diff --git a/.github/workflows/solve.yml b/.github/workflows/solve.yml
index 9a9991b1..617995b1 100644
--- a/.github/workflows/solve.yml
+++ b/.github/workflows/solve.yml
@@ -6,7 +6,7 @@ on:
       filter:
         description: "Instance filter regex"
         required: false
-        default: "django-1[67]"
+        default: "marshmallow"
       dataset:
         description: "Dataset name"
         required: true
@@ -27,7 +27,7 @@ on:
       split:
         description: "Dataset split"
         required: true
-        default: test
+        default: dev
         type: choice
         options:
           - dev
@@ -195,6 +195,7 @@ jobs:
           conda activate swe-bench
           export PYTHONPATH=$PYTHONPATH:$(pwd)
           python solver/solve.py \
+            --keep \
             --instances "${SWE_DATASET:-princeton-nlp/SWE-bench_Lite}" \
             --split ${SWE_SPLIT:-dev} \
             --filter "${SWE_FILTER:-marshmallow-1343}" \
@@ -241,6 +242,12 @@ jobs:
           path: output_${{ matrix.index }}.tar.xz
           compression-level: 0
 
+      - name: Setup upterm session
+        uses: owenthereal/action-upterm@v1
+        with:
+          limit-access-to-actor: true # Restrict to the user who triggered the workflow
+          wait-timeout-minutes: 5
+
   report:
     needs:
       - solve
@@ -301,3 +308,4 @@ jobs:
             logs/
             predictions.jsonl
             results.csv
+