Hooks?

[ErikBjare]

One potential way to deal with linting etc is to add a system with hooks that tools can register with.

These hooks could do things like:

• Generating dynamic system prompt on every message: #151
• Implement memory (as done by https://pypi.org/project/simplemind/)
• Interrupt generation on detected tooluse (as currently hard-coded in our llm code)
• Before running tools: Given an assistant message, react to or transform the message. #152
  • Would work for patch/save-optimization
• After saving/patching a file: lint/formatting.
  • Maybe shouldn't be a hook?
• Call out to expert sub-agents to help with problems?

...what else?

Not sure if making a "hook" concept like this is the way to go.

Maybe adopting something like https://pypi.org/project/simplemind/ (https://github.com/ErikBjare/gptme/issues/233) makes sense long-term, if simple and pliable enough.

[ErikBjare]

Experimented with a lot of the linting/checking stuff in https://github.com/ErikBjare/gptme/pull/199, without going the hook approach.