fix(vfs): add error on vfs materialize fail

Author: Sg312

apps/sim/lib/copilot/request/tools/executor.ts

@@ -399,20 +399,31 @@ export async function executeToolAndReport(
     },
     async (otelSpan) => {
       const startedAt = Date.now()
-      const completion = await executeToolAndReportInner(toolCall, context, execContext, options)
-      const durationMs = Date.now() - startedAt
-      otelSpan.setAttribute(TraceAttr.ToolOutcome, completion.status)
-      otelSpan.setAttribute(TraceAttr.ToolDurationMs, durationMs)
-      if (completion.message) {
-        otelSpan.setAttribute(
-          TraceAttr.ToolOutcomeMessage,
-          String(completion.message).slice(0, 500)
-        )
+      try {
+        const completion = await executeToolAndReportInner(toolCall, context, execContext, options)
+        const durationMs = Date.now() - startedAt
+        otelSpan.setAttribute(TraceAttr.ToolOutcome, completion.status)
+        otelSpan.setAttribute(TraceAttr.ToolDurationMs, durationMs)
+        if (completion.message) {
+          otelSpan.setAttribute(
+            TraceAttr.ToolOutcomeMessage,
+            String(completion.message).slice(0, 500)
+          )
+        }
+        // Durable Grafana signal for "which Sim tool is slowest" (executor=sim);
+        // pairs with the Go executor-boundary metric (U15) as one series set.
+        recordSimToolMetric(toolCall.name, completion.status, durationMs)
+        return completion
+      } catch (err) {
+        // executeToolAndReportInner threw (infra/unexpected error, not a normal
+        // 'error' completion). Still stamp the span + record the dispatch so
+        // copilot.tool.* isn't silently biased toward successful calls.
+        const durationMs = Date.now() - startedAt
+        otelSpan.setAttribute(TraceAttr.ToolOutcome, 'error')
+        otelSpan.setAttribute(TraceAttr.ToolDurationMs, durationMs)
+        recordSimToolMetric(toolCall.name, 'error', durationMs)
+        throw err
       }
-      // Durable Grafana signal for "which Sim tool is slowest" (executor=sim);
-      // pairs with the Go executor-boundary metric (U15) as one series set.
-      recordSimToolMetric(toolCall.name, completion.status, durationMs)
-      return completion
     }
   )
 }

apps/sim/lib/copilot/vfs/workspace-vfs.ts

@@ -473,20 +473,25 @@ export class WorkspaceVFS {
             markSpanForError(span, err)
             throw err
           } finally {
+            // Record on success AND failure: a mid-phase failure (e.g. a DB
+            // timeout) still belongs in copilot.vfs.materialize.duration, else
+            // p50/p99 skew toward successes only. phaseMs holds whatever phases
+            // completed before the failure.
+            for (const [phase, ms] of Object.entries(phaseMs)) {
+              recordVfsMaterialize(phase, ms)
+            }
+            recordVfsMaterialize('total', Date.now() - start)
             span.end()
           }
         }
       )
 
-    const totalMs = Date.now() - start
     // Durable Grafana signal for "how long does VFS materialize" — total plus
     // per-phase (bounded phase set). getOrMaterializeVFS runs per VFS tool call
     // with no cross-request cache, so this reveals whether materialize is the
-    // bottleneck (observability only; not a fix).
-    for (const [phase, ms] of Object.entries(phaseMs)) {
-      recordVfsMaterialize(phase, ms)
-    }
-    recordVfsMaterialize('total', totalMs)
+    // bottleneck (observability only; not a fix). Recorded inside the span's
+    // finally above so a failed materialize is captured too, not just successes.
+    const totalMs = Date.now() - start
 
     logger.info('VFS materialized', {
       workspaceId,