apply_diff: *** End of File hunks corrupt or mis-locate edits when the input ends in a newline #4660

Please read this first

• Have you read the docs? Yes.
• Have you searched for related issues? Yes. fix(apply_diff): drop all trailing blank lines in diff normalization #3263, fix(apply-diff): accept the stacked @@ headers the tool description asks for #4360, fix(apply-diff): apply stacked anchors sequentially #4369, and fix(apply-diff): reject partially matched stacked anchors #4431 touch apply_diff, but none covers *** End of File.

Describe the bug

apply_diff mishandles *** End of File hunks whenever the input ends in a newline, which is nearly every real source file. Two symptoms, one cause.

1. An append at end of file inserts a blank line and drops the trailing newline.

apply_diff("a\nb\n", "@@\n+c\n*** End of File")
# returns 'a\nb\n\nc'
# expected 'a\nb\nc\n'

2. An EOF hunk with repeated context silently patches the first match instead of the last.

apply_diff("x\nfoo\ny\nfoo\n", " foo\n+added\n*** End of File")
# returns 'x\nfoo\nadded\ny\nfoo\n'   (after the FIRST foo)
# expected 'x\nfoo\ny\nfoo\nadded\n'  (after the LAST foo)

Both are silent.

Neither reproduces when the input lacks a trailing newline. apply_diff("a\nb", ...) returns 'a\nb\nc' and apply_diff("x\nfoo\ny\nfoo", ...) appends after the last foo, both correct. The trailing newline is the trigger.

Mechanism

_parse_update_diff and _apply_chunks split the input with input.split("\n"). For text ending in a newline, this yields a phantom empty final element: "a\nb\n" becomes ["a", "b", ""].

_find_context (src/agents/apply_diff.py:333-341) anchors an EOF hunk at len(lines) - len(context) and counts that phantom element as a line:

• With empty context, the anchor resolves to len(lines), i.e. after the phantom "". The insert lands past the empty element, so the output gains a blank line and loses its final newline.
• With non-empty context, the end-anchored window always includes the phantom "", so the comparison can never match. The code falls through to _find_context_core(lines, context, start), a forward search that returns the first occurrence. The fuzz + 10000 penalty that marks this as a degraded match is accumulated into parser.fuzz and then discarded. apply_diff never surfaces it, so the wrong-location edit is invisible.

Debug information

• Agents SDK version: v0.22.0, and main at 5f9f4f0
• Python version: 3.10.16 (repo default via uv)
• Operating system: macOS (darwin, arm64)
• Model and model provider: n/a — apply_diff is pure text manipulation, no model or network involved
• Does the issue reproduce with the latest Agents SDK release? Yes, and on main.
• Does the issue occur consistently or intermittently? Consistently, deterministically.

Repro steps

from agents import apply_diff

# 1. append at EOF
print(repr(apply_diff("a\nb\n", "@@\n+c\n*** End of File")))
# 'a\nb\n\nc'      expected 'a\nb\nc\n'

# 2. ambiguous context with an EOF marker
print(repr(apply_diff("x\nfoo\ny\nfoo\n", " foo\n+added\n*** End of File")))
# 'x\nfoo\nadded\ny\nfoo\n'   expected 'x\nfoo\ny\nfoo\nadded\n'

# controls: same diffs, no trailing newline in the input, both correct
print(repr(apply_diff("a\nb", "@@\n+c\n*** End of File")))          # 'a\nb\nc'
print(repr(apply_diff("x\nfoo\ny\nfoo", " foo\n+added\n*** End of File")))  # '...foo\nadded'

Expected behavior

An *** End of File hunk should treat the phantom trailing element as the end of the file rather than as a line:

1. An append at EOF yields 'a\nb\nc\n', preserving the trailing newline and adding no blank line.
2. Repeated context resolves to the last occurrence, which is what the marker is for.

tests/test_apply_diff.py currently has no *** End of File coverage, which is why this went unnoticed.

The JS SDK carries the same logic in packages/agents-core/src/utils/applyDiff.ts (findContext, lines 327-333) and looks affected the same way, though I have only verified the Python side.

I have a fix with regression tests and will open a PR shortly.