made sure exp 3 prompts interface with the existing model interface m…

.gitignore

@@ -12,6 +12,12 @@ build/
 venv/
 env/
 
+# IDE configurations
+.vscode/
+.idea/
+*.sublime-project
+*.sublime-workspace
+
 # Test / tooling caches
 .pytest_cache/
 .mypy_cache/

interface/config.py

@@ -10,7 +10,7 @@ class ExperimentConfig:
 
     prompting: Literal["minimal", "standard", "verbose"] = "standard"
     observation: Literal["text_only", "image_text", "image_only"] = "image_text"
-    context_window: Literal["current", "last3"] = "last3"
+    context_window: Literal["current", "last3"] = "current"
     querying: Literal["step_by_step", "subgoal", "full_trajectory"] = "step_by_step"
     chat_history: Literal["stateless", "rolling", "full"] = "stateless"
     chat_turns_max: int = 3

interface/coords.py

@@ -4,6 +4,7 @@
 
 from gridworld.backends.base import GridState
 from gridworld.task_spec import Position, TaskSpecification
+from prompting_experiments.prompt_templates import observation as observation_templates
 
 FACING_ORDER = ["NORTH", "EAST", "SOUTH", "WEST"]
 
@@ -126,29 +127,42 @@ def describe_cell(
     cols: int,
 ) -> str:
     if row < 1 or row > rows or col < 1 or col > cols:
-        return "out of bounds"
+        return observation_templates.CELL_OUT_OF_BOUNDS
     if (row, col) in walls:
-        return "wall"
+        return observation_templates.CELL_WALL
     if (row, col) == goal:
-        return f"GOAL ({row},{col})"
+        return observation_templates.CELL_GOAL.format(row=row, col=col)
 
     key_color = key_at_cell(task_spec, state, row, col)
     if key_color:
-        return f"{key_color} key ({row},{col})"
+        return observation_templates.CELL_KEY.format(
+            key_color=key_color,
+            row=row,
+            col=col,
+        )
 
     for door in task_spec.mechanisms.doors:
         if to_row_col(door.position) == (row, col):
             status = "open" if door.id in state.open_doors else door.initial_state
-            return f"{status} {door.requires_key} door ({row},{col})"
+            return observation_templates.CELL_DOOR.format(
+                status=status,
+                requires_key=door.requires_key,
+                row=row,
+                col=col,
+            )
 
     for gate in task_spec.mechanisms.gates:
         if to_row_col(gate.position) == (row, col):
             cur = "open" if gate.id in state.open_gates else gate.initial_state
-            return f"{cur} gate ({row},{col})"
+            return observation_templates.CELL_GATE.format(state=cur, row=row, col=col)
 
     for switch in task_spec.mechanisms.switches:
         if to_row_col(switch.position) == (row, col):
             on_off = "on" if switch.id in state.active_switches else switch.initial_state
-            return f"switch ({on_off}) ({row},{col})"
+            return observation_templates.CELL_SWITCH.format(
+                state=on_off,
+                row=row,
+                col=col,
+            )
 
-    return f"open ({row},{col})"
+    return observation_templates.CELL_OPEN.format(row=row, col=col)

interface/feedback.py

@@ -15,6 +15,7 @@
     switch_at_cell,
     switches_controlling_gate,
 )
+from prompting_experiments.prompt_templates import feedback as feedback_templates
 
 
 def infer_step_outcome(
@@ -35,13 +36,17 @@ def infer_step_outcome(
         door = next((d for d in task_spec.mechanisms.doors if d.id == door_id), None)
         color = door.requires_key if door else "matching"
         if action == "MOVE_FORWARD" and prev_pos != curr_pos:
-            return "OPENED", f"Opened {color} door {door_id} and moved to {curr_pos}."
-        return "OPENED", f"Opened {color} door {door_id}."
+            return "OPENED", feedback_templates.OPENED_AND_MOVED.format(
+                color=color,
+                door_id=door_id,
+                position=curr_pos,
+            )
+        return "OPENED", feedback_templates.OPENED_DOOR.format(color=color, door_id=door_id)
 
     if action in ("TURN_LEFT", "TURN_RIGHT"):
         if prev.agent_direction != curr.agent_direction:
-            return "TURNED", f"Now facing {agent_facing(curr)}."
-        return "NOTHING", f"{action} had no effect."
+            return "TURNED", feedback_templates.NOW_FACING.format(facing=agent_facing(curr))
+        return "NOTHING", feedback_templates.ACTION_NO_EFFECT.format(action=action)
 
     if action == "MOVE_FORWARD":
         if prev_pos == curr_pos:
@@ -50,9 +55,10 @@ def infer_step_outcome(
             if key_color:
                 return (
                     "BLOCKED",
-                    f"MOVE_FORWARD blocked by a {key_color} key at {fwd}. "
-                    "Keys occupy their cell; you cannot walk onto them. "
-                    "Face the key and use PICKUP from your current cell.",
+                    feedback_templates.MOVE_BLOCKED_BY_KEY.format(
+                        key_color=key_color,
+                        position=fwd,
+                    ),
                 )
             gate = gate_at_cell(task_spec, prev, fwd[0], fwd[1])
             if gate and not gate["open"]:
@@ -61,33 +67,39 @@ def infer_step_outcome(
                     switch_list = ", ".join(controllers)
                     return (
                         "BLOCKED",
-                        f"MOVE_FORWARD blocked by closed gate {gate['id']} at {fwd}. "
-                        f"Activate switch(es) {switch_list} to open it.",
+                        feedback_templates.MOVE_BLOCKED_BY_GATE_WITH_SWITCHES.format(
+                            gate_id=gate["id"],
+                            position=fwd,
+                            switches=switch_list,
+                        ),
                     )
                 return (
                     "BLOCKED",
-                    f"MOVE_FORWARD blocked by closed gate {gate['id']} at {fwd}.",
+                    feedback_templates.MOVE_BLOCKED_BY_GATE.format(
+                        gate_id=gate["id"],
+                        position=fwd,
+                    ),
                 )
-            return "BLOCKED", "MOVE_FORWARD blocked by wall or closed door/gate."
+            return "BLOCKED", feedback_templates.MOVE_BLOCKED_GENERIC
         if terminated and reward > 0 and curr_pos == goal:
-            return "DONE", f"Reached goal at {goal}."
-        return "MOVED", f"Moved to {curr_pos}."
+            return "DONE", feedback_templates.REACHED_GOAL.format(goal=goal)
+        return "MOVED", feedback_templates.MOVED_TO.format(position=curr_pos)
 
     if action == "PICKUP":
         if (
             prev.agent_carrying != curr.agent_carrying
             or len(curr.collected_keys) > len(prev.collected_keys)
         ):
             carried = curr.agent_carrying or "a"
-            return "PICKUP", f"Picked up {carried} key."
-        return "NOTHING", "Nothing to pick up here."
+            return "PICKUP", feedback_templates.PICKED_UP_KEY.format(key_color=carried)
+        return "NOTHING", feedback_templates.NOTHING_TO_PICK_UP
 
     if action == "TOGGLE":
         if (
             prev.active_switches != curr.active_switches
             or prev.open_gates != curr.open_gates
         ):
-            return "TOGGLED", "Toggled switch or gate state changed."
+            return "TOGGLED", feedback_templates.TOGGLED_STATE_CHANGED
         fwd = forward_cell(prev)
         switch_ahead = switch_at_cell(task_spec, fwd[0], fwd[1])
         switch_here = switch_at_cell(task_spec, prev_pos[0], prev_pos[1])
@@ -96,34 +108,34 @@ def infer_step_outcome(
             if switch_ahead["switch_type"] == "hold":
                 return (
                     "NOTHING",
-                    f"TOGGLE had no effect. MOVE_FORWARD onto the switch at {fwd} "
-                    "(hold switches activate while you stand on them).",
+                    feedback_templates.TOGGLE_HOLD_SWITCH_HINT.format(position=fwd),
                 )
             return (
                 "NOTHING",
-                f"TOGGLE had no effect. MOVE_FORWARD onto the switch at {fwd}, then TOGGLE.",
+                feedback_templates.TOGGLE_SWITCH_HINT.format(position=fwd),
             )
         if gate_ahead and not gate_ahead["open"]:
             controllers = switches_controlling_gate(task_spec, str(gate_ahead["id"]))
             if controllers:
                 switch_list = ", ".join(controllers)
                 return (
                     "NOTHING",
-                    "Gates cannot be toggled directly. "
-                    f"Activate switch(es) {switch_list} instead.",
+                    feedback_templates.GATE_TOGGLE_WITH_SWITCHES.format(
+                        switches=switch_list,
+                    ),
                 )
-            return "NOTHING", "Gates cannot be toggled directly. Activate a linked switch instead."
+            return "NOTHING", feedback_templates.GATE_TOGGLE_GENERIC
         return (
             "NOTHING",
-            "TOGGLE had no effect. Stand on a switch and TOGGLE, or use PICKUP/keys for doors.",
+            feedback_templates.TOGGLE_NO_EFFECT,
         )
 
     if action == "DONE":
         if terminated and reward > 0 and curr_pos == goal:
-            return "DONE", f"Task complete at {goal}."
-        return "WRONG_DONE", f"DONE called but not at goal {goal}."
+            return "DONE", feedback_templates.TASK_COMPLETE.format(goal=goal)
+        return "WRONG_DONE", feedback_templates.WRONG_DONE.format(goal=goal)
 
-    return "INVALID", f"Unknown or unsupported action {action}."
+    return "INVALID", feedback_templates.UNKNOWN_ACTION.format(action=action)
 
 
 def format_step_feedback(
@@ -139,23 +151,27 @@ def format_step_feedback(
     )
     prev_pos = agent_row_col(prev)
     if event_type == "BLOCKED":
-        return f"BLOCKED — {action}: {event_message} You remain at {prev_pos}.", event_type
+        return feedback_templates.BLOCKED_FEEDBACK.format(action=action, message=event_message, position=prev_pos), event_type
     if event_type == "TURNED":
-        return f"TURNED — {action}: {event_message}", event_type
+        return feedback_templates.TURNED_FEEDBACK.format(action=action, message=event_message), event_type
     if event_type == "MOVED":
-        return f"MOVED — {action}: {event_message}", event_type
+        return feedback_templates.MOVED_FEEDBACK.format(action=action, message=event_message), event_type
     if event_type == "DONE":
-        return f"SUCCESS — {action}: {event_message}", event_type
+        return feedback_templates.SUCCESS_FEEDBACK.format(action=action, message=event_message), event_type
     if event_type == "PICKUP":
-        return f"PICKUP — {action}: {event_message}", event_type
+        return feedback_templates.PICKUP_FEEDBACK.format(action=action, message=event_message), event_type
     if event_type == "NOTHING":
-        return f"NOTHING — {action}: {event_message} You remain at {prev_pos}.", event_type
+        return feedback_templates.NOTHING_FEEDBACK.format(action=action, message=event_message, position=prev_pos), event_type
     if event_type == "OPENED":
-        return f"OPENED — {action}: {event_message}", event_type
+        return feedback_templates.OPENED_FEEDBACK.format(action=action, message=event_message), event_type
     if event_type == "TOGGLED":
-        return f"TOGGLED — {action}: {event_message}", event_type
+        return feedback_templates.TOGGLED_FEEDBACK.format(action=action, message=event_message), event_type
     if event_type == "WRONG_DONE":
-        return f"WRONG DONE — {action}: {event_message} You remain at {prev_pos}.", event_type
+        return feedback_templates.WRONG_DONE_FEEDBACK.format(action=action, message=event_message, position=prev_pos), event_type
     if event_type == "INVALID":
-        return f"INVALID — {action}: {event_message} You remain at {prev_pos}.", event_type
-    return f"{event_type} — {action}: {event_message}", event_type
+        return feedback_templates.INVALID_FEEDBACK.format(action=action, message=event_message, position=prev_pos), event_type
+    return feedback_templates.DEFAULT_FEEDBACK.format(
+        event_type=event_type,
+        action=action,
+        message=event_message,
+    ), event_type

interface/observation.py

@@ -19,6 +19,7 @@
 from gridworld.task_spec import TaskSpecification
 
 from interface.renderer import render_user_observation_text, rgb_to_image_block
+from prompting_experiments.prompt_templates import observation as observation_templates
 
 ObservationMode = Literal["text_only", "image_text", "image_only"]
 ContextWindow = Literal["current", "last3"]
@@ -51,11 +52,17 @@ def history_text(
     if not recs:
         return ""
 
-    lines = ["Recent history (last 3 steps, oldest first):"]
+    lines = [observation_templates.RECENT_HISTORY_HEADER]
     for rec in recs:
         row, col = rec["position_after"]
         lines.append(
-            f"  ({int(row)}, {int(col)}) facing {rec['facing_after']} -> {rec['action']} -> {rec['prompt_feedback']}"
+            observation_templates.RECENT_HISTORY_STEP.format(
+                row=int(row),
+                col=int(col),
+                facing=rec["facing_after"],
+                action=rec["action"],
+                feedback=rec["prompt_feedback"],
+            )
         )
     return "\n".join(lines)
 
@@ -78,16 +85,22 @@ def history_content_blocks(
             continue
         blocks.append(rgb_to_image_block(rgb))
         if observation == "image_only":
-            blocks.append({"type": "text", "text": f"Action: {rec['action']}\n\n"})
+            blocks.append(
+                {
+                    "type": "text",
+                    "text": observation_templates.IMAGE_HISTORY_ACTION.format(
+                        action=rec["action"]
+                    ),
+                }
+            )
 
     if not blocks:
         return []
 
     intro = (
-        "Recent steps (oldest first). Each image is the maze view from which the "
-        "following action was chosen; infer pose and environment state from the image.\n\n"
+        observation_templates.IMAGE_ONLY_HISTORY_INTRO
         if observation == "image_only"
-        else "Recent step views (oldest first):\n\n"
+        else observation_templates.IMAGE_TEXT_HISTORY_INTRO
     )
     return [{"type": "text", "text": intro}] + blocks