diff --git a/README.md b/README.md
index f65b80f..2142dad 100644
--- a/README.md
+++ b/README.md
@@ -64,3 +64,28 @@ jobs:
 
 See [`ci.yml`](.github/workflows/ci.yml) to see a full lint and test workflow
 using these actions (including uploading output artifacts from the tests).
+
+## Caveats
+
+* GDB doesn't seem to support separate output streams for `stdout` and `stderr`,
+  so all test output to `stderr` will end up combined with `stdout` and both will be
+  printed to the runner's `stdout`. If you know a workaround for this that doesn't
+  require patching + building GDB itself please open an issue about it!
+
+* Since the custom test runner runs as part of `cargo test`, it won't be able to
+  find a `3dsx` that hasn't built yet. `cargo-3ds` doesn't build `3dsx` executables until
+  _after_ the cargo command it runs internally, so this means that tests can't depend
+  on any features of the `3dsx` (like embedded romFS). A workaround for this is to
+  simply build the tests as a separate step before running them, after which the
+  runner will be able to find the `3dsx`.
+
+* Doctests require a bit of extra setup to work with the runner, since they don't
+  use the crate's `#![test_runner]`. To write doctests, add the following to the
+  beginning of the doctest (or `fn main()` if the test defines it):
+
+  ```rust
+  let _runner = test_runner::GdbRunner::default();
+  ```
+
+  The runner must remain in scope for the duration of the test in order for
+  the test output to be printed.
diff --git a/run-tests/docker/entrypoint.sh b/run-tests/docker/entrypoint.sh
index 3e2ecbf..72e1860 100755
--- a/run-tests/docker/entrypoint.sh
+++ b/run-tests/docker/entrypoint.sh
@@ -1,10 +1,12 @@
 #!/bin/bash
 
-# Uncomment for debugging the action itself. Maybe consider a job summary or
-# grouping the output, to keep this stuff visible but make it simpler to use:
+# TODO: Maybe consider a job summary or grouping the output, to keep xtrace output
+# visible but make it simpler to use:
 # https://docs.github.com/en/actions/using-workflows/workflow-commands-for-github-actions
+# set -o xtrace
 
-# set -x
+set -o nounset
+set -o pipefail
 
 function cleanup_jobs() {
     # shellcheck disable=SC2317 # Unreachable because it's only used in trap
@@ -21,7 +23,8 @@ function cleanup_jobs() {
 trap cleanup_jobs EXIT
 
 EXE_ELF=$1
-EXE_3DSX="$(dirname "$EXE")/$(basename "$EXE" .elf).3dsx"
+EXE_NOEXT="$(dirname "$EXE_ELF")/$(basename "$EXE_ELF" .elf)"
+EXE_3DSX="$EXE_NOEXT.3dsx"
 
 EXE_TO_RUN="$EXE_ELF"
 if [ -f "$EXE_3DSX" ]; then
@@ -29,7 +32,7 @@ if [ -f "$EXE_3DSX" ]; then
     EXE_TO_RUN="$EXE_3DSX"
 fi
 
-VIDEO_OUT="$(dirname "$EXE_ELF")/$(basename "$EXE_ELF" .elf)_capture.webm"
+VIDEO_OUT="${EXE_NOEXT}_capture.webm"
 
 CITRA_LOG_DIR=~/.local/share/citra-emu/log
 CITRA_OUT="$CITRA_LOG_DIR/citra_output.txt"
@@ -54,7 +57,7 @@ cleanup_jobs
 CITRA_LOG="$CITRA_LOG_DIR/citra_log.txt"
 
 for f in "$CITRA_LOG" "$CITRA_OUT"; do
-    OUT="$(dirname "$EXE_ELF")/$(basename "$EXE_ELF" .elf)_$(basename "$f")"
+    OUT="${EXE_NOEXT}_$(basename "$f")"
     if test -f "$f"; then
         cp "$f" "$OUT"
         if [ $STATUS -ne 0 ]; then