Review Calls

Review Calls

When this applies

• User says "review calls {client}", "call review {client}", or "how are calls going for {client}"
• User wants a high-level audit of a client's recent AI call performance
• User wants to see onboarding call history alongside live call metrics

Not the same as avoca-vapi-call-resolve — that skill resolves specific call IDs for debugging. This skill is a higher-level review/audit across recent calls.

Prerequisites

• avoca-client.json must exist with a non-empty avoca_client_id
• context.md should have the company name in the heading

MCP servers and tools

1. Attention (user-attention): search_calls, ask_attention, get_call_details
2. Supabase (user-Supabase): execute_sql — project_id wmizcewjcybhvkpwpmim
3. VAPI API (user-vapi-api): get_call — for transcript and tool-call traces on sampled calls

Workflow

Step 1 — Resolve client

1. Match the client to their clients/{Name}-{ID}/ folder.
2. Read avoca-client.json — extract avoca_client_id (this is the team_id for Supabase).
3. Read context.md — extract company name from the # {Name} — {ID} heading.
4. If no avoca_client_id, stop and tell the user the health check can't run without a team ID.

Step 2 — Search Attention (onboarding calls)

Search for Ivan's recorded onboarding/testing/go-live calls with the client.

• Tool: search_calls
• Args: title: "{company name}", from_date: 30 days ago (ISO 8601), size: 25

These are Ivan's Zoom/Meet calls — kickoffs, testing sessions, go-lives.

Step 3 — Batch analyze onboarding calls

If Attention returned call IDs, run ask_attention with all IDs and this prompt:

"For each call: summarize key decisions, action items, and any unresolved blockers. Note whether each is a kickoff, testing, go-live, or follow-up call."

This returns a single AI-generated analysis across all calls — much faster than calling get_call_details in a loop.

If Attention returned no calls or is unavailable, skip this step.

Step 4 — Query Avoca calls (AI-handled)

Run against Supabase to get the client's recent AI-handled calls:

SELECT
  id,
  vapi_call_id,
  ai_or_human,
  status,
  created_at,
  duration_seconds,
  customer_phone,
  call_reason,
  summary,
  emergency
FROM public.calls
WHERE team_id = {team_id}
  AND created_at > NOW() - INTERVAL '14 days'
ORDER BY created_at DESC
LIMIT 50;

From the results, compute:

• Total calls in the period
• AI-handled count and percentage
• Outcome breakdown: count calls by status (booked, message_taken, transferred, etc.)
• Average duration (from duration_seconds)
• Emergency flagged count

Step 5 — Sample deep dive (VAPI transcripts)

Pick up to 5 representative calls from the Supabase results that have a vapi_call_id. Prioritize:

1. Calls with unusual outcomes (transferred, failed, emergency)
2. Short-duration calls (possible hang-ups or errors)
3. Recent calls (most relevant to current state)

For each sampled call, use VAPI API get_call with callId set to the vapi_call_id. Extract:

• Full transcript (formatted as dialogue with speaker labels)
• Tool calls: which tools were called, whether they succeeded or failed
• End reason
• Cost

Assess each sampled call for:

• Booking success: Did the bot complete the booking flow?
• Tool failures: Did findCustomerDNS, getBookingInformationByStAtAdd, or other tools error?
• Flow correctness: Did the bot follow the expected module sequence?
• Customer experience: Any awkward loops, repeated questions, or confusion?

Step 6 — Generate report

Save to: clients/{Name}-{ID}/logs/{YYYY-MM-DD}-call-review.md

If a file with that name exists, append a disambiguator (e.g., -2).

# {Client Name} — Call Review ({start date} to {end date})

**Team ID:** {team_id}
**Generated:** {timestamp}

---

## Onboarding calls (Attention)

| Date | Call title | Type | Key outcomes | Open items |
|------|-----------|------|--------------|------------|
| 4/10 | Sykes Go-Live | Go-live | Went live, first call booked in 1hr | ZIP 20743 missing, fee mapping |

{ask_attention analysis summary — key decisions and unresolved items across all calls}

---

## AI call performance (last 14 days)

| Metric | Value |
|--------|-------|
| Total calls | {n} |
| AI-handled | {n} ({%}) |
| Booked | {n} |
| Message taken | {n} |
| Transferred | {n} |
| Other | {n} |
| Avg duration | {n}s |
| Emergency flagged | {n} |

---

## Sample call analysis

### Call 1 — {short summary}

- **Avoca ID:** `{id}`
- **VAPI ID:** `{vapi_call_id}`
- **Time:** {created_at}
- **Duration:** {duration}s
- **Outcome:** {booked / transferred / message_taken}
- **End reason:** {endedReason}

**Tool calls:**

| Tool | Result | Notes |
|------|--------|-------|
| findCustomerDNS | ✅ Success | Found existing customer |
| getBookingInformationByStAtAdd | ❌ Failed | Timeout on first attempt, succeeded on retry |

**Issues:** {any problems observed}

**Transcript excerpt:**
> (include only the key exchange — 5-10 turns max, not the full transcript)

---

{repeat for each sampled call}

## Patterns & recommendations

- {Pattern 1: e.g., "3 of 5 sampled calls had tool retry on getBookingInformationByStAtAdd — may indicate slow ST API response times"}
- {Pattern 2: e.g., "Bot asked for equipment age on a plumbing call — HVAC-only question leaking into other trades"}
- {Recommendation: e.g., "Consider adding ZIP 20743 to service areas — 2 calls transferred unnecessarily"}

Step 7 — Chat summary

Show a concise version in chat (not the full report):

Reviewed {Client Name} calls — {n} AI calls in the last 14 days ({booking_rate}% booked). Sampled {n} calls: {brief findings}. Full report: clients/{Name}-{ID}/logs/{date}-call-review.md

Customization

The user can specify:

• Date range: "review calls for OHA last 7 days" → adjust the SQL INTERVAL and Attention from_date
• Sample size: "review calls for OHA, check 10 calls" → increase the VAPI sample
• Focus area: "review calls for OHA, focus on transfers" → filter Supabase results to transferred calls and prioritize those for sampling

Error handling

• If Attention MCP is unavailable, skip onboarding calls section and note it.
• If Supabase returns no calls, report "No AI calls found in the last 14 days" and skip the sample analysis.
• If VAPI get_call fails for a sampled call, note the failure and continue with remaining samples.
• Never block the report because one source failed — output whatever data is available.

Safety

• Read-only: never write to Supabase or VAPI.
• Only interpolate validated team IDs into SQL — no arbitrary user input in queries.
• Large transcripts: truncate to key exchanges in the report, not full dumps.