SandboxApplyPatchTool splits patch lines with str.splitlines(), so a form feed or in a line breaks Add/Update File
维护者通常 1 天内回复
@abhishekacharya200 已经在做这个了。
开始于 2026年10月11日。
评估
调研方向
从 src/agents/sandbox/capabilities/tools/apply_patch_tool.py 中的 _parse_apply_patch_input 开始,将其分行方式与在 LF 处结束行的 Lark grammar 进行比较。然后检查 agents/apply_diff.py,它也按 "\n" 分割。验证 Add File 内容是否保留列出的字符,以及 Update File 复现案例是否将 const n = 1 改为 const n = 2。
由索引模型根据 Issue 内容生成。
描述
Please read this first
- Have you read the docs? Agents SDK docs
- Have you searched for related issues? Others may have faced similar issues.
Describe the bug
SandboxApplyPatchTool parses the freeform patch with raw_input.splitlines() (_parse_apply_patch_input in src/agents/sandbox/capabilities/tools/apply_patch_tool.py). str.splitlines() also breaks at \v, \f, \x1c-\x1e, \x85, \u2028, \u2029 and a bare \r. The tool's own Lark grammar only ends a line at LF: add_line: "+" /(.*)/ LF and change_line: ("+" | "-" | " ") /(.*)/ LF. agents/apply_diff.py also splits on "\n".
So a patch the grammar accepts is rejected when any line contains one of those characters. For example, a form feed between sections of a Python or C file, or a \u2028 inside a JS string or pasted text:
*** Add File:fails withInvalid Add File line: <rest of the line>, and the file is not created.*** Update File:turns the rest of the line into an unprefixed diff line, andapply_difffails withInvalid Line: <rest of the line>. This happens even when the character is only in a context line.
Splitting the envelope on "\n", as the grammar and apply_diff do, makes both cases work. In the repro below, the same hunk applies cleanly once it is split that way.
Debug information
- Agents SDK version:
mainat 125efa0 (0.23.1) - Python version: 3.12.3
- Operating system: Linux
- Model and model provider: n/a (the parser is called directly)
- Does the issue reproduce with the latest Agents SDK release? Yes. v0.23.1 has the same
splitlines()call and grammar; I ran the repro onmain(125efa0). - Does the issue occur consistently or intermittently? Consistently
'+page one\x0cpage two' -> Invalid Add File line: page two
'+const s = "a\u2028b";' -> Invalid Add File line: b";
'+x\x85y' -> Invalid Add File line: y
parsed diff: '@@\n const s = "a\nb";\n-const n = 1;\n+const n = 2;\n'
apply_diff -> Invalid Line: b";
'const s = "a\u2028b";\nconst n = 2;\n'
Repro steps
from agents.apply_diff import apply_diff
from agents.sandbox.capabilities.tools import SandboxApplyPatchTool
from agents.testing import scripted_sandbox_session
tool = SandboxApplyPatchTool(session=scripted_sandbox_session())
# 1. Add File: the grammar's add_line is "+" /(.*)/ LF, so these lines are valid input.
for line in ["+page one\fpage two", '+const s = "a\u2028b";', "+x\x85y"]:
patch = f"*** Begin Patch\n*** Add File: notes.txt\n{line}\n*** End Patch\n"
try:
tool.parse_custom_input(patch)
print(ascii(line), "-> parsed")
except ValueError as e:
print(ascii(line), "->", e)
# 2. Update File whose context line contains U+2028.
original = 'const s = "a\u2028b";\nconst n = 1;\n'
patch = (
"*** Begin Patch\n*** Update File: app.js\n@@\n"
' const s = "a\u2028b";\n-const n = 1;\n+const n = 2;\n*** End Patch\n'
)
(op,) = tool.parse_custom_input(patch)
print("parsed diff:", ascii(op.diff))
try:
apply_diff(original, op.diff)
except ValueError as e:
print("apply_diff ->", e)
# The same hunk, split on "\n" only, applies cleanly.
print(ascii(apply_diff(original, '@@\n const s = "a\u2028b";\n-const n = 1;\n+const n = 2;\n')))
Expected behavior
The three Add File patches create files whose content keeps the form feed, \u2028 and \x85. The Update File patch changes const n = 1; to const n = 2;, which is what apply_diff produces when the hunk is split on "\n" only.
- 主要语言
- Python
- 星标
- 30.4k
- 派生
- 4.9k
- 平均合并
- 16 小时 32 分钟
- 30 天内合并 PR
- 85
环境准备
- 没有 Dockerfile 或 Docker Compose 文件
- 有 Pull Request 模板
- 阅读贡献指南
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。
openai/openai-agents-python 的其他 Issue
-
难度 2/5 1-3 小时 新手友好度 62/100
openai/openai-agents-python#5477 ·
维护者通常 1 天内回复
-
Voice: default sentence splitter never splits Chinese/Japanese, so TTS waits for the whole reply未关闭
难度 2/5 1-3 小时 新手友好度 78/100
openai/openai-agents-python#5475 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 70/100
openai/openai-agents-python#5472 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 78/100
openai/openai-agents-python#5471 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 82/100
openai/openai-agents-python#5468 · 1 条评论 ·
维护者通常 1 天内回复
查看 openai/openai-agents-python 的全部 Issue
相似的 Issue
-
难度 2/5 1-3 小时 新手友好度 72/100
维护者通常 1 天内回复
-
enhancement good first issue
难度 2/5 1-3 小时 新手友好度 78/100
elviskahoro/sdk-python-attio#144 · 1 条评论 ·
维护者通常 1 天内回复
-
难度 2/5 1-3 小时 新手友好度 84/100
维护者通常 1 天内回复
-
RenyiELBO raises TypeError when model has no sample sites可能已有人在做 @alibb007 今天认领。 未关闭
难度 1/5 1 小时以内 新手友好度 85/100
维护者通常 1 天内回复
-
[Bug]: Four evaluation scripts unpack init_memmachine_params's five return values into three names未关闭
难度 1/5 1 小时以内 新手友好度 88/100
MemMachine/MemMachine#1911 ·
维护者通常 3 天内回复