Handle diff code fences

gptdiff/applydiff.py

@@ -9,6 +9,21 @@
 import hashlib
 from collections import defaultdict
 
+
+def _strip_diff_fence(diff_text: str) -> str:
+    """Remove wrapping triple backticks from a diff block."""
+    stripped = diff_text.strip()
+    if stripped.startswith("```") and stripped.endswith("```"):
+        lines = stripped.splitlines()
+        if lines[0].startswith("```"):
+            lines = lines[1:]
+        if lines and lines[-1].startswith("```"):
+            lines = lines[:-1]
+        if lines and lines[0].strip().lower() == "diff":
+            lines = lines[1:]
+        return "\n".join(lines)
+    return diff_text
+
 def apply_diff(project_dir, diff_text):
     """
     Applies a unified diff (as generated by git diff) to the files in project_dir
@@ -112,6 +127,7 @@ def apply_patch_to_file(file_path, patch):
         return True
 
     # Parse the diff into per-file patches.
+    diff_text = _strip_diff_fence(diff_text)
     file_patches = parse_diff_per_file(diff_text)
     if not file_patches:
         print("No file patches found in diff.")
@@ -192,6 +208,8 @@ def parse_diff_per_file(diff_text):
         Uses 'b/' prefix detection from git diffs to determine target paths
         This doesn't work all the time and needs to be revised with stronger models
     """
+    diff_text = _strip_diff_fence(diff_text)
+
     def dedup_diffs(diffs):
         groups = defaultdict(list)
         for key, value in diffs:

tests/test_multidiff.py

@@ -41,9 +41,15 @@ def test_fail_diff_through_call_llm(monkeypatch):
     def dummy_call_llm(api_key, base_url, model, messages, max_tokens, budget_tokens, temperature):
         return DummyResponse(diff_str, prompt_tokens=10, completion_tokens=20, total_tokens=30)
 
-    # Patch call_llm in the gptdiff module with our dummy function.
+    # Patch call_llm and tiktoken.get_encoding to avoid network access.
     monkeypatch.setattr("gptdiff.gptdiff.call_llm", dummy_call_llm)
 
+    class DummyEnc:
+        def encode(self, text):
+            return [0] * len(text)
+
+    monkeypatch.setattr("tiktoken.get_encoding", lambda name: DummyEnc())
+
     # generate_diff calls call_llm_for_diff internally, which now uses our dummy_call_llm.
     result = generate_diff("dummy environment", "dummy goal", model="test-model")
 

tests/test_parse_diff_per_file.py

@@ -169,5 +169,22 @@ def test_parse_diff_per_file_unconventional_header():
     assert "+++ game.js" in patch, "Expected patch to include '+++ game.js'"
     assert "+let player" in patch, "Expected patch to include added lines"
 
+
+def test_parse_diff_with_code_fence():
+    diff_text = """```diff
+diff --git a/file.txt b/file.txt
+--- a/file.txt
++++ b/file.txt
+@@ -1 +1 @@
+-old
++new
+```"""
+
+    result = parse_diff_per_file(diff_text)
+    assert len(result) == 1
+    file_path, patch = result[0]
+    assert file_path == "file.txt"
+    assert "+new" in patch
+
 if __name__ == '__main__':
     unittest.main()