bug fixes, better generalization

baseline_scores.json

@@ -1,8 +1,8 @@
 {
   "scores": {
-    "workflow_A": 0.7,
+    "workflow_A": 0.137,
     "workflow_B": 0.563,
-    "workflow_C": 0.625
+    "workflow_C": 0.522
   },
-  "average": 0.6293
+  "average": 0.4073
 }

server/app.py

@@ -173,13 +173,19 @@ async def step(body: Dict[str, Any] = Body(...)):
         {"action": {"app": "...", "operation": "...", "args": {...}}, "timeout_s": 15}
     and direct format:
         {"app": "...", "operation": "...", "args": {...}}
+    Invalid or empty actions return a -0.05 penalty observation rather than HTTP 400.
     """
     action_data = body.get("action", body)
     try:
         action = OrgOSAction(**action_data)
         obs    = env.step(action)
-    except (TypeError, KeyError, Exception) as exc:
-        raise HTTPException(status_code=400, detail=str(exc))
+    except Exception as exc:
+        # Return a graceful error observation so the inference loop can continue
+        obs = env._build_obs(
+            reward=-0.05,
+            done=False,
+            message=f"Invalid action format: {exc}. Use {{\"app\": \"...\", \"operation\": \"...\", \"args\": {{...}}}}",
+        )
     return StepResponse(observation=obs, reward=obs.reward, done=obs.done)
 
 

server/apps/workday.py

@@ -92,11 +92,13 @@ class WorkdayApp(BaseApp):
 
     def _op_list_employees(self, department: Optional[str] = None,
                            status: Optional[str] = None,
+                           territory: Optional[str] = None,
                            limit: int = 10) -> Dict:
         matching = [
             r for r in self._records.values()
             if (department is None or r.get("department") == department)
             and (status is None or r.get("status") == status)
+            and (territory is None or r.get("territory") == territory)
         ][:limit]
         drifted = [self._to_agent_view(r) for r in matching]
         keep = ["employee_id", "name",
@@ -106,9 +108,10 @@ class WorkdayApp(BaseApp):
                 "department", "territory"]
         compact = [{k: v for k, v in r.items() if k in keep and v is not None}
                    for r in drifted]
+        filters = [f for f in [department, territory, status] if f]
         return {"success": True, "data": compact,
                 "message": f"Found {len(compact)} employees"
-                           + (f" in {department}" if department else "")}
+                           + (f" ({', '.join(filters)})" if filters else "")}
 
     def _op_provision_access(self, employee_id: str, app_name: str,
                              **kwargs) -> Dict:

server/environment.py

@@ -87,11 +87,14 @@ class OrgOSEnvironment:
         old_score    = self._last_score
         extra_penalty = 0.0
 
+        # Check max-steps first — applies regardless of action outcome
+        at_max = self._step_count >= self.MAX_STEPS[self._workflow_id]
+
         # 1. Validate app exists
         if action.app not in self._apps:
             return self._build_obs(
                 reward=-0.05,
-                done=False,
+                done=at_max,
                 message=f"Unknown app '{action.app}'. Valid apps: {list(self._apps)}",
             )
 
@@ -104,7 +107,7 @@ class OrgOSEnvironment:
             extra_penalty    = rule_penalty
             return self._build_obs(
                 reward=extra_penalty,
-                done=False,
+                done=at_max,
                 message=f"Rule violation: {reason}",
             )
 
@@ -116,7 +119,7 @@ class OrgOSEnvironment:
             self._efficiency -= 0.02
             return self._build_obs(
                 reward=-0.20,
-                done=False,
+                done=at_max,
                 message=(
                     f"Stale schema: field '{result['schema_error']}' is no longer valid. "
                     "Check schema_hints for the current field name. "
@@ -128,7 +131,7 @@ class OrgOSEnvironment:
             self._efficiency -= 0.02   # penalize failed/no-op actions
             return self._build_obs(
                 reward=-0.01,
-                done=False,
+                done=at_max,
                 message=result.get("message", "Operation failed"),
             )
 

server/workflow_engine.py

@@ -20,27 +20,27 @@ class WorkflowStep:
 # ---------------------------------------------------------------------------
 WORKFLOW_A_STEPS = [
     WorkflowStep(
-        "A1", "Acknowledge the incoming Zendesk ticket (ZD-001)",
+        "A1", "Find and acknowledge the new P1 support ticket in Zendesk",
         "zendesk", "acknowledge_ticket",
         lambda apps: apps["zendesk"].ticket_acknowledged(),
     ),
     WorkflowStep(
-        "A2", "Escalate to Jira — create a new issue linked to ZD-001",
+        "A2", "Create a new Jira issue linked to that Zendesk ticket",
         "jira", "create_issue",
         lambda apps: apps["jira"].has_linked_issue(),
     ),
     WorkflowStep(
-        "A3", "Verify the customer's account status in Salesforce (ACME-001)",
+        "A3", "Verify the customer's account status in Salesforce",
         "salesforce", "get_account",
         lambda apps: apps["salesforce"].account_checked(),
     ),
     WorkflowStep(
-        "A4", "Assign the Jira issue to an engineer (JIRA-001)",
+        "A4", "Assign the pre-existing Jira bug for this customer to an engineer",
         "jira", "assign_owner",
         lambda apps: apps["jira"].issue_assigned(),
     ),
     WorkflowStep(
-        "A5", "Log the SLA compliance event in Workday",
+        "A5", "Log the SLA compliance event in Workday using the ticket ID",
         "workday", "log_sla_event",
         lambda apps: apps["workday"].sla_logged(),
     ),
@@ -52,7 +52,7 @@ WORKFLOW_A_STEPS = [
 # ---------------------------------------------------------------------------
 WORKFLOW_B_STEPS = [
     WorkflowStep(
-        "B1", "Create the new employee's onboarding record in Workday (EMP-NEW-001)",
+        "B1", "Find the pending new hire in Workday and create their onboarding record",
         "workday", "create_onboarding_task",
         lambda apps: apps["workday"].employee_created(),
     ),
@@ -62,7 +62,7 @@ WORKFLOW_B_STEPS = [
         lambda apps: apps["workday"].access_provisioned("jira"),
     ),
     WorkflowStep(
-        "B3", "Assign the new employee to the correct Salesforce territory team",
+        "B3", "Assign the new employee to a Salesforce territory account (west region)",
         "salesforce", "assign_account_owner",
         lambda apps: apps["salesforce"].team_assigned(),
     ),
@@ -79,22 +79,22 @@ WORKFLOW_B_STEPS = [
 # ---------------------------------------------------------------------------
 WORKFLOW_C_STEPS = [
     WorkflowStep(
-        "C1", "Flag at-risk account ACME-003 as churn risk in Salesforce",
+        "C1", "Identify and flag the at-risk account as churn risk in Salesforce",
         "salesforce", "flag_churn_risk",
         lambda apps: apps["salesforce"].churn_flagged(),
     ),
     WorkflowStep(
-        "C2", "Query recent support ticket volume for ACME-003 in Zendesk",
+        "C2", "Query recent support tickets for the at-risk account in Zendesk",
         "zendesk", "get_ticket",
         lambda apps: apps["zendesk"].support_queried("ACME-003"),
     ),
     WorkflowStep(
-        "C3", "Check outstanding Jira bugs linked to ACME-003",
+        "C3", "List open Jira bugs linked to the at-risk account",
         "jira", "list_issues",
         lambda apps: apps["jira"].bugs_checked(),
     ),
     WorkflowStep(
-        "C4", "Assign an intervention owner to ACME-003 in Salesforce",
+        "C4", "Assign an intervention owner to the at-risk account in Salesforce",
         "salesforce", "assign_account_owner",
         lambda apps: apps["salesforce"].intervention_assigned(),
     ),