Rework how we locate chunks in code

bugbug/tools/code_review/__init__.py

@@ -30,8 +30,7 @@
 
 # Exceptions (backward compatibility)
 from bugbug.tools.core.exceptions import (
-    FileNotInPatchError,
-    HunkNotInPatchError,
+    CommentNotLocatedError,
     LargeDiffError,
     ModelResultError,
 )
@@ -56,8 +55,7 @@
     "InlineComment",
     "ReviewRequest",
     # Exceptions
-    "FileNotInPatchError",
-    "HunkNotInPatchError",
+    "CommentNotLocatedError",
     "LargeDiffError",
     "ModelResultError",
     # Base classes

bugbug/tools/code_review/agent.py

@@ -49,9 +49,11 @@
 )
 from bugbug.tools.code_review.utils import (
     convert_generated_comments_to_inline,
+    find_line_text,
     format_patch_set,
 )
 from bugbug.tools.core.exceptions import (
+    CommentNotLocatedError,
     LargeDiffError,
     RecursionLimitError,
 )
@@ -358,12 +360,31 @@ def _get_comment_examples(self, patch, created_before: datetime | None = None):
             for example in comment_examples:
                 example["comment"]["explanation"] = "THE JUSTIFICATION GOES HERE"
 
-        def format_comment(comment):
+        def format_comment(example):
             # TODO: change the schema that we expect the model to return so we
             # can remove this function.
+            comment = example["comment"]
+            filename = comment["filename"]
+            raw_hunk = example.get("hunk") or example.get("raw_hunk")
+            existing_code = None
+            if raw_hunk:
+                try:
+                    wrapped = TEMPLATE_PATCH_FROM_HUNK.format(
+                        filename=filename, raw_hunk=raw_hunk
+                    )
+                    patched_file = PatchSet.from_string(wrapped)[0]
+                    existing_code = find_line_text(patched_file, comment["start_line"])
+                except (CommentNotLocatedError, IndexError):
+                    logger.warning(
+                        "Could not recover existing_code for example comment on %s:%s",
+                        filename,
+                        comment["start_line"],
+                    )
+            if existing_code is None:
+                return None
             return {
-                "file": comment["filename"],
-                "code_line": comment["start_line"],
+                "file": filename,
+                "existing_code": existing_code,
                 "comment": comment["content"],
             }
 
@@ -375,10 +396,14 @@ def generate_formatted_patch_from_raw_hunk(raw_hunk, filename):
             return format_patch_set(patch_set)
 
         if not self.show_patch_example:
-            return json.dumps(
-                [format_comment(example["comment"]) for example in comment_examples],
-                indent=2,
-            )
+            formatted_comments = [
+                comment
+                for comment in (
+                    format_comment(example) for example in comment_examples
+                )
+                if comment is not None
+            ]
+            return json.dumps(formatted_comments, indent=2)
 
         return "\n\n".join(
             TEMPLATE_COMMENT_EXAMPLE.format(
@@ -387,7 +412,11 @@ def generate_formatted_patch_from_raw_hunk(raw_hunk, filename):
                     example["raw_hunk"], example["comment"]["filename"]
                 ),
                 comments=json.dumps(
-                    [format_comment(example["comment"])],
+                    [
+                        comment
+                        for comment in [format_comment(example)]
+                        if comment is not None
+                    ],
                     indent=2,
                 ),
             )

bugbug/tools/code_review/data_types.py

@@ -12,7 +12,14 @@ class GeneratedReviewComment(BaseModel):
     """A review comment generated by the code review agent."""
 
     file: str = Field(description="The path to the file the comment applies to.")
-    code_line: int = Field(description="The line number that the comment refers to.")
+    existing_code: str = Field(
+        description=(
+            "The exact source line(s) the comment is about, copied verbatim "
+            "from the patch (leading '+'/'-'/' ' markers may be included or "
+            "omitted). Used to locate the comment in the diff, so do not "
+            "paraphrase, reformat, or truncate it."
+        )
+    )
     comment: str = Field(description="The review comment.")
     explanation: str = Field(
         description="A brief rationale for the comment, including how confident you are and why."

bugbug/tools/code_review/prompts.py

@@ -44,6 +44,7 @@
 - Use directive language: "Fix", "Remove", "Change", "Add"
 - NEVER use these banned phrases: "maybe", "might want to", "consider", "possibly", "could be", "you may want to"
 - Focus strictly on code-related concerns
+- For `existing_code`, copy the exact source line(s) the comment is about verbatim from the patch — do not paraphrase, reformat, or guess a line number instead
 
 ## What NOT to Include
 
@@ -72,7 +73,7 @@
 - A large change that has no natural seam and must land atomically is acceptable — return an empty list rather than suggesting an impractical split.
 - If you do comment, name concrete seams: the distinct concerns, or the stages of a large cohesive change (e.g. land the data-model change separately from the call-site updates).
 - Briefly tell the author *why*: larger patches get less thorough review and empirically introduce more bugs and regressions, so smaller patches are easier to review and safer to land.
-- Anchor the comment to a representative changed line (a line that begins with `+`). Set `file` to that file's path and `code_line` to that line's number.
+- Anchor the comment to a representative changed line (a line that begins with `+`). Set `file` to that file's path and `existing_code` to that line's exact text, copied verbatim.
 - Use direct, declarative language. NEVER use these banned phrases: "maybe", "might want to", "consider", "possibly", "could be", "you may want to".
 
 Here is a summary of the patch: