Input tokens

your question system/developer instructions chat history included in the context file contents pasted or attached repo/folder code the tool chooses to include error logs terminal output tool results

Yes. Input tokens are everything the model receives for that request, including:

your question system/developer instructions chat history included in the context file contents pasted or attached repo/folder code the tool chooses to include error logs terminal output tool results

So if you say:

Look at my whole repo and find the bug

the input may include a lot of files, summaries, search results, and code chunks.

If you say:

Only look at backend/auth/middleware.go and backend/db/users.sql

then the input is much smaller and usually more accurate.

Best way to reduce tokens

Be specific about scope:

Look only at this file.

or:

Look only in cmd/server and internal/auth.

or:

Do not scan node_modules, dist, vendor, .git, build, coverage, generated files, or lock files.

For repo work, this is usually best:

Here is the error. First inspect only the files likely involved. Ask before expanding scope.

Even better, give:

the exact error the command you ran the file/function involved the expected behavior the actual behavior

Simple rule

Bigger scope = more input tokens = more cost/slower/more noise Smaller scope = fewer input tokens = faster/cheaper/often better

So yes: asking it to inspect a specific file, directory, function, or error path is the right way to keep tokens down.