diff --git a/.github/workflows/dart-test.yaml b/.github/workflows/dart-test.yaml
index ecfc6a93..63b3ed55 100644
--- a/.github/workflows/dart-test.yaml
+++ b/.github/workflows/dart-test.yaml
@@ -49,7 +49,6 @@ jobs:
           - ubuntu-22.04-arm     # linux-arm64
           - macos-latest         # darwin-arm64
           - macos-15-intel       # darwin-amd64
-          - windows-latest       # windows-amd64
           # windows-11-arm is out until kcl-lang/kcl publishes a
           # kcl-lib-<version>-windows-arm64.zip: kcl-lang/kcl#2219 adds the
           # target to that repo's release matrix, but no released version has
diff --git a/dart/README.md b/dart/README.md
index 34f7cb06..c774dba0 100644
--- a/dart/README.md
+++ b/dart/README.md
@@ -125,7 +125,10 @@ try {
 ```
 
 `exec_program` and `validate_code` are **not thread-safe**, mirroring the
-spec.
+spec. Call them from a single isolate: the shared runtime keeps process-global
+state, and overlapping calls can lose a panic payload — the KCL error then
+comes back with an empty message. The test suite runs one file at a time for
+the same reason; see [`dart_test.yaml`](dart_test.yaml).
 
 ## API reference
 
diff --git a/dart/dart_test.yaml b/dart/dart_test.yaml
new file mode 100644
index 00000000..512ac536
--- /dev/null
+++ b/dart/dart_test.yaml
@@ -0,0 +1,17 @@
+# Run one test file at a time.
+#
+# `dart test` normally loads and runs test *files* concurrently, each in its
+# own isolate — but every isolate shares the one loaded `libkcl`, and the KCL
+# runtime keeps process-global state its own C API declares not thread-safe.
+# The sharpest edge is the panic hook: the Rust side takes, replaces and
+# restores the *process-global* `std::panic` hook around every evaluation
+# (`FastRunner::run` in kcl_runner) and records the panic payload into a
+# thread-local. Two overlapping evaluations can clobber each other's hook, so
+# the panicking thread's payload never reaches its own record and the KCL error
+# comes back with an empty message — the plugin tests in kcl_lib_test.dart
+# then fail on `expect(errMessage, contains(...))`, with the real text showing
+# up only in the runtime's stderr.
+#
+# Serial execution costs ~0.1s on this package, so the suite stays honest
+# about what it exercises instead of racing the runtime.
+concurrency: 1