Blind Scroll Loop
After finding a large file, repeatedly scrolling without search/goto wastes the episode — re-invoke search for the next symbol instead of linear hunting.
A stateful file viewer enables efficient location
when the agent uses find → open → search_file → goto. The complementary
failure mode is a blind scroll loop: after landing on a class (or other
anchor), the agent hunts a nested symbol (kind, a method, …) by issuing
scroll_down for dozens of turns while observations collapse to omitted
windows and the narrative forgets it already found the class.
Harness mitigations under ACI design principles:
- Prompt tips / demos that prefer
search_file/gotoover scroll for symbol lookup in multi-thousand-line files. - Detect repetitive scroll-without-search streaks and inject a system warning or force a search tool (failure-derived prompt tips).
- Collapse scroll observations aggressively (collapsed observations) so the loop is both expensive and information-poor — make search the cheaper path.
This is an instance of compound mistake amplification inside agent trajectory phases localization: early search success does not prevent mid-episode navigation regression. Pair with issue-grounded localization so the agent reaches the right file before scroll discipline matters.
Sources
- SWE-agent (Yang et al.), pp. 76–90