Watch & React Workflow
The Watch & React workflow is for situations where you want to monitor nodes over time rather than analyze a specific recorded moment. Instead of recording and reviewing, you set up watches and the agent observes the game as it runs.
When to use watches instead of recordings
| Use recordings when... | Use watches when... |
|---|---|
| The bug already happened | You want to catch the bug as it happens |
| You need frame-precise timing | Coarse timing is sufficient |
| You want to scrub backward in time | You want real-time monitoring |
| The bug is intermittent and hard to trigger | The bug is happening continuously |
| Post-mortem analysis | Live investigation |
The basic loop
1. spatial_snapshot — get current frame number
2. spatial_watch (create) — register nodes to monitor
3. [game runs, things happen]
4. spatial_delta — check what changed
5. [repeat 3-4 as needed]
6. spatial_watch (delete) — clean upExample: Monitoring player health during combat
You are debugging a health regen system that seems to not work correctly during certain combat conditions.
Step 1: Set up the watches
Step 2: Check changes during combat
Step 3: Narrow down the cause
The agent found that the in_combat flag never resets — likely a missing call to set it to false when the last enemy dies. The fix is in the combat state manager.
Example: Watching enemy patrol behavior
You suspect enemies are teleporting slightly during patrol waypoint transitions.
Watching multiple nodes simultaneously
Set up watches on related nodes to catch interactions:
{ "action": "create", "node": "Player", "track": ["position", "health"] }
{ "action": "create", "node": "Boss", "track": ["position", "velocity", "phase", "health"] }
{ "action": "create", "node": "Boss/AttackHitbox", "track": ["monitoring"] }
{ "action": "create", "node": "BossArena/HazardTiles", "track": ["monitoring"] }Now a single spatial_delta call shows changes across all four nodes — you can see the Boss switch phase at a certain health threshold, the AttackHitbox become active, and the HazardTiles respond, all in one response.
Delta poll frequency
You control how often to poll. The agent does not poll automatically — you ask for a delta when you want it.
Typical patterns:
- Quick check: poll once after a specific event ("I just died, check the delta")
- Periodic: poll every 10-30 seconds during a long session
- Event-triggered: poll after specific in-game events ("I just finished wave 3, check it")
There is no overhead from not polling — the watch just accumulates data. Poll whenever you want; all changes since the last poll are included.
Cleaning up
{ "action": "clear" }Always clear watches when moving to a new investigation. Stale watches from a previous investigation add noise to subsequent deltas.
Or delete a specific watch:
{ "action": "delete", "watch_id": "w_c8d9e0" }