security: warn on model swap; correct docs and CLI credential guidance

docs/sleep/README.md

@@ -17,7 +17,7 @@ normal agent requests.
 One "night":
 
 ```
-harvest Claude Code / Codex / Cursor transcripts → mine recurring tasks → replay in isolated model calls
+harvest Claude Code / Codex / Cursor transcripts → mine recurring tasks → replay via the configured backend (isolation varies by backend; mock/handoff make no network calls)
    → consolidate (reflect → bounded edit → GATE on real held-out tasks)
    → stage proposal → (you) adopt
 ```

scripts/train.py

@@ -383,8 +383,24 @@ def parse_args() -> argparse.Namespace:
 
 def load_config(args: argparse.Namespace) -> dict:
     """Load config with _base_ inheritance, then apply CLI overrides."""
+    import warnings
     from skillopt.config import load_config as _load, flatten_config, is_structured
 
+    # F08: Warn when API keys are supplied on the CLI — prefer env vars or managed identity.
+    for _cli_key in (
+        "azure_api_key", "azure_openai_api_key",
+        "optimizer_azure_openai_api_key", "target_azure_openai_api_key",
+        "minimax_api_key",
+    ):
+        if getattr(args, _cli_key, None):
+            warnings.warn(
+                f"--{_cli_key} is deprecated: provide credentials via the "
+                f"AZURE_OPENAI_API_KEY environment variable or set "
+                f"--azure_openai_auth_mode=managed_identity instead.",
+                DeprecationWarning,
+                stacklevel=2,
+            )
+
     cfg = _load(args.config, overrides=args.cfg_options)
     structured = is_structured(cfg)
 

skillopt_sleep/cycle.py

@@ -30,6 +30,34 @@
 from skillopt_sleep.types import SessionDigest, SleepReport, TaskRecord
 
 
+# ── Model-swap detection (F16) ───────────────────────────────
+def _make_model_key(cfg: SleepConfig) -> str:
+    """Stable string identifying the backend+model combination for this cycle."""
+    return "{}::{}".format(
+        cfg.get("backend", "mock"),
+        cfg.get("model", ""),
+    )
+
+
+def _check_model_change(cfg: SleepConfig, state: SleepState) -> None:
+    """Warn when the backend/model has changed since the last night.
+
+    Skill text is backend-specific; adopting edits from a different model's
+    reflections into a new model's skill file can cause regressions.
+    This is advisory only — the cycle continues either way.
+    """
+    current_key = _make_model_key(cfg)
+    prior_key = state.last_model_key
+    if prior_key and prior_key != current_key:
+        print(
+            f"[sleep] WARNING: model changed since last night "
+            f"(was {prior_key!r}, now {current_key!r}). "
+            "Learned skill text may not transfer cleanly. "
+            "Consider starting from a fresh skill document.",
+            file=sys.stderr,
+        )
+
+
 @dataclass
 class CycleOutcome:
     report: SleepReport
@@ -127,6 +155,7 @@ def run_sleep_cycle(
     """
     cfg = cfg or load_config()
     state = SleepState.load(cfg.state_path)
+    _check_model_change(cfg, state)  # F16: warn if model changed between nights
     night = state.begin_night(clock)
     project = _project_paths(cfg)
     started = _now_iso(clock)
@@ -409,6 +438,7 @@ def run_sleep_cycle(
             "baseline": result.baseline_score, "candidate": result.candidate_score,
             "n_tasks": len(tasks), "staging": staging_dir,
         })
+        state.set_last_model_key(_make_model_key(cfg))  # F16: track model for next night
         # ── 6. adopt (opt-in) ────────────────────────────────────────────
         if cfg.get("auto_adopt") and result.accepted:
             adopted_paths = adopt_staging(staging_dir)

skillopt_sleep/state.py

@@ -29,6 +29,7 @@ def _now_iso(clock: Optional[float] = None) -> str:
     "slow_memory": "",      # cross-night consolidated lessons (meta-skill analogue)
     "history": [],          # list of per-night summaries
     "task_archive": [],     # capped list of past mined tasks (for associative recall)
+    "last_model_key": "",   # "backend::model" string used in the last successful night (F16)
 }
 
 
@@ -94,3 +95,11 @@ def add_to_archive(self, task_dicts: list, cap: int = 300) -> None:
         arc.extend(task_dicts)
         if len(arc) > cap:
             self.data["task_archive"] = arc[-cap:]
+
+    # ── model-swap tracking (F16) ─────────────────────────────────────────
+    @property
+    def last_model_key(self) -> str:
+        return str(self.data.get("last_model_key", ""))
+
+    def set_last_model_key(self, key: str) -> None:
+        self.data["last_model_key"] = key

tests/test_model_change_warning.py

@@ -0,0 +1,66 @@
+"""Tests for the F16 model-change warning and F08 CLI credential warning."""
+from __future__ import annotations
+
+import argparse
+import warnings
+
+from skillopt_sleep.cycle import _check_model_change, _make_model_key
+from skillopt_sleep.config import load_config
+from skillopt_sleep.state import SleepState
+
+
+def _cfg():
+    cfg = load_config()
+    return cfg
+
+
+def test_last_model_key_roundtrips(tmp_path) -> None:
+    state = SleepState.load(str(tmp_path / "state.json"))
+    assert state.last_model_key == ""
+    state.set_last_model_key("anthropic::claude")
+    assert state.last_model_key == "anthropic::claude"
+
+
+def test_warns_when_model_changed(tmp_path, capsys) -> None:
+    cfg = _cfg()
+    state = SleepState.load(str(tmp_path / "state.json"))
+    state.set_last_model_key("some-other::model")
+    _check_model_change(cfg, state)
+    err = capsys.readouterr().err
+    assert "model changed since last night" in err
+
+
+def test_no_warning_on_first_night(tmp_path, capsys) -> None:
+    cfg = _cfg()
+    state = SleepState.load(str(tmp_path / "state.json"))  # last_model_key == ""
+    _check_model_change(cfg, state)
+    assert "model changed" not in capsys.readouterr().err
+
+
+def test_no_warning_when_model_same(tmp_path, capsys) -> None:
+    cfg = _cfg()
+    state = SleepState.load(str(tmp_path / "state.json"))
+    state.set_last_model_key(_make_model_key(cfg))
+    _check_model_change(cfg, state)
+    assert "model changed" not in capsys.readouterr().err
+
+
+def test_cli_api_key_emits_deprecation_warning() -> None:
+    from scripts.train import load_config as train_load_config
+
+    args = argparse.Namespace(
+        config="configs/_base_/default.yaml",
+        cfg_options=None,
+        azure_openai_api_key="sk-secret-value",
+    )
+    with warnings.catch_warnings(record=True) as caught:
+        warnings.simplefilter("always")
+        try:
+            train_load_config(args)
+        except Exception:
+            pass  # config loading may fail; we only assert the warning fired
+    assert any(
+        issubclass(w.category, DeprecationWarning)
+        and "azure_openai_api_key" in str(w.message)
+        for w in caught
+    )