feat(discovery): source available actions from scenario.action_set (interact reaches the agent)

proteus/game/runtime/_session_core.py

@@ -155,7 +155,7 @@ def build_observation(
     parts.append("Now:")
     parts.append(render_ascii(scenario, game))
     parts.append(legend_text(legend))
-    parts.append(f"Available actions: [{', '.join(_ACTIONS)}]")
+    parts.append(f"Available actions: [{', '.join(scenario.action_set)}]")
     return "\n".join(parts)
 
 

proteus/game/runtime/interactive.py

@@ -74,7 +74,7 @@ class InteractiveSession:
             "play_turns": self._play_turns,
             "grid": core.grid_to_list(self._game.current_grid()),
             "legend": {str(k): v for k, v in self._scenario.legend().items()},
-            "actions": list(core._ACTIONS),
+            "actions": list(self._scenario.action_set),
             "outcome": None,
             "cut_frames": self._cut_grids if played == 0 else None,
             "review": None,
@@ -103,9 +103,9 @@ class InteractiveSession:
     def step(self, action: str, probe_answer: str = "") -> dict:
         if self._is_done():
             raise core.SessionFinishedError("session already finished")
-        if action not in core._ACTIONS:
+        if action not in self._scenario.action_set:
             raise ValueError(
-                f"invalid action {action!r}; choose one of {core._ACTIONS}"
+                f"invalid action {action!r}; choose one of {self._scenario.action_set}"
             )
         turn_idx = len(self._turns) + 1
         observation = core.build_observation(

proteus/game/runtime/session.py

@@ -92,7 +92,7 @@ class SessionRunner:
                     probe_thinking_tokens=probe.thinking_tokens,
                 )
 
-            result = self._agent.act(observation, list(core._ACTIONS), system_prompt)
+            result = self._agent.act(observation, list(scenario.action_set), system_prompt)
             turns.append(core.make_turn_trace(
                 scenario, game,
                 turn_idx=turn_idx,

tests/runtime/test_errand_discovery.py

@@ -64,3 +64,21 @@ def test_make_turn_trace_no_self_report_is_none():
     tt = core.make_turn_trace(scen, game, turn_idx=1, observation="o",
                               action="stay", raw_text="ACTION: stay")
     assert tt.self_belief is None and tt.self_correct is None
+
+
+def test_observation_lists_scenario_action_set():
+    scen, game = _errand()
+    obs = core.build_observation(scen, game, cut_frames=[], turn_idx=1)
+    assert "interact" in obs  # errand_runner exposes interact in Available actions
+    # template stays at the 5 movement actions (regression)
+    t = get_scenario("template")()
+    tgame = MotiveGridGame(t, random.Random(0), Difficulty.EASY, max_steps=10)
+    tobs = core.build_observation(t, tgame, cut_frames=[], turn_idx=1)
+    assert "interact" not in tobs
+
+
+def test_interactive_accepts_interact_for_errand():
+    from proteus.game.runtime.interactive import InteractiveSession
+    s = InteractiveSession("errand_runner", seed=7, play_turns=5, use_default_memory=False)
+    assert "interact" in s.state()["actions"]
+    s.step("interact")  # must not raise "invalid action"