fix(config): handle corrupted config.toml in credential functions (#481)

* fix(config): handle corrupted config.toml in get_credentials and set_credentials

get_credentials() called toml.load() without exception handling, causing
a raw TOMLDecodeError to crash any code that imports the runpod package
when ~/.runpod/config.toml is malformed. Return None on parse failure,
matching the existing pattern in check_credentials().

* fix(test): use module-scoped patches and mock filesystem calls

Address review feedback:
- Patch os.path.exists via module path for consistency with existing tests
- Patch os.makedirs and Path.touch in set_credentials tests for hermeticity

* chore: remove broken cleanup-endpoints workflow

Fails with 403 on the Runpod GraphQL API.

* fix(e2e): replace flash undeploy with direct GraphQL endpoint deletion

flash undeploy looks up endpoints in .runpod/resources.pkl which does
not exist in CI. Use the Runpod GraphQL API to query endpoints by name
and delete them directly.

* fix(e2e): make endpoint cleanup non-fatal

The RUNPOD_API_KEY in CI lacks GraphQL API access, causing 403 on
cleanup. Catch the error so test results are not masked by teardown
failures.

* fix(e2e): use flash undeploy --all --force for cleanup

Author: deanq

.github/workflows/cleanup-endpoints.yml

@@ -31,11 +31,15 @@ def set_credentials(api_key: str, profile: str = "default", overwrite=False) ->
     Path(CREDENTIAL_FILE).touch(exist_ok=True)
 
     if not overwrite:
-        with open(CREDENTIAL_FILE, "rb") as cred_file:
-            if profile in toml.load(cred_file):
-                raise ValueError(
-                    "Profile already exists. Use `update_credentials` instead."
-                )
+        try:
+            with open(CREDENTIAL_FILE, "rb") as cred_file:
+                existing = toml.load(cred_file)
+        except (TypeError, ValueError):
+            existing = {}
+        if profile in existing:
+            raise ValueError(
+                "Profile already exists. Use `update_credentials` instead."
+            )
 
     with open(CREDENTIAL_FILE, "w", encoding="UTF-8") as cred_file:
         cred_file.write("[" + profile + "]\n")
@@ -72,12 +76,18 @@ def check_credentials(profile: str = "default"):
 def get_credentials(profile="default"):
     """
     Returns the credentials for the specified profile from ~/.runpod/config.toml
+
+    Returns None if the file does not exist, is not valid TOML, or does not
+    contain the requested profile.
     """
     if not os.path.exists(CREDENTIAL_FILE):
         return None
 
-    with open(CREDENTIAL_FILE, "rb") as cred_file:
-        credentials = toml.load(cred_file)
+    try:
+        with open(CREDENTIAL_FILE, "rb") as cred_file:
+            credentials = toml.load(cred_file)
+    except (TypeError, ValueError):
+        return None
 
     if profile not in credentials:
         return None

runpod/cli/groups/config/functions.py

@@ -57,22 +57,17 @@ def endpoints(require_api_key, test_cases):
         log.info("Endpoint ready: name=%s image=%s template.dockerArgs=%s", ep.name, ep.image, ep.template.dockerArgs if ep.template else "N/A")
     yield eps
 
-    # Undeploy only the endpoints provisioned by this test run.
-    # Uses by-name undeploy to avoid tearing down unrelated endpoints
-    # sharing the same API key (parallel CI runs, developer endpoints).
-    endpoint_names = [ep.name for ep in eps.values()]
-    log.info("Cleaning up %d provisioned endpoints: %s", len(endpoint_names), endpoint_names)
-    for name in endpoint_names:
-        try:
-            result = subprocess.run(
-                ["flash", "undeploy", name, "--force"],
-                capture_output=True,
-                text=True,
-                timeout=60,
-            )
-            if result.returncode == 0:
-                log.info("Undeployed %s", name)
-            else:
-                log.warning("flash undeploy %s failed (rc=%d): %s", name, result.returncode, result.stderr)
-        except Exception:
-            log.exception("Failed to undeploy %s", name)
+    log.info("Cleaning up all provisioned endpoints")
+    try:
+        result = subprocess.run(
+            ["flash", "undeploy", "--all", "--force"],
+            capture_output=True,
+            text=True,
+            timeout=120,
+        )
+        if result.returncode == 0:
+            log.info("Undeployed all endpoints")
+        else:
+            log.warning("flash undeploy --all --force failed (rc=%d): %s", result.returncode, result.stderr)
+    except Exception:
+        log.exception("Failed to undeploy endpoints")

tests/e2e/conftest.py

@@ -97,3 +97,53 @@ def test_get_credentials_non_existent_profile(
         assert result is None
         assert mock_open_call.called
         assert mock_exists.called
+
+    @patch("runpod.cli.groups.config.functions.os.path.exists", return_value=True)
+    @patch(
+        "runpod.cli.groups.config.functions.toml.load",
+        side_effect=ValueError("Invalid value"),
+    )
+    @patch("builtins.open", new_callable=mock_open)
+    def test_get_credentials_corrupted_toml(
+        self, _mock_open_call, _mock_toml_load, _mock_exists
+    ):
+        """get_credentials returns None when config.toml contains invalid TOML."""
+        result = functions.get_credentials("default")
+        assert result is None
+
+    @patch("runpod.cli.groups.config.functions.os.path.exists", return_value=True)
+    @patch(
+        "runpod.cli.groups.config.functions.toml.load",
+        side_effect=TypeError("bad type"),
+    )
+    @patch("builtins.open", new_callable=mock_open)
+    def test_get_credentials_type_error(
+        self, _mock_open_call, _mock_toml_load, _mock_exists
+    ):
+        """get_credentials returns None on TypeError from corrupted file."""
+        result = functions.get_credentials("default")
+        assert result is None
+
+    @patch("runpod.cli.groups.config.functions.Path.touch")
+    @patch("runpod.cli.groups.config.functions.os.makedirs")
+    @patch("runpod.cli.groups.config.functions.toml.load")
+    @patch("builtins.open", new_callable=mock_open())
+    def test_set_credentials_corrupted_toml_allows_overwrite(
+        self, _mock_file, mock_toml_load, _mock_makedirs, _mock_touch
+    ):
+        """set_credentials with overwrite=True ignores corrupted existing file."""
+        mock_toml_load.side_effect = ValueError("Invalid TOML")
+        # overwrite=True skips the toml.load check entirely
+        functions.set_credentials("NEW_KEY", overwrite=True)
+
+    @patch("runpod.cli.groups.config.functions.Path.touch")
+    @patch("runpod.cli.groups.config.functions.os.makedirs")
+    @patch("runpod.cli.groups.config.functions.toml.load")
+    @patch("builtins.open", new_callable=mock_open())
+    def test_set_credentials_corrupted_toml_no_overwrite(
+        self, _mock_file, mock_toml_load, _mock_makedirs, _mock_touch
+    ):
+        """set_credentials without overwrite treats corrupted file as empty."""
+        mock_toml_load.side_effect = ValueError("Invalid TOML")
+        # Should not raise — corrupted file is treated as having no profiles
+        functions.set_credentials("NEW_KEY", overwrite=False)