feat(mcp-core): Surface intentional/unintentional tags in find_dropped_events

packages/mcp-core/src/skillDefinitions.json

@@ -19,7 +19,7 @@
       },
       {
         "name": "find_dropped_events",
-        "description": "Find events dropped before they were stored in Sentry — ground-truth data-fidelity information about what was and wasn't captured.\n\nEvents can be dropped client-side in the SDK (sample_rate, before_send) or\nserver-side at ingest (rate limited, over quota, filtered, invalid, abuse/spike\nprotection, cardinality limited). Accepted-only views (searches, aggregates,\ncharts) can't show this, so the data may be incomplete in ways they don't reveal\n— for example, a flat or spiky chart caused entirely by drops.\n\nUse this tool when you need to:\n- Explain why a chart is flat, lower than expected, or doesn't match what the user is sending\n- Confirm the data you need is actually in Sentry (not dropped) before trusting a query, aggregate, or dashboard\n- Attribute a volume anomaly to a specific drop reason (quota, spike protection, sampling, filters)\n- Tell the user why their data is missing and what to do about it (raise quota, fix sampling, etc.)\n\nReturns dropped event volume bucketed over time, with the drop `outcome` and `reason`\nfor each bucket, plus the accepted volume per bucket so you can compute the dropped share.\n\n<examples>\nfind_dropped_events(organizationSlug='my-org', dataset='spans', projectSlug='my-project')\nfind_dropped_events(organizationSlug='my-org', dataset='logs', statsPeriod='30d')\nfind_dropped_events(organizationSlug='my-org', dataset='errors', outcome='rate_limited')\n</examples>\n\n<hints>\n- This is independent of any search query — it reports drops for the whole project/time range.\n- `outcome` is the drop kind (e.g. rate_limited, filtered); `reason` is the sub-cause (e.g. key_quota, sample_rate).\n- Pass `outcome` and/or `reason` to scope the dropped side to one classification; the accepted volume is always returned in full.\n- An empty `droppedEvents` list means no drops in the window — the data can be trusted.\n</hints>",
+        "description": "Find events dropped before they were stored in Sentry — ground-truth data-fidelity information about what was and wasn't captured.\n\nEvents can be dropped client-side in the SDK (sample_rate, before_send) or\nserver-side at ingest (rate limited, over quota, filtered, invalid, abuse/spike\nprotection, cardinality limited). Accepted-only views (searches, aggregates,\ncharts) can't show this, so the data may be incomplete in ways they don't reveal\n— for example, a flat or spiky chart caused entirely by drops.\n\nUse this tool when you need to:\n- Explain why a chart is flat, lower than expected, or doesn't match what the user is sending\n- Confirm the data you need is actually in Sentry (not dropped) before trusting a query, aggregate, or dashboard\n- Attribute a volume anomaly to a specific drop reason (quota, spike protection, sampling, filters)\n- Tell the user why their data is missing and what to do about it (raise quota, fix sampling, etc.)\n\nReturns dropped event volume bucketed over time, with the drop `outcome` and `reason`\nfor each bucket, plus the accepted volume per bucket so you can compute the dropped share.\n\n<examples>\nfind_dropped_events(organizationSlug='my-org', dataset='spans', projectSlug='my-project')\nfind_dropped_events(organizationSlug='my-org', dataset='logs', statsPeriod='30d')\nfind_dropped_events(organizationSlug='my-org', dataset='errors', outcome='rate_limited')\n</examples>\n\n<hints>\n- This is independent of any search query — it reports drops for the whole project/time range.\n- `outcome` is the drop kind (e.g. rate_limited, filtered); `reason` is the sub-cause (e.g. key_quota, sample_rate).\n- Each bucket's `tags` classify the drop: `intentional` (user-configured, e.g. sampling or filters) vs `unintentional` (not intended, e.g. quota or abuse).\n- Pass `outcome` and/or `reason` to scope the dropped side to one classification; the accepted volume is always returned in full.\n- An empty `droppedEvents` list means no drops in the window — the data can be trusted.\n</hints>",
         "requiredScopes": ["event:read"]
       },
       {

packages/mcp-core/src/toolDefinitions.json

@@ -2291,7 +2291,7 @@
   },
   {
     "name": "find_dropped_events",
-    "description": "Find events dropped before they were stored in Sentry — ground-truth data-fidelity information about what was and wasn't captured.\n\nEvents can be dropped client-side in the SDK (sample_rate, before_send) or\nserver-side at ingest (rate limited, over quota, filtered, invalid, abuse/spike\nprotection, cardinality limited). Accepted-only views (searches, aggregates,\ncharts) can't show this, so the data may be incomplete in ways they don't reveal\n— for example, a flat or spiky chart caused entirely by drops.\n\nUse this tool when you need to:\n- Explain why a chart is flat, lower than expected, or doesn't match what the user is sending\n- Confirm the data you need is actually in Sentry (not dropped) before trusting a query, aggregate, or dashboard\n- Attribute a volume anomaly to a specific drop reason (quota, spike protection, sampling, filters)\n- Tell the user why their data is missing and what to do about it (raise quota, fix sampling, etc.)\n\nReturns dropped event volume bucketed over time, with the drop `outcome` and `reason`\nfor each bucket, plus the accepted volume per bucket so you can compute the dropped share.\n\n<examples>\nfind_dropped_events(organizationSlug='my-org', dataset='spans', projectSlug='my-project')\nfind_dropped_events(organizationSlug='my-org', dataset='logs', statsPeriod='30d')\nfind_dropped_events(organizationSlug='my-org', dataset='errors', outcome='rate_limited')\n</examples>\n\n<hints>\n- This is independent of any search query — it reports drops for the whole project/time range.\n- `outcome` is the drop kind (e.g. rate_limited, filtered); `reason` is the sub-cause (e.g. key_quota, sample_rate).\n- Pass `outcome` and/or `reason` to scope the dropped side to one classification; the accepted volume is always returned in full.\n- An empty `droppedEvents` list means no drops in the window — the data can be trusted.\n</hints>",
+    "description": "Find events dropped before they were stored in Sentry — ground-truth data-fidelity information about what was and wasn't captured.\n\nEvents can be dropped client-side in the SDK (sample_rate, before_send) or\nserver-side at ingest (rate limited, over quota, filtered, invalid, abuse/spike\nprotection, cardinality limited). Accepted-only views (searches, aggregates,\ncharts) can't show this, so the data may be incomplete in ways they don't reveal\n— for example, a flat or spiky chart caused entirely by drops.\n\nUse this tool when you need to:\n- Explain why a chart is flat, lower than expected, or doesn't match what the user is sending\n- Confirm the data you need is actually in Sentry (not dropped) before trusting a query, aggregate, or dashboard\n- Attribute a volume anomaly to a specific drop reason (quota, spike protection, sampling, filters)\n- Tell the user why their data is missing and what to do about it (raise quota, fix sampling, etc.)\n\nReturns dropped event volume bucketed over time, with the drop `outcome` and `reason`\nfor each bucket, plus the accepted volume per bucket so you can compute the dropped share.\n\n<examples>\nfind_dropped_events(organizationSlug='my-org', dataset='spans', projectSlug='my-project')\nfind_dropped_events(organizationSlug='my-org', dataset='logs', statsPeriod='30d')\nfind_dropped_events(organizationSlug='my-org', dataset='errors', outcome='rate_limited')\n</examples>\n\n<hints>\n- This is independent of any search query — it reports drops for the whole project/time range.\n- `outcome` is the drop kind (e.g. rate_limited, filtered); `reason` is the sub-cause (e.g. key_quota, sample_rate).\n- Each bucket's `tags` classify the drop: `intentional` (user-configured, e.g. sampling or filters) vs `unintentional` (not intended, e.g. quota or abuse).\n- Pass `outcome` and/or `reason` to scope the dropped side to one classification; the accepted volume is always returned in full.\n- An empty `droppedEvents` list means no drops in the window — the data can be trusted.\n</hints>",
     "inputSchema": {
       "type": "object",
       "properties": {
@@ -2443,6 +2443,13 @@
               },
               "count": {
                 "type": "number"
+              },
+              "tags": {
+                "default": [],
+                "type": "array",
+                "items": {
+                  "type": "string"
+                }
               }
             },
             "required": [
@@ -2451,7 +2458,8 @@
               "category",
               "start",
               "end",
-              "count"
+              "count",
+              "tags"
             ],
             "additionalProperties": false
           }
@@ -2478,6 +2486,13 @@
               },
               "count": {
                 "type": "number"
+              },
+              "tags": {
+                "default": [],
+                "type": "array",
+                "items": {
+                  "type": "string"
+                }
               }
             },
             "required": [
@@ -2486,7 +2501,8 @@
               "category",
               "start",
               "end",
-              "count"
+              "count",
+              "tags"
             ],
             "additionalProperties": false
           }

packages/mcp-core/src/tools/catalog/find-dropped-events.test.ts

@@ -33,6 +33,7 @@ const DROPPED_EVENTS_RESPONSE = {
       start: 1_700_000_000_000,
       end: 1_700_003_600_000,
       count: 400,
+      tags: ["unintentional"],
     },
   ],
   acceptedEvents: [
@@ -44,6 +45,7 @@ const DROPPED_EVENTS_RESPONSE = {
       start: 1_700_000_000_000,
       end: 1_700_003_600_000,
       count: 1000,
+      tags: [],
     },
   ],
 };
@@ -89,6 +91,7 @@ describe("find_dropped_events", () => {
             "outcome": "accepted",
             "reason": "accepted",
             "start": 1700000000000,
+            "tags": [],
           },
         ],
         "dataset": "spans",
@@ -100,6 +103,9 @@ describe("find_dropped_events", () => {
             "outcome": "rate_limited",
             "reason": "key_quota",
             "start": 1700000000000,
+            "tags": [
+              "unintentional",
+            ],
           },
         ],
         "interval": 3600000,

packages/mcp-core/src/tools/catalog/find-dropped-events.ts

@@ -28,6 +28,7 @@ const droppedBucketSchema = z.object({
   start: z.number(),
   end: z.number(),
   count: z.number(),
+  tags: z.array(z.string()).default([]),
 });
 
 export const findDroppedEventsOutputSchema = z.object({
@@ -68,6 +69,7 @@ export default defineTool({
     "<hints>",
     "- This is independent of any search query — it reports drops for the whole project/time range.",
     "- `outcome` is the drop kind (e.g. rate_limited, filtered); `reason` is the sub-cause (e.g. key_quota, sample_rate).",
+    "- Each bucket's `tags` classify the drop: `intentional` (user-configured, e.g. sampling or filters) vs `unintentional` (not intended, e.g. quota or abuse).",
     "- Pass `outcome` and/or `reason` to scope the dropped side to one classification; the accepted volume is always returned in full.",
     "- An empty `droppedEvents` list means no drops in the window — the data can be trusted.",
     "</hints>",
@@ -169,13 +171,15 @@ export default defineTool({
       start: number;
       end: number;
       count: number;
+      tags?: string[];
     }) => ({
       outcome: bucket.outcome,
       reason: bucket.reason,
       category: bucket.category,
       start: bucket.start,
       end: bucket.end,
       count: bucket.count,
+      tags: bucket.tags ?? [],
     });
 
     return structuredResult({