Ag Lag Diagnosis

AG IDE Lag Diagnosis & Fix

When to trigger: User reports AG IDE edits stuck on "Working...", Agent Manager freezing, or general IDE sluggishness that resolves when switching to Editor view.

Quick Diagnosis (Run All)

Step 1: Find the latest log session

ls -lt ~/Library/Application\ Support/Antigravity/logs/ | head -3
# Use the most recent directory, e.g. 20260306T175017
# Set AG_LOG variable for subsequent commands

Step 2: Check streamAgentStateUpdates polling storm

# Total error count
grep -c "streamAgentStateUpdates error" "$AG_LOG/agent-window-console.log"

# Which conversations are polling (sorted by count)
grep "streamAgentStateUpdates error" "$AG_LOG/agent-window-console.log" \
  | awk -F'error for ' '{print $2}' | awk -F: '{print $1}' \
  | sort | uniq -c | sort -rn

Diagnosis: If any single conversation has 100+ errors, it is in a polling death loop (~2s interval, no backoff). This is the #1 cause of lag.

Step 3: Check VERY LONG TASK renderer blocking

grep "VERY LONG TASK" "$AG_LOG/rendererPerf.log" | tail -20

Diagnosis: Tasks >200ms block the UI. Tasks >500ms cause visible freezes. Tasks >1000ms cause multi-second hangs.

Step 4: Check oversized conversation .pb files (remote workstation)

ssh workstation 'find ~/.gemini/antigravity/conversations -name "*.pb" -size +4M \
  -printf "%s %f\n" | sort -rn | head -20'

Diagnosis: Files >4MB exceed gRPC message limit and trigger update is too large errors, causing infinite spin.

Step 5: Check process resource usage

ps axo pid,rss,pcpu,comm | grep -i antigravity | sort -k2 -rn | head -10

Diagnosis: Renderer processes >500MB RSS indicate memory leaks. CPU >20% on a Renderer indicates active polling storm.

Fixes (In Order of Effectiveness)

Fix 1: Archive oversized conversation .pb files

This is the most impactful fix. Move .pb files >4MB to an archive directory (preserves history, can restore later).

# Use the provided script
bash ag-lag-diagnosis/scripts/archive_large_conversations.sh

Or manually on the remote workstation:

mkdir -p ~/.gemini/antigravity/conversations_archive
find ~/.gemini/antigravity/conversations -name "*.pb" -size +4M \
  -exec mv {} ~/.gemini/antigravity/conversations_archive/ \;

IMPORTANT: Before archiving, summarize conversation contents for user confirmation:

for f in $(find ~/.gemini/antigravity/conversations -name "*.pb" -size +4M -printf "%f\n"); do
  cid=${f%.pb}
  echo "=== $cid ==="
  head -5 ~/.gemini/antigravity/brain/$cid/task.md 2>/dev/null || echo "(no task.md)"
done

After archiving, restart AG IDE.

Fix 2: Restart AG IDE

Clears all orphan stream subscriptions and resets memory. This alone fixes polling storms temporarily, but they will recur if oversized .pb files remain.

Fix 3: Enable GPU rasterization (macOS only)

Edit ~/.antigravity/argv.json and add:

{
    "enable-gpu-rasterization": true
}

Offloads rendering from CPU main thread to GPU, reducing VERY LONG TASK frequency.

Fix 4: Reduce open workspaces

Each workspace window creates an independent cascadeId and starts polling. Close workspaces not actively used for agent operations.

Fix 5: Restore archived conversations (if needed)

mv ~/.gemini/antigravity/conversations_archive/<conversation_id>.pb \
  ~/.gemini/antigravity/conversations/

Root Cause Summary

Anti-Patterns

❌ Only restarting AG IDE without archiving .pb files
   → Polling storms WILL recur within hours. Archive >4MB files FIRST, then restart

❌ Deleting .pb files instead of archiving them
   → Move to conversations_archive/, don't delete. User may want to restore later

❌ Diagnosing lag without checking ALL 5 steps
   → Multiple causes are often active simultaneously. Run ALL diagnosis steps before fixing

❌ Archiving .pb files without user confirmation
   → ALWAYS show conversation summaries (task.md) and get user approval before moving files

❌ Assuming lag is a network issue
   → AG lag is almost always caused by local polling storms or oversized .pb files, not network

❌ Running diagnosis on the remote workstation when logs are on macOS
   → AG logs are in ~/Library/... on macOS, but .pb files are on the remote workstation. Check both

Mandatory Rules

1. Confirm before archiving: Always show conversation summaries before moving .pb files
2. Archive, don't delete: Move .pb files to conversations_archive/, never rm
3. Restart after archive: AG IDE must be restarted after archiving for changes to take effect

Prevention

• Archive conversations >4MB weekly
• Restart AG IDE daily
• Keep open workspace count to ≤3
• Monitor with: grep -c "streamAgentStateUpdates error" $AG_LOG/agent-window-console.log