feat(mcp): pre-execution lint scanner and namespace pre-injection for catalog code

src/tallyman_core/summary_stats.py

@@ -58,6 +58,16 @@
     "isinstance",
     "issubclass",
     "type",
+    # Exception types — needed for try/except Exception: in compute()
+    "Exception",
+    "ValueError",
+    "TypeError",
+    "KeyError",
+    "AttributeError",
+    "IndexError",
+    "RuntimeError",
+    "StopIteration",
+    "NotImplementedError",
 )
 _NAME_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
 

src/tallyman_mcp/server.py

@@ -203,31 +203,36 @@ def catalog_run(code: str, prompt: str = "") -> dict:
     """Execute a xorq expression script and persist it as a catalog entry.
 
     The provided code MUST bind a top-level variable named `expr` to a xorq/ibis
-    expression. Imports allowed; all imports happen in a fresh module scope.
+    expression. The four canonical names are PRE-INJECTED — no import statement
+    needed for basic usage:
 
-    PREFERRED pattern:
+        ibis          # xorq.vendor.ibis — ibis._, ibis.cases, ibis.window,
+                      #   ibis.literal, ibis.coalesce, ibis.desc, ibis.schema
+        xo            # xorq.api — xo.memtable, xo.deferred_read_parquet, xo.connect
+        from_catalog  # read a named catalog entry or content hash
+        from_project  # read a raw file under <project>/data/
 
-        from tallyman_xorq.io import from_project, from_catalog
-        import xorq.vendor.ibis as ibis
-        t = from_catalog("alias_name")            # named catalog entry (alias or hash)
+    MINIMAL PATTERN (no imports needed):
+
+        t = from_catalog("alias_name")            # named entry
         t = from_project("file.parquet")          # raw file under <project>/data/
         expr = t.filter(t.col > 0).group_by("region").aggregate(n=t.count())
+        expr = t.mutate(pk=ibis._.a + "_" + ibis._.b)   # ibis._ available pre-injected
+
+    Explicit `import xorq.api as xo` or `import xorq.vendor.ibis as ibis` still work
+    and are fine for clarity — they just overwrite the pre-injected binding with the
+    same module.
+
+    FORBIDDEN IMPORTS — the linter rejects these before any code runs:
+
+        import ibis           # imports system ibis (no datafusion backend); use
+                              #   the pre-injected ibis or `import xorq.vendor.ibis as ibis`
+        from ibis import X    # same problem — use xorq.vendor.ibis
+        import xorq as xo     # bare xorq has no API methods; use `import xorq.api as xo`
+        import xorq           # same — use `import xorq.api as xo`
 
-    THREE NAMESPACES — mixing these up is the #1 build failure:
-
-        import xorq.api as xo            # backends + deferred reads:
-                                         # xo.memtable, xo.deferred_read_parquet, xo.connect
-        import xorq.vendor.ibis as ibis  # the expression API: ibis._, ibis.cases,
-                                         # ibis.window, ibis.literal, ibis.coalesce, ibis.desc
-        from tallyman_xorq.io import from_project, from_catalog   # reading data
-
-        t.mutate(pk=ibis._.a + "_" + ibis._.b)   # deferred string concat via ibis._
-        # NEVER `import xorq` then `xorq.<fn>`: bare xorq.read_parquet / xorq.memtable /
-        #   xorq._ all raise AttributeError. The api module is `xo` (xorq.api).
-        # NEVER bare `import ibis` / `from ibis ...`: it builds but fails at save time
-        #   with a vendored-Expr class error. Always `import xorq.vendor.ibis as ibis`.
-        # Math is a COLUMN METHOD: col.sin(), col.log(), col.sqrt() — not ibis.sin(col).
-        # Read data only via from_project / from_catalog (no xo.read_parquet / ibis.read_parquet).
+    Math is a COLUMN METHOD: col.sin(), col.log(), col.sqrt() — not ibis.sin(col).
+    Read data only via from_project / from_catalog (no xo.read_parquet / ibis.read_parquet).
 
     ONLY BACKEND — xorq's built-in datafusion; there is NO duckdb. Do not use
     `ibis.duckdb`, a duckdb connection, `.sql()`, or `con.register()`. Build

src/tallyman_xorq/__init__.py

@@ -2,6 +2,7 @@
     BuildError,
     BuildResult,
     build_and_persist,
+    lint_catalog_code,
     list_entries,
     load_entry,
     read_prompts,
@@ -47,6 +48,7 @@
     "key_diff",
     "key_diff_polars",
     "key_diff_xorq",
+    "lint_catalog_code",
     "list_entries",
     "load_entry",
     "project_path",

src/tallyman_xorq/build.py

@@ -1,5 +1,6 @@
 from __future__ import annotations
 
+import ast
 import importlib.util
 import json
 import re
@@ -73,6 +74,51 @@ def _user_imports_bare_ibis(code: str) -> bool:
     return False
 
 
+def lint_catalog_code(code: str) -> list[str]:
+    """AST-based pre-execution scan. Returns actionable error strings, empty = clean.
+
+    Catches import mistakes that cause cryptic runtime errors. Runs before any
+    code is executed so the model gets a clear rejection on the first bad call
+    rather than after exec blows up with an AttributeError or IbisError.
+    """
+    try:
+        tree = ast.parse(code)
+    except SyntaxError as e:
+        return [f"syntax error: {e}"]
+
+    errors: list[str] = []
+    for node in ast.walk(tree):
+        if isinstance(node, ast.Import):
+            for alias in node.names:
+                if alias.name == "ibis":
+                    errors.append(
+                        "`import ibis` imports the system ibis package which has no datafusion "
+                        "backend here and will shadow the pre-injected xorq.vendor.ibis. "
+                        "ibis is already available without any import — just use `ibis.window(...)`, "
+                        "`ibis.literal(...)`, `ibis._` etc. directly. If you need an explicit import "
+                        "for clarity, write `import xorq.vendor.ibis as ibis`."
+                    )
+                elif alias.name == "xorq":
+                    stmt = f"`import xorq as {alias.asname}`" if alias.asname else "`import xorq`"
+                    errors.append(
+                        f"{stmt} — the bare `xorq` module has no API methods "
+                        "(connect, memtable, deferred_read_parquet live on `xorq.api`). "
+                        "Use `import xorq.api as xo`, or just use the pre-injected `xo` "
+                        "directly without any import statement."
+                    )
+        elif isinstance(node, ast.ImportFrom):
+            module = node.module or ""
+            if module == "ibis" or module.startswith("ibis."):
+                names = ", ".join(a.name for a in node.names)
+                errors.append(
+                    f"`from {module} import {names}` imports from the system ibis package. "
+                    "Use `import xorq.vendor.ibis as ibis` instead, or just use the pre-injected "
+                    "`ibis` without any import statement."
+                )
+
+    return errors
+
+
 # Read helpers the model reaches for on the wrong namespace. The project-aware
 # way is from_project/from_catalog; xo.deferred_read_parquet handles a raw path.
 _READ_FNS = frozenset(
@@ -210,6 +256,27 @@ def _ibis_import_hint(exc_msg: str, code: str = "") -> str:
     return "\n\nHint: " + " ".join(hints)
 
 
+def _inject_prelude(module: object) -> None:
+    """Pre-inject the four canonical names into a module before exec_module runs.
+
+    ibis, xo, from_catalog, from_project are available without any import
+    statement. Explicit `import xorq.api as xo` / `import xorq.vendor.ibis as ibis`
+    still work — they just overwrite the same module object.
+    """
+    try:
+        import xorq.api as _xo  # noqa: PLC0415
+        import xorq.vendor.ibis as _ibis  # noqa: PLC0415
+
+        from tallyman_xorq.io import from_catalog as _from_catalog  # noqa: PLC0415
+        from tallyman_xorq.io import from_project as _from_project  # noqa: PLC0415
+    except ImportError:
+        return
+    module.ibis = _ibis  # type: ignore[attr-defined]
+    module.xo = _xo  # type: ignore[attr-defined]
+    module.from_catalog = _from_catalog  # type: ignore[attr-defined]
+    module.from_project = _from_project  # type: ignore[attr-defined]
+
+
 def _import_script(code: str) -> tuple[object, Path]:
     """Write code to a temp file, import it, return (module, temp_path).
 
@@ -221,6 +288,7 @@ def _import_script(code: str) -> tuple[object, Path]:
     if spec is None or spec.loader is None:
         raise BuildError(f"Could not load module spec from {tmp}")
     module = importlib.util.module_from_spec(spec)
+    _inject_prelude(module)
     sys.modules[spec.name] = module
     try:
         spec.loader.exec_module(module)
@@ -241,6 +309,12 @@ def build_and_persist(
     The user code must bind a variable named `expr_name` (default "expr") to an
     ibis/xorq expression. Imports happen in a fresh module scope.
     """
+    lint_errors = lint_catalog_code(code)
+    if lint_errors:
+        raise BuildError(
+            "pre-execution lint:\n" + "\n".join(f"  - {e}" for e in lint_errors)
+        )
+
     from xorq.ibis_yaml.compiler import build_expr, load_expr
 
     from tallyman_core.paths import compute_cache_dir

tests/test_build.py

@@ -105,7 +105,87 @@ def test_build_no_hint_on_unrelated_error(project: str, orders_parquet: Path):
 # hint only caught bare `import ibis`, the Expr class mismatch, and `xorq._`.
 # ---------------------------------------------------------------------------
 
-from tallyman_xorq.build import _ibis_import_hint  # noqa: E402
+from tallyman_xorq.build import _ibis_import_hint, lint_catalog_code  # noqa: E402
+
+# ---------------------------------------------------------------------------
+# lint_catalog_code — pre-execution static scanner
+# ---------------------------------------------------------------------------
+
+
+def test_lint_rejects_import_ibis():
+    errors = lint_catalog_code("import ibis\nexpr = ibis.table('x')")
+    assert errors
+    assert any("xorq.vendor.ibis" in e for e in errors)
+
+
+def test_lint_rejects_from_ibis_import():
+    errors = lint_catalog_code("from ibis import window\nexpr = None")
+    assert errors
+    assert any("xorq.vendor.ibis" in e for e in errors)
+
+
+def test_lint_rejects_from_ibis_submodule():
+    errors = lint_catalog_code("from ibis.expr.types import Column\nexpr = None")
+    assert errors
+
+
+def test_lint_rejects_bare_import_xorq():
+    errors = lint_catalog_code("import xorq\nexpr = xorq.connect()")
+    assert errors
+    assert any("xorq.api" in e for e in errors)
+
+
+def test_lint_rejects_import_xorq_as_xo():
+    errors = lint_catalog_code("import xorq as xo\nexpr = xo.connect()")
+    assert errors
+    assert any("xorq.api" in e for e in errors)
+
+
+def test_lint_accepts_correct_imports():
+    code = """
+import xorq.api as xo
+import xorq.vendor.ibis as ibis
+from tallyman_xorq.io import from_catalog, from_project
+t = from_catalog("foo")
+expr = t.filter(t.x > 0)
+"""
+    assert lint_catalog_code(code) == []
+
+
+def test_lint_accepts_no_imports():
+    code = "t = from_catalog('foo')\nexpr = t.filter(t.x > 0)"
+    assert lint_catalog_code(code) == []
+
+
+def test_lint_returns_syntax_error():
+    errors = lint_catalog_code("def broken(\nexpr = None")
+    assert errors
+    assert any("syntax" in e for e in errors)
+
+
+def test_build_lint_fires_before_exec(project: str):
+    # `import ibis` is caught by the linter — no code should execute.
+    code = "import ibis\nexpr = ibis.table('x', schema={'a': 'int64'})"
+    with pytest.raises(BuildError, match="pre-execution lint"):
+        build_and_persist(project, code)
+
+
+def test_preinjected_names_available(project: str, orders_parquet: Path):
+    # ibis, xo, from_project are pre-injected — no import statements needed.
+    code = f"""
+t = xo.deferred_read_parquet({str(orders_parquet)!r})
+expr = t.group_by("region").aggregate(
+    total=t.price.sum(),
+    label=ibis.literal("ok"),
+)
+"""
+    res = build_and_persist(project, code)
+    assert res.row_count == 4
+
+
+# ---------------------------------------------------------------------------
+# Namespace / column-name hints
+# ---------------------------------------------------------------------------
 
 
 def test_hint_bare_xorq_attribute_points_to_api():