diff --git a/.github/workflows/ci.yml b/.github/workflows/ci.yml
index c8bd0991..73bcd5b9 100644
--- a/.github/workflows/ci.yml
+++ b/.github/workflows/ci.yml
@@ -117,6 +117,81 @@ jobs:
       - name: Test
         run: vp run --filter helmcode test --shard ${{ matrix.shard }}/${{ strategy.job-total }}
 
+  # THROWAWAY: isolates each real-subprocess-spawning ProviderRegistry suite
+  # on its own runner, one file per job, to find out which one (and where in
+  # it) is hanging "Test Server 2" on every run since PR #48's
+  # opencodeRuntime.ts change (detached spawn + process-group kill
+  # finalizer). Delete these jobs once the probe answers that.
+  probe_opencode_hang:
+    name: Probe OpenCode Hang (${{ matrix.file }})
+    runs-on: ubuntu-latest
+    timeout-minutes: 8
+    strategy:
+      fail-fast: false
+      matrix:
+        file:
+          - ProviderRegistry
+    steps:
+      - name: Checkout
+        uses: actions/checkout@v6
+        with:
+          sparse-checkout: |
+            /*
+            !/.repos/
+          sparse-checkout-cone-mode: false
+
+      - name: Setup Vite+
+        uses: voidzero-dev/setup-vp@v1
+        with:
+          node-version-file: package.json
+          cache: true
+          run-install: true
+
+      # `timeout`'s SIGTERM only reaches vp's direct process, not a
+      # detached:true grandchild (opencodeRuntime.ts spawns those) — the last
+      # run sat past 300s with the step still "in_progress", suggesting an
+      # orphaned child is holding the log pipe open. So: run in the
+      # background, poll for it dying on our own, and if it doesn't, snapshot
+      # `ps -ef` (proof of a surviving orphan) and force-kill everything
+      # before exiting 0 — the step must always complete so the log upload
+      # step actually runs.
+      - name: Probe
+        run: |
+          set +e
+          vp run --filter helmcode test \
+            src/provider/Layers/${{ matrix.file }}.test.ts \
+            -- --reporter=verbose > /tmp/probe.log 2>&1 &
+          VP_PID=$!
+          for i in $(seq 1 60); do
+            sleep 5
+            kill -0 "$VP_PID" 2>/dev/null || break
+          done
+          if kill -0 "$VP_PID" 2>/dev/null; then
+            {
+              echo "=== TIMED OUT after ~300s, vp (pid $VP_PID) still alive ==="
+              echo "=== process tree snapshot ==="
+              ps -ef
+            } | tee -a /tmp/probe.log
+            pkill -9 -P "$VP_PID" 2>/dev/null
+            kill -9 "$VP_PID" 2>/dev/null
+          else
+            wait "$VP_PID"
+            echo "probe exited with $?" | tee -a /tmp/probe.log
+          fi
+          {
+            echo "=== final process snapshot before job teardown ==="
+            ps -ef
+          } | tee -a /tmp/probe.log
+          exit 0
+
+      - name: Upload probe log
+        if: always()
+        uses: actions/upload-artifact@v4
+        with:
+          name: probe-log-${{ matrix.file }}
+          path: /tmp/probe.log
+          if-no-files-found: ignore
+
   # Split out of Check and Test: both paid a Rust toolchain install for checks
   # that take a couple seconds, on the critical path of every PR.
   rust: