diff --git a/.bazelrc b/.bazelrc
new file mode 100644
index 0000000..04615fa
--- /dev/null
+++ b/.bazelrc
@@ -0,0 +1,3 @@
+# //tests:version_test asserts on the Python interpreter version, so enable
+# toolchains to ensure we get the right one (avoid bazelbuild/bazel#4815).
+build --incompatible_use_python_toolchains
diff --git a/WORKSPACE b/WORKSPACE
index 09df09f..0ddafb5 100644
--- a/WORKSPACE
+++ b/WORKSPACE
@@ -1,5 +1,13 @@
 workspace(name = "subpar")
 
+# These version vars aren't consumed by anything but are good to know for
+# documentation / compatibility purposes.
+#
+# Because of use of `PyInfo` provider.
+MIN_BAZEL_VERSION = "0.23.0"
+# Because tests require --incompatible_use_python_toolchains.
+MIN_BAZEL_VERSION_FOR_TESTS = "0.25.0"
+
 # Used by integration tests
 local_repository(
     name = "test_workspace",
diff --git a/compiler/BUILD b/compiler/BUILD
index 321b692..e69b9f3 100644
--- a/compiler/BUILD
+++ b/compiler/BUILD
@@ -30,7 +30,7 @@ py_library(
 py_binary(
     name = "compiler",
     srcs = ["compiler.py"],
-    default_python_version = "PY2",
+    python_version = "PY2",
     main = "compiler.py",
     srcs_version = "PY2AND3",
     deps = [":compiler_lib"],
diff --git a/compiler/cli.py b/compiler/cli.py
index 083f931..a721c13 100644
--- a/compiler/cli.py
+++ b/compiler/cli.py
@@ -53,20 +53,23 @@ def make_command_line_parser():
         help='Root directory of all relative paths in manifest file.',
         default=os.getcwd())
     parser.add_argument(
-        '--outputpar',
+        '--output_par',
         help='Filename of generated par file.',
         required=True)
     parser.add_argument(
         '--stub_file',
         help='Read imports and interpreter path from the specified stub file',
         required=True)
+    parser.add_argument(
+        '--interpreter',
+        help='Interpreter to use instead of determining it from the stub file')
     # The default timestamp is "Jan 1 1980 00:00:00 utc", which is the
     # earliest time that can be stored in a zip file.
     #
     # "Seconds since Unix epoch" was chosen to be compatible with
     # the SOURCE_DATE_EPOCH standard
     #
-    # Numeric calue is from running this:
+    # Numeric value is from running this:
     #   "date --date='Jan 1 1980 00:00:00 utc' --utc +%s"
     parser.add_argument(
         '--timestamp',
@@ -84,50 +87,73 @@ def make_command_line_parser():
         'directory at the start of execution.',
         type=bool_from_string,
         required=True)
+    parser.add_argument(
+        '--import_root',
+        help='Path to add to sys.path, may be repeated to provide multiple roots.',
+        action='append',
+        default=[],
+        dest='import_roots'),
+    parser.add_argument(
+        '--no_remove',
+        help='Keep extracted files after program finishes if --zip_safe is ' +
+        'enabled.',
+        type=bool_from_string,
+        required=True)
     return parser
 
 
 def parse_stub(stub_filename):
-    """Parse the imports and interpreter path from a py_binary() stub.
+    """Parse interpreter path from a py_binary() stub.
 
     We assume the stub is utf-8 encoded.
 
-    TODO(b/29227737): Remove this once we can access imports from skylark.
+    TODO(bazelbuild/bazel#7805): Remove this once we can access the py_runtime from Starlark.
 
-    Returns (list of relative paths, path to Python interpreter)
+    Returns path to Python interpreter
     """
 
-    # Find the list of import roots
-    imports_regex = re.compile(r'''^  python_imports = '([^']*)'$''')
+    # Find the interpreter
     interpreter_regex = re.compile(r'''^PYTHON_BINARY = '([^']*)'$''')
-    import_roots = None
     interpreter = None
     with io.open(stub_filename, 'rt', encoding='utf8') as stub_file:
         for line in stub_file:
-            importers_match = imports_regex.match(line)
-            if importers_match:
-                import_roots = importers_match.group(1).split(':')
-                # Filter out empty paths
-                import_roots = [x for x in import_roots if x]
             interpreter_match = interpreter_regex.match(line)
             if interpreter_match:
                 interpreter = interpreter_match.group(1)
-    if import_roots is None or not interpreter:
+    if not interpreter:
         raise error.Error('Failed to parse stub file [%s]' % stub_filename)
 
-    # Find the Python interpreter, matching the search logic in
-    # stub_template.txt
+    # Determine the Python interpreter, checking for default toolchain.
+    #
+    # This somewhat mirrors the logic in python_stub_template.txt, but we don't support
+    # relative paths (i.e., in-workspace interpreters). This is because the interpreter
+    # will be used in the .par file's shebang, and putting a relative path in a shebang
+    # is extremely brittle and non-relocatable. (The reason the standard py_binary rule
+    # can use an in-workspace interpreter is that its stub script runs in a separate
+    # process and has a shebang referencing the system interpreter). As a special case,
+    # if the Python target is using the autodetecting Python toolchain, which is
+    # technically an in-workspace runtime, we rewrite it to "/usr/bin/env python[2|3]"
+    # rather than fail.
     if interpreter.startswith('//'):
         raise error.Error('Python interpreter must not be a label [%s]' %
                           stub_filename)
     elif interpreter.startswith('/'):
         pass
+    elif interpreter == 'bazel_tools/tools/python/py3wrapper.sh': # Default toolchain
+        # Replace default toolchain python3 wrapper with default python3 on path
+        interpreter = '/usr/bin/env python3'
+    elif interpreter == 'bazel_tools/tools/python/py2wrapper.sh': # Default toolchain
+        # Replace default toolchain python2 wrapper with default python2 on path
+        interpreter = '/usr/bin/env python2'
     elif '/' in interpreter:
-        pass
+        raise error.Error(
+            'par files require a Python runtime that is ' +
+            'installed on the system, not defined inside the workspace. Use ' +
+            'a `py_runtime` with an absolute path, not a label.')
     else:
         interpreter = '/usr/bin/env %s' % interpreter
 
-    return (import_roots, interpreter)
+    return interpreter
 
 
 def main(argv):
@@ -135,17 +161,21 @@ def main(argv):
     parser = make_command_line_parser()
     args = parser.parse_args(argv[1:])
 
-    # Parse information from stub file that's too hard to compute in Skylark
-    import_roots, interpreter = parse_stub(args.stub_file)
+    # Parse interpreter from stub file that's not available in Starlark
+    interpreter = parse_stub(args.stub_file)
+
+    if args.interpreter:
+        interpreter = args.interpreter
 
     par = python_archive.PythonArchive(
         main_filename=args.main_filename,
-        import_roots=import_roots,
+        import_roots=args.import_roots,
         interpreter=interpreter,
-        output_filename=args.outputpar,
+        output_filename=args.output_par,
         manifest_filename=args.manifest_file,
         manifest_root=args.manifest_root,
         timestamp=args.timestamp,
         zip_safe=args.zip_safe,
+        no_remove=args.no_remove,
     )
     par.create()
diff --git a/compiler/cli_test.py b/compiler/cli_test.py
index 56a6f18..bd375e6 100644
--- a/compiler/cli_test.py
+++ b/compiler/cli_test.py
@@ -35,45 +35,58 @@ def test_make_command_line_parser(self):
         args = parser.parse_args([
             '--manifest_file=bar',
             '--manifest_root=bazz',
-            '--outputpar=baz',
+            '--output_par=baz',
             '--stub_file=quux',
             '--zip_safe=False',
+            '--no_remove=False',
+            '--import_root=root1',
+            '--import_root=root2',
             'foo',
         ])
         self.assertEqual(args.manifest_file, 'bar')
+        self.assertEqual(args.manifest_root, 'bazz')
+        self.assertEqual(args.output_par, 'baz')
+        self.assertEqual(args.stub_file, 'quux')
+        self.assertEqual(args.zip_safe, False)
+        self.assertEqual(args.import_roots, ['root1', 'root2'])
+        self.assertEqual(args.main_filename, 'foo')
+
+    def test_make_command_line_parser_for_interprerter(self):
+        parser = cli.make_command_line_parser()
+        args = parser.parse_args([
+            '--manifest_file=bar',
+            '--manifest_root=bazz',
+            '--output_par=baz',
+            '--stub_file=quux',
+            '--zip_safe=False',
+            '--no_remove=False',
+            '--interpreter=foobar',
+            'foo',
+        ])
+        self.assertEqual(args.interpreter, 'foobar')
 
     def test_stub(self):
         valid_cases = [
-            # Empty list
+            # Absolute path to interpreter
             [b"""
-  python_imports = ''
 PYTHON_BINARY = '/usr/bin/python'
 """,
-             ([], '/usr/bin/python')],
-            # Single import
-            [b"""
-  python_imports = 'myworkspace/spam/eggs'
-PYTHON_BINARY = '/usr/bin/python'
-""",
-             (['myworkspace/spam/eggs'], '/usr/bin/python')],
-            # Multiple imports
+             '/usr/bin/python'],
+            # Search for interpreter on $PATH
             [b"""
-  python_imports = 'myworkspace/spam/eggs:otherworkspace'
-PYTHON_BINARY = '/usr/bin/python'
+PYTHON_BINARY = 'python'
 """,
-             (['myworkspace/spam/eggs', 'otherworkspace'], '/usr/bin/python')],
-            # Relative path to interpreter
+             '/usr/bin/env python'],
+            # Default toolchain wrapped python3 interpreter
             [b"""
-  python_imports = ''
-PYTHON_BINARY = 'mydir/python'
+PYTHON_BINARY = 'bazel_tools/tools/python/py3wrapper.sh'
 """,
-             ([], 'mydir/python')],
-            # Search for interpreter on $PATH
+             '/usr/bin/env python3'],
+            # Default toolchain wrapped python2 interpreter
             [b"""
-  python_imports = ''
-PYTHON_BINARY = 'python'
+PYTHON_BINARY = 'bazel_tools/tools/python/py2wrapper.sh'
 """,
-             ([], '/usr/bin/env python')],
+             '/usr/bin/env python2'],
         ]
         for content, expected in valid_cases:
             with test_utils.temp_file(content) as stub_file:
@@ -81,15 +94,13 @@ def test_stub(self):
                 self.assertEqual(actual, expected)
 
         invalid_cases = [
+            # No interpreter
             b'',
             b'\n\n',
-            # No interpreter
-            b"  python_imports = 'myworkspace/spam/eggs'",
-            # No imports
-            b"PYTHON_BINARY = 'python'\n",
+            # Relative interpreter path
+            b"PYTHON_BINARY = 'mydir/python'",
             # Interpreter is label
             b"""
-  python_imports = ''
 PYTHON_BINARY = '//mypackage:python'
 """,
             ]
diff --git a/compiler/python_archive.py b/compiler/python_archive.py
index 8a119fd..a4a9093 100755
--- a/compiler/python_archive.py
+++ b/compiler/python_archive.py
@@ -29,6 +29,8 @@
 """
 
 from datetime import datetime
+import contextlib
+import errno
 import io
 import logging
 import os
@@ -46,7 +48,8 @@
 _boilerplate_template = """\
 # Boilerplate added by subpar/compiler/python_archive.py
 from %(runtime_package)s import support as _
-_.setup(import_roots=%(import_roots)s, zip_safe=%(zip_safe)s)
+_.setup(import_roots=%(import_roots)s, zip_safe=%(zip_safe)s,
+    no_remove=%(no_remove)s)
 del _
 # End boilerplate
 """
@@ -100,6 +103,7 @@ def __init__(self,
                  output_filename,
                  timestamp,
                  zip_safe,
+                 no_remove,
                  ):
         self.main_filename = main_filename
 
@@ -112,6 +116,7 @@ def __init__(self,
         t = datetime.utcfromtimestamp(timestamp)
         self.timestamp_tuple = t.timetuple()[0:6]
         self.zip_safe = zip_safe
+        self.no_remove = no_remove
 
         self.compression = zipfile.ZIP_DEFLATED
 
@@ -170,6 +175,7 @@ def generate_boilerplate(self, import_roots):
             'runtime_package': _runtime_package,
             'import_roots': str(import_roots),
             'zip_safe': self.zip_safe,
+            'no_remove': self.no_remove,
         }
         return boilerplate_contents.encode('ascii').decode('ascii')
 
@@ -274,7 +280,7 @@ def write_zip_data(self, temp_parfile, stored_resources):
         """
 
         logging.debug('Storing Files...')
-        with zipfile.ZipFile(temp_parfile, 'w', self.compression) as z:
+        with contextlib.closing(zipfile.ZipFile(temp_parfile, 'w', self.compression)) as z:
             items = sorted(stored_resources.items())
             for relative_path, resource in items:
                 assert resource.zipinfo.filename == relative_path
@@ -289,9 +295,14 @@ def create_final_from_temp(self, temp_parfile_name):
 
 
 def remove_if_present(filename):
-    """Delete any existing file"""
-    if os.path.exists(filename):
+    """Delete a file if it exists"""
+    try:
+        # Remove atomically
         os.remove(filename)
+    except OSError as exc:
+        # Ignore failure if file does not exist
+        if exc.errno != errno.ENOENT:
+            raise
 
 
 def fetch_support_file(name, timestamp_tuple):
diff --git a/compiler/python_archive_test.py b/compiler/python_archive_test.py
index 3fd5ec2..05400d7 100644
--- a/compiler/python_archive_test.py
+++ b/compiler/python_archive_test.py
@@ -14,6 +14,7 @@
 
 import os
 import subprocess
+import sys
 import time
 import unittest
 import zipfile
@@ -45,11 +46,12 @@ def setUp(self):
         if not os.path.exists(self.output_dir):
             os.makedirs(self.output_dir)
         self.output_filename = os.path.join(self.output_dir, 'output.par')
-        self.interpreter = '/usr/bin/python2'
+        self.interpreter = sys.executable
         self.import_roots = []
         self.date_time_tuple = (1980, 1, 1, 0, 0, 0)
         self.timestamp = 315532800
         self.zip_safe = True
+        self.no_remove = False
 
     def _construct(self, manifest_filename=None):
         return python_archive.PythonArchive(
@@ -61,6 +63,7 @@ def _construct(self, manifest_filename=None):
             output_filename=self.output_filename,
             timestamp=self.timestamp,
             zip_safe=self.zip_safe,
+            no_remove=self.no_remove,
         )
 
     def test_create_manifest_not_found(self):
diff --git a/docs/BUILD b/docs/BUILD
index f72a6ea..555d460 100644
--- a/docs/BUILD
+++ b/docs/BUILD
@@ -1,9 +1,10 @@
 # To regenerate html docs, run:
-#   ./update_docs.sh
+#   ../update_docs.sh
 
-load("@io_bazel_skydoc//skylark:skylark.bzl", "skylark_doc", "skylark_library")
+load("@bazel_skylib//:bzl_library.bzl", "bzl_library")
+load("@io_bazel_skydoc//skylark:skylark.bzl", "skylark_doc")
 
-skylark_library(
+bzl_library(
     name = "subpar-rules",
     srcs = [
         "@subpar//:debug.bzl",
@@ -16,7 +17,7 @@ skylark_doc(
     format = "markdown",
     overview = True,
     site_root = ".",
-    deps = ["subpar-rules"],
+    srcs = [":subpar-rules"],
 )
 
 skylark_doc(
@@ -24,5 +25,5 @@ skylark_doc(
     format = "html",
     overview = True,
     site_root = ".",
-    deps = ["subpar-rules"],
+    srcs = [":subpar-rules"],
 )
diff --git a/docs/WORKSPACE b/docs/WORKSPACE
index 77a5e97..23bddbd 100644
--- a/docs/WORKSPACE
+++ b/docs/WORKSPACE
@@ -1,28 +1,40 @@
 workspace(name = "subpar_docs")
 
+load("@bazel_tools//tools/build_defs/repo:git.bzl", "git_repository")
+
 local_repository(
     name = "subpar",
     path = "..",
 )
 
-# Dependency of skydoc
-git_repository(
-    name = "io_bazel_rules_sass",
-    remote = "https://github.com/bazelbuild/rules_sass.git",
-    tag = "0.0.3",
-)
-
-load("@io_bazel_rules_sass//sass:sass.bzl", "sass_repositories")
+# Skydoc / stardoc and dependencies.
 
-sass_repositories()
-
-# Used to create documentation
 git_repository(
     name = "io_bazel_skydoc",
     remote = "https://github.com/bazelbuild/skydoc.git",
-    commit = "b36d22cc4436a7a7933e36a111bfc00fd494b9fb",
+    commit = "13063139c7c2bca7c725f382fa1e78bdfe93c887",  # 2019-02-27
 )
 
-load("@io_bazel_skydoc//skylark:skylark.bzl", "skydoc_repositories")
+# This initialization comes from Stardoc's WORKSPACE file as of commit
+# 13063139c7c2bca7c725f382fa1e78bdfe93c887.
 
+load("@io_bazel_skydoc//:setup.bzl", "skydoc_repositories")
 skydoc_repositories()
+
+load("@io_bazel_rules_sass//:package.bzl", "rules_sass_dependencies")
+rules_sass_dependencies()
+
+load("@build_bazel_rules_nodejs//:defs.bzl", "node_repositories")
+node_repositories()
+
+load("@io_bazel_rules_sass//:defs.bzl", "sass_repositories")
+sass_repositories()
+
+# This definition is also needed for (legacy) Skydoc.
+# TODO(#93): Migrate to Stardoc.
+
+git_repository(
+    name = "com_google_protobuf",
+    remote = "https://github.com/protocolbuffers/protobuf.git",
+    commit = "7b28271a61a3da0a37f6fda399b0c4c86464e5b3",  # 2018-11-16
+)
diff --git a/docs/index.html b/docs/index.html
index f46d709..d685f9e 100644
--- a/docs/index.html
+++ b/docs/index.html
@@ -70,7 +70,7 @@ <h2>Rule sets</h2>
   </ul>
 </nav>
 
-<h2><a href="./debug.html">debug Rules</a></h2>
+<h2><a href="./debug.html" id="debug">debug Rules</a></h2>
 
 <h3>Macros</h3>
 <table class="overview-table">
@@ -92,7 +92,7 @@ <h3>Macros</h3>
     </tr>
   </tbody>
 </table>
-<h2><a href="./subpar.html">Build self-contained python executables.</a></h2>
+<h2><a href="./subpar.html" id="subpar">Build self-contained python executables.</a></h2>
 
 <h3>Rules</h3>
 <table class="overview-table">
diff --git a/docs/index.md b/docs/index.md
index a62b873..f4f0c2b 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -10,7 +10,7 @@
   </ul>
 </nav>
 
-<h2><a href="./debug.html">debug Rules</a></h2>
+<h2><a href="./debug.html" id="debug">debug Rules</a></h2>
 
 <h3>Macros</h3>
 <table class="overview-table">
@@ -32,7 +32,7 @@
     </tr>
   </tbody>
 </table>
-<h2><a href="./subpar.html">Build self-contained python executables.</a></h2>
+<h2><a href="./subpar.html" id="subpar">Build self-contained python executables.</a></h2>
 
 <h3>Rules</h3>
 <table class="overview-table">
diff --git a/docs/main.css b/docs/main.css
index f506e12..89f66fe 100755
--- a/docs/main.css
+++ b/docs/main.css
@@ -1,3 +1 @@
-body{background-color:#fafafa}pre,code{font-family:'Liberation Mono', Consolas, Monaco, 'Andale Mono', monospace}pre{background-color:#eee;padding:20px;overflow-x:auto;word-wrap:normal}pre code{overflow-wrap:normal;white-space:pre}code{display:inline-block;font-size:90%;white-space:pre-wrap}.mdl-layout__drawer{background-color:#fff}.mdl-layout__drawer .mdl-layout-title{border-bottom:1px solid #e0e0e0;padding-left:24px}.drawer-nav ul{list-style:none;padding-left:0}.drawer-nav ul li{display:block;padding:0}.drawer-nav ul li ul li a{padding-left:44px;font-weight:400}.drawer-nav ul li a{display:block;flex-shrink:0;padding:15px 0 15px 22px;margin:0;font-weight:600;color:#757575;line-height:1em;text-decoration:none;cursor:pointer}.drawer-nav ul li a:active,.drawer-nav ul li a:hover{background-color:#f0f0f0}.drawer-nav ul li.active a{color:#4caf50;font-weight:500}h1.page-title{font-size:34px;font-weight:400;line-height:40px;margin-bottom:30px;color:#4caf50}p.lead{font-size:20px;line-height:32px}table{border-collapse:collapse;border-spacing:0;background-color:#fff;table-layout:auto}table thead th{background-color:#fafafa;border:1px solid #eee;color:#757575;padding:12px 12px 12px 24px;vertical-align:top}table tbody td{border:1px solid #eee;padding:12px 12px 12px 24px;vertical-align:top}table.params-table{width:100%}table.params-table col.col-param{width:25%}table.params-table col.col-description{width:75%}table.overview-table{width:100%}table.overview-table col.col-name{width:25%}table.overview-table col.col-description{width:75%}table.overview-table td p{margin:0}hr{margin-top:80px;margin-bottom:80px}nav.toc{border-left:5px solid #4caf50;padding-left:20px;margin-bottom:48px}nav.toc h1,nav.toc h2{font-size:15px;line-height:16px;padding-bottom:12px;margin-bottom:0;font-weight:400;color:#757575}nav.toc ul{list-style:none;margin-top:0;padding-left:0}nav.toc ul li{font-size:20px;line-height:40px}nav.toc ul li a{color:#4caf50}.page-content{margin-left:auto;margin-right:auto;padding-top:60px;padding-bottom:60px;width:760px}.page-content a{text-decoration:none}.page-content h1{font-size:34px;font-weight:400;line-height:40px;margin-bottom:30px;color:#4caf50}.page-content h2{font-size:24px;font-weight:400;line-height:32px;margin-bottom:30px;color:#4caf50}.page-content h3{font-size:20px;font-weight:400;line-height:32px;margin-bottom:30px;color:#4caf50}@media (max-width: 768px){.page-content{width:360px}}@media (min-width: 768px){.page-content{width:760px}}@media (min-width: 1476px){.page-content{width:1160px}}.mdl-mini-footer{padding-left:40px}
-
-/*# sourceMappingURL=main.css.map */
\ No newline at end of file
+body{background-color:#fafafa}pre,code{font-family:"Liberation Mono",Consolas,Monaco,"Andale Mono",monospace}pre{background-color:#eee;padding:20px;overflow-x:auto;word-wrap:normal}pre code{overflow-wrap:normal;white-space:pre}code{display:inline-block;font-size:90%;white-space:pre-wrap}.mdl-layout__drawer{background-color:#fff}.mdl-layout__drawer .mdl-layout-title{border-bottom:1px solid #e0e0e0;padding-left:24px}.drawer-nav ul{list-style:none;padding-left:0}.drawer-nav ul li{display:block;padding:0}.drawer-nav ul li ul li a{padding-left:44px;font-weight:400}.drawer-nav ul li a{display:block;flex-shrink:0;padding:15px 0 15px 22px;margin:0;font-weight:600;color:#757575;line-height:1em;text-decoration:none;cursor:pointer}.drawer-nav ul li a:active,.drawer-nav ul li a:hover{background-color:#f0f0f0}.drawer-nav ul li.active a{color:#4caf50;font-weight:500}h1.page-title{font-size:34px;font-weight:400;line-height:40px;margin-bottom:30px;color:#4caf50}p.lead{font-size:20px;line-height:32px}table{border-collapse:collapse;border-spacing:0;background-color:#fff;table-layout:auto}table thead th{background-color:#fafafa;border:1px solid #eee;color:#757575;padding:12px 12px 12px 24px;vertical-align:top}table tbody td{border:1px solid #eee;padding:12px 12px 12px 24px;vertical-align:top}table.params-table{width:100%}table.params-table col.col-param{width:25%}table.params-table col.col-description{width:75%}table.overview-table{width:100%}table.overview-table col.col-name{width:25%}table.overview-table col.col-description{width:75%}table.overview-table td p{margin:0}hr{margin-top:40px;margin-bottom:40px}nav.toc{border-left:5px solid #4caf50;padding-left:20px;margin-bottom:48px}nav.toc h1,nav.toc h2{font-size:15px;line-height:16px;padding-bottom:12px;margin-bottom:0;font-weight:400;color:#757575}nav.toc ul{list-style:none;margin-top:0;padding-left:0}nav.toc ul li{font-size:20px;line-height:40px}nav.toc ul li a{color:#4caf50}.page-content{margin-left:auto;margin-right:auto;padding-top:60px;padding-bottom:60px;width:760px}.page-content a{text-decoration:none}.page-content h1{font-size:34px;font-weight:400;line-height:40px;margin-bottom:30px;color:#4caf50}.page-content h2{font-size:24px;font-weight:400;line-height:32px;margin-bottom:30px;color:#4caf50}.page-content h3{font-size:20px;font-weight:400;line-height:32px;margin-bottom:30px;color:#4caf50}@media(max-width: 768px){.page-content{width:360px}}@media(min-width: 768px){.page-content{width:760px}}@media(min-width: 1476px){.page-content{width:1160px}}.mdl-mini-footer{padding-left:40px}/*# sourceMappingURL=main.css.map */
diff --git a/docs/subpar.html b/docs/subpar.html
index 1009446..d510172 100644
--- a/docs/subpar.html
+++ b/docs/subpar.html
@@ -77,7 +77,7 @@ <h2>Macros</h2>
 
           <h2 id="par_binary">par_binary</h2>
 
-          <pre>par_binary(<a href="#par_binary.name">name</a>)</pre>
+          <pre>par_binary(<a href="#par_binary.name">name</a>, <a href="#par_binary.**kwargs">**kwargs</a>)</pre>
 
           <p>An executable Python program.</p>
 <p>par_binary() is a drop-in replacement for py_binary() that also
@@ -106,13 +106,20 @@ <h3 id="par_binary_args">Attributes</h3>
         <p>A unique name for this rule.</p>
       </td>
     </tr>
+    <tr id="par_binary.**kwargs">
+      <td><code>**kwargs</code></td>
+      <td>
+        <p><code>Unknown; Optional</code></p>
+        
+      </td>
+    </tr>
   </tbody>
 </table>
           <hr>
 
           <h2 id="par_test">par_test</h2>
 
-          <pre>par_test(<a href="#par_test.name">name</a>)</pre>
+          <pre>par_test(<a href="#par_test.name">name</a>, <a href="#par_test.**kwargs">**kwargs</a>)</pre>
 
           <p>An executable Python test.</p>
 <p>Just like par_binary, but for py_test instead of py_binary.  Useful if you
@@ -134,13 +141,20 @@ <h3 id="par_test_args">Attributes</h3>
         <p>A unique name for this rule.</p>
       </td>
     </tr>
+    <tr id="par_test.**kwargs">
+      <td><code>**kwargs</code></td>
+      <td>
+        <p><code>Unknown; Optional</code></p>
+        
+      </td>
+    </tr>
   </tbody>
 </table>
           <hr>
 
           <h2 id="parfile">parfile</h2>
 
-          <pre>parfile(<a href="#parfile.name">name</a>, <a href="#parfile.src">src</a>, <a href="#parfile.compiler">compiler</a>, <a href="#parfile.default_python_version">default_python_version</a>, <a href="#parfile.imports">imports</a>, <a href="#parfile.main">main</a>, <a href="#parfile.zip_safe">zip_safe</a>)</pre>
+          <pre>parfile(<a href="#parfile.name">name</a>, <a href="#parfile.src">src</a>, <a href="#parfile.compiler">compiler</a>, <a href="#parfile.compiler_args">compiler_args</a>, <a href="#parfile.default_python_version">default_python_version</a>, <a href="#parfile.imports">imports</a>, <a href="#parfile.main">main</a>, <a href="#parfile.no_remove">no_remove</a>, <a href="#parfile.zip_safe">zip_safe</a>)</pre>
 
           <p>A self-contained, single-file Python program, with a .par file extension.</p>
 <p>You probably want to use par_binary() instead of this.</p>
@@ -177,6 +191,13 @@ <h3 id="parfile_args">Attributes</h3>
         <p>Internal use only.</p>
       </td>
     </tr>
+    <tr id="parfile.compiler_args">
+      <td><code>compiler_args</code></td>
+      <td>
+        <p><code>List of strings; Optional; Default is []</code></p>
+        
+      </td>
+    </tr>
     <tr id="parfile.default_python_version">
       <td><code>default_python_version</code></td>
       <td>
@@ -202,6 +223,14 @@ <h3 id="parfile_args">Attributes</h3>
 <p>See <a href="http://www.bazel.io/docs/be/python.html#py_binary.main">py_binary.main</a></p>
       </td>
     </tr>
+    <tr id="parfile.no_remove">
+      <td><code>no_remove</code></td>
+      <td>
+        <p><code>Boolean; Optional; Default is False</code></p>
+        <p>Whether to keep the extracted temporary directory after the
+program finishes, if zip_safe is enabled.</p>
+      </td>
+    </tr>
     <tr id="parfile.zip_safe">
       <td><code>zip_safe</code></td>
       <td>
@@ -218,7 +247,7 @@ <h3 id="parfile_args">Attributes</h3>
 
           <h2 id="parfile_test">parfile_test</h2>
 
-          <pre>parfile_test(<a href="#parfile_test.name">name</a>, <a href="#parfile_test.src">src</a>, <a href="#parfile_test.compiler">compiler</a>, <a href="#parfile_test.default_python_version">default_python_version</a>, <a href="#parfile_test.imports">imports</a>, <a href="#parfile_test.main">main</a>, <a href="#parfile_test.zip_safe">zip_safe</a>)</pre>
+          <pre>parfile_test(<a href="#parfile_test.name">name</a>, <a href="#parfile_test.src">src</a>, <a href="#parfile_test.compiler">compiler</a>, <a href="#parfile_test.compiler_args">compiler_args</a>, <a href="#parfile_test.default_python_version">default_python_version</a>, <a href="#parfile_test.imports">imports</a>, <a href="#parfile_test.main">main</a>, <a href="#parfile_test.no_remove">no_remove</a>, <a href="#parfile_test.zip_safe">zip_safe</a>)</pre>
 
           <p>Identical to par_binary, but the rule is marked as being a test.</p>
 <p>You probably want to use par_test() instead of this.</p>
@@ -253,6 +282,13 @@ <h3 id="parfile_test_args">Attributes</h3>
         <p>Internal use only.</p>
       </td>
     </tr>
+    <tr id="parfile_test.compiler_args">
+      <td><code>compiler_args</code></td>
+      <td>
+        <p><code>List of strings; Optional; Default is []</code></p>
+        
+      </td>
+    </tr>
     <tr id="parfile_test.default_python_version">
       <td><code>default_python_version</code></td>
       <td>
@@ -278,6 +314,14 @@ <h3 id="parfile_test_args">Attributes</h3>
 <p>See <a href="http://www.bazel.io/docs/be/python.html#py_binary.main">py_binary.main</a></p>
       </td>
     </tr>
+    <tr id="parfile_test.no_remove">
+      <td><code>no_remove</code></td>
+      <td>
+        <p><code>Boolean; Optional; Default is False</code></p>
+        <p>Whether to keep the extracted temporary directory after the
+program finishes, if zip_safe is enabled.</p>
+      </td>
+    </tr>
     <tr id="parfile_test.zip_safe">
       <td><code>zip_safe</code></td>
       <td>
diff --git a/docs/subpar.md b/docs/subpar.md
index 1d71527..ddde9df 100644
--- a/docs/subpar.md
+++ b/docs/subpar.md
@@ -21,7 +21,7 @@ Documentation generated by Skydoc
 ## par_binary
 
 <pre>
-par_binary(<a href="#par_binary.name">name</a>)
+par_binary(<a href="#par_binary.name">name</a>, <a href="#par_binary.**kwargs">**kwargs</a>)
 </pre>
 
 An executable Python program.
@@ -56,13 +56,20 @@ for arguments and usage.
         <p>A unique name for this rule.</p>
       </td>
     </tr>
+    <tr id="par_binary.**kwargs">
+      <td><code>**kwargs</code></td>
+      <td>
+        <p><code>Unknown; Optional</code></p>
+        
+      </td>
+    </tr>
   </tbody>
 </table>
 <a name="par_test"></a>
 ## par_test
 
 <pre>
-par_test(<a href="#par_test.name">name</a>)
+par_test(<a href="#par_test.name">name</a>, <a href="#par_test.**kwargs">**kwargs</a>)
 </pre>
 
 An executable Python test.
@@ -88,13 +95,20 @@ specifically need to test a module's behaviour when used in a .par binary.
         <p>A unique name for this rule.</p>
       </td>
     </tr>
+    <tr id="par_test.**kwargs">
+      <td><code>**kwargs</code></td>
+      <td>
+        <p><code>Unknown; Optional</code></p>
+        
+      </td>
+    </tr>
   </tbody>
 </table>
 <a name="parfile"></a>
 ## parfile
 
 <pre>
-parfile(<a href="#parfile.name">name</a>, <a href="#parfile.src">src</a>, <a href="#parfile.compiler">compiler</a>, <a href="#parfile.default_python_version">default_python_version</a>, <a href="#parfile.imports">imports</a>, <a href="#parfile.main">main</a>, <a href="#parfile.zip_safe">zip_safe</a>)
+parfile(<a href="#parfile.name">name</a>, <a href="#parfile.src">src</a>, <a href="#parfile.compiler">compiler</a>, <a href="#parfile.compiler_args">compiler_args</a>, <a href="#parfile.default_python_version">default_python_version</a>, <a href="#parfile.imports">imports</a>, <a href="#parfile.main">main</a>, <a href="#parfile.no_remove">no_remove</a>, <a href="#parfile.zip_safe">zip_safe</a>)
 </pre>
 
 A self-contained, single-file Python program, with a .par file extension.
@@ -136,6 +150,13 @@ is a bug, don't use or depend on it.
         <p>Internal use only.</p>
       </td>
     </tr>
+    <tr id="parfile.compiler_args">
+      <td><code>compiler_args</code></td>
+      <td>
+        <p><code>List of strings; Optional; Default is []</code></p>
+        
+      </td>
+    </tr>
     <tr id="parfile.default_python_version">
       <td><code>default_python_version</code></td>
       <td>
@@ -161,6 +182,14 @@ the application.</p>
 <p>See <a href="http://www.bazel.io/docs/be/python.html#py_binary.main">py_binary.main</a></p>
       </td>
     </tr>
+    <tr id="parfile.no_remove">
+      <td><code>no_remove</code></td>
+      <td>
+        <p><code>Boolean; Optional; Default is False</code></p>
+        <p>Whether to keep the extracted temporary directory after the
+program finishes, if zip_safe is enabled.</p>
+      </td>
+    </tr>
     <tr id="parfile.zip_safe">
       <td><code>zip_safe</code></td>
       <td>
@@ -177,7 +206,7 @@ par file executes.</p>
 ## parfile_test
 
 <pre>
-parfile_test(<a href="#parfile_test.name">name</a>, <a href="#parfile_test.src">src</a>, <a href="#parfile_test.compiler">compiler</a>, <a href="#parfile_test.default_python_version">default_python_version</a>, <a href="#parfile_test.imports">imports</a>, <a href="#parfile_test.main">main</a>, <a href="#parfile_test.zip_safe">zip_safe</a>)
+parfile_test(<a href="#parfile_test.name">name</a>, <a href="#parfile_test.src">src</a>, <a href="#parfile_test.compiler">compiler</a>, <a href="#parfile_test.compiler_args">compiler_args</a>, <a href="#parfile_test.default_python_version">default_python_version</a>, <a href="#parfile_test.imports">imports</a>, <a href="#parfile_test.main">main</a>, <a href="#parfile_test.no_remove">no_remove</a>, <a href="#parfile_test.zip_safe">zip_safe</a>)
 </pre>
 
 Identical to par_binary, but the rule is marked as being a test.
@@ -216,6 +245,13 @@ You probably want to use par_test() instead of this.
         <p>Internal use only.</p>
       </td>
     </tr>
+    <tr id="parfile_test.compiler_args">
+      <td><code>compiler_args</code></td>
+      <td>
+        <p><code>List of strings; Optional; Default is []</code></p>
+        
+      </td>
+    </tr>
     <tr id="parfile_test.default_python_version">
       <td><code>default_python_version</code></td>
       <td>
@@ -241,6 +277,14 @@ the application.</p>
 <p>See <a href="http://www.bazel.io/docs/be/python.html#py_binary.main">py_binary.main</a></p>
       </td>
     </tr>
+    <tr id="parfile_test.no_remove">
+      <td><code>no_remove</code></td>
+      <td>
+        <p><code>Boolean; Optional; Default is False</code></p>
+        <p>Whether to keep the extracted temporary directory after the
+program finishes, if zip_safe is enabled.</p>
+      </td>
+    </tr>
     <tr id="parfile_test.zip_safe">
       <td><code>zip_safe</code></td>
       <td>
diff --git a/run_tests.sh b/run_tests.sh
index 1dc7403..acc5410 100755
--- a/run_tests.sh
+++ b/run_tests.sh
@@ -16,8 +16,85 @@
 
 set -euo pipefail
 
+function die {
+  echo "$*" 1>&2
+  exit 1
+}
+
+# This sets up the toolchain hook to run tests for the given version of Python
+# whose interpreter is located at the given absolute path.
+#
+# $1 may be either "PY2" or "PY3". If it is PY2, then the Python 2 runtime is
+# set to the path in $2, and the Python 3 runtime is set to the default value
+# given by $PYTHON3. If $1 is PY3, then $2 is the Python 3 runtime and the
+# Python 2 runtime is given by $PYTHON2.
+#
+# The PYVER constant is also set to $1. It is consumed at loading time by
+# //tests:BUILD.
+#
+# Note that even though tests are only run for one version of Python at a time,
+# we still need to provide both runtimes in the toolchain for the sake of
+# tools. In particular, the par compiler itself requires PY2, even if the par
+# that we are compiling uses PY3.
+function set_toolchain_hook {
+  pyver=$1
+  if [ $pyver == "PY3" ]; then
+    py2_path="$PYTHON2"
+    py3_path="$2"
+  else
+    py2_path="$2"
+    py3_path="$PYTHON3"
+  fi
+
+  cat > toolchain_test_hook.bzl << EOF
+load("@bazel_tools//tools/python:toolchain.bzl", "py_runtime_pair")
+
+PYVER = "$pyver"
+
+def define_toolchain_for_testing():
+    native.py_runtime(
+        name = "py2_runtime",
+        interpreter_path = "$py2_path",
+        python_version = "PY2",
+    )
+
+    native.py_runtime(
+        name = "py3_runtime",
+        interpreter_path = "$py3_path",
+        python_version = "PY3",
+    )
+
+    py_runtime_pair(
+        name = "runtime_pair_for_testing",
+        py2_runtime = ":py2_runtime",
+        py3_runtime = ":py3_runtime",
+        visibility = ["//visibility:public"],
+    )
+
+    native.toolchain(
+        name = "toolchain_for_testing",
+        toolchain = ":runtime_pair_for_testing",
+        toolchain_type = "@bazel_tools//tools/python:toolchain_type",
+        visibility = ["//visibility:public"],
+    )
+EOF
+}
+
+# Clear the toolchain hook back to its original no-op contents.
+#
+# If the test exits abnormally and this function isn't run, we may be left with
+# a modified version of this file in our source tree.
+function clear_toolchain_hook {
+  cat > toolchain_test_hook.bzl << EOF
+PYVER = "PY3"
+
+def define_toolchain_for_testing():
+    pass
+EOF
+}
+
 # Find various tools in environment
-PYTHON2=$(which python)
+PYTHON2=$(which python||true)
 if [ -z "${PYTHON2}" ]; then
   PYTHON2=$(which python2)
 fi
@@ -29,22 +106,30 @@ fi
 #   https://www.python.org/dev/peps/pep-0394/
 if [ -n "${PYTHON2}" ]; then
   VERSION=$("${PYTHON2}" -V 2>&1)
-  if ! expr match "${VERSION}" "Python 2." >/dev/null ; then
+  if ! expr "${VERSION}" : "Python 2." >/dev/null ; then
     PYTHON2=
   fi
 fi
 
-PYTHON3=$(which python3)
+PYTHON3=$(which python3||true)
 
-VIRTUALENV=$(which virtualenv)
+VIRTUALENV=$(which virtualenv) || die "virtualenv not installed"
 VIRTUALENVDIR=$(dirname $0)/.env
 # Virtualenv `activate` needs $PS1 set
 PS1='$ '
 
-# Must have at least one Python interpreter to test
-if [ -z "${PYTHON2}" -a -z "${PYTHON3}" ]; then
-  echo "ERROR: Could not find Python 2 or 3 interpreter on $PATH" 1>&2
-  exit 1
+# Must have both Python interpreters to test.
+if [ -z "${PYTHON2}" ]; then
+  die "Could not find Python 2 on $PATH"
+fi
+if [ -z "${PYTHON3}" ]; then
+  die "Could not find Python 3 on $PATH"
+fi
+
+# Must be able to locate toolchain_test_hook.bzl. This will fail if cwd is not
+# the root of the subpar workspace.
+if [ ! -f "toolchain_test_hook.bzl" ]; then
+  die "Could not locate toolchain_test_hook.bzl (are we in the workspace root?)"
 fi
 
 # Run test matrix
@@ -53,15 +138,19 @@ for PYTHON_INTERPRETER in "${PYTHON2}" "${PYTHON3}"; do
     continue;
   fi
 
+  BAZEL_TEST="bazel test --test_output=errors \
+--incompatible_use_python_toolchains \
+--extra_toolchains=//tests:toolchain_for_testing"
   if [ "${PYTHON_INTERPRETER}" = "${PYTHON3}" ]; then
-    BAZEL_TEST="bazel test --define subpar_test_python_version=3"
+    PYVER="PY3"
   else
-    BAZEL_TEST="bazel test"
+    PYVER="PY2"
   fi
 
   echo "Testing ${PYTHON_INTERPRETER}"
   bazel clean
-  ${BAZEL_TEST} --python_path="${PYTHON_INTERPRETER}" --test_output=errors //...
+  set_toolchain_hook "$PYVER" "$PYTHON_INTERPRETER"
+  ${BAZEL_TEST} //...
 
   if [ -n "${VIRTUALENV}" ]; then
     echo "Testing bare virtualenv"
@@ -72,7 +161,8 @@ for PYTHON_INTERPRETER in "${PYTHON2}" "${PYTHON3}"; do
       "${VIRTUALENVDIR}"
     source "${VIRTUALENVDIR}"/bin/activate
     bazel clean
-    ${BAZEL_TEST} --python_path=$(which python) --test_output=errors //...
+    set_toolchain_hook $PYVER $(which python)
+    ${BAZEL_TEST} //...
     deactivate
 
     for REQUIREMENTS in tests/requirements-test-*.txt; do
@@ -84,8 +174,11 @@ for PYTHON_INTERPRETER in "${PYTHON2}" "${PYTHON3}"; do
       source "${VIRTUALENVDIR}"/bin/activate
       pip install -r "${REQUIREMENTS}"
       bazel clean
-      ${BAZEL_TEST} --python_path=$(which python) --test_output=errors //...
+      set_toolchain_hook $PYVER $(which python)
+      ${BAZEL_TEST} //...
       deactivate
     done
   fi
 done
+
+clear_toolchain_hook
diff --git a/runtime/support.py b/runtime/support.py
index e8ec253..b5bcb84 100644
--- a/runtime/support.py
+++ b/runtime/support.py
@@ -95,7 +95,7 @@ def _find_archive():
     return archive_path
 
 
-def _extract_files(archive_path):
+def _extract_files(archive_path, no_remove):
     """Extract the contents of this .par file to disk.
 
     This creates a temporary directory, and registers an atexit
@@ -107,9 +107,10 @@ def _extract_files(archive_path):
     """
     extract_dir = tempfile.mkdtemp()
 
-    def _extract_files_cleanup():
-        shutil.rmtree(extract_dir, ignore_errors=True)
-    atexit.register(_extract_files_cleanup)
+    if not no_remove:
+        def _extract_files_cleanup():
+            shutil.rmtree(extract_dir, ignore_errors=True)
+        atexit.register(_extract_files_cleanup)
     _log('# extracting %s to %s' % (archive_path, extract_dir))
 
     zip_file = zipfile.ZipFile(archive_path, mode='r')
@@ -282,7 +283,7 @@ def _initialize_import_path(import_roots, import_prefix):
     _log('# adding %s to sys.path' % full_roots)
 
 
-def setup(import_roots, zip_safe):
+def setup(import_roots, zip_safe, no_remove):
     """Initialize subpar run-time support
 
     Args:
@@ -309,7 +310,7 @@ def setup(import_roots, zip_safe):
 
     # Extract files to disk if necessary
     if not zip_safe:
-        extract_dir = _extract_files(archive_path)
+        extract_dir = _extract_files(archive_path, no_remove)
         # sys.path[0] is the name of the executing .par file.  Point
         # it to the extract directory instead, so that Python searches
         # there for imports.
diff --git a/runtime/support_test.py b/runtime/support_test.py
index 565adb1..8dda700 100644
--- a/runtime/support_test.py
+++ b/runtime/support_test.py
@@ -101,7 +101,7 @@ def test__find_archive(self):
 
     def test__extract_files(self):
         # Extract zipfile
-        extract_path = support._extract_files(self.zipfile_name)
+        extract_path = support._extract_files(self.zipfile_name, False)
 
         # Check results
         self.assertTrue(os.path.isdir(extract_path))
@@ -140,7 +140,7 @@ def test_setup(self):
             mock_sys_path[0] = self.zipfile_name
             sys.path = mock_sys_path
             success = support.setup(import_roots=['some_root', 'another_root'],
-                                    zip_safe=True)
+                                    zip_safe=True, no_remove=False)
             self.assertTrue(success)
         finally:
             sys.path = old_sys_path
@@ -165,7 +165,8 @@ def test_setup__extract(self):
             mock_sys_path = list(sys.path)
             mock_sys_path[0] = self.zipfile_name
             sys.path = mock_sys_path
-            success = support.setup(import_roots=['some_root'], zip_safe=False)
+            success = support.setup(import_roots=['some_root'], zip_safe=False,
+                                    no_remove=False)
             self.assertTrue(success)
         finally:
             sys.path = old_sys_path
diff --git a/subpar.bzl b/subpar.bzl
index f12230d..11e3b18 100644
--- a/subpar.bzl
+++ b/subpar.bzl
@@ -14,107 +14,121 @@
 
 """Build self-contained python executables."""
 
-load("//:debug.bzl", "dump")
-
-DEFAULT_COMPILER = '//compiler:compiler.par'
+DEFAULT_COMPILER = "//compiler:compiler.par"
 
 def _parfile_impl(ctx):
     """Implementation of parfile() rule"""
+
     # Find the main entry point
     py_files = ctx.files.main
     if len(py_files) == 0:
-        fail('Expected exactly one .py file, found none', 'main')
+        fail("Expected exactly one .py file, found none", "main")
     elif len(py_files) > 1:
-        fail('Expected exactly one .py file, found these: [%s]' % py_files, 'main')
+        fail("Expected exactly one .py file, found these: [%s]" % py_files, "main")
     main_py_file = py_files[0]
-    if main_py_file not in ctx.attr.src.data_runfiles.files:
-        fail('Main entry point [%s] not listed in srcs' % main_py_file, 'main')
+    if main_py_file not in ctx.attr.src.data_runfiles.files.to_list():
+        fail("Main entry point [%s] not listed in srcs" % main_py_file, "main")
 
     # Find the list of things that must be built before this thing is built
     # TODO: also handle ctx.attr.src.data_runfiles.symlinks
-    inputs = list(ctx.attr.src.default_runfiles.files)
+    inputs = ctx.attr.src.default_runfiles.files.to_list()
 
     # Make a manifest of files to store in the .par file.  The
     # runfiles manifest is not quite right, so we make our own.
     sources_map = {}
+
     # First, add the zero-length __init__.py files
-    for empty in ctx.attr.src.default_runfiles.empty_filenames:
+    for empty in ctx.attr.src.default_runfiles.empty_filenames.to_list():
         stored_path = _prepend_workspace(empty, ctx)
-        local_path = ''
+        local_path = ""
         sources_map[stored_path] = local_path
+
     # Now add the regular (source and generated) files
     for input_file in inputs:
         stored_path = _prepend_workspace(input_file.short_path, ctx)
         local_path = input_file.path
         sources_map[stored_path] = local_path
+
     # Now make a nice sorted list
     sources_lines = []
-    for k,v in sorted(sources_map.items()):
-        sources_lines.append('%s %s' % (k, v))
-    sources_content = '\n'.join(sources_lines) + '\n'
+    for k, v in sorted(sources_map.items()):
+        sources_lines.append("%s %s" % (k, v))
+    sources_content = "\n".join(sources_lines) + "\n"
 
     # Write the list to the manifest file
-    sources_file = ctx.new_file(ctx.label.name + '_SOURCES')
-    ctx.file_action(
-        output=sources_file,
-        content=sources_content,
-        executable=False)
+    sources_file = ctx.actions.declare_file(ctx.label.name + "_SOURCES")
+    ctx.actions.write(
+        output = sources_file,
+        content = sources_content,
+        is_executable = False,
+    )
 
     # Find the list of directories to add to sys.path
-    # TODO(b/29227737): Use 'imports' provider from Bazel
-    stub_file = ctx.attr.src.files_to_run.executable.path
+    import_roots = ctx.attr.src[PyInfo].imports.to_list()
 
     # Inputs to the action, but don't actually get stored in the .par file
     extra_inputs = [
         sources_file,
-        ctx.attr.src.files_to_run.executable,
         ctx.attr.src.files_to_run.runfiles_manifest,
-        ]
+    ]
 
     zip_safe = ctx.attr.zip_safe
+    no_remove = ctx.attr.no_remove
 
     # Assemble command line for .par compiler
-    args = [
-        '--manifest_file', sources_file.path,
-        '--outputpar', ctx.outputs.executable.path,
-        '--stub_file', stub_file,
-        '--zip_safe', str(zip_safe),
-        main_py_file.path,
+    args = ctx.attr.compiler_args + [
+        "--manifest_file",
+        sources_file.path,
+        "--output_par",
+        ctx.outputs.executable.path,
+        "--stub_file",
+        ctx.attr.src.files_to_run.executable.path,
+        "--zip_safe",
+        str(zip_safe),
+        '--no_remove',
+        str(no_remove),
     ]
-    ctx.action(
-        inputs=inputs + extra_inputs,
-        outputs=[ctx.outputs.executable],
-        progress_message='Building par file %s' % ctx.label,
-        executable=ctx.executable.compiler,
-        arguments=args,
-        mnemonic='PythonCompile',
+    for import_root in import_roots:
+        args.extend(['--import_root', import_root])
+    args.append(main_py_file.path)
+
+    # Run compiler
+    ctx.actions.run(
+        inputs = inputs + extra_inputs,
+        tools = [ctx.attr.src.files_to_run.executable],
+        outputs = [ctx.outputs.executable],
+        progress_message = "Building par file %s" % ctx.label,
+        executable = ctx.executable.compiler,
+        arguments = args,
+        mnemonic = "PythonCompile",
+        use_default_shell_env = True,
     )
 
     # .par file itself has no runfiles and no providers
-    return struct()
+    return []
 
 def _prepend_workspace(path, ctx):
     """Given a path, prepend the workspace name as the parent directory"""
+
     # It feels like there should be an easier, less fragile way.
-    if path.startswith('../'):
+    if path.startswith("../"):
         # External workspace, for example
         # '../protobuf/python/google/protobuf/any_pb2.py'
-        stored_path = path[len('../'):]
-    elif path.startswith('external/'):
+        stored_path = path[len("../"):]
+    elif path.startswith("external/"):
         # External workspace, for example
         # 'external/protobuf/python/__init__.py'
-        stored_path = path[len('external/'):]
+        stored_path = path[len("external/"):]
     else:
         # Main workspace, for example 'mypackage/main.py'
-        stored_path = ctx.workspace_name + '/' + path
+        stored_path = ctx.workspace_name + "/" + path
     return stored_path
 
 parfile_attrs = {
     "src": attr.label(mandatory = True),
     "main": attr.label(
         mandatory = True,
-        allow_files = True,
-        single_file = True,
+        allow_single_file = True,
     ),
     "imports": attr.string_list(default = []),
     "default_python_version": attr.string(mandatory = True),
@@ -123,7 +137,9 @@ parfile_attrs = {
         executable = True,
         cfg = "host",
     ),
-    "zip_safe": attr.bool(default=True),
+    "compiler_args": attr.string_list(default = []),
+    "zip_safe": attr.bool(default = True),
+    "no_remove": attr.bool(default = False),
 }
 
 # Rule to create a parfile given a py_binary() as input
@@ -159,6 +175,9 @@ Args:
             extracted to a temporary directory on disk each time the
             par file executes.
 
+  no_remove: Whether to keep the extracted temporary directory after the
+             program finishes, if zip_safe is enabled.
+
 TODO(b/27502830): A directory foo.par.runfiles is also created. This
 is a bug, don't use or depend on it.
 """
@@ -189,25 +208,29 @@ def par_binary(name, **kwargs):
     See [py_binary](http://www.bazel.io/docs/be/python.html#py_binary)
     for arguments and usage.
     """
-    compiler = kwargs.pop('compiler', None)
-    zip_safe = kwargs.pop('zip_safe', True)
-    native.py_binary(name=name, **kwargs)
-
-    main = kwargs.get('main', name + '.py')
-    imports = kwargs.get('imports')
-    default_python_version = kwargs.get('default_python_version', 'PY2')
-    visibility = kwargs.get('visibility')
-    testonly = kwargs.get('testonly', False)
+    compiler = kwargs.pop("compiler", None)
+    compiler_args = kwargs.pop("compiler_args", [])
+    zip_safe = kwargs.pop("zip_safe", True)
+    no_remove = kwargs.pop('no_remove', False)
+    native.py_binary(name = name, **kwargs)
+
+    main = kwargs.get("main", name + ".py")
+    imports = kwargs.get("imports")
+    default_python_version = kwargs.get("default_python_version", "PY2")
+    visibility = kwargs.get("visibility")
+    testonly = kwargs.get("testonly", False)
     parfile(
-        compiler=compiler,
-        default_python_version=default_python_version,
-        imports=imports,
-        main=main,
-        name=name + '.par',
-        src=name,
-        testonly=testonly,
-        visibility=visibility,
-        zip_safe=zip_safe,
+        compiler = compiler,
+        compiler_args = compiler_args,
+        default_python_version = default_python_version,
+        imports = imports,
+        main = main,
+        name = name + ".par",
+        src = name,
+        testonly = testonly,
+        visibility = visibility,
+        zip_safe = zip_safe,
+        no_remove=no_remove,
     )
 
 def par_test(name, **kwargs):
@@ -216,23 +239,25 @@ def par_test(name, **kwargs):
     Just like par_binary, but for py_test instead of py_binary.  Useful if you
     specifically need to test a module's behaviour when used in a .par binary.
     """
-    compiler = kwargs.pop('compiler', None)
-    zip_safe = kwargs.pop('zip_safe', True)
-    native.py_test(name=name, **kwargs)
-
-    main = kwargs.get('main', name + '.py')
-    imports = kwargs.get('imports')
-    default_python_version = kwargs.get('default_python_version', 'PY2')
-    visibility = kwargs.get('visibility')
-    testonly = kwargs.get('testonly', True)
+    compiler = kwargs.pop("compiler", None)
+    zip_safe = kwargs.pop("zip_safe", True)
+    no_remove = kwargs.pop('no_remove', False)
+    native.py_test(name = name, **kwargs)
+
+    main = kwargs.get("main", name + ".py")
+    imports = kwargs.get("imports")
+    default_python_version = kwargs.get("default_python_version", "PY2")
+    visibility = kwargs.get("visibility")
+    testonly = kwargs.get("testonly", True)
     parfile_test(
-        compiler=compiler,
-        default_python_version=default_python_version,
-        imports=imports,
-        main=main,
-        name=name + '.par',
-        src=name,
-        testonly=testonly,
-        visibility=visibility,
-        zip_safe=zip_safe,
+        compiler = compiler,
+        default_python_version = default_python_version,
+        imports = imports,
+        main = main,
+        name = name + ".par",
+        src = name,
+        testonly = testonly,
+        visibility = visibility,
+        zip_safe = zip_safe,
+        no_remove=no_remove,
     )
diff --git a/tests/BUILD b/tests/BUILD
index 0e72f9d..00030fd 100644
--- a/tests/BUILD
+++ b/tests/BUILD
@@ -11,14 +11,13 @@ exports_files([
 ])
 
 load("//:subpar.bzl", "par_binary")
+load("//:toolchain_test_hook.bzl", "PYVER", "define_toolchain_for_testing")
 
-# Configurations
-config_setting(
-    name = "subpar_test_python_version_3",
-    define_values = {
-        "subpar_test_python_version": "3",
-    },
-)
+# Creates the target //tests:toolchain_for_testing.
+#
+# This hook is used by run_tests.sh to inject a Python toolchain into the build
+# during tests. When not running tests this hook is a no-op.
+define_toolchain_for_testing()
 
 # Utility targets
 sh_library(
@@ -65,14 +64,11 @@ par_binary(
 
 par_binary(
     name = "package_f/f",
-    srcs = select({
-        "subpar_test_python_version_3": ["package_f/f_PY3.py"],
-        "//conditions:default": ["package_f/f_PY2.py"],
-    }),
-    main = select({
-        "subpar_test_python_version_3": "package_f/f_PY3.py",
-        "//conditions:default": "package_f/f_PY2.py",
-    }),
+    # We need to use the PYVER constant defined by the test hook, rather than
+    # select(), because python_version is not a configurable attribute.
+    srcs = ["package_f/f_PY3.py"] if PYVER == "PY3" else ["package_f/f_PY2.py"],
+    main = "package_f/f_PY3.py" if PYVER == "PY3" else "package_f/f_PY2.py",
+    python_version = "PY3" if PYVER == "PY3" else "PY2"
 )
 
 par_binary(
@@ -94,6 +90,7 @@ par_binary(
     main = "package_extract/extract.py",
     srcs_version = "PY2AND3",
     zip_safe = False,
+    no_remove = False,
 )
 
 par_binary(
@@ -151,16 +148,12 @@ par_binary(
         args = [
             "--par",
             "%s.par" % path,
-        ] + select({
-            "subpar_test_python_version_3": ["%s_PY3_filelist.txt" % path],
-            "//conditions:default": ["%s_PY2_filelist.txt" % path],
-        }),
+            ("%s_PY3_filelist.txt" % path) if PYVER == "PY3" else ("%s_PY2_filelist.txt" % path)
+        ],
         data = [
             "%s.par" % label,
-        ] + select({
-            "subpar_test_python_version_3": ["%s_PY3_filelist.txt" % label],
-            "//conditions:default": ["%s_PY2_filelist.txt" % label],
-        }),
+            ("%s_PY3_filelist.txt" % label) if PYVER == "PY3" else ("%s_PY2_filelist.txt" % label)
+        ],
     ),
 ) for name, label, path in [
     ("basic", "//tests/package_a:a", "tests/package_a/a"),
diff --git a/tests/package_f/f_PY3.py b/tests/package_f/f_PY3.py
index c5ac97d..5b42e49 100644
--- a/tests/package_f/f_PY3.py
+++ b/tests/package_f/f_PY3.py
@@ -22,7 +22,7 @@
 
 def main():
     assert sys.version_info.major == 3, sys.version
-    print('In f_PY2.py main()')
+    print('In f_PY3.py main()')
 
 
 if __name__ == '__main__':
diff --git a/tests/test_harness.sh b/tests/test_harness.sh
index b3a14d0..7c4b380 100755
--- a/tests/test_harness.sh
+++ b/tests/test_harness.sh
@@ -40,8 +40,12 @@ TMP_EXECUTABLE="$TEST_TMPDIR"/$(basename "$EXECUTABLE")
 
 # Compare list of files in zipfile with expected list
 if [ "$PAR" -eq 1 ]; then
+  # Exclude runfiles of the autodetecting toolchain. The test is still brittle
+  # with respect to runfiles introduced by any other toolchain. When tests are
+  # invoked by run_tests.sh, a custom toolchain with no runfiles is used.
   diff \
-    <(unzip -l "$EXECUTABLE" | awk '{print $NF}' | head -n -2 | tail -n +4) \
+    <(unzip -l -q -q "$EXECUTABLE" | awk '{print $NF}' \
+        | grep -v 'bazel_tools/tools/python/py.wrapper\.sh') \
     "$FILELIST" \
     || die 'FATAL: zipfile contents do not match expected'
 fi
diff --git a/toolchain_test_hook.bzl b/toolchain_test_hook.bzl
new file mode 100644
index 0000000..c2e3d7d
--- /dev/null
+++ b/toolchain_test_hook.bzl
@@ -0,0 +1,4 @@
+PYVER = "PY3"
+
+def define_toolchain_for_testing():
+    pass