I would assume the context requirements are the same, i.e. whatever knowledge is needed to complete the task. There is no novel technique that can avoid this.
No novel technique can avoid this, but certain techniques can make it worse.
Can you elaborate? Do you mean the model + harness combo reading more files than is necessary?
Yes, reading more files than necessary. Was curious about how their model handles a series of edits over longer tasks, but then tested it using openrouter myself, it worked quite well.