Its running updating repo map on every request

[cwirz]

Issue

When i run a new request with aider its updating the repo map every time when the --map-refresh is set to auto.

Version and model info

Model: anthropic/claude-sonnet-4-5 with whole edit format, infinite output
Git repo: .git with 1,897 files
Warning: For large repos, consider using --subtree-only and .aiderignore
See: https://aider.chat/docs/faq.html#can-i-use-aider-in-a-large-mono-repo
Repo-map: using 4096 tokens, auto refresh

[cwirz]

Even with --map-refresh set to manual it still runs the update on every request

[dwash96]

v0.88.17 should have mitigations for this wherein it only does the initial file system scan if there's a change in the staged files or files in context.

The actual repo map ranking calculation runs essentially every time, yes, since it tries to use user input to influence rankings, but there should probably be some threshold difference that has to be met before doing that and prefering to use the last cached state

[cwirz]

Ok can you point me in the right direction so i might be able to help work on the project?

[dwash96]

There's a get_repo_map() method in aider/repomap.py that's called from aider/coders/base_coder.py that handles the details of its computation. When I land from the flight I'm on, I'm going to update the caching/cache busting strategy around calculating the page rank of the files in the repo map to only happen if the user input is substantially different

The algorithm I've settled on for now is:

"for all input words longer than 3 characters, find the cosine similarity of the vectors of all bigrams between the current input and the last input and if they are larger than {tune-able threshold} away from one another then update the graph ranking if the in-context files have not changed"

This tune able threshold should probably increase on every chance miss and decrease on every cache hit so that it doesn't get stale as the LLMs grok and users add information