fix: MCP server loads embedding config from .java-codebase-rag.yml (#238)

The MCP server now reads embedding.model and embedding.device from
.java-codebase-rag.yml at startup and applies them to os.environ,
matching the behavior of CLI commands (init, reprocess, etc.).

Previously, the MCP server ignored YAML config and always fell back to
the HuggingFace hub default, causing search tool failures when offline or
behind corporate proxies.

Changes:
- server.py: Call resolve_operator_config() in main() and apply to os.environ
- tests: Add tests for MCP server YAML config loading and precedence

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

Author: HumanBean17

server.py

@@ -16,7 +16,7 @@
     emit_lance_cocoindex_finish,
     emit_lance_cocoindex_start,
 )
-from java_codebase_rag.config import emit_legacy_env_hints_if_present, resolved_sbert_model_for_process_env
+from java_codebase_rag.config import emit_legacy_env_hints_if_present, resolved_sbert_model_for_process_env, resolve_operator_config
 from kuzu_queries import KuzuGraph, resolve_kuzu_path
 from mcp.server.fastmcp import FastMCP
 from pydantic import BaseModel, Field
@@ -570,6 +570,13 @@ async def resolve(
 
 def main() -> None:
     emit_legacy_env_hints_if_present()
+
+    # Load YAML config and apply embedding settings to environment
+    # This ensures SBERT_MODEL and SBERT_DEVICE from .java-codebase-rag.yml are available
+    # before any tool handler runs (same behavior as CLI path)
+    cfg = resolve_operator_config(source_root=_project_root())
+    cfg.apply_to_os_environ()
+
     asyncio.run(create_mcp_server().run_stdio_async())
 
 

tests/test_java_codebase_rag_cli.py

@@ -909,3 +909,65 @@ def test_cli_unknown_subcommand_returns_error(tmp_path, monkeypatch) -> None:
     env = _base_env(tmp_path)
     proc = _run_cli(["bogus"], env=env)
     assert proc.returncode == 2
+
+
+def test_mcp_server_loads_yaml_config_at_startup(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """MCP server main() loads YAML config and applies to os.environ (issue #238).
+
+    Verifies that main() calls resolve_operator_config with the correct source_root
+    and applies the result to os.environ. Uses mocks to avoid loading real models
+    or leaking env state (e.g. SBERT_DEVICE=cuda) to subsequent tests.
+    """
+    import server as server_mod
+    from unittest.mock import MagicMock
+
+    fake_cfg = MagicMock()
+    fake_cfg.apply_to_os_environ = MagicMock()
+
+    monkeypatch.setenv("JAVA_CODEBASE_RAG_SOURCE_ROOT", str(tmp_path))
+    monkeypatch.setattr(server_mod, "resolve_operator_config", MagicMock(return_value=fake_cfg))
+
+    def fake_asyncio_run(awaitable, *, debug=None):
+        return None
+
+    monkeypatch.setattr("asyncio.run", fake_asyncio_run)
+
+    server_mod.main()
+
+    # resolve_operator_config should have been called with the project root
+    server_mod.resolve_operator_config.assert_called_once_with(source_root=server_mod._project_root())
+    # apply_to_os_environ should have been called to set env vars
+    fake_cfg.apply_to_os_environ.assert_called_once()
+
+
+def test_mcp_server_yaml_config_precedence_env_over_yaml(
+    tmp_path: Path, monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    """MCP server passes _project_root() to resolve_operator_config (issue #238).
+
+    Precedence (env > YAML > default) is already tested by
+    test_embedding_model_precedence_cli_over_env_over_yaml_over_default.
+    This test verifies that main() delegates to resolve_operator_config
+    with the correct source root, which handles precedence internally.
+    """
+    import server as server_mod
+    from unittest.mock import MagicMock, call
+
+    fake_cfg = MagicMock()
+    fake_cfg.apply_to_os_environ = MagicMock()
+
+    # Set source root so _project_root() returns it
+    monkeypatch.setenv("JAVA_CODEBASE_RAG_SOURCE_ROOT", str(tmp_path))
+    monkeypatch.setattr(server_mod, "resolve_operator_config", MagicMock(return_value=fake_cfg))
+
+    def fake_asyncio_run(awaitable, *, debug=None):
+        return None
+
+    monkeypatch.setattr("asyncio.run", fake_asyncio_run)
+
+    server_mod.main()
+
+    server_mod.resolve_operator_config.assert_called_once()
+    assert server_mod.resolve_operator_config.call_args.kwargs["source_root"] == server_mod._project_root()