CHG AR Agent -- Knowledge Base & Test Harness

Automated accounts receivable agent for Clean Health Group (CHG). This repo contains the agent's complete knowledge base (identity, skills, commands) and a local test harness for validating behaviour against structured scenarios.

The agent runs on Manus AI, orchestrated by Make.com. It receives student payment data as JSON, applies week-by-week escalation logic, classifies inbound student replies, and returns structured JSON to a human approval queue. Every draft communication requires human sign-off before anything is sent.

How It Works

BigQuery  -->  Make.com  -->  Manus AI  -->  Make.com  -->  BigQuery
  (data)      (orchestration)   (logic)     (routing)     (state + logs)

• BigQuery -- single source of truth for student records and escalation state
• Make.com -- polls BigQuery, enriches data (aXcelerate notes, Gmail threads), calls Manus, writes results back
• Manus AI -- applies the AR logic defined in this repo's knowledge base files
• aXcelerate -- student management system (AIF, VAST, CH instances)
• Gmail -- ar@cleanhealth.edu.au, monitored for inbound student replies

Manus never reads from or writes to BigQuery directly. Make.com handles all data movement.

Repo Structure

CHG_AR_Agent/
  01_Identity/                  -- Who the agent is
    Soul.md                        Core personality, guardrails, tone
    Context.md                     System architecture, data flow, field definitions
    Policy_Reference.md            CHG AR policies and escalation rules
    Brand_Details.md               Brand-specific details (AIF, VAST, CH)
    Product_Code.md                Course/product code reference

  02_Skills/                    -- Operational knowledge
    Comms_Templates.md             Outbound email templates by escalation week
    AR_Inbound_Classification.md   6-type inbound email classification system
    Inbound_Reply_Templates.md     Reply templates for each inbound type
    Student_FAQ_Reference.md       Balance field guide + common student questions
    Edge_Cases_Guide.md            Routing exceptions, bad faith, flag conflicts
    Correction_Log.md              Log of AR Supervisor corrections (placeholder)

  03_Commands/                  -- Executable commands + schemas
    Primary_Role_Command.md        Top-level routing (outbound vs inbound)
    AR_Outbound_Command.md         Outbound escalation pipeline (Steps 0-6)
    AR_Inbound_Command.md          Inbound classification pipeline (Steps 0-4b)
    outbound_output_schema.json    JSON schema for outbound responses
    inbound_output_schema.json     JSON schema for inbound responses

scenarios/                      -- 23 JSON test scenarios
test_agent.py                   -- Local test runner (Anthropic API)

Agent Modes

Outbound (Phase B)

Processes overdue student accounts through a week-by-week escalation sequence. Two interaction modes:

• Mode 1 (first_interaction) -- No prior state. Make.com enriches with aXcelerate notes and Gmail inbox search. Agent classifies across both sources and returns seed flags.
• Mode 2 (subsequent_interaction) -- Existing state row. Flags come structured from BigQuery. Agent escalates based on current week and flag state.

Inbound (Phase C)

Classifies incoming student emails into 6 types (hardship, complaints, promises to pay, general queries, payment queries, unclassifiable) and drafts appropriate replies. Make.com sets conversation_active: true on receipt, and the agent determines when it auto-clears vs requires manual clearing.

Key Flags

Flag	Purpose
dnc_flag	Hard stop on outbound only. Covers DNC requests, complaints, disputes, chargebacks, legal matters. Inbound still processed. AR manager sets/clears.
conversation_active	Pauses outbound while a conversation is in progress. Auto-clears on resolution for some types, manual for others.
hardship_flag	Triggers hardship proposal workflow instead of standard escalation.
promise_to_pay / promise_to_pay_date	Pauses outbound while a future payment date is active.
escalation_paused	Manual override by AR team.

Test Scenarios

23 scenarios covering outbound escalation weeks 1-6+, inbound classification types, edge cases (DNC, hardship, broken promises, bad faith, CH non-accredited routing), and cross-mode interactions.

#	Scenario	Mode
01	Week 1 AIF first failure	Outbound
02	Week 2 VAST no response	Outbound
03	Week 4 AIF broken promise	Outbound
04	Week 5 VAST urgent	Outbound
05	Week 6 AIF final notice	Outbound
06	Hard stop -- DNC flag	Outbound
07	Hardship (VAST)	Outbound
08	Conversation active -- pause	Outbound
09	CH non-accredited routing	Outbound
10	Week 6+ AIF bad faith	Outbound
11	Inbound hardship	Inbound
12	Inbound complaint	Inbound
13	Inbound hardship + DNC active	Inbound
14	Inbound hardship acceptance	Inbound
15	Outbound resume after pause	Outbound
16	Inbound promise (vague)	Inbound
17	Inbound general query	Inbound
18	Inbound unclassifiable bounce	Inbound
19	Inbound unclassifiable short ack	Inbound
20	Inbound multi-signal hardship+partial	Inbound
21	Inbound payment query	Inbound
22	Inbound promise (explicit)	Inbound
23	Hardship payment failure	Inbound

Running Tests Locally

# Activate virtual environment
& venv\Scripts\Activate.ps1

# Install dependencies (first time)
pip install -r requirements.txt

# Run test scenarios
python test_agent.py

Requires a .env file with your Anthropic API key. The test runner uses claude-opus-4-6 to simulate Manus locally.

Design Principles

• Human-in-the-loop -- 100% of draft communications go through an approval queue
• Agent never sends directly -- no emails, no logging, no escalation without human sign-off
• Single source of truth -- BigQuery owns all state; Manus is stateless per call
• Flag-driven logic -- escalation behaviour determined by structured flags, not free-text parsing
• Separation of concerns -- Manus handles logic, Make.com handles data movement