From ce7afd48d4e645b76d7d4195e1d868051081de36 Mon Sep 17 00:00:00 2001
From: Diamond Hands Dig <98722191+33hodl@users.noreply.github.com>
Date: Tue, 15 Sep 2026 11:48:10 +0000
Subject: [PATCH] fix(checkpoints): safe restore must find the ledger under the
 walked project key
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

record_agent_write keys the agent-write ledger by the marker-walk result of
get_working_dir_for_path(), while safe restore reads the ledger by the exact
hash of the directory named in the restore.  The two keys diverge whenever
the walk lands somewhere other than that directory:

- a markerless project dir under an ancestor that carries a generic project
  marker (e.g. /tmp with a stray package.json) -> the ledger is filed under
  the ancestor's hash and restore reads an empty ledger;
- restoring a repo subdir while the write was recorded at the repo root.

An empty ledger makes restore() silently degrade to a full restore, which
overwrites files the user hand-edited after Hermes' last write — the exact
outcome safe restore exists to prevent.

safe_restore_plan now falls back to the walked key when the exact-key ledger
is empty, so the recorded protection is honored regardless of where the walk
landed.  Verified: 9 previously failing TestSafeRestore tests now pass, plus
a new regression test that seeds an ancestor marker and asserts the user
hand-edit survives the restore.
---
 tests/tools/test_checkpoint_manager.py | 21 +++++++++++++++++++++
 tools/checkpoint_manager.py            | 13 +++++++++++++
 2 files changed, 34 insertions(+)

diff --git a/tests/tools/test_checkpoint_manager.py b/tests/tools/test_checkpoint_manager.py
index f36f97e112..b87bbcf970 100644
--- a/tests/tools/test_checkpoint_manager.py
+++ b/tests/tools/test_checkpoint_manager.py
@@ -338,6 +338,27 @@ class TestSafeRestore:
         assert "README.md" in result["skipped_user_edits"]
         assert "main.py" in result["restored_files"]
 
+    def test_safe_restore_finds_ledger_under_walked_key(self, mgr, work_dir, tmp_path):
+        base = self._checkpoint(mgr, work_dir)
+
+        # A generic project marker in an ancestor dir (e.g. a stray
+        # package.json in /tmp) makes record_agent_write's marker walk key
+        # the ledger to the ancestor, while restore reads the exact dir's
+        # hash.  Safe restore must still find the ledger, or it silently
+        # degrades to a full restore and overwrites user edits.
+        (tmp_path / "package.json").write_text("{}\n")
+
+        (work_dir / "main.py").write_text("agent version\n")
+        (work_dir / "README.md").write_text("user hand edit\n")
+        mgr.record_agent_write(str(work_dir / "main.py"))
+
+        result = mgr.restore(str(work_dir), base, safe=True)
+        assert result["success"] is True
+        assert (work_dir / "main.py").read_text() == "print('hello')\n"
+        assert (work_dir / "README.md").read_text() == "user hand edit\n"
+        assert result["restored_files"] == ["main.py"]
+        assert "README.md" in result["skipped_user_edits"]
+
     def test_safe_restore_skips_file_user_edited_after_agent(self, mgr, work_dir):
         base = self._checkpoint(mgr, work_dir)
 
diff --git a/tools/checkpoint_manager.py b/tools/checkpoint_manager.py
index 790e5124b8..bee7ab758d 100644
--- a/tools/checkpoint_manager.py
+++ b/tools/checkpoint_manager.py
@@ -638,6 +638,19 @@ class CheckpointManager:
             return {"success": False, "error": f"Could not compute changed files: {err}"}
 
         ledger = _load_ledger(p.store, p.dir_hash)
+        if not ledger:
+            # record_agent_write keys the ledger by the marker-walk result
+            # (get_working_dir_for_path), which can diverge from the dir this
+            # restore names — e.g. a markerless project dir under an ancestor
+            # that carries a generic project marker, or a restore of a repo
+            # subdir while the write was recorded at the repo root.  Fall back
+            # to the walked key so safe restore never silently degrades to a
+            # full restore (which overwrites user edits) just because the
+            # ledger landed under a sibling key.
+            walked = self.get_working_dir_for_path(working_dir)
+            walk_key = _project_hash(walked)
+            if walk_key != p.dir_hash:
+                ledger = _load_ledger(p.store, walk_key)
         if not ledger:
             return {"success": True, "restore": [], "skipped": [], "ledger_empty": True}
         out: Dict[str, List[str]] = {"restore": [], "skipped": []}