Merge pull request #277 from bogdanbaciu21/fix-57-unchanged-staging

fix(sleep): stage only documents changed by an accepted cycle

Author: Yif-Yang

skillopt_sleep/cycle.py

@@ -988,8 +988,24 @@ def run_sleep_cycle(
     adopted_paths: List[str] = []
     if not dry_run:
         _progress(cfg, "staging start")
-        proposed_skill = result.new_skill if (cfg.get("evolve_skill") and result.accepted) else None
-        proposed_memory = result.new_memory if (cfg.get("evolve_memory") and result.accepted) else None
+        proposed_skill = (
+            result.new_skill
+            if (
+                cfg.get("evolve_skill")
+                and result.accepted
+                and result.new_skill != skill
+            )
+            else None
+        )
+        proposed_memory = (
+            result.new_memory
+            if (
+                cfg.get("evolve_memory")
+                and result.accepted
+                and result.new_memory != memory
+            )
+            else None
+        )
         skill_proposals, skip_notes = _skill_proposals_from_groups(
             cfg,
             group_outcomes,

tests/test_sleep_engine.py

@@ -1564,6 +1564,114 @@ def test_cycle_pins_the_exact_managed_skill_and_memory_bytes_it_read(self):
                 os.path.realpath(memory_path),
             )
 
+    def _assert_only_changed_documents_are_staged(
+        self, new_skill, new_memory, expect_skill, expect_memory
+    ):
+        from skillopt_sleep.consolidate import ConsolidationResult
+
+        skill = "# managed baseline\nrule\n"
+        memory = "# memory baseline\npreference\n"
+        with tempfile.TemporaryDirectory() as proj, tempfile.TemporaryDirectory() as home:
+            target = os.path.join(proj, ".agents", "skills", "taste", "SKILL.md")
+            memory_path = os.path.join(proj, "CLAUDE.md")
+            os.makedirs(os.path.dirname(target), exist_ok=True)
+            with open(target, "w", encoding="utf-8") as handle:
+                handle.write(skill)
+            with open(memory_path, "w", encoding="utf-8") as handle:
+                handle.write(memory)
+            cfg = load_config(
+                invoked_project=proj,
+                projects="invoked",
+                backend="mock",
+                claude_home=os.path.join(home, ".claude"),
+                target_skill_path=target,
+                auto_adopt=False,
+            )
+            applied = []
+            if new_skill != skill:
+                applied.append(EditRecord("skill", "add", "sharpened rule"))
+            if new_memory != memory:
+                applied.append(EditRecord("memory", "add", "learned preference"))
+            result = ConsolidationResult(
+                accepted=True,
+                gate_action="accept_new_best",
+                baseline_score=0.1,
+                candidate_score=0.2,
+                new_skill=new_skill,
+                new_memory=new_memory,
+                applied_edits=applied,
+                rejected_edits=[],
+                holdout_baseline=0.1,
+                holdout_candidate=0.2,
+            )
+            tasks = assign_splits(
+                researcher_persona(), holdout_fraction=0.34, seed=42
+            )
+
+            with mock.patch(
+                "skillopt_sleep.cycle.dream_consolidate",
+                return_value=result,
+            ):
+                outcome = run_sleep_cycle(cfg, seed_tasks=tasks)
+
+            # Staging never edits the live documents; adoption stays explicit.
+            with open(target, encoding="utf-8") as handle:
+                self.assertEqual(handle.read(), skill)
+            with open(memory_path, encoding="utf-8") as handle:
+                self.assertEqual(handle.read(), memory)
+
+            with open(
+                os.path.join(outcome.staging_dir, "manifest.json"),
+                encoding="utf-8",
+            ) as handle:
+                manifest = json.load(handle)
+            # Manifest flags and artifact presence have to agree; a flag without
+            # its file (or a file without its flag) would break adoption.
+            self.assertEqual(manifest["has_managed_skill"], expect_skill)
+            self.assertEqual(manifest["has_managed_memory"], expect_memory)
+            self.assertEqual(
+                os.path.exists(
+                    os.path.join(outcome.staging_dir, "proposed_SKILL.md")
+                ),
+                expect_skill,
+            )
+            self.assertEqual(
+                os.path.exists(
+                    os.path.join(outcome.staging_dir, "proposed_CLAUDE.md")
+                ),
+                expect_memory,
+            )
+
+    def test_cycle_stages_only_documents_that_changed(self):
+        # The staging contract is byte/text equality, not semantic or whitespace
+        # normalized comparison: an accepted cycle proposes a document only when it
+        # actually rewrote it. Covered for every shape an accepted result can take,
+        # so a symmetric regression on the skill side cannot hide behind the
+        # memory-only case.
+        skill = "# managed baseline\nrule\n"
+        memory = "# memory baseline\npreference\n"
+        new_skill = skill + "prefer the shortest reproduction\n"
+        new_memory = memory + "learned preference\n"
+        cases = (
+            ("neither_changed", skill, memory, False, False),
+            ("skill_only", new_skill, memory, True, False),
+            ("memory_only", skill, new_memory, False, True),
+            ("both_changed", new_skill, new_memory, True, True),
+            # Whitespace-only is a real change under a byte-equality contract, so it
+            # is a positive case. If this ever fails, the comparison has started
+            # normalizing and the documented contract has silently moved.
+            ("whitespace_only_skill", skill + "\n", memory, True, False),
+            ("whitespace_only_memory", skill, memory + "  \n", False, True),
+        )
+        for name, candidate_skill, candidate_memory, expect_skill, expect_memory in cases:
+            with self.subTest(case=name):
+                self._assert_only_changed_documents_are_staged(
+                    candidate_skill,
+                    candidate_memory,
+                    expect_skill,
+                    expect_memory,
+                )
+
     def test_managed_skill_change_during_consolidation_refuses_the_night(self):
         from skillopt_sleep.consolidate import ConsolidationResult
         from skillopt_sleep.staging import StagingError, latest_staging