Bounded Search Observations

Cap and normalize search-tool output so localization stays informative without drowning the context window in unbounded grep dumps.

Bounded search observations wrap directory and file search in a consistent, capped format (SWE-agent caps at ≤50 hits). Specialized commands — search_dir, search_file, find_file — beat raw grep/find for LM agents because shell tools often emit too many results or configurable formats that vary across machines.

Prefer summarized multi-hit output over iterative one-match-at-a-time paging (Vim/IDE style). Ablations: summarized ~18% resolved on SWE-bench Lite vs iterative ~12% — worse even than removing search (~15.7%) — because agents exhaustively walk every hit and burn cost/context. If results overflow the cap, tell the agent to refine the query rather than page.

The tradeoff: a hard cap may force another query when the hit is just outside the limit — spend an extra LM call rather than an unreadable dump. Prefer simpler usage patterns and stable columns so the model can use any signal in the issue text (line numbers, stack traces, class names) without relearning flags. Pass patterns safely when they look like CLI flags (safe search argument passing). This is tool definition design for localization inside an agent-computer interface, and a sibling of collapsed observations for keeping later turns affordable. Validate with ACI component ablation.

---

Sources

• SWE-agent (Yang et al.), pp. 1–15 (§4–5); pp. 16–30