fix(cli): stop the memory back-out from looping on the picker

raven/cli/onboard_commands.py

@@ -2687,20 +2687,27 @@ def _everos_section(section: str) -> dict[str, Any]:
     return load_everos_config().get(section, {}) or {}
 
 
+def _everos_role_configured(section: str) -> bool:
+    """True iff the user really configured this EverOS role.
+
+    Sole criterion for "configured", shared by every caller: model AND api_key.
+    The shipped everos.toml template seeds each section's model name with an
+    empty api_key, so a model alone also holds on a fresh install — two callers
+    disagreeing on this made the wizard's Back loop on itself forever.
+    """
+    sec = _everos_section(section)
+    return bool(sec.get("model") and sec.get("api_key"))
+
+
 def _memory_enabled() -> bool:
     """True iff EverOS memory is both selected AND usable on disk.
 
-    "Usable" requires both required models (llm and embedding) with api_key
-    in the EverOS toml. The seed/schema default sets model names but leaves
-    api_key empty, so checking model alone would mis-report a fresh,
-    unconfigured install as "enabled".
+    "Usable" requires both required roles (llm and embedding) configured.
     """
     data = _load_raw_config()
     if (data.get("memory") or {}).get("backend") != "everos":
         return False
-    llm = _everos_section("llm")
-    emb = _everos_section("embedding")
-    return bool(llm.get("model") and llm.get("api_key") and emb.get("model") and emb.get("api_key"))
+    return _everos_role_configured("llm") and _everos_role_configured("embedding")
 
 
 # Providers whose main model can be reused as the EverOS memory LLM: they
@@ -3691,8 +3698,7 @@ def _config_everos_role(
     )
 
     while True:  # role-menu loop — a back-out of the source picker returns here
-        sec = _everos_section(section)
-        current = sec.get("model") if sec.get("api_key") else None
+        current = _everos_section(section).get("model") if _everos_role_configured(section) else None
         if current:
             choices = [
                 questionary.Choice(_t(f"Keep current: {current}", f"沿用当前:{current}"), value="keep"),
@@ -3740,10 +3746,10 @@ def _config_everos_role(
             recommendation=role.get("recommendation"),
         )
         if result is _BACK:
-            if optional or _everos_section(section).get("model"):
-                # Optional roles offer Skip; a required role that already has a
-                # value on disk falls back to its keep/reconfigure menu. Either
-                # way, re-show the role menu rather than forcing the give-up exit.
+            if optional or _everos_role_configured(section):
+                # Optional roles offer Skip; a required role already configured
+                # falls back to its keep/reconfigure menu. Either way, re-show
+                # the role menu rather than forcing the give-up exit.
                 continue
             # A required role with nothing configured has no Skip, so backing out
             # of the picker would loop forever. Offer a bounded exit: keep trying,

tests/test_cli_onboard_commands.py

@@ -1281,6 +1281,51 @@ def ask(self):
     assert everos["rerank"]["base_url"] == "https://api.deepinfra.com/v1/inference"
 
 
+def test_memory_seeded_role_is_not_configured(tmp_env: Path, everos_isolated: Path) -> None:
+    """A seeded model with an empty api_key does not count as configured."""
+    from raven.config.update_everos import set_everos_section
+
+    assert onboard_commands._everos_role_configured("llm") is False
+    set_everos_section("llm", {"model": "openai/gpt-4.1-mini", "api_key": ""})
+    assert onboard_commands._everos_role_configured("llm") is False
+    set_everos_section("llm", {"api_key": "sk-real"})
+    assert onboard_commands._everos_role_configured("llm") is True
+
+
+def test_memory_required_role_back_reaches_give_up_menu(
+    tmp_env: Path, everos_isolated: Path, monkeypatch: pytest.MonkeyPatch
+) -> None:
+    """Backing out of the picker must offer the give-up exit even when the
+    shipped everos.toml template already seeded a model with an empty api_key —
+    otherwise Back re-asks the provider picker forever."""
+    from raven.config.update_everos import set_everos_section
+
+    set_everos_section(
+        "llm",
+        {"model": "openai/gpt-4.1-mini", "api_key": "", "base_url": "https://openrouter.ai/api/v1"},
+    )
+
+    import questionary
+
+    asked: list[str] = []
+
+    class _FQ:
+        def __init__(self, *a: object, **kw: object) -> None:
+            values = [getattr(c, "value", None) for c in kw.get("choices", [])]  # type: ignore[union-attr]
+            self._can_abort = "abort" in values
+
+        def ask(self) -> object:
+            asked.append("give-up" if self._can_abort else "picker")
+            assert len(asked) <= 4, f"Back re-asked the picker instead of offering an exit: {asked}"
+            return "abort" if self._can_abort else onboard_commands._BACK
+
+    monkeypatch.setattr(questionary, "select", _FQ)
+
+    out = onboard_commands._config_everos_role(section="llm", main_model=None, non_interactive=False, warnings=[])
+    assert out is onboard_commands._ABORT_EVEROS
+    assert asked == ["picker", "give-up"]
+
+
 def test_model_openai_compatible_heuristic(tmp_env: Path) -> None:
     """Compat heuristic gates whether the memory LLM can reuse the main model."""
     f = onboard_commands._model_is_openai_compatible