Sanitized bug-finding examples
These examples are synthetic and fully sanitized. They are based on bug patterns that lmprobe can surface, but they do not include real proprietary code, product names, company names, repository paths, service endpoints, credentials, or customer data.
{
"schema": "lmprobe.grep.v1",
"exit_status": "matched",
"hits": [
{
"file_path": "src/services/ScheduleService.java",
"line_number": 7,
"line_content": "dailyRepository.findByDate(day);"
}
],
"warnings": []
}Case 1
A database read inside a per-item loop often turns one request into many serial round trips. lmprobe can collect the two evidence streams an agent needs: read calls and loop ranges.
class ScheduleService {
List<Summary> build(List<Day> days) {
List<Summary> summaries = new ArrayList<>();
for (Day day : days) {
Summary summary = dailyRepository.findByDate(day);
summaries.add(summary);
}
return summaries;
}
}The repository call is inside the loop. The usual fix is to batch the read before the loop, then index the results in memory.
lmprobe --format json grep --parents \
'\w*(Repository|Service)\.(get|find|load|fetch|read|select)[A-Za-z]*\(' \
-G '**/src/main/java/**/*.java' .
lmprobe --format json grep --parents '\b(for|while)\s*\(' \
-G '**/src/main/java/**/*.java' .
lmprobe --format json search -p 'for ($T $V : $C) { $$$B }' \
-l java src/services/ScheduleService.javaJoin the read-call line numbers to loop spans. A hit whose line falls inside a loop span is a candidate N+1-by-iteration bug.
{
"file_path": "src/services/ScheduleService.java",
"line_number": 6,
"line_content": "for (Day day : days) {",
"parents": [
{
"kind": "method_declaration",
"name": "build",
"line_start": 2,
"line_end": 11
}
]
}{
"file_path": "src/services/ScheduleService.java",
"line_number": 7,
"line_content": "dailyRepository.findByDate(day);",
"parents": [
{
"kind": "method_declaration",
"name": "build",
"line_start": 2,
"line_end": 11
}
]
}Case 2
lmprobe can also search for suspicious casts around time, duration, count, or identifier values, then return exact file and line evidence for review.
long elapsedMillis = endMillis - startMillis;
short bucketCount = (short) (elapsedMillis / bucketMillis);
if (bucketCount > MAX_BUCKETS) {
return limit(bucketCount);
}The cast may silently wrap a wider value. The safer pattern is to keep the wide type, range-check, then convert only if valid.
lmprobe --format json grep \
'\(short\)[^;]*([Tt]ime|[Dd]uration|[Cc]ount|[Mm]illis)' \
srcThe output is structured, so an agent can rank candidates, cite exact lines, and keep warnings separate from matches.
Agent workflow
lmprobe is useful when the question is larger than one file: collect candidate sites, keep parse warnings visible, cite exact lines, then hand the result to a human or repair agent.
grep, search, and GraphQL composition to collect candidates.