Summary
Running swebench-infer on SWE-Bench Lite with a Docker workspace fails before agent execution starts.
The Docker image appears to build successfully, but the benchmark runner rejects it because the built image tags do not match the expected image tag.
Environment
- Docker: 29.5.0
- Benchmarks commit:
2d635c0
- SDK submodule commit:
3e0a3a09
- Python: 3.12
Reproduction
Clone and initialize
git clone https://github.com/OpenHands/benchmarks.git
cd benchmarks
make build
Verify versions
git rev-parse --short HEAD
# 2d635c0
git submodule status
# 3e0a3a0915b369c7e2057c77722e98585855d30a vendor/software-agent-sdk (v1.14.0-186-g3e0a3a09)
Run SWE-Bench Lite
uv run swebench-infer .llm_config/example.json \
--dataset princeton-nlp/SWE-bench_Lite \
--split test \
--workspace docker \
--max-iterations 100 \
--num-workers 1 \
--n-limit 1
The failure occurred on:
scikit-learn__scikit-learn-25500
Expected behavior
The workspace image is built and the agent starts solving the selected SWE-Bench instance.
Actual behavior
The image build completes, but the benchmark runner fails during image validation.
Error message
Built image tags ['ghcr.io/openhands/eval-agent-server:3e0a3a0-sweb.eval.x86_64.scikit-learn_1776_scikit-learn-25500-source-minimal', 'ghcr.io/openhands/eval-agent-server:3e0a3a0-sweb.eval.x86_64.scikit-learn_1776_sciki_tag_latest-435039b0ab34-source-minimal'] do not include expected tag ghcr.io/openhands/eval-agent-server:3e0a3a0-35d813f-sweb.eval.x86_64.scikit-learn_1776_scikit-learn-25500-source-minimal
Traceback
Traceback (most recent call last):
File "/home/khidki/FedAgent/openhands_setup/benchmarks/benchmarks/utils/evaluation.py", line 1069, in _execute_single_attempt
workspace = self.prepare_workspace(
^^^^^^^^^^^^^^^^^^^^^^^
File "/home/khidki/FedAgent/openhands_setup/benchmarks/benchmarks/swebench/run_infer.py", line 190, in prepare_workspace
built = ensure_local_image(
^^^^^^^^^^^^^^^^^^^
File "/home/khidki/FedAgent/openhands_setup/benchmarks/benchmarks/utils/build_utils.py", line 517, in ensure_local_image
raise RuntimeError(
RuntimeError: Built image tags ['ghcr.io/openhands/eval-agent-server:3e0a3a0-sweb.eval.x86_64.scikit-learn_1776_scikit-learn-25500-source-minimal', 'ghcr.io/openhands/eval-agent-server:3e0a3a0-sweb.eval.x86_64.scikit-learn_1776_sciki_tag_latest-435039b0ab34-source-minimal'] do not include expected tag ghcr.io/openhands/eval-agent-server:3e0a3a0-35d813f-sweb.eval.x86_64.scikit-learn_1776_scikit-learn-25500-source-minimal
Notes
The built image tags contain the SDK commit hash (3e0a3a0), while the expected tag additionally includes 35d813f.
Happy to provide additional logs if needed.
Summary
Running
swebench-inferon SWE-Bench Lite with a Docker workspace fails before agent execution starts.The Docker image appears to build successfully, but the benchmark runner rejects it because the built image tags do not match the expected image tag.
Environment
2d635c03e0a3a09Reproduction
Clone and initialize
git clone https://github.com/OpenHands/benchmarks.git cd benchmarks make buildVerify versions
Run SWE-Bench Lite
uv run swebench-infer .llm_config/example.json \ --dataset princeton-nlp/SWE-bench_Lite \ --split test \ --workspace docker \ --max-iterations 100 \ --num-workers 1 \ --n-limit 1The failure occurred on:
Expected behavior
The workspace image is built and the agent starts solving the selected SWE-Bench instance.
Actual behavior
The image build completes, but the benchmark runner fails during image validation.
Error message
Traceback
Notes
The built image tags contain the SDK commit hash (
3e0a3a0), while the expected tag additionally includes35d813f.Happy to provide additional logs if needed.