Towards an autonomous PR planner

[irgolic]

Some thoughts on what an improved pull request planning agent might look like, and how to implement it.

There should be another repo called EmbeddingsRepo, that stores embedded texts and searches among them. It should allow multiple implementations under the same interface:

• storing the embeddings in a file in the repo (preferably not pickle, which can execute arbitrary code at unpickle-time), loading them all into memory and computing cosine distance over the vectors
• using a vector database provider like Pinecone, wherein the distance calculation is done in a separate abstraction layer

The files in the repo should be summarized, the summary embedded and cached by the EmbeddingsRepo. This should be triggerable either:

• by a separate github workflow (i.e., on main push, any changed files are recomputed)
• at the start of the AutoPR run, before giving the brain_agent control

Then, an autonomous version of a pull_request_agent should be implemented; its action set should be:

• semantic query over repo, searching through file summary embeddings
• Ctrl+Shift+F style global find query, allowing it to control context window (default ~10 lines)
• plan a commit
• finish (after a commit has been planned)

[sionsmith]

Following this thread, would the idea be to be basically store the embeddings for each repo? So that there is a summary available of each repo or file?