diff --git a/cdk/stacks/pipeline_stack.py b/cdk/stacks/pipeline_stack.py
index 8b3a36e..aa38888 100644
--- a/cdk/stacks/pipeline_stack.py
+++ b/cdk/stacks/pipeline_stack.py
@@ -64,6 +64,11 @@ from constructs import Construct
 GLUE_DATABASE = "apm_wo_analysis"
 GLUE_TABLE = "apm_wo_snapshots"
 ATHENA_WORKGROUP = "apm-wo-analysis"
+# AWS-managed SDK-for-pandas layer (awswrangler 3.16.1, py3.12, arm64). Provides
+# awswrangler/pandas/pyarrow/numpy pre-stripped to fit the Lambda size limit.
+AWSSDKPANDAS_LAYER_ARN = (
+    "arn:aws:lambda:us-east-1:336392948345:layer:AWSSDKPandas-Python312-Arm64:27"
+)
 ANTHROPIC_SECRET = "apm-wo-analysis/anthropic-api-key"
 SLACK_SECRET = "apm-wo-analysis/slack-credentials"
 GRAFANA_URL_PARAM = "/apm-wo-analysis/grafana-base-url"
@@ -231,6 +236,15 @@ class PipelineStack(Stack):
             timeout=Duration.seconds(120),
             log_group=classifier_logs,
             environment={"APM_HAIKU_FALLBACK": "on"},
+            # awswrangler/pandas/pyarrow/numpy come from the AWS-managed
+            # SDK-for-pandas layer (pre-stripped to fit the 250 MB unzipped
+            # limit, which bundling them ourselves blows). The function package
+            # only bundles openpyxl; boto3 is in the runtime, urllib is stdlib.
+            layers=[
+                lambda_.LayerVersion.from_layer_version_arn(
+                    self, "PandasLayer", AWSSDKPANDAS_LAYER_ARN
+                )
+            ],
             code=lambda_.Code.from_asset(
                 os.path.join(LAMBDAS_DIR, "classifier"),
                 bundling=BundlingOptions(
diff --git a/lambdas/classifier/requirements.txt b/lambdas/classifier/requirements.txt
index 4f3a94d..2ac82aa 100644
--- a/lambdas/classifier/requirements.txt
+++ b/lambdas/classifier/requirements.txt
@@ -1,3 +1,5 @@
-awswrangler>=3.9.0
+# awswrangler + pandas/pyarrow/numpy come from the AWS-managed SDK-for-pandas
+# Lambda layer (see pipeline_stack.py) — bundling them here blows the 250 MB
+# unzipped limit. The Haiku fallback uses stdlib urllib, so no anthropic SDK.
+# Only openpyxl (xlsx parsing) is bundled into the function package.
 openpyxl>=3.1.0
-anthropic>=0.40.0
diff --git a/lambdas/slack_post/blockkit.py b/lambdas/slack_post/blockkit.py
index 2823976..3cab1bf 100644
--- a/lambdas/slack_post/blockkit.py
+++ b/lambdas/slack_post/blockkit.py
@@ -282,8 +282,11 @@ def build_daily_summary(
     for cat, count in drill_cats:
         short_label = cat.replace(" Escalation", " Esc.").replace("Awaiting ", "")
         short_label = short_label[:20]  # button text kept concise
+        # action_id must be UNIQUE per message (Slack rejects duplicates), so
+        # qualify it with the category; the interactions handler matches on the
+        # "drill_category:" prefix and reads the filter value from `value`.
         button_elements.append(
-            _button(f"{short_label} ({count})", "drill_category", cat)
+            _button(f"{short_label} ({count})", f"drill_category:{cat}", cat)
         )
 
     # Dashboard link button always present (no action_id — url button).
diff --git a/lambdas/slack_post/interactions.py b/lambdas/slack_post/interactions.py
index 647f3fe..b664fda 100644
--- a/lambdas/slack_post/interactions.py
+++ b/lambdas/slack_post/interactions.py
@@ -49,7 +49,10 @@ def handler(event, context):
         return {"statusCode": 200, "body": ""}
 
     action = (payload.get("actions") or [{}])[0]
-    field = _FILTER_FIELD.get(action.get("action_id"))
+    # action_id is qualified for Slack uniqueness, e.g. "drill_category:Report /
+    # Docs Needed" — match on the prefix before ":"; the filter value is in `value`.
+    action_kind = (action.get("action_id") or "").split(":", 1)[0]
+    field = _FILTER_FIELD.get(action_kind)
     value = action.get("value")
     trigger_id = payload.get("trigger_id")
     if not field or not value or not trigger_id:
diff --git a/tests/test_blockkit.py b/tests/test_blockkit.py
index 02eb1d7..6a2fb1d 100644
--- a/tests/test_blockkit.py
+++ b/tests/test_blockkit.py
@@ -151,6 +151,20 @@ class TestDailySummaryStructure:
         blocks = build_daily_summary(TODAY_SUMMARY, YESTERDAY_SUMMARY, DASHBOARD_URL)
         assert any(b.get("type") == "header" for b in blocks)
 
+    def test_action_ids_are_unique(self):
+        # Slack rejects a message with duplicate action_ids across its elements
+        # (regression: all drill buttons once shared action_id "drill_category").
+        blocks = build_daily_summary(TODAY_SUMMARY, YESTERDAY_SUMMARY, DASHBOARD_URL)
+        action_ids = [
+            el["action_id"]
+            for b in blocks
+            for el in b.get("elements", [])
+            if isinstance(el, dict) and "action_id" in el
+        ]
+        assert len(action_ids) == len(set(action_ids)), (
+            f"duplicate action_id(s): {action_ids}"
+        )
+
     def test_header_contains_date(self):
         blocks = build_daily_summary(TODAY_SUMMARY, YESTERDAY_SUMMARY, DASHBOARD_URL)
         header = next(b for b in blocks if b.get("type") == "header")
@@ -191,7 +205,8 @@ class TestDailySummaryStructure:
             if block.get("type") != "actions":
                 continue
             for elem in block.get("elements", []):
-                if elem.get("action_id") == "drill_category":
+                # action_id is qualified for uniqueness: "drill_category:<cat>".
+                if str(elem.get("action_id", "")).startswith("drill_category:"):
                     drill_found = True
                     assert "value" in elem
         assert drill_found, "expected at least one drill_category action button"