fix(versioning): 修复首改文件 diff 基准丢失导致全文件计新增

根因:track_edit 在编辑前备份的 v1(修改前内容)只存在于该消息的
中间行,make_snapshot 用最终行(v2 修改后内容)整体替换该行后,
v1 不再被任何快照引用。对会话中首次被跟踪的文件,diff 配对找不到
上一条消息里的基准,按 added 处理并把全文件计为新增(+479 -0)。
26f31a09 修的是重复行配对,未覆盖这一数据丢失路径,故跨平台复现。

修复(不改数据模型,纯读取侧恢复):
- 新增 _resolve_diff_base_backup:优先用上一条消息快照里的备份;
  缺失时利用备份名确定性(sha256(绝对路径)@vN)与版本逐文件连续的
  不变量,回退到 v(N-1) 作为消息前状态(新建文件 v1 不落盘,自然
  仍判 added);首次出现即 v1 说明消息未改动该文件,与自身比较得 0。
- get_diff_stats / get_file_diff_stats / get_file_patch_lines 三处
  统一走该解析;旧 state 数据(v1 备份文件仍在)透明愈合,亦兼容
  MAX_SNAPSHOTS 逐出旧行的场景。
- 新增 5 个回归测试(生产顺序:track_edit 先于写入);修正 2 个旧
  测试的模拟顺序使其与生产一致(原顺序先写后 track,把新建文件
  误模拟成 v1 首现)。

验证:11/11 通过;旧代码跑新测试 4/5 失败(复现);用事故对话
conv_20260731_114736_581 真实数据只读复算,#1 由 +479/-0 added
修正为 +4/-1 modified(与实际编辑一致),#2 保持 +0/-0;全量测试
失败/错误数与 HEAD 基线完全相同(2+10,均为预存在/环境问题)。
This commit is contained in:
JOJO 2026-07-31 04:39:35 +00:00
parent 1760ca66f3
commit 77be044c44
2 changed files with 194 additions and 10 deletions

View File

@ -25,6 +25,29 @@ Concurrency: mutations (track_edit / make_snapshot) run inside a per-state-file
process lock as reload -> mutate -> save critical sections, so short-lived
manager instances (one per edit / per task end) can never lose each other's
rows. Readers only see whole files because writes go through os.replace.
Diff bases (version fallback)
-----------------------------
The diff shown for a message compares its final snapshot against the previous
message's snapshot. A file first tracked in the target message is absent from
every earlier snapshot: make_snapshot replaces the message's intermediate
track_edit row (which held the pre-edit backup) with the final row, so the
pre-edit state can no longer be found via snapshot pairing.
It is recovered through two invariants of the backup store:
- backup file names are deterministic: sha256 of the resolved absolute path,
suffixed with @v<version>;
- backup versions are sequential per file: track_edit creates v1 (pre-edit),
each make_snapshot bumps at most one version, and backup files are never
garbage-collected.
Hence for a first-appearance backup at version N >= 2, version N-1 is exactly
the pre-message state (for a file created by the message no v1 exists on disk,
so the base stays "missing" and the file correctly shows as added). A
first-appearance version-1 backup means the message left the file unchanged
since track_edit. This also heals state files written before the fix and
survives MAX_SNAPSHOTS eviction of older rows.
"""
from __future__ import annotations
@ -474,6 +497,40 @@ class ShallowVersioningManager:
break
return target_snapshot, previous_snapshot
def _resolve_diff_base_backup(
self,
tracking_path: str,
target_backup: FileBackup,
previous_snapshot: Optional[ShallowSnapshot],
) -> Optional[FileBackup]:
"""Resolve the backup representing the file's state BEFORE the target message.
Normal case: the previous message snapshot references the file, so its
backup is the diff base. A file first tracked in the target message is
absent from every earlier snapshot (its intermediate track_edit row was
replaced by make_snapshot); the pre-message state is then recovered via
the version fallback documented in the module docstring: version N-1 of
the deterministic backup name, used only when that backup file still
exists on disk. A first-appearance version-1 backup means the file was
left unchanged since track_edit, so it is diffed against itself.
"""
if previous_snapshot is not None:
previous = previous_snapshot.tracked_file_backups.get(tracking_path)
if previous is not None:
return previous
if target_backup.version >= 2:
resolved = self._resolve_file_path(tracking_path)
candidate_name = self._backup_file_name(str(resolved), target_backup.version - 1)
if self._backup_path(candidate_name).exists():
return FileBackup(
backup_file_name=candidate_name,
version=target_backup.version - 1,
backup_time=target_backup.backup_time,
)
elif target_backup.version == 1 and target_backup.backup_file_name is not None:
return target_backup
return None
def get_diff_stats(self, message_id: str) -> Optional[Dict[str, Any]]:
"""Compute insertions/deletions/files_changed between the target snapshot and the previous snapshot."""
target_snapshot, previous_snapshot = self._find_snapshot_pair(message_id)
@ -489,7 +546,7 @@ class ShallowVersioningManager:
if not target_backup:
continue
previous_backup = previous_snapshot.tracked_file_backups.get(tracking_path) if previous_snapshot else None
previous_backup = self._resolve_diff_base_backup(tracking_path, target_backup, previous_snapshot)
previous_lines = self._read_backup_lines(previous_backup)
target_lines = self._read_backup_lines(target_backup)
@ -525,7 +582,7 @@ class ShallowVersioningManager:
if not target_backup:
continue
previous_backup = previous_snapshot.tracked_file_backups.get(tracking_path) if previous_snapshot else None
previous_backup = self._resolve_diff_base_backup(tracking_path, target_backup, previous_snapshot)
previous_lines = self._read_backup_lines(previous_backup)
target_lines = self._read_backup_lines(target_backup)
@ -558,7 +615,7 @@ class ShallowVersioningManager:
if not target_backup:
return {"lines": [], "truncated": False}
previous_backup = previous_snapshot.tracked_file_backups.get(tracking_path) if previous_snapshot else None
previous_backup = self._resolve_diff_base_backup(tracking_path, target_backup, previous_snapshot)
previous_lines = self._read_backup_lines(previous_backup)
target_lines = self._read_backup_lines(target_backup)

View File

@ -49,18 +49,21 @@ class ShallowVersioningPairingTest(unittest.TestCase):
# realistic multi-instance flow (the original bug scenario)
# ------------------------------------------------------------------
def test_diff_pairs_with_previous_message_across_instances(self) -> None:
# Production order: track_edit runs BEFORE the tool applies the edit
# (tools_execution backs up first), so for newly created files
# track_edit sees a missing file (v1=None) and make_snapshot stores v2.
# message 1: AI creates two new files (each track_edit runs in its own instance)
self._write("a.txt", "a1\na2\n")
_new_manager(self.project, self.data).track_edit("a.txt", "msg_1")
self._write("b.txt", "b1\nb2\n")
self._write("a.txt", "a1\na2\n")
_new_manager(self.project, self.data).track_edit("b.txt", "msg_1")
self._write("b.txt", "b1\nb2\n")
_new_manager(self.project, self.data).make_snapshot("msg_1")
# message 2: modify a.txt, create c.txt
_new_manager(self.project, self.data).track_edit("a.txt", "msg_2") # early-return: already tracked
self._write("a.txt", "a1\na2 changed\na3\n")
_new_manager(self.project, self.data).track_edit("a.txt", "msg_2")
self._write("c.txt", "c1\n")
_new_manager(self.project, self.data).track_edit("c.txt", "msg_2")
self._write("c.txt", "c1\n")
_new_manager(self.project, self.data).make_snapshot("msg_2")
# state file must hold exactly one row per message
@ -130,15 +133,16 @@ class ShallowVersioningPairingTest(unittest.TestCase):
# accidentally correct pre-fix; lazy view was not)
# ------------------------------------------------------------------
def test_single_new_file_message(self) -> None:
self._write("a.txt", "a1\n")
# Production order: track_edit before the edit is applied.
_new_manager(self.project, self.data).track_edit("a.txt", "msg_1")
self._write("a.txt", "a1\n")
_new_manager(self.project, self.data).make_snapshot("msg_1")
# msg_2 edits a.txt and tracks one new file b.txt
_new_manager(self.project, self.data).track_edit("a.txt", "msg_2") # early-return: already tracked
self._write("a.txt", "a1\na2\n")
_new_manager(self.project, self.data).track_edit("a.txt", "msg_2")
self._write("b.txt", "b1\nb2\nb3\n")
_new_manager(self.project, self.data).track_edit("b.txt", "msg_2")
self._write("b.txt", "b1\nb2\nb3\n")
_new_manager(self.project, self.data).make_snapshot("msg_2")
reader = _new_manager(self.project, self.data)
@ -201,5 +205,128 @@ class ShallowVersioningPairingTest(unittest.TestCase):
self.assertIsNone(reader.get_snapshot_by_seq(4))
class ShallowVersioningFirstEditDiffBaseTest(unittest.TestCase):
"""Regression tests for the diff base of files first tracked in a message.
Production order is: file exists on disk -> track_edit backs up the
pre-edit content (v1) -> the edit is applied -> make_snapshot stores the
post-edit backup (v2), replacing the intermediate track_edit row. The
pre-edit backup then survives only on disk; diffs must recover it via the
version fallback instead of showing the whole file as added (+all -0).
"""
def setUp(self) -> None:
self.tmp = Path(tempfile.mkdtemp(prefix="shallow_ver_base_test_"))
self.project = self.tmp / "project"
self.data = self.tmp / "data"
self.project.mkdir(parents=True)
self.data.mkdir(parents=True)
def tearDown(self) -> None:
shutil.rmtree(self.tmp, ignore_errors=True)
def _write(self, rel: str, content: str) -> None:
path = self.project / rel
path.parent.mkdir(parents=True, exist_ok=True)
path.write_text(content, encoding="utf-8")
# ------------------------------------------------------------------
# the reported bug: first message edits a pre-existing file
# ------------------------------------------------------------------
def test_first_message_edit_of_preexisting_file_shows_real_diff(self) -> None:
self._write("main.py", "line1\nline2\nline3\n")
_new_manager(self.project, self.data).track_edit("main.py", "msg_1")
self._write("main.py", "line1\nline2 changed\nline3\nline4\n")
_new_manager(self.project, self.data).make_snapshot("msg_1")
# the intermediate track_edit row was replaced: only the final row
# remains, referencing v2; the pre-edit v1 survives only on disk.
rows = _read_state_rows(self.data)
self.assertEqual(len(rows), 1)
self.assertEqual(rows[0]["tracked_file_backups"]["main.py"]["version"], 2)
reader = _new_manager(self.project, self.data)
files = {f["path"]: f for f in reader.get_file_diff_stats("msg_1")}
self.assertEqual(set(files), {"main.py"})
self.assertEqual(files["main.py"]["status"], "modified")
self.assertEqual((files["main.py"]["insertions"], files["main.py"]["deletions"]), (2, 1))
stats = reader.get_diff_stats("msg_1") or {}
self.assertEqual((stats.get("insertions"), stats.get("deletions")), (2, 1))
patch = reader.get_file_patch_lines("main.py", "msg_1")
adds = [l["content"] for l in patch["lines"] if l["type"] == "add"]
removes = [l["content"] for l in patch["lines"] if l["type"] == "remove"]
self.assertEqual(removes, ["line2"])
self.assertEqual(adds, ["line2 changed", "line4"])
# ------------------------------------------------------------------
# a file created by the first message must still show as added
# ------------------------------------------------------------------
def test_first_message_new_file_still_added(self) -> None:
_new_manager(self.project, self.data).track_edit("new.txt", "msg_1") # absent at track time
self._write("new.txt", "n1\nn2\n")
_new_manager(self.project, self.data).make_snapshot("msg_1")
reader = _new_manager(self.project, self.data)
files = {f["path"]: f for f in reader.get_file_diff_stats("msg_1")}
self.assertEqual(set(files), {"new.txt"})
self.assertEqual(files["new.txt"]["status"], "added")
self.assertEqual((files["new.txt"]["insertions"], files["new.txt"]["deletions"]), (2, 0))
# ------------------------------------------------------------------
# an edit that leaves content identical yields no changes (v1 kept)
# ------------------------------------------------------------------
def test_first_message_unchanged_edit_shows_no_changes(self) -> None:
self._write("same.txt", "s1\ns2\n")
_new_manager(self.project, self.data).track_edit("same.txt", "msg_1")
_new_manager(self.project, self.data).make_snapshot("msg_1")
rows = _read_state_rows(self.data)
self.assertEqual(rows[0]["tracked_file_backups"]["same.txt"]["version"], 1)
reader = _new_manager(self.project, self.data)
self.assertEqual(reader.get_file_diff_stats("msg_1"), [])
self.assertFalse(reader.has_any_changes("msg_1"))
# ------------------------------------------------------------------
# first message deletes a pre-existing file -> full deletions
# ------------------------------------------------------------------
def test_first_message_delete_shows_deletions(self) -> None:
self._write("gone.txt", "g1\ng2\ng3\n")
_new_manager(self.project, self.data).track_edit("gone.txt", "msg_1")
(self.project / "gone.txt").unlink()
_new_manager(self.project, self.data).make_snapshot("msg_1")
reader = _new_manager(self.project, self.data)
files = {f["path"]: f for f in reader.get_file_diff_stats("msg_1")}
self.assertEqual(set(files), {"gone.txt"})
self.assertEqual(files["gone.txt"]["status"], "deleted")
self.assertEqual((files["gone.txt"]["insertions"], files["gone.txt"]["deletions"]), (0, 3))
# ------------------------------------------------------------------
# the normal previous-snapshot path keeps winning for later messages
# ------------------------------------------------------------------
def test_second_message_diff_uses_previous_snapshot_row(self) -> None:
self._write("a.txt", "a1\n")
_new_manager(self.project, self.data).track_edit("a.txt", "msg_1")
self._write("a.txt", "a1\na2\n")
_new_manager(self.project, self.data).make_snapshot("msg_1")
# already tracked: track_edit early-returns in production, so a plain
# content change followed by make_snapshot is the realistic flow.
self._write("a.txt", "a1\na2\na3\n")
_new_manager(self.project, self.data).make_snapshot("msg_2")
reader = _new_manager(self.project, self.data)
files = {f["path"]: f for f in reader.get_file_diff_stats("msg_2")}
self.assertEqual(set(files), {"a.txt"})
self.assertEqual(files["a.txt"]["status"], "modified")
self.assertEqual((files["a.txt"]["insertions"], files["a.txt"]["deletions"]), (1, 0))
# msg_1 diff still resolves through the fallback (pre-existing file)
files1 = {f["path"]: f for f in reader.get_file_diff_stats("msg_1")}
self.assertEqual(files1["a.txt"]["status"], "modified")
self.assertEqual((files1["a.txt"]["insertions"], files1["a.txt"]["deletions"]), (1, 0))
if __name__ == "__main__":
unittest.main()