feat(autonomy): planner gate for non-Claude providers at level off (P3.T2)

apps/backend/agents/planner.py

@@ -9,9 +9,11 @@
 from pathlib import Path
 
 from analysis.prevention_scanner import PreventionScanner
+from core.autonomy_level import AutonomyLevel, resolve_autonomy_settings
 from core.providers import create_engine_provider
 from core.providers.base import SessionConfig
 from core.providers.config import ProviderConfig, get_provider_config
+from core.providers.exceptions import ProviderError
 from implementation_plan import ImplementationPlan
 from phase_config import get_phase_model, get_phase_thinking_budget
 from phase_event import ExecutionPhase, emit_phase
@@ -60,6 +62,28 @@
 logger = logging.getLogger(__name__)
 
 
+def _assert_planner_autonomy_allows(provider_name: str) -> None:
+    """Block non-Claude planner sessions when autonomy is fully off (P3·T2).
+
+    ``AUTO_CODE_AUTONOMY=off`` promises "no autonomous non-Claude execution";
+    the planner writes implementation_plan.json, so the session must be
+    refused before it exists rather than degraded afterwards.
+
+    Raises:
+        ProviderError: When the resolved level is ``off`` and the effective
+            provider is not Claude.
+    """
+    if provider_name == "claude":
+        return
+    autonomy = resolve_autonomy_settings()
+    if autonomy.level is AutonomyLevel.OFF:
+        raise ProviderError(
+            f"AUTO_CODE_AUTONOMY=off blocks the non-Claude planner "
+            f"(provider {provider_name!r}). Set AUTO_CODE_AUTONOMY to claude, "
+            f"safe, or bold — or switch the planner provider to claude."
+        )
+
+
 def create_planner_session(
     project_dir: Path,
     spec_dir: Path,
@@ -108,6 +132,9 @@ def create_planner_session(
         )
 
     provider = create_engine_provider(config)
+    # Gate on the effective provider (after runner routing), not the raw
+    # config value, so a route to a different engine is still enforced.
+    _assert_planner_autonomy_allows(provider.name)
 
     # For Claude provider, pass provider-specific kwargs
     if provider.name == "claude":

docs/strategy/roadmap.md

@@ -82,7 +82,7 @@
 | # | Задача | Файлы | Критерий приёмки |
 |---|---|---|---|
 | T1 | ✅ **Уже реализовано.** Coder уважает уровень автономности | `agents/coder.py:1282–1296` + `agents/runtime/direct_api_autonomy.py:124` | `AUTO_CODE_AUTONOMY=claude` блокирует автономный direct-API coder; `safe` — через gate; `bold` — мимо gate (хардкод `=True` остался только в смоук-тесте `cli/provider_smoke_commands.py`) |
-| T2 | Гейт для planner на не-Claude провайдерах | `agents/planner.py` | При `off` не-Claude planner блокируется |
+| T2 | ✅ Гейт для planner на не-Claude провайдерах | `agents/planner.py` (`_assert_planner_autonomy_allows` в `create_planner_session`) | При `off` не-Claude planner блокируется `ProviderError` до создания сессии; Claude работает на любом уровне |
 | T3 | Фабрика принимает `autonomy_settings` и логирует уровень | `agents/runtime/adapters/__init__.py` (`create_runtime_session`) | Уровень автономности виден в логе сессии |
 | T4 | Один тумблер в UI + вывод уровня в JSON | `cli/runtime_commands.py`, настройки фронта | UI показывает и задаёт один селектор автономности |
 | T5 | Доки: вести с `AUTO_CODE_AUTONOMY`, 30+ переменных — в приложение | `guides/QUICK-START.md`, `docs/` | Quickstart: «поставь `safe` — готово» |
@@ -145,6 +145,7 @@
 - ✅ **P1·T4** — QA-экран отчёта доверия в task overview (reader + IPC + `VerificationReportPanel`, i18n en/fr).
 - ✅ **P3·T4** — тумблер автономности в настройках → инжектит `AUTO_CODE_AUTONOMY` в окружение сборки (явный env приоритетнее).
 - ✅ **P5·T2** — агрегация стоимости подключена: `save_token_stats()` пишет запись в `cost_report.json` спеки (роль+фаза+$) и обновляет проектный `.auto-claude/model_usage_summary.json` (`cost_by_phase` добавлен; попутно починены сериализация summary и `+00:00Z`-таймстампы записей).
+- ✅ **P3·T2** — гейт планнера: `AUTO_CODE_AUTONOMY=off` блокирует не-Claude planner (`ProviderError` до создания сессии, после runner-роутинга); Claude-путь не затронут.
 
 Отчёт доверия теперь несёт: **вердикт · тесты · дифф · out-of-scope · confidence · uncertainty** — и виден в UI.
 

tests/test_planner_autonomy_gate.py

@@ -0,0 +1,90 @@
+"""Tests for the planner autonomy gate (P3.T2).
+
+``AUTO_CODE_AUTONOMY=off`` must block non-Claude planner sessions before the
+session exists; Claude keeps working at every level, and claude/safe/bold
+keep non-Claude planners available.
+"""
+
+import sys
+from pathlib import Path
+
+import pytest
+
+sys.path.insert(0, str(Path(__file__).parent.parent / "apps" / "backend"))
+
+import agents.planner as planner_mod  # noqa: E402
+from agents.planner import (  # noqa: E402
+    _assert_planner_autonomy_allows,
+    create_planner_session,
+)
+from core.autonomy_level import AUTONOMY_LEVEL_ENV  # noqa: E402
+from core.providers.exceptions import ProviderError  # noqa: E402
+
+
+class TestAutonomyGateHelper:
+    def test_off_blocks_non_claude(self, monkeypatch):
+        monkeypatch.setenv(AUTONOMY_LEVEL_ENV, "off")
+        with pytest.raises(ProviderError, match="AUTO_CODE_AUTONOMY=off"):
+            _assert_planner_autonomy_allows("openai")
+
+    def test_off_keeps_claude_available(self, monkeypatch):
+        monkeypatch.setenv(AUTONOMY_LEVEL_ENV, "off")
+        _assert_planner_autonomy_allows("claude")  # must not raise
+
+    @pytest.mark.parametrize("level", ["claude", "safe", "bold"])
+    def test_other_levels_allow_non_claude(self, monkeypatch, level):
+        monkeypatch.setenv(AUTONOMY_LEVEL_ENV, level)
+        _assert_planner_autonomy_allows("openai")  # must not raise
+
+    def test_default_level_allows_non_claude(self, monkeypatch):
+        monkeypatch.delenv(AUTONOMY_LEVEL_ENV, raising=False)
+        _assert_planner_autonomy_allows("openai")  # default is claude, not off
+
+    def test_error_names_the_provider(self, monkeypatch):
+        monkeypatch.setenv(AUTONOMY_LEVEL_ENV, "off")
+        with pytest.raises(ProviderError, match="'ollama'"):
+            _assert_planner_autonomy_allows("ollama")
+
+
+class _FakeProvider:
+    """Provider stub: session creation must never be reached when gated."""
+
+    def __init__(self, name: str):
+        self.name = name
+        self.create_session_calls = 0
+
+    def create_session(self, *args, **kwargs):
+        self.create_session_calls += 1
+        return object()
+
+
+class TestCreatePlannerSessionGate:
+    def test_off_blocks_before_session_creation(self, monkeypatch, tmp_path):
+        monkeypatch.setenv(AUTONOMY_LEVEL_ENV, "off")
+        fake = _FakeProvider("openai")
+        monkeypatch.setattr(planner_mod, "create_engine_provider", lambda config: fake)
+
+        with pytest.raises(ProviderError, match="AUTO_CODE_AUTONOMY=off"):
+            create_planner_session(tmp_path, tmp_path / "spec")
+
+        assert fake.create_session_calls == 0
+
+    def test_safe_level_creates_non_claude_session(self, monkeypatch, tmp_path):
+        monkeypatch.setenv(AUTONOMY_LEVEL_ENV, "safe")
+        fake = _FakeProvider("openai")
+        monkeypatch.setattr(planner_mod, "create_engine_provider", lambda config: fake)
+
+        session = create_planner_session(tmp_path, tmp_path / "spec")
+
+        assert session is not None
+        assert fake.create_session_calls == 1
+
+    def test_off_still_creates_claude_session(self, monkeypatch, tmp_path):
+        monkeypatch.setenv(AUTONOMY_LEVEL_ENV, "off")
+        fake = _FakeProvider("claude")
+        monkeypatch.setattr(planner_mod, "create_engine_provider", lambda config: fake)
+
+        session = create_planner_session(tmp_path, tmp_path / "spec")
+
+        assert session is not None
+        assert fake.create_session_calls == 1