fix(mcp): split comma-separated tags in search_notes tags param

src/basic_memory/mcp/tools/search.py

@@ -11,7 +11,7 @@
 from pydantic import AliasChoices, BeforeValidator, Field
 
 from basic_memory.config import ConfigManager, has_cloud_credentials
-from basic_memory.utils import build_canonical_permalink, coerce_dict, coerce_list
+from basic_memory.utils import build_canonical_permalink, coerce_dict, coerce_list, parse_tags
 from basic_memory.mcp.async_client import (
     _explicit_routing,
     _force_local_mode,
@@ -676,9 +676,13 @@ async def search_notes(
         Dict[str, Any] | None,
         BeforeValidator(coerce_dict),
     ] = None,
+    # parse_tags, not coerce_list: tags="a,b" must split into ["a", "b"] to match the
+    # tag: query shorthand below and write_note's documented tags convention (#910).
+    # coerce_list would wrap the comma string as the single literal tag ["a,b"],
+    # which matches nothing.
     tags: Annotated[
         List[str] | None,
-        BeforeValidator(coerce_list),
+        BeforeValidator(parse_tags),
     ] = None,
     status: Optional[str] = None,
     min_similarity: Annotated[
@@ -795,7 +799,9 @@ async def search_notes(
                    observations whose category matches exactly.
         after_date: Optional date filter for recent content (e.g., "1 week", "2d", "2024-01-01")
         metadata_filters: Optional structured frontmatter filters (e.g., {"status": "in-progress"})
-        tags: Optional tag filter (frontmatter tags); shorthand for metadata_filters["tags"]
+        tags: Optional tag filter (frontmatter tags); shorthand for metadata_filters["tags"].
+              Accepts a list (["a", "b"]) or a comma-separated string ("a,b"), matching the
+              write_note tags convention and the tag: query shorthand.
         status: Optional status filter (frontmatter status); shorthand for metadata_filters["status"]
         min_similarity: Optional float to override the global semantic_min_similarity threshold
                        for this query. E.g., 0.0 to see all vector results, or 0.8 for high precision.

tests/mcp/test_tool_search.py

@@ -1,11 +1,15 @@
 """Tests for search MCP tools."""
 
+import inspect
+
 import pytest
 from contextlib import asynccontextmanager
 from datetime import datetime, timedelta
 from types import SimpleNamespace
 from typing import cast
 
+from pydantic import TypeAdapter
+
 from basic_memory.mcp.tools import write_note
 from basic_memory.mcp.tools.search import (
     search_notes,
@@ -1622,6 +1626,74 @@ async def fake_resolve(client, query, project, context):
     assert captured_payload["text"] == "authentication"
 
 
+# --- Tests for comma-separated tags parameter (#910) ----------------------------
+
+
+def test_search_notes_tags_annotation_splits_comma_strings():
+    """The tags parameter annotation must parse every documented input form (#910).
+
+    Direct function calls bypass the BeforeValidator, so validate through the same
+    Annotated metadata pydantic applies on the MCP path. coerce_list wrapped a bare
+    comma string as the single literal tag ["a,b"]; parse_tags splits it like the
+    tag: query shorthand and write_note's tags convention.
+    """
+    annotation = inspect.signature(search_notes).parameters["tags"].annotation
+    adapter = TypeAdapter(annotation)
+
+    real_list = adapter.validate_python(["a", "b"])
+    comma_string = adapter.validate_python("a,b")
+    json_string = adapter.validate_python('["a", "b"]')
+    single_string = adapter.validate_python("a")
+
+    assert real_list == ["a", "b"]
+    # The comma string and the real list must behave identically (the #910 bug).
+    assert comma_string == real_list
+    assert json_string == real_list
+    assert single_string == ["a"]
+
+
+@pytest.mark.asyncio
+async def test_search_notes_tags_comma_string_filters_via_mcp(mcp, client, test_project):
+    """tags="alpha,beta" through the real MCP layer must match like a real list (#910)."""
+    from fastmcp import Client
+
+    async with Client(mcp) as mcp_client:
+        await mcp_client.call_tool(
+            "write_note",
+            {
+                "project": test_project.name,
+                "title": "Tag Split Note",
+                "directory": "test",
+                "content": "# Tag Split Note\nTagSplitToken body",
+                "tags": ["alpha", "beta"],
+            },
+        )
+
+        async def found(tags_value: object) -> bool:
+            result = await mcp_client.call_tool(
+                "search_notes",
+                {
+                    "project": test_project.name,
+                    "query": "TagSplitToken",
+                    "search_type": "text",
+                    "tags": tags_value,
+                },
+            )
+            return "Tag Split Note" in result.content[0].text
+
+        as_list = await found(["alpha", "beta"])
+        as_comma_string = await found("alpha,beta")
+        as_json_string = await found('["alpha", "beta"]')
+        as_single_string = await found("alpha")
+
+        assert as_list, "real-list tags must match (sanity)"
+        assert as_comma_string == as_list, "comma string must behave like the real list"
+        assert as_json_string == as_list
+        assert as_single_string == as_list
+        # Negative control: the filter is actually applied, not silently dropped.
+        assert not await found("gamma")
+
+
 # --- Tests for text output format (#641) -----------------------------------