SWE-smith tasks expose the pre-bug code through git history (`git checkout HEAD~1 -- .`)
维护者通常 2 天内回复
评估
调研方向
Start in debug_gym/gym/envs/swe_smith.py, where the bug patch is applied and committed (around lines 100-112), then read setup_terminal in swe_bench.py, which the issue says keeps older commits. Check that the gold patch and eval() still work on the working tree and index after the history is replaced. Done when git checkout HEAD~1 -- . no longer restores the pre-bug code, and tests/gym/envs/test_swe_smith.py (around lines 103-110) is updated and passing.
由索引模型根据 Issue 内容生成。
描述
What happens
SWESmithEnv commits the bug patch on top of the code from before the bug, and the gold patch is the bug patch reversed. So for a patch that only changes tracked files, HEAD~1 holds exactly the code the gold patch produces. With the bash tool from configs/swesmith.yaml, an agent can run git checkout HEAD~1 -- . and submit. In every task I ran, that scored exactly what the gold patch scores, without any debugging.
Code at 3f72ecc400a11752fc68cf08428225b6ba01a99e:
swe_bench.py#L129-L134:SWEBenchEnv.setup_terminal(inherited) commits the setup and removes the remote "so the agent won't see newer commits". Older commits stay.swe_smith.py#L105-L112: applies the bug patch, then runsgit commit -am 'Applying bug patch for <instance_id>'.swe_smith.py#L100-L103: the gold patch is the bug patch applied withgit apply --reverse.swe_smith.py#L149-L155:eval()restores the test files from the index, then runs the tests in the same container.git checkout HEAD~1 -- .updates the index too, so that reset does not undo it, andcalculate_scoregrades the same code the gold patch would produce.test_swe_smith.py#L103-L110asserts thatgit show HEADequals the bug patch.
In the container, git log --oneline shows:
a53785d Applying bug patch for john-kurkowski__tldextract.3d1bf184.combine_file__1vnuqpt4
50ed423 Initial commit
The image's .git also keeps the commit objects of the SWE-smith task branches (Bug Patch / Remove F2P Tests) after git remote remove origin. That is 202 commit objects in the tldextract image and 1,882 in the oauthlib image, with 2 reachable. So amending the commit or deleting refs would not be enough on its own.
Reproduction
The tasks are built through debug_gym.experiment.create_env with the shipped configs/swesmith.yaml (Docker terminal; bash, view, edit, submit). They are driven by scripted tool calls through env.step. Results on 3f72ecc, Docker 29.1.3:
| Task | no-op, then submit | env.apply_gold_patch(), then submit |
bash git checkout HEAD~1 -- ., then submit |
|---|---|---|---|
john-kurkowski__tldextract.3d1bf184.combine_file__1vnuqpt4 |
0/39 | 39/39, resolved | 39/39, resolved |
oauthlib__oauthlib.1fd52536.combine_file__09vlzwgc |
0/18 | 18/18, resolved | 18/18, resolved |
seatgeek__thefuzz.8a05a3ee.combine_file__2uoca06x |
0/25 | 25/25, resolved | 25/25, resolved |
iterative__dvc.1d6ea681.pr_7514 (the patch also adds a file) |
0/7 | 7/7, resolved | 7/7, resolved |
kurtmckee__feedparser.cad965a3.combine_file__115z1mk8 |
0/2226 | 2223/2226, unresolved | 2223/2226, unresolved |
Minimal manual version (any task from the table above where the gold patch resolves):
env = create_env(dict(config, uuid=str(uuid.uuid4())), task_data, logger) # config = configs/swesmith.yaml
env.reset()
env.step(ToolCall(id="1", name="bash", arguments={"command": "git checkout HEAD~1 -- ."}))
info = env.step(ToolCall(id="2", name="submit", arguments={}))
assert info.resolved
Scope
At the dataset revision SWESmithEnv.load_dataset pins (699b5340), there are 50,103 tasks after the excluded list. In 50,090 of them, the bug patch only modifies or deletes tracked files, so git checkout HEAD~1 -- . restores the same tracked files as the gold patch. The other 13 also add files, which git commit -am leaves untracked. The one I tested (dvc above) still scored like the gold patch.
Impact
- RL:
git logshows the bug commit ("Applying bug patch for ..."), and one bash command then earns the gold patch's reward. A policy that finds this is rewarded for restoring code from git history instead of for debugging. - Eval: an agent that looks at git history can be graded resolved without fixing the bug, so resolve rates can be inflated with no sign of it in the score.
Suggested fix
After applying the bug patch, replace .git with a new repository whose only commit is the buggy code. It should track the same files as before, plus any file the bug patch adds. Once the old object store is gone, HEAD~1, the old commit objects and the task-branch objects go with it. The gold patch and eval() work on the working tree and index, so they keep working. A PR with this change follows.
R2EGymEnv has a related TODO: remove .git/ entirely? at r2egym.py#L240. This issue covers SWE-smith only.
- 主要语言
- Python
- 星标
- 308
- 派生
- 47
- 平均合并
- 2 天 5 小时
- 30 天内合并 PR
- 9
环境准备
在浏览器里用你自己的 GitHub 账号启动这个项目的开发容器。
- 没有 Dockerfile 或 Docker Compose 文件
- 没有 Pull Request 模板
- 没有贡献指南
从这里开始
- 先读完整个 Issue,再读项目的贡献指南。
- 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
- Fork 仓库,在一个分支上完成修改。
- 提交 Pull Request,并在描述里引用这个 Issue 编号。
microsoft/debug-gym 的其他 Issue
-
难度 5/5 一周以上 新手友好度 25/100
维护者通常 2 天内回复
-
难度 1/5 1 小时以内 新手友好度 45/100
维护者通常 2 天内回复
-
难度 3/5 1-2 天 新手友好度 42/100
维护者通常 2 天内回复
-
难度 3/5 1-2 天 新手友好度 25/100
维护者通常 2 天内回复
-
难度 2/5 1-3 小时 新手友好度 45/100
microsoft/debug-gym#304 · 1 条评论 ·
维护者通常 2 天内回复
查看 microsoft/debug-gym 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 66/100
维护者通常 1 天内回复
-
[Bug] vllm-sr config migrate drops providers.models[].deployment可能已有人在做 @Bevisy 今天认领。 未关闭needs-acceptance wg/developer-experience-ecosystem
难度 2/5 1-3 小时 新手友好度 86/100
vllm-project/semantic-router#5996 ·
维护者通常 2 天内回复
-
bugfix low-priority plugin:abstract plugin:leyline small-effort source:discussion
难度 2/5 1-3 小时 新手友好度 66/100
athola/claude-night-market#1132 ·
-
难度 2/5 1-3 小时 新手友好度 77/100
WeblateOrg/language-data#2773 ·
维护者通常 1 天内回复
-
kind/bug priority/medium qa_automation
难度 2/5 1-3 小时 新手友好度 78/100
yugabyte/yugabyte-db#37324 ·
维护者通常 1 天内回复