Fix #87: enforce intent→behavior accountability in re-reasoning

extropy/simulation/reasoning.py

@@ -330,8 +330,12 @@ def build_pass1_prompt(
         prompt_parts.extend(
             [
                 "",
-                f'Last time I said I intended to: "{context.prior_action_intent}". '
-                "Has anything changed?",
+                "## Follow Through Check",
+                "",
+                f'Last time I said I intended to: "{context.prior_action_intent}".',
+                "Did I actually do it? Be explicit:",
+                "- If yes: what exactly did I do, and what happened?",
+                "- If no: what stopped me (time, money, logistics, fear, competing priorities)?",
             ]
         )
 
@@ -444,6 +448,10 @@ def build_pass1_schema() -> dict[str, Any]:
                 "type": "string",
                 "description": "A single sentence capturing your core reaction (for your own memory).",
             },
+            "action_intent": {
+                "type": "string",
+                "description": "The concrete action you intend to take next. Use an empty string only if you truly have no plan.",
+            },
             "sentiment": {
                 "type": "number",
                 "minimum": -1.0,
@@ -487,6 +495,7 @@ def build_pass1_schema() -> dict[str, Any]:
             "private_thought",
             "public_statement",
             "reasoning_summary",
+            "action_intent",
             "sentiment",
             "conviction",
             "will_share",
@@ -622,6 +631,10 @@ def build_merged_schema(outcomes: OutcomeConfig) -> dict[str, Any]:
             "type": "string",
             "description": "A single sentence capturing your core reaction (for your own memory).",
         },
+        "action_intent": {
+            "type": "string",
+            "description": "The concrete action you intend to take next. Use an empty string only if you truly have no plan.",
+        },
         "sentiment": {
             "type": "number",
             "minimum": -1.0,
@@ -664,6 +677,7 @@ def build_merged_schema(outcomes: OutcomeConfig) -> dict[str, Any]:
         "reasoning",
         "public_statement",
         "reasoning_summary",
+        "action_intent",
         "sentiment",
         "conviction",
         "will_share",
@@ -754,6 +768,14 @@ def _sentiment_to_tone(sentiment: float) -> str:
         return "strongly opposed"
 
 
+def _normalize_action_intent(value: Any) -> str | None:
+    """Normalize action intent to None when empty."""
+    if value is None:
+        return None
+    text = str(value).strip()
+    return text or None
+
+
 # =============================================================================
 # Two-pass reasoning (async)
 # =============================================================================
@@ -844,6 +866,7 @@ async def _reason_agent_two_pass_async(
     conviction_score = pass1_response.get("conviction")
     will_share = pass1_response.get("will_share", False)
     actions = pass1_response.get("actions", [])
+    pass1_action_intent = _normalize_action_intent(pass1_response.get("action_intent"))
 
     # Map conviction score (0-100) to float via bucketing
     conviction_float = score_to_conviction_float(conviction_score)
@@ -909,14 +932,19 @@ async def _reason_agent_two_pass_async(
     # Merge sentiment into outcomes for backwards compat
     if sentiment is not None:
         outcomes["sentiment"] = sentiment
+    action_intent = pass1_action_intent or _normalize_action_intent(
+        outcomes.get("action_intent")
+    )
+    if action_intent is not None:
+        outcomes["action_intent"] = action_intent
 
     return ReasoningResponse(
         position=position,
         sentiment=sentiment,
         conviction=conviction_float,
         public_statement=public_statement,
         reasoning_summary=reasoning_summary,
-        action_intent=outcomes.get("action_intent"),
+        action_intent=action_intent,
         will_share=will_share,
         reasoning=reasoning,
         outcomes=outcomes,
@@ -1014,6 +1042,7 @@ async def _reason_agent_merged_async(
     conviction_score = response.get("conviction")
     will_share = response.get("will_share", False)
     actions = response.get("actions", [])
+    action_intent = _normalize_action_intent(response.get("action_intent"))
 
     conviction_float = score_to_conviction_float(conviction_score)
 
@@ -1027,6 +1056,7 @@ async def _reason_agent_merged_async(
         "reasoning",
         "public_statement",
         "reasoning_summary",
+        "action_intent",
         "sentiment",
         "conviction",
         "will_share",
@@ -1037,14 +1067,16 @@ async def _reason_agent_merged_async(
     # Merge sentiment into outcomes for backwards compat
     if sentiment is not None:
         outcomes["sentiment"] = sentiment
+    if action_intent is not None:
+        outcomes["action_intent"] = action_intent
 
     return ReasoningResponse(
         position=position,
         sentiment=sentiment,
         conviction=conviction_float,
         public_statement=public_statement,
         reasoning_summary=reasoning_summary,
-        action_intent=outcomes.get("action_intent"),
+        action_intent=action_intent,
         will_share=will_share,
         reasoning=reasoning,
         outcomes=outcomes,
@@ -1138,6 +1170,7 @@ def reason_agent(
     sentiment = pass1_response.get("sentiment")
     conviction_score = pass1_response.get("conviction")
     will_share = pass1_response.get("will_share", False)
+    pass1_action_intent = _normalize_action_intent(pass1_response.get("action_intent"))
     conviction_float = score_to_conviction_float(conviction_score)
 
     # === Pass 2: Classification ===
@@ -1181,6 +1214,11 @@ def reason_agent(
 
     if sentiment is not None:
         outcomes["sentiment"] = sentiment
+    action_intent = pass1_action_intent or _normalize_action_intent(
+        outcomes.get("action_intent")
+    )
+    if action_intent is not None:
+        outcomes["action_intent"] = action_intent
 
     logger.info(
         f"[REASON] Agent {context.agent_id} - SUCCESS: position={position}, "
@@ -1193,7 +1231,7 @@ def reason_agent(
         conviction=conviction_float,
         public_statement=public_statement,
         reasoning_summary=reasoning_summary,
-        action_intent=outcomes.get("action_intent"),
+        action_intent=action_intent,
         will_share=will_share,
         reasoning=reasoning,
         outcomes=outcomes,

tests/test_reasoning_execution.py

@@ -0,0 +1,135 @@
+"""Execution tests for reasoning internals that affect memory/accountability."""
+
+import asyncio
+from datetime import datetime
+from unittest.mock import patch
+
+from extropy.core.llm import TokenUsage
+from extropy.core.models import ExposureRecord, ReasoningContext, SimulationRunConfig
+from extropy.core.models.scenario import (
+    Event,
+    EventType,
+    ExposureChannel,
+    ExposureRule,
+    InteractionConfig,
+    InteractionType,
+    OutcomeConfig,
+    OutcomeDefinition,
+    OutcomeType,
+    ScenarioMeta,
+    ScenarioSpec,
+    SeedExposure,
+    SimulationConfig,
+    SpreadConfig,
+    TimestepUnit,
+)
+from extropy.simulation.reasoning import _reason_agent_two_pass_async
+
+
+def _make_context() -> ReasoningContext:
+    return ReasoningContext(
+        agent_id="a1",
+        persona="I am a parent with limited time.",
+        event_content="School board policy update",
+        exposure_history=[
+            ExposureRecord(
+                timestep=0,
+                channel="broadcast",
+                content="Policy update",
+                credibility=0.9,
+                source_agent_id=None,
+            )
+        ],
+        agent_name="Taylor",
+    )
+
+
+def _make_scenario_no_pass2() -> ScenarioSpec:
+    return ScenarioSpec(
+        meta=ScenarioMeta(
+            name="reasoning_exec",
+            description="Reasoning execution test",
+            population_spec="test.yaml",
+            study_db="study.db",
+            population_id="default",
+            network_id="default",
+            created_at=datetime(2024, 1, 1),
+        ),
+        event=Event(
+            type=EventType.NEWS,
+            content="Policy update",
+            source="Board",
+            credibility=0.8,
+            ambiguity=0.3,
+            emotional_valence=0.0,
+        ),
+        seed_exposure=SeedExposure(
+            channels=[
+                ExposureChannel(
+                    name="broadcast",
+                    description="Broadcast",
+                    reach="broadcast",
+                )
+            ],
+            rules=[
+                ExposureRule(
+                    channel="broadcast",
+                    timestep=0,
+                    when="true",
+                    probability=1.0,
+                )
+            ],
+        ),
+        interaction=InteractionConfig(
+            primary_model=InteractionType.PASSIVE_OBSERVATION,
+            description="Observe updates",
+        ),
+        spread=SpreadConfig(share_probability=0.4),
+        outcomes=OutcomeConfig(
+            suggested_outcomes=[
+                OutcomeDefinition(
+                    name="notes",
+                    description="Open notes only",
+                    type=OutcomeType.OPEN_ENDED,
+                    required=False,
+                )
+            ]
+        ),
+        simulation=SimulationConfig(max_timesteps=2, timestep_unit=TimestepUnit.DAY),
+    )
+
+
+def test_action_intent_is_captured_without_pass2_outcome():
+    context = _make_context()
+    scenario = _make_scenario_no_pass2()
+    config = SimulationRunConfig(
+        scenario_path="test.yaml",
+        output_dir="test/",
+        max_retries=1,
+    )
+
+    async def _mock_simple_call_async(*args, **kwargs):
+        return (
+            {
+                "reasoning": "I need to show up and speak.",
+                "private_thought": "If I don't go, nothing changes.",
+                "public_statement": "I should attend the meeting.",
+                "reasoning_summary": "I need to attend and speak.",
+                "action_intent": "Attend the next board meeting",
+                "sentiment": -0.2,
+                "conviction": 75,
+                "will_share": True,
+                "actions": [],
+            },
+            TokenUsage(),
+        )
+
+    with patch(
+        "extropy.simulation.reasoning.simple_call_async",
+        new=_mock_simple_call_async,
+    ):
+        response = asyncio.run(_reason_agent_two_pass_async(context, scenario, config))
+
+    assert response is not None
+    assert response.action_intent == "Attend the next board meeting"
+    assert response.outcomes.get("action_intent") == "Attend the next board meeting"

tests/test_reasoning_prompts.py

@@ -368,6 +368,8 @@ def test_intent_accountability(self):
         prompt = build_pass1_prompt(context, scenario)
         assert "cancel my subscription" in prompt
         assert "intended to" in prompt.lower()
+        assert "did i actually do it" in prompt.lower()
+        assert "what stopped me" in prompt.lower()
 
     def test_no_intent_on_first_reasoning(self):
         """No intent accountability section when prior_action_intent is None."""
@@ -545,6 +547,7 @@ def test_has_required_fields(self):
             "reasoning",
             "public_statement",
             "reasoning_summary",
+            "action_intent",
             "sentiment",
             "conviction",
             "will_share",