Quartz 4

SKILL

5 min read

card-action-handler — Aurora Reasoning for Card Actions

When a user clicks [Send via OpenPhone], [Email via Gmail], [Send via SalesMsg], [Group SMS], or [Hand off to Aurora] in the OpenPhone Call Intelligence tab card, the action gateway invokes this skill to generate Aurora’s reasoning + draft + alternatives. Result renders in the iframe modal where user reviews, edits, and confirms.

When this skill fires

Endpoint /api/ctie/aurora/draft (port 18790) calls this skill. Trigger surfaces:

  • Iframe modal load (initial draft)
  • “Regenerate” button click in the iframe modal
  • Direct API call from any other surface that wants Aurora’s reasoning on a CTIE call

NEVER fires for:

  • HubSpot task creation (CTIE task content already exists, no Aurora needed)
  • Native HubSpot composer launches (HubSpot’s own UI, not Aurora-mediated)
  • Note creation, meeting scheduling (HubSpot native, not Aurora-mediated)

Inputs (from /api/ctie/aurora/draft POST body)

{
  "action_type": "sms" | "email" | "salesmsg" | "group-sms" | "handoff",
  "call_id": "AC...",                         // OpenPhone external_call_id
  "context": {
    "contact_name": "Robert",
    "contact_phone": "+15623203611",
    "line_name": "BetterFiles",
    "rep_name": "Henry Hill III",
    "temperature": "WARM",
    "role": "SELLER_LEAD",
    "confidence": 78,
    "situation": "Robert under contract, two tenants signed cash-for-keys...",
    "ctie_suggested_sms": "Hey Robert, send the lease docs and we move fast.",
    "score_deal": 6,
    "score_relationship": 70,
    "thread_history_summary": "Last 3 SMS replies came on this line within 12min",
    "deal_stage": "EVALUATING",
    "previous_sends": 0
  }
}

Outputs

{
  "channel_recommendation": "openphone-sms",
  "channel_reasoning": "Contact's last 3 replies came on SMS within 12 min — fastest reply channel for them. Email avg reply: 2 days.",
  "primary_draft": {
    "label": "Aurora's recommended",
    "subject": null,                         // only for email
    "body": "Hey Robert, getting buyer feedback on the tenanted side. Can you send the lease docs and any eviction filings? Want to move fast if numbers work.",
    "tone": "casual + action-oriented"
  },
  "alternatives": [
    { "label": "More formal", "subject": null, "body": "...", "tone": "formal" },
    { "label": "Higher urgency", "subject": null, "body": "...", "tone": "urgent" }
  ],
  "compliance_precheck": {
    "tcpa_window_ok": true,
    "suppression_status": "clean",
    "dedup_window_ok": true,
    "notes": "Within 8am-9pm contact local time. No opt-out flags. No identical message in last 7d."
  },
  "model_used": "gemma-3-27b",
  "ms_elapsed": 2340
}

Model routing

  • Primary: Gemma 3 27B via Portkey config pc-gemma-card-action (route to local M1 Ultra Ollama once available, or OpenRouter google/gemma-2-27b-it until then).
  • Fallback: Sonnet via Portkey pc-anthropic-sonnet if Gemma errors or response time > 8s.
  • NEVER use: Opus (cost prohibitive for per-click usage).

Cost budget

  • Per skill invocation target: $0.001 or less via Gemma.
  • Sonnet fallback: ~$0.003 per invocation (still acceptable).
  • Hard ceiling: any skill invocation exceeding $0.01 is logged as anomaly.

Compliance integration

This skill does NOT execute the send. It only generates the draft + reasoning. The actual send goes through:

  1. Compliance gate (scripts/lib/compliance-gate.js) — runs at executor level, not here
  2. Channel executor (openphone-sms.js, gmail-client.js, salesmsg-send.js) — sends the message
  3. Audit log (scripts/lib/action-audit.js) — records the send

If the skill’s compliance_precheck flags any issue, the iframe modal SHOULD warn the user before they click Send. The actual gate enforcement is at executor level — defense in depth.

Constraints (Gemma quirks)

  • Gemma 4 fails on multi-step tool chains — this skill makes ZERO tool calls, just text generation
  • Gemma may hallucinate phone numbers, contact names, or HubSpot IDs — the skill MUST output content the user can edit before send
  • Gemma’s training cutoff is Jan 2025 — don’t ask it about current events / weather / market conditions
  • The user ALWAYS reviews and edits before sending — Gemma is the drafter, human is the gatekeeper

Prompt template (sent to Gemma)

You are Aurora, an AI sales assistant for a real estate wholesaling team.
A team member just clicked "{action_type}" on a call card with a contact.
Your job: generate the most effective draft message and explain your reasoning.

CALL CONTEXT:
- Contact: {contact_name} ({contact_phone})
- Rep handling: {rep_name} via {line_name}
- Call temperature: {temperature} | Role: {role} | Confidence: {confidence}%
- Situation: {situation}
- Last reply pattern: {thread_history_summary}
- Deal stage: {deal_stage}
- CTIE suggested SMS (baseline): {ctie_suggested_sms}

YOUR TASK:
1. Recommend the best channel (sms / email / salesmsg) with reasoning grounded in the context above.
2. Write a primary draft (under 320 chars for SMS; full email body for email).
3. Write 2 alternative drafts: one more formal, one more urgent.
4. Flag any compliance concerns (TCPA quiet hours, recent send to same contact).

OUTPUT (strict JSON):
{
  "channel_recommendation": "...",
  "channel_reasoning": "...",
  "primary_draft": { "label": "Aurora's recommended", "subject": null, "body": "...", "tone": "..." },
  "alternatives": [
    { "label": "More formal", "subject": null, "body": "...", "tone": "formal" },
    { "label": "Higher urgency", "subject": null, "body": "...", "tone": "urgent" }
  ],
  "compliance_precheck": { "tcpa_window_ok": true, "suppression_status": "clean", "dedup_window_ok": true, "notes": "..." }
}

NEVER:
- Make up the contact's address, phone, or any factual detail not in the context
- Use the contact's first name unless it's in the context
- Promise things outside RERI's scope (financing, legal advice, etc.)
- Send anything that violates TCPA (after-hours, opt-out ignored)

Failure modes

  • Gemma returns malformed JSON → skill catches, returns CTIE suggested_sms as primary draft with reasoning=“Aurora unavailable, using CTIE baseline”
  • Gemma timeout (>8s) → fallback to Sonnet
  • Both fail → return CTIE suggested_sms with reasoning=“Both Gemma and Sonnet unavailable”

Audit + observability

Every skill invocation writes:

  1. tool_calls row (per CHOKEPOINT-1) with model + tokens + cost
  2. action_audit_log.reasoning field (the channel_reasoning + compliance_precheck) — persists the Aurora output even if the user never clicks Send

This makes “what would Aurora have said?” queryable for audit and training data.

Graph View