# openenv.yaml
name: Invoice Exception Handler
version: "1.0.0"
description: |
  An agent learning environment simulating accounts payable exception handling.
  The agent acts as an AP analyst: investigates flagged invoices, applies business
  rules, detects fraud signals, makes decisions, and closes cases with an audit trail.

authors:
  - name: Moahmmed Yusuf, Nadella Harshith
    email: [yusufindian09@gmail.com] [nadellaharshith4@gmail.com]

license: MIT

tasks:
  - id: task1_price_variance
    name: Price Variance Exception
    difficulty: easy
    description: |
      Office stationery invoice arrives 3.08% above PO. Company tolerance policy
      allows +/-2% auto-approval. Agent must detect the variance, verify through
      the tolerance rule, confirm verbal approval with procurement, and approve
      with a PO amendment request.
    max_steps: 18
    optimal_score: 1.0
    min_passing_score: 0.60

  - id: task2_duplicate_tax
    name: Duplicate Invoice with Tax Error
    difficulty: medium
    description: |
      Logistics supplier submits INV-2024-891, a duplicate of paid INV-2024-819
      (digit transposition: 891 vs 819). Original invoice had wrong GST rate (15%
      vs correct 18%) — company overpaid 3,240 INR. New invoice has correct rate.
      Agent must detect the duplicate, identify the tax error in the original,
      and partially approve only the 3,240 INR tax correction.
    max_steps: 20
    optimal_score: 1.0
    min_passing_score: 0.50

  - id: task3_compound_fraud
    name: Compound Fraud Signals
    difficulty: hard
    description: |
      IT equipment supplier invoice with four simultaneous fraud signals: bank
      account changed via BEC attack (lookalike email domain), GSTIN belongs to
      a different entity, 2 of 15 laptops not yet received, and unit price 8.65%
      above PO. Agent must find all signals, use the correct communication channel
      (phone, not email — which would contact the fraudster), and escalate to legal
      and security.
    max_steps: 25
    optimal_score: 1.0
    min_passing_score: 0.40

observation_space:
  type: object
  description: EnvironmentState Pydantic model
  fields:
    task_id:             {type: string}
    step_number:         {type: integer}
    case_status:         {type: string, enum: [open, in_review, decided, routed, closed]}
    purchase_order:      {type: object, description: "PO with line items and terms"}
    invoice:             {type: object, description: "Supplier invoice with line items and tax"}
    grn:                 {type: object, description: "Goods receipt — what actually arrived"}
    supplier_master:     {type: object, description: "Verified supplier record"}
    exception_flag:      {type: object, description: "Why the system flagged this invoice"}
    inspections:         {type: array, description: "Fields the agent has inspected"}
    checks_run:          {type: array, description: "Validation checks completed"}
    queries:             {type: array, description: "Internal and supplier queries"}
    rules_applied:       {type: array, description: "Business rules applied"}
    decision:            {type: string, nullable: true}
    routed_to:           {type: array}
    available_actions:   {type: array}
    available_checks:    {type: array}
    available_rules:     {type: array}
    knowledge_base:      {type: array}
    cumulative_reward:   {type: number}

action_space:
  type: object
  description: Action with type and params
  actions:
    inspect_field:
      params: {document: string, field: string}
    cross_check:
      params: {field: string, doc_a: string, doc_b: string}
    run_check:
      params: {check_name: string}
    query_supplier:
      params: {question: string, channel: string}
    query_internal:
      params: {department: string, question: string}
    apply_rule:
      params: {rule_id: string}
    make_decision:
      params: {decision: string, reason: string}
    route_to:
      params: {team: string, notes: string}
    close_case:
      params: {summary: string}

reward:
  range: [-1.0, 1.0]
  description: |
    Shaped reward at every step. Relevant inspections: +0.01 to +0.14.
    Diagnostics revealing issues: +0.08 to +0.18. Correct fixes: +0.08 to +0.30.
    Wrong decision on fraud: -0.15 to -0.40. Repeat actions: -0.02 to -0.05.
    SLA breach: -0.10.

grading:
  method: task_grader
  scores:
    - score
    - diagnosis_score
    - investigation_score
    - decision_score
    - routing_score
    - closure_score
    - efficiency_score

api:
  reset:
    signature: "reset(task_id: str | None = None) -> EnvironmentState"
  step:
    signature: "step(action: Action | dict) -> StepResult"
  state:
    signature: "state() -> EnvironmentState"
  grade:
    signature: "grade() -> Dict[str, float]"

http_endpoints:
  - path: /reset
    method: POST
    description: Reset environment, returns EnvironmentState JSON
  - path: /step
    method: POST
    description: Execute action, returns StepResult JSON
  - path: /state
    method: GET
    description: Current state, returns EnvironmentState JSON
  - path: /grade
    method: POST
    description: Grade current episode
  - path: /health
    method: GET
    description: Health check

dependencies:
  python: ">=3.10"
  packages:
    - pydantic>=2.7
    - fastapi>=0.111
    - uvicorn>=0.29
    - gradio>=4.36
    - openai>=1.35
    - pyyaml>=6.0

docker:
  port: 7860
  health_check: /health