You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
if(filtered.length>maxFiles){log.warn(`Repository has ${filtered.length} files; only the first ${maxFiles} were collected. Use --max-files to raise the limit.`);}
I'd like to take this one. I'll add --max-files to teamai import (default stays 200), pass it through to the codebase extractor, and open a PR against main.
Thanks @hongchuwudi — that was a remarkably fast turnaround, PR #1196 landed the same day.
Confirming from the reporter side: the truncation warning is exactly what we needed.
Our case — importing 9 Scala/Java repos, 4 of which exceed 200 code files (434 / 391 / 372 / 394). Before the fix the graphs were silently cut to 200 files (~50% coverage), and the missing half produced broken dependency edges. That is arguably worse than having no graph at all: downstream queries (recall / AI Q&A) return confidently wrong answers rather than admitting ignorance.
--max-files + the warning together cover both halves of the problem — the escape hatch and the visibility. We will pick it up once it ships in a release (still on 0.28.0-beta.5 here, which predates the merge).
For reference, the effect of raising the limit on one repo: DataTools went from 200 files / 616 nodes / 157 edges to 434 files / 1150 nodes / 775 edges.
Activity
hongchuwudi commentedon Oct 17, 2026
I'd like to take this one. I'll add
--max-filestoteamai import(default stays 200), pass it through to the codebase extractor, and open a PR againstmain.feat(import): expose --max-files and warn on truncated code collection (
GDUFZXP commentedon Oct 18, 2026
Thanks @hongchuwudi — that was a remarkably fast turnaround, PR #1196 landed the same day.
Confirming from the reporter side: the truncation warning is exactly what we needed.
Our case — importing 9 Scala/Java repos, 4 of which exceed 200 code files (434 / 391 / 372 / 394). Before the fix the graphs were silently cut to 200 files (~50% coverage), and the missing half produced broken dependency edges. That is arguably worse than having no graph at all: downstream queries (recall / AI Q&A) return confidently wrong answers rather than admitting ignorance.
--max-files+ the warning together cover both halves of the problem — the escape hatch and the visibility. We will pick it up once it ships in a release (still on 0.28.0-beta.5 here, which predates the merge).For reference, the effect of raising the limit on one repo: DataTools went from 200 files / 616 nodes / 157 edges to 434 files / 1150 nodes / 775 edges.
Thanks again.