feat: add new healthcare domain

data/tau2/domains/healthcare/user_db.json

@@ -0,0 +1,43 @@
+{
+  "patient_device": {
+    "insurance_card": {
+      "provider": "BlueCross",
+      "policy_number": "BC123456789",
+      "group_number": "GRP001",
+      "member_name": "Default Patient",
+      "copay_info": "Primary Care: $20, Specialist: $35, Urgent Care: $50"
+    },
+    "current_symptoms": [],
+    "current_temperature": null,
+    "medications_at_home": [],
+    "calendar_availability": [],
+    "portal_info": null,
+    "has_blood_pressure_monitor": true,
+    "latest_bp_reading": {
+      "systolic": 128,
+      "diastolic": 82
+    },
+    "has_glucose_meter": true,
+    "latest_glucose_reading": 105,
+    "glucose_measurement_time": "Fasting (8am)",
+    "has_pulse_oximeter": true,
+    "latest_spo2_reading": 98,
+    "latest_heart_rate": 72,
+    "current_pain": null,
+    "confirmed_appointments": [],
+    "consents_provided": [],
+    "acknowledged_instructions": [],
+    "notification_preferences": [],
+    "pharmacy_transfer_requests": [],
+    "uploaded_photos": []
+  },
+  "surroundings": {
+    "patient_id": "patient_001",
+    "full_name": "Default Patient",
+    "date_of_birth": "1990-01-01",
+    "location": "home",
+    "has_internet_access": true,
+    "payment_methods_available": ["credit_card"],
+    "emergency_contact": null
+  }
+}

src/tau2/data_model/message.py

@@ -58,6 +58,10 @@ class ToolCall(BaseModel):
         "assistant",
         description="The requestor of the tool call.",
     )
+    compare_args: Optional[list[str]] = Field(
+        default=None,
+        description="List of argument keys to compare. If None, compare all. If empty list, compare only tool name.",
+    )
 
     def __str__(self) -> str:
         lines = [f"ToolCall (from {self.requestor})"]

src/tau2/domains/healthcare/README.md

@@ -0,0 +1,297 @@
+# Healthcare Domain
+
+The healthcare domain simulates a customer service environment where agents help patients with appointments, prescriptions, insurance verification, test results, and chronic condition monitoring.
+
+## Overview
+
+The healthcare domain tests agent capabilities in:
+
+- **Workflow compliance**: Identity verification first, insurance checks before booking
+- **Clinical safety**: Specific thresholds determine when to escalate to medical staff (fever ≥103°F, BP ≥180/120, etc.)
+- **Bidirectional coordination**: Agent and patient both have tools - agent guides, patient performs actions
+- **Mixed evaluation**: Tasks check both correct outcomes (ENV_ASSERTION) and safe procedures (ACTION)
+- **Patient personas**: Easy, None (neutral), and Hard personas with different health literacy levels
+
+## Domain at a Glance
+
+| Metric | Value |
+|--------|-------|
+| **Base Tasks** | 70 (152 full, 37 small) |
+| **Intents** | 8 (appointment, prescription, monitoring, telehealth, test results, triage) |
+| **Personas** | Easy (24), None (23), Hard (23) |
+| **Evaluation** | Mixed (ENV + ACTION) for critical workflows |
+| **Patient Records** | 3 with comprehensive medical history |
+| **Agent Tools** | 18 (5 evaluated in tasks) |
+| **User Tools** | 20 (evaluated via ENV_ASSERTION) |
+
+### Task Distribution by Intent
+
+**Base Set (tasks.json - 70 tasks):**
+- `appointment_scheduling` (27) - Book/cancel/reschedule with insurance verification
+- `telehealth_setup` (18) - Set up remote care with consent and instructions
+- `chronic_monitoring` (15) - Monitor vitals for diabetes, hypertension, COPD
+- `patient_mistake` (4) - Handle patient confusion gracefully
+- `urgent_triage` (3) - Triage urgent symptoms appropriately (fever, pain, breathing)
+- `critical_triage` (3) - Escalate critical conditions immediately (≥103°F, ≥180/120, <90% O2)
+
+**Full Set (tasks_full.json - 152 tasks):**
+Includes all base tasks plus additional complexity variations:
+- `appointment_scheduling` (71) - Extended scenarios with more edge cases
+- `telehealth_setup` (35) - Additional consent and setup variations
+- `chronic_monitoring` (23) - More vital sign combinations
+- `patient_mistake` (8) - More confusion scenarios
+- `urgent_triage` (7) - Additional symptom presentations
+- `critical_triage` (3) - Same critical escalation tests
+- `prescription_refill` (2) - Basic refill scenarios
+- `test_results_access` (3) - Lab results review scenarios
+
+**Small Set (tasks_small.json - 37 tasks):**
+Single-subtask tasks for quick evaluation, one per (intent × persona) combination.
+
+## Architecture
+
+The domain uses a bidirectional setup where both agent and patient have their own tools:
+
+```
+┌─────────────────────────────────────────────────────────┐
+│                  HealthcareEnvironment                  │
+├────────────────────────┬────────────────────────────────┤
+│   Agent Side           │   Patient Side                 │
+│   (HealthcareTools)    │   (HealthcareUserTools)        │
+├────────────────────────┼────────────────────────────────┤
+│ • get_patient_details  │ • check_insurance_card         │
+│ • verify_insurance     │ • check_calendar               │
+│ • book_appointment     │ • measure_blood_pressure       │
+│ • check_test_results   │ • measure_blood_glucose        │
+│ • transfer_to_nurse    │ • provide_consent              │
+│ • 20+ more tools       │ • 15+ more tools               │
+└────────────────────────┴────────────────────────────────┘
+           ↓                           ↓
+    Agent performs              Patient performs
+    system actions              real-world actions
+```
+
+The agent can't directly access the patient's insurance card or vital signs. Instead, it must ask the patient to check these using user-side tools. This tests realistic coordination between agent requests and patient actions.
+
+## Getting Started
+
+### Quick Run
+
+```bash
+# View domain info
+tau2 domain healthcare
+
+# Test with 5 random tasks
+tau2 run --domain healthcare \
+         --agent-llm claude-sonnet-4-5-20250929 \
+         --user-llm claude-sonnet-4-5-20250929 \
+         --num-tasks 5
+
+# Try tasks interactively (will prompt for domain and task selection)
+tau2 play
+```
+
+## Example: Appointment Scheduling Flow
+
+**Scenario**: Patient needs routine checkup appointment
+
+**Required Workflow** (identity → insurance → availability → book):
+
+1. **Agent**: `get_patient_details("Sarah Johnson", "1985-03-15")` → Verify identity
+   - Returns: `patient_id="patient_001"` (needed for subsequent calls)
+
+2. **Agent asks**: "Can you check your insurance card?"
+   - **Patient**: `check_insurance_card()` → Returns: "BlueCross BlueShield, BC123456"
+
+3. **Agent**: `verify_insurance_coverage(patient_id="patient_001", procedure_type="routine_checkup")` → Check coverage
+   - Returns: Copay $20, covered
+
+4. **Agent**: `check_available_time_slots(doctor_id="doc_001", date="2024-05-20")` → Find slots
+   - Returns: Available times [09:00, 14:00, 16:00]
+
+5. **Agent asks**: "Can you check your calendar for May 20th at 2 PM?"
+   - **Patient**: `check_calendar()` → Returns: "May 20th is available"
+
+6. **Agent**: `book_appointment(patient_id="patient_001", doctor_id="doc_001", appointment_type="routine_checkup", date="2024-05-20", time="14:00", reason="Annual checkup")` → Book it
+
+**Evaluation**:
+- **ENV_ASSERTION**: ✓ Appointment exists in database with correct details
+- **ACTION**: ✓ Correct sequence (identity first, then insurance, then availability, then book)
+- Both must pass for reward = 1.0
+
+## Policies
+
+### Workflow Requirements
+
+**Multi-Step Pattern** (all workflows follow this hierarchy):
+1. **Identity Verification** - Always verify patient identity first using `get_patient_details(full_name, date_of_birth)`
+2. **Assessment** - Gather necessary information (insurance, symptoms, vitals)
+3. **Verification** - Confirm availability, eligibility, or status
+4. **Action** - Execute the requested operation only after all prerequisites met
+
+**Mandatory Sequences:**
+- Appointment booking: identity → insurance → availability → book
+- Prescription refills: identity → medication check → prescription verification → insurance → refill
+- Chronic monitoring: identity → measure vitals → assess → schedule/transfer
+- Telehealth setup: identity → consent → emergency contact → instructions
+
+### Clinical Safety Thresholds
+
+**Critical Values** (immediate transfer to nurse):
+| Vital Sign | Transfer Threshold | Booking Range |
+|------------|-------------------|---------------|
+| **Fever** | ≥103°F | 100-102.9°F |
+| **Pain** | ≥7/10 (severe) | 1-6/10 (mild-moderate) |
+| **Blood Pressure** | ≥180/120 mmHg | 130-179/80-119 mmHg |
+| **Blood Glucose** | <70 or >250 mg/dL | 100-250 mg/dL |
+| **Oxygen Saturation** | <90% | 90-95% |
+
+**Decision Rule:**
+- At or above transfer threshold → `transfer_to_nurse()` immediately
+- Within booking range → `book_appointment()` for follow-up
+- Normal readings → Routine follow-up scheduling
+
+### Communication Guidelines
+
+- State appointment details (date, time, doctor, specialty, copay) clearly after booking
+- Explain next steps explicitly (when to arrive, what to bring)
+- Transfer to nurse for clinical questions (interpreting test results, medication advice)
+- Transfer to human agent for administrative issues (billing disputes, system errors)
+
+## Tools Overview
+
+### Agent-Side (HealthcareTools)
+
+**Identity & Records:**
+- `get_patient_details(full_name, date_of_birth)` - Verify and retrieve patient record
+- `get_chronic_conditions(patient_id)` - View patient's chronic conditions
+- `get_vital_signs_history(patient_id)` - Review past vital readings
+
+**Appointments:**
+- `book_appointment(...)` - Schedule appointment
+- `cancel_appointment(appointment_id)` - Cancel existing appointment
+- `check_available_time_slots(doctor_id, date)` - Find available slots
+- `list_available_doctors(specialty)` - Find doctors by specialty
+
+**Insurance & Billing:**
+- `verify_insurance_coverage(patient_id, procedure_type)` - Check coverage
+- `calculate_cost(patient_id, appointment_type)` - Calculate copay
+
+**Clinical:**
+- `check_test_results(patient_id)` - Access lab results
+- `get_prescription_details(prescription_id)` - View prescription info
+- `request_prescription_refill(patient_id, prescription_id)` - Refill prescription
+- `transfer_to_nurse()` - Escalate to clinical staff
+
+### Patient-Side (HealthcareUserTools)
+
+**Information Access:**
+- `check_insurance_card()` - View insurance provider and policy number
+- `check_calendar()` - Check personal availability
+- `check_medication_bottle()` - Read prescription number from bottle
+
+**Vital Measurements:**
+- `measure_blood_pressure()` - Take BP reading
+- `measure_blood_glucose()` - Check blood sugar
+- `measure_oxygen_saturation()` - Measure O2 saturation
+- `take_temperature()` - Measure temperature
+- `check_symptoms()` - Describe current symptoms
+
+**Consent & Actions:**
+- `provide_consent(consent_type)` - Give consent for telehealth/billing/data
+- `acknowledge_instructions(instruction_type)` - Acknowledge medical instructions
+- `update_emergency_contact()` - Update emergency contact info
+
+
+## Adding Tasks
+
+To add new healthcare tasks:
+
+1. **Create task intent** in `src/tau2/domains/healthcare/tasks/<intent>_issues.py`
+   - Follow existing patterns in `appointment_issues.py`, `prescription_issues.py`, etc.
+   - Define `create_<intent>_tasks()` function that returns list of tasks
+
+2. **Define evaluation functions** in `src/tau2/domains/healthcare/tasks/evaluation_functions.py`
+   - Add `is_<intent>_fixed(env)` to check if issue is resolved
+   - Add `get_<intent>_env_assertions(...)` if using ENV_ASSERTION checks
+   - Centralize reusable evaluation logic here
+
+3. **Set evaluation mode**:
+   - Use `["ENV_ASSERTION", "ACTION"]` for workflows requiring specific order (appointments, prescriptions)
+   - Use `["ENV_ASSERTION"]` for outcome-only checks (consent, emergency contact updates)
+   - Identity verification must be first action if using ACTION mode
+   - Insurance verification before booking/refilling if using ACTION mode
+
+4. **Register in task manager** in `src/tau2/domains/healthcare/tasks/create_tasks.py`
+   - Import your task creation function
+   - Add to the appropriate TaskManager
+   - Run `python -m src.tau2.domains.healthcare.tasks.create_tasks` to regenerate task files
+
+5. **Update splits** in `data/tau2/domains/healthcare/split_tasks.json` if adding to train/test sets
+
+## Testing
+
+Run the comprehensive test suite:
+
+```bash
+# Test agent-side tools (39 tests)
+pytest tests/test_domains/test_healthcare/test_tools_healthcare.py -v
+
+# Test patient-side tools (27 tests)
+pytest tests/test_domains/test_healthcare/test_user_tools_healthcare.py -v
+
+# Test both (66 tests total)
+pytest tests/test_domains/test_healthcare/ -v
+```
+
+All tests should pass (100% pass rate required).
+
+**Note**: You may see a pytest config warning about `asyncio_default_fixture_loop_scope` - this is harmless and can be ignored (the healthcare tests don't use async).
+
+## File Structure
+
+```
+src/tau2/domains/healthcare/
+├── __init__.py                        # Domain exports
+├── README.md                          # This file
+├── data_model.py                      # Patient, Doctor, Appointment, Prescription models
+├── user_data_model.py                 # PatientDevice, PatientSurroundings
+├── environment.py                     # HealthcareEnvironment + factory
+├── tools.py                           # Agent-side tools (18 tools)
+├── user_tools.py                      # Patient-side tools (20 tools)
+├── utils.py                           # Path constants
+└── tasks/
+    ├── __init__.py
+    ├── const.py                       # Personas and tool grounding
+    ├── create_tasks.py                # Task generation pipeline
+    ├── manager.py                     # TaskManager class
+    ├── utils.py                       # Task composition utilities
+    ├── evaluation_functions.py        # Centralized evaluation logic
+    ├── appointment_issues.py          # Appointment scheduling tasks
+    ├── prescription_issues.py         # Prescription refill tasks
+    ├── chronic_monitoring_issues.py   # Vital signs monitoring tasks
+    ├── telehealth_issues.py           # Telehealth setup tasks
+    ├── test_results_issues.py         # Lab results access tasks
+    ├── urgent_triage_issues.py        # Urgent symptom triage tasks
+    ├── critical_triage_issues.py      # Critical condition escalation
+    └── patient_mistake_issues.py      # Patient confusion handling
+
+data/tau2/domains/healthcare/
+├── db.json                            # Patient database (3 patients)
+├── user_db.json                       # Patient device state
+├── policy.md                          # Agent policy (512 lines)
+├── tasks.json                         # Main task set (70 tasks)
+├── tasks_full.json                    # All tasks (152 tasks)
+├── tasks_small.json                   # Single-intent tasks (37 tasks)
+└── split_tasks.json                   # Train/dev/test splits
+
+tests/test_domains/test_healthcare/
+├── test_tools_healthcare.py           # Agent tool tests (39 tests)
+└── test_user_tools_healthcare.py      # Patient tool tests (27 tests)
+```
+
+## Additional Documentation
+
+- **Agent Policy**: Full policy with clinical thresholds and workflow requirements at `data/tau2/domains/healthcare/policy.md`
+- **Data Models**: Detailed schema documentation in `src/tau2/domains/healthcare/data_model.py`
+- **Task Generation**: Implementation details in `src/tau2/domains/healthcare/tasks/`

src/tau2/domains/healthcare/__init__.py

@@ -0,0 +1 @@
+# Copyright Sierra