fix: align step_reward with grade_episode, pin deps, update docs, clean inference

- grader.py: step_reward now applies skip-inv penalties matching grade_episode
- environment.py: passes investigation_done to grade() for honest step rewards
- requirements.txt: pinned to exact versions (fastapi, uvicorn, pydantic, etc.)
- TESTING.md: updated to v2.0.0, 20 tasks (4/6/6/4 across 4 difficulty tiers)
- inference.py: removed dead investigation_hints block (hints always empty)
- server/Dockerfile: synced with root Dockerfile (PYTHONPATH, ENV vars)

TESTING.md

@@ -30,7 +30,7 @@ curl -s http://localhost:8000/health
 
 **Expected output**
 ```json
-{"status": "ok", "environment": "payops_env", "version": "1.0.0"}
+{"status": "ok", "environment": "payops_env", "version": "2.0.0"}
 ```
 
 **Failure indicator:** Connection refused, or any field missing / wrong value.
@@ -63,7 +63,7 @@ curl -s http://localhost:8000/schema | python3 -m json.tool
 
 ## T-03  Tasks Endpoint
 
-**Goal:** Confirm all 12 tasks are returned with the correct difficulty distribution.
+**Goal:** Confirm all 20 tasks are returned with the correct difficulty distribution.
 
 ```bash
 curl -s http://localhost:8000/tasks | python3 -c "
@@ -75,30 +75,41 @@ c = Counter(t['difficulty'] for t in d['tasks'])
 print('By difficulty:', dict(c))
 print()
 for t in d['tasks']:
-    print(f\"  {t['task_id']:12} [{t['difficulty']:6}] correct={t['correct_action']}\")
+    print(f\"  {t['task_id']:12} [{t['difficulty']:8}] correct={t['correct_action']}\")
 "
 ```
 
 **Expected output**
 ```
-Total tasks: 12
-By difficulty: {'easy': 4, 'medium': 4, 'hard': 4}
-
-  EASY-001     [easy  ] correct=approve
-  EASY-002     [easy  ] correct=reject
-  EASY-003     [easy  ] correct=approve
-  EASY-004     [easy  ] correct=flag
-  MED-001      [medium] correct=escalate
-  MED-002      [medium] correct=hold
-  MED-003      [medium] correct=flag
-  MED-004      [medium] correct=flag
-  HARD-001     [hard  ] correct=escalate
-  HARD-002     [hard  ] correct=reject
-  HARD-003     [hard  ] correct=reject
-  HARD-004     [hard  ] correct=approve
-```
-
-**Failure indicator:** count != 12, missing difficulty tier, wrong correct_action.
+Total tasks: 20
+By difficulty: {'easy': 4, 'medium': 6, 'hard': 6, 'critical': 4}
+
+  EASY-001     [easy    ] correct=approve
+  EASY-002     [easy    ] correct=reject
+  EASY-003     [easy    ] correct=approve
+  EASY-004     [easy    ] correct=flag
+  MED-001      [medium  ] correct=escalate
+  MED-002      [medium  ] correct=hold
+  MED-003      [medium  ] correct=flag
+  MED-004      [medium  ] correct=flag
+  MED-005      [medium  ] correct=hold
+  MED-006      [medium  ] correct=escalate
+  HARD-001     [hard    ] correct=escalate
+  HARD-002     [hard    ] correct=reject
+  HARD-003     [hard    ] correct=reject
+  HARD-004     [hard    ] correct=approve
+  HARD-005     [hard    ] correct=escalate
+  HARD-006     [hard    ] correct=flag
+  CRIT-001     [critical] correct=approve
+  CRIT-002     [critical] correct=reject
+  CRIT-003     [critical] correct=escalate
+  CRIT-004     [critical] correct=reject
+```
+
+> Note: correct_action values for jitter-variant tasks (EASY-004, MED-001/003/004/006,
+> HARD-001/006, CRIT-001/003/004) may differ per episode seed — the above shows default values.
+
+**Failure indicator:** count != 20, missing difficulty tier, wrong correct_action.
 
 ---
 

environment.py

@@ -257,7 +257,12 @@ class PayOpsEnvironment:
         sar_used = task_id in self._sar_filed
         inspected_already = "inspect" in self._used_inv.get(task_id, set())
 
-        reward = grade(action_type, task, inspected_already=inspected_already)
+        investigation_done = bool(self._used_inv.get(task_id, set()))
+        reward = grade(
+            action_type, task,
+            inspected_already=inspected_already,
+            investigation_done=investigation_done,
+        )
         self._state.cumulative_reward = round(
             self._state.cumulative_reward + reward, 4
         )

grader.py

@@ -123,15 +123,25 @@ def step_reward(
     action_type: str,
     task: PayOpsTask,
     inspected_already: bool = False,
+    investigation_done: bool = True,
 ) -> float:
     """
     Single-step reward used by the real-time environment (step_async).
-    Investigation bonuses are now INVESTIGATION_BONUS (0.20).
+
+    ``investigation_done`` must be False when the agent issued zero
+    investigation sub-actions for this task — the same skip-investigation
+    penalty applied by grade_episode is then applied here so the per-step
+    reward the agent sees during training matches the final episode score.
     """
     if _is_investigation(action_type):
         return 0.0 if inspected_already else INVESTIGATION_BONUS
 
-    return _base_terminal_reward(action_type, task)
+    base = _base_terminal_reward(action_type, task)
+    requires_inv = getattr(task, "requires_investigation", set())
+    if requires_inv and not investigation_done and task.difficulty in ("hard", "critical"):
+        correct = action_type == task.correct_action
+        base = base * (SKIP_INV_CORRECT_MULTIPLIER if correct else SKIP_INVESTIGATION_MULTIPLIER)
+    return base
 
 
 # ---------------------------------------------------------------------------
@@ -356,7 +366,16 @@ def grade_episode(
 # Convenience wrapper used by the environment
 # ---------------------------------------------------------------------------
 
-def grade(action_type: str, task: PayOpsTask, inspected_already: bool = False) -> float:
+def grade(
+    action_type: str,
+    task: PayOpsTask,
+    inspected_already: bool = False,
+    investigation_done: bool = True,
+) -> float:
     """Single-step reward used inside environment.step_async."""
-    return step_reward(action_type, task, inspected_already=inspected_already)
+    return step_reward(
+        action_type, task,
+        inspected_already=inspected_already,
+        investigation_done=investigation_done,
+    )
 

inference.py

@@ -157,10 +157,6 @@ def build_observation_text(obs: dict) -> str:
         f"Budget Remaining: {obs.get('budget_remaining', 5.0):.2f}",
         f"Step           : {obs.get('chain_step', 1)}/{obs.get('chain_total', 1)}",
     ]
-    # Show investigation hints prominently when present
-    hints = obs.get('investigation_hints', [])
-    if hints:
-        lines.append(f"INVESTIGATION RECOMMENDED: {', '.join(hints)} (use these before deciding for bonus reward + better evidence)")
     # Append revealed intel if present
     for field, label in [
         ("inspection_notes", "Inspection Notes"),

requirements.txt

@@ -1,6 +1,6 @@
-openenv-core>=0.2.0
-fastapi>=0.115.0
-uvicorn>=0.24.0
-pydantic>=2.0.0
-openai>=1.0.0
-httpx>=0.27.0
+openenv-core==0.2.1
+fastapi==0.135.1
+uvicorn==0.42.0
+pydantic==2.12.5
+openai==2.29.0
+httpx==0.28.1

server/Dockerfile

@@ -25,15 +25,15 @@ RUN pip install --no-cache-dir -r requirements.txt
 COPY . /app/payops_env
 
 # Put parent directory on PYTHONPATH so `payops_env` is importable
-ENV PYTHONPATH="/app:${PYTHONPATH}"
+ENV PYTHONPATH="/app"
 
 # HuggingFace Spaces requires port 7860; default to 7860
 ENV PORT=7860
 
 # Runtime env vars (overridden at deploy time via HF Space secrets)
+# HF_TOKEN / OPENAI_API_KEY are injected at runtime — not baked into the image
 ENV API_BASE_URL="" \
     MODEL_NAME="" \
-    HF_TOKEN="" \
     PAYOPS_BASE_URL="http://localhost:7860"
 
 # Switch to non-root user