Dynamic system prompt

[ErikBjare]

• Current date & training cutoff (model-dependent): https://github.com/ErikBjare/gptme/issues/137
  • This seems easy to do, just looks for a date/cutoff-block in the system prompt and replace it if present on each call.
• Tool selection
  • Currently, we load default tools or those set by --tools, but it only impacts the system prompt on the first run.
  • [x] ~~We could change --tools to only be valid for new conversations (lame?)~~
    • Would still need us to figure out which tools the conversation was using when loaded.
  • [ ] We could store the loaded tools for the conversation somewhere, so that we can at least resume with the correct toolset.
  • [ ] We could let the user override --tools and have it modify the system prompt.

Vaguely related to (similar issues around mutability of conversations): https://github.com/ErikBjare/gptme/issues/150

[kennethreitz]

Curious to see how this develops.

[ErikBjare]

This has been fixed a long time ago in various PRs. Can now be achieved via hooks or gptme.toml config such as files and context_cmd, as well as more experimental things like lesson retrieval and "fresh context" stuff.

Closing.