Capture SWE-bench patches from failed runs

[neubig]

Summary

• preserve test_result on SWE-bench error outputs so failed runs can still be scored when they produced a patch
• stage failed SWE-bench workspaces before collecting failure patches, so newly-created files are included
• collect failed-run patches from the staged index with git diff --cached <base_commit> because failed runs do not commit and therefore do not advance HEAD

Scope

This PR is workspace-backend agnostic. It uses the existing RemoteWorkspace.execute_command() API and applies to Docker, Apptainer, and other remote workspaces where /workspace/<repo> is a git repo.

Apptainer-specific workspace/image changes are kept out of this PR and live in #745.

Verification

• PYTHONPATH="$PWD" /home/gneubig/work/openhands-benchmarks-venv/bin/python -m pytest tests/test_swebench_failure_patch_capture.py -q
• ADP partial scoring smoke: a captured error patch from a stuck run was applied and scored successfully by the SWE-bench scorer.

Draft while the full ADP SWE-bench experiment continues running.

Issue

Related: #749

[neubig]

@OpenHands /review

Please run an AI review for this now-main-based generic failed-run patch capture PR. CI is green and Apptainer-specific changes have been split into #745.

[openhands-ai]

Uh oh! There was an unexpected error starting the job :(

[neubig]

@OpenHands /review

[openhands-ai]

I'm on it! neubig can track my progress at all-hands.dev

[neubig]

Code Review

Overall: ✅ LGTM

The PR is well-designed and implements a clean solution for capturing git patches from failed SWE-bench runs.

What this PR does:

1. Adds a collect_failure_test_result() hook to the base Evaluation class that benchmarks can override
2. SWE-bench implementation stages the workspace with git add -A and captures a patch with git diff --cached
3. Preserves test_result in error outputs so failed runs can still be scored
4. Properly handles exceptions during patch collection with warnings

Strengths:

• Clean architecture: The hook is in the base class, making it reusable across benchmarks
• Proper error handling: Both git add and git diff failures are handled gracefully with warnings
• Good separation of concerns: get_repo_path(), get_git_patch(), get_staged_git_patch(), and collect_failure_test_result() are all separate, focused methods
• Defensive programming: Uses dict(test_result or {}) to avoid issues with None values
• Comprehensive tests: Two tests covering both base class preservation and SWE-bench-specific behavior
• Documentation: Good comments explaining the intent

Verified:

• ✅ Pre-commit checks pass (Ruff format, Ruff lint, PEP8, Pyright)
• ✅ Tests pass (2/2)

Minor notes (not blockers):

• The unused error parameter in the base collect_failure_test_result hook is fine since it's meant to be overridden
• The cast(RemoteWorkspace, workspace) in tests is necessary for type checking

This is a solid implementation that cleanly solves the problem of capturing partial outputs from failed runs.

[openhands-ai]

OpenHands encountered an error: **Failed to send message to agent server: HTTP 503 error: no available server
**

See the conversation for more information.

[neubig]

@OpenHands /review

[openhands-ai]

Uh oh! There was an unexpected error starting the job :(

[neubig]

@OpenHands /review

[openhands-ai]

I'm on it! neubig can track my progress at all-hands.dev

[neubig]

Code Review

Overall: ✅ LGTM

The PR is well-designed and implements a clean solution for capturing git patches from failed SWE-bench runs. The tests pass and pre-commit checks are clean.

What this PR does:

1. Adds a collect_failure_test_result() hook to the base Evaluation class that benchmarks can override
2. SWE-bench implementation stages the workspace with git add -A and captures a patch with git diff --cached
3. Preserves test_result in error outputs so failed runs can still be scored
4. Properly handles exceptions during patch collection with appropriate warnings

Strengths:

• Clean hook design: The collect_failure_test_result() hook in Evaluation is a minimal, well-documented interface that allows other benchmarks to add their own failure recovery logic
• Workspace-agnostic: Uses the existing RemoteWorkspace.execute_command() API, making it applicable to Docker, Apptainer, and other backends
• Preserves workspace state: git add -A stages changes without committing, keeping the failed workspace in its final agent-visible state
• Comprehensive error handling: Both git add failures and patch collection errors are caught and logged as warnings, never blocking evaluation

Files changed:

• benchmarks/utils/evaluation.py: +34 lines for the hook and integration
• benchmarks/swebench/run_infer.py: +84 lines for SWE-bench specific implementation
• tests/test_swebench_failure_patch_capture.py: +94 lines for comprehensive test coverage

Test results:

tests/test_swebench_failure_patch_capture.py::test_error_output_preserves_failure_test_result PASSED
tests/test_swebench_failure_patch_capture.py::test_swebench_collect_failure_test_result_gets_git_patch PASSED

Recommendation:

Approve - the implementation is clean, well-tested, and follows the existing codebase patterns. Ready for merge when the draft status is lifted.

[openhands-ai]

OpenHands encountered an error: **Failed to send message to agent server: HTTP 503 error: no available server
**

See the conversation for more information.