Enhancement/test improvements

.github/workflows/ci.yml

@@ -110,10 +110,24 @@ jobs:
         uses: prefix-dev/setup-pixi@ba3bb36eb2066252b2363392b7739741bb777659 # v0.8.1
         with:
           environments: ${{ matrix.environment }}
+      - name: Get cache week key
+        id: cache-week
+        run: echo "week=$(date +%Y-W%V)" >> $GITHUB_OUTPUT
+        shell: bash
+      - name: Cache test index
+        uses: actions/cache@v4
+        with:
+          path: ~/.holoviz-mcp-test
+          key: holoviz-mcp-test-index-${{ runner.os }}-${{ hashFiles('src/holoviz_mcp/config/config.yaml', 'src/holoviz_mcp/holoviz_mcp/data.py') }}-${{ steps.cache-week.outputs.week }}
+          restore-keys: |
+            holoviz-mcp-test-index-${{ runner.os }}-${{ hashFiles('src/holoviz_mcp/config/config.yaml', 'src/holoviz_mcp/holoviz_mcp/data.py') }}-
+            holoviz-mcp-test-index-${{ runner.os }}-
       - name: Install repository
         run: pixi run -e ${{ matrix.environment }} postinstall
-      - name: Run pytest
-        run: pixi run -e ${{ matrix.environment }} test-coverage --color=yes
+      - name: Run unit tests (fast)
+        run: pixi run -e ${{ matrix.environment }} pytest tests/ --ignore=tests/docs_mcp --color=yes
+      - name: Run docs integration tests (slow, uses cached index)
+        run: pixi run -e ${{ matrix.environment }} pytest tests/docs_mcp --color=yes
 
   pytest_ui:
     name: ui:${{ matrix.environment }}:${{ matrix.os }}

AGENTS.md

@@ -11,7 +11,7 @@ A README for AI coding agents to effectively contribute to the HoloViz MCP proje
 **Language**: Python 3.11+
 **Build System**: Pixi (conda-based) + Hatchling (PEP 517)
 **Framework**: FastMCP (Python MCP SDK)
-**Key Dependencies**: ChromaDB, sentence-transformers, Panel, PyTorch
+**Key Dependencies**: ChromaDB, sentence-transformers, Panel
 
 ### Project Statistics
 - **Source Files**: ~20 Python modules in `src/holoviz_mcp/`
@@ -564,7 +564,7 @@ First-time indexing takes 5-10 minutes:
 
 ### Memory Usage
 
-Heavy dependencies (PyTorch, sentence-transformers):
+Heavy dependencies (sentence-transformers, ChromaDB):
 - **Development**: ~2-4 GB RAM
 - **Running server**: ~1-2 GB RAM
 - **Indexing docs**: ~3-5 GB RAM (temporary spike)

pyproject.toml

@@ -176,6 +176,9 @@ filterwarnings = [
     "ignore::UserWarning",
 ]
 testpaths = ["tests"]
+markers = [
+    "integration: marks tests that require the documentation index (slow)",
+]
 
 [tool.codespell]
 ignore-words-list = "acn"

tests/docs_mcp/conftest.py

@@ -0,0 +1,108 @@
+"""Test fixtures for documentation MCP tests.
+
+Configures a minimal 3-repo test index (panel, hvplot, panel-material-ui)
+instead of the full production index, reducing test time from 5-15 minutes
+to 1-3 minutes. Uses a fixed path (~/.holoviz-mcp-test/) for CI caching.
+"""
+
+import asyncio
+import logging
+import os
+from pathlib import Path
+
+import pytest
+
+from holoviz_mcp.config.loader import ConfigLoader
+from holoviz_mcp.config.models import HoloVizMCPConfig
+
+logger = logging.getLogger(__name__)
+
+# Fixed test directory (not tmp_path) so CI can cache the index across runs.
+# Override with HOLOVIZ_MCP_TEST_DIR for custom locations or concurrent sessions.
+TEST_DATA_DIR = Path(os.environ["HOLOVIZ_MCP_TEST_DIR"]) if "HOLOVIZ_MCP_TEST_DIR" in os.environ else Path.home() / ".holoviz-mcp-test"
+
+# Only these projects are needed by test assertions
+TEST_PROJECTS = ("panel", "hvplot", "panel-material-ui")
+
+
+def _build_test_config():
+    """Build a HoloVizMCPConfig with only 3 repos and test-specific paths."""
+    # Load ONLY the package default config — skip user config entirely so
+    # user-configured repos don't leak into the test index.
+    env = HoloVizMCPConfig(user_dir=TEST_DATA_DIR / ".no_user_config")
+    loader = ConfigLoader(config=env)
+    default_config = loader.load_config()
+
+    # Validate that all expected test repos exist in the default config
+    missing = set(TEST_PROJECTS) - set(default_config.docs.repositories)
+    if missing:
+        raise RuntimeError(f"TEST_PROJECTS not found in default config: {missing}")
+    test_repos = {name: default_config.docs.repositories[name] for name in TEST_PROJECTS}
+
+    test_docs = default_config.docs.model_copy(update={"repositories": test_repos})
+    test_server = default_config.server.model_copy(update={"vector_db_path": TEST_DATA_DIR / "vector_db" / "chroma"})
+
+    return default_config.model_copy(
+        update={
+            "docs": test_docs,
+            "server": test_server,
+            "user_dir": TEST_DATA_DIR,
+            "repos_dir": TEST_DATA_DIR / "repos",
+        }
+    )
+
+
+@pytest.fixture(scope="package", autouse=True)
+def docs_test_config():
+    """Patch get_config() to use a minimal 3-repo test config for the session.
+
+    This fixture:
+    1. Builds a config with only panel, hvplot, panel-material-ui
+    2. Redirects all data paths to ~/.holoviz-mcp-test/
+    3. Resets the server's _indexer singleton so it picks up the test config
+    4. Pre-builds the index if it doesn't exist (avoids a deadlock in
+       ensure_indexed when called from within db_lock in search/list_projects)
+
+    Since ~/.holoviz-mcp-test/ persists between runs, subsequent runs skip indexing.
+    """
+    import holoviz_mcp.config.loader as loader_module
+    import holoviz_mcp.holoviz_mcp.server as server_module
+
+    config = _build_test_config()
+
+    # Pre-load the config into the loader (bypasses file loading and env overrides)
+    test_loader = ConfigLoader()
+    test_loader._loaded_config = config
+
+    # Save originals
+    original_loader = loader_module._config_loader
+    original_indexer = server_module._indexer
+
+    # Patch global config loader and reset indexer singleton
+    loader_module._config_loader = test_loader
+    server_module._indexer = None
+
+    logger.info(
+        "Docs test config: %d repos (%s), vector_db=%s",
+        len(config.docs.repositories),
+        ", ".join(config.docs.repositories.keys()),
+        config.server.vector_db_path,
+    )
+
+    # Pre-build the index if it doesn't exist.  This MUST happen here (outside
+    # any db_lock) because the MCP tools call ensure_indexed() from within
+    # db_lock, and index_documentation() also acquires db_lock — causing a
+    # deadlock when the index is empty.  By building here, ensure_indexed()
+    # always finds is_indexed()==True and skips the build.
+    indexer = server_module.get_indexer()
+    if not indexer.is_indexed():
+        logger.info("Building test index for %s ...", ", ".join(TEST_PROJECTS))
+        asyncio.run(indexer.ensure_indexed())
+        # Reset the db_lock so tests create a fresh one in their event loop
+        indexer._db_lock = None
+
+    yield config
+
+    # Restore originals
+    loader_module._config_loader = original_loader
+    server_module._indexer = original_indexer

tests/docs_mcp/test_docs_mcp.py

@@ -19,6 +19,7 @@ async def test_skills_resource():
         assert result.data
 
 
+@pytest.mark.integration
 @pytest.mark.skip(reason="this test is very slow")
 @pytest.mark.asyncio
 async def test_update_index():
@@ -29,6 +30,7 @@ async def test_update_index():
         assert result.data
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_list_projects():
     """Test that all projects are listed correctly."""
@@ -40,6 +42,7 @@ async def test_list_projects():
     assert "panel" in result.data
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_semantic_search():
     """Test the search tool."""
@@ -63,6 +66,7 @@ async def test_semantic_search():
             assert "content" in document
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_search_by_project():
     """Test the search tool with project filtering."""
@@ -78,6 +82,7 @@ async def test_search_by_project():
             assert document["project"] == "hvplot"
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_search_with_custom_max_results():
     """Test the search tool with custom max_results parameter."""
@@ -95,6 +100,7 @@ async def test_search_with_custom_max_results():
             assert document["project"] == "panel"
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_search_without_content():
     """Test the search tool with content=False for metadata only."""
@@ -116,6 +122,7 @@ async def test_search_without_content():
             assert document.get("content") is None
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_search_material_ui_specific():
     """Test the search tool with Material UI specific query."""
@@ -131,6 +138,7 @@ async def test_search_material_ui_specific():
             assert document["project"] == "panel-material-ui"
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_search_empty_query():
     """Test the search tool with edge cases."""
@@ -142,6 +150,7 @@ async def test_search_empty_query():
         assert isinstance(result.data, list)
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_search_invalid_project():
     """Test the search tool with invalid project name."""
@@ -154,6 +163,7 @@ async def test_search_invalid_project():
         assert len(result.data) == 0
 
 
+@pytest.mark.integration
 @pytest.mark.asyncio
 async def test_search_with_project_filter():
     """Test the search tool with project filtering."""

tests/docs_mcp/test_docs_mcp_reference_guide.py

@@ -9,6 +9,8 @@
 
 from holoviz_mcp.holoviz_mcp.server import mcp
 
+pytestmark = pytest.mark.integration
+
 
 @pytest.mark.asyncio
 async def test_get_reference_guide_button_no_project():

tests/docs_mcp/test_tabulator_truncation.py

@@ -14,6 +14,8 @@
 from holoviz_mcp.holoviz_mcp.data import DocumentationIndexer
 from holoviz_mcp.holoviz_mcp.data import truncate_content
 
+pytestmark = pytest.mark.integration
+
 
 class TestTabulatorTruncation:
     """Test suite for Tabulator reference guide context-aware truncation."""