactive mode. For example, currently both /prompts:MAKE-WHOLESALE-COMMIT and /prompts:MAKE-ATOMIC-COMMITS both live in-context all the time.

Version: OpenAI Codex (v0.63.0) on WSL Ubuntu 24.04

Additional information

_No response_