Repository navigation
Snapshot git add runs indefinitely on worktrees with large non-code files #18072
Description
Description
When opencode is used in a workspace that contains large non-code files (media files, datasets, ML models, etc.), the snapshot system's git add . can run indefinitely, consuming excessive CPU and memory.
The snapshot system maintains a separate git repo to track worktree state. It runs git add . against the full worktree on every agentic step (via the add effect in snapshot/index.ts). When the worktree contains large directories with binary/media files, this operation never completes in a reasonable time.
Steps to reproduce
- Open opencode in a workspace containing a large directory (e.g. 500MB+ of media files, datasets, or ML model weights)
- Start a session and interact with the assistant
- Observe
git add .andgit pack-objectsprocesses consuming high CPU/memory indefinitely
Observed behavior
git add .ran for 3+ hours at ~90% CPU- The subsequent hourly
git gc(snapshot cleanup) consumed ~3.7GB of RAM trying to repack the bloated snapshot repo - The system became sluggish and required manual process killing
Expected behavior
The snapshot system should handle large worktrees gracefully — either by timing out and excluding large/binary files, or by respecting size limits.
Environment
- OS: Linux (Ubuntu)
- opencode version: v1.2.27
Activity
github-actions commentedon Mar 18, 2026
This issue might be a duplicate of existing issues. Please check:
watcher.ignoreconfig, causing the samegit add .performance degradation on large files/directoriesihubanov commentedon Mar 18, 2026
Thanks for flagging these — #8887 and #3182 are definitely the same root cause.
#8887 proposes respecting the existing
watcher.ignoreconfig in the snapshot module, which makes sense as a proactive fix. Our PR (#18069) takes a complementary approach with a runtime safety net: a 15-second timeout ongit add .that triggers automatic exclude generation only when the operation is actually slow. This handles cases where users haven't configuredwatcher.ignore(which seems common given #3182).Ideally both approaches could be combined — respect
watcher.ignoreupfront (as #8887 suggests) AND have a timeout fallback for unconfigured workspaces. Happy to update the PR to incorporate thewatcher.ignoreintegration from #8887 if the maintainers prefer a combined fix.Linking: this also fixes #8887 and #3182.
00sapo commentedon Mar 18, 2026
Any temporary workaround?
thdxr commentedon Mar 25, 2026
This should be fixed by commit 0a80ef4 which adds a 2MB file size limit for snapshotting. Large files are now skipped during git add operations.