Telephone.org IVR Dashboard

Conversation

Debug

DSL:

🔧 Kazoo Pivot Output

Start a simulation to see Kazoo Pivot JSON output here.

Each step will show the JSON that would be sent to Kazoo.

Select a flow to edit

📍 Node Location

📊 Available Variables

Select a node to see available variables

🔗 Connections

⚠️ Warnings

✓ No issues

🔗 Endpoints Used

Load a flow to see endpoints

🎯 Prediction Rules

Load a flow to see prediction rules

🤖 AI Flow Assistant

AI flow assistant coming soon. Will propose node additions, modifications, and rerouting.

⚙️ Flow Settings (click to edit)

Start Node

Description

🚨 Failover Callflow ID

Transfer here on errors

🔍 Analyze Flow:

💾 Presets:

Traces all reachable flows from selected root

Loading analysis...

📊 Flow Statistics

Click a flow to see stats

🌳 Branch Visualizer

All types

🗺️ Flow Connection Map

🔌 Endpoint Dependency Map

Run analysis to see endpoint usage

💾 Saved Presets (click to expand)

🎯 Test Navigator & Node Tools (click to expand)

🎯 Test Navigator

Jump to any node with pre-filled session data

Preset:

📊 Coverage Progress

📝 Node Verification (click to expand)

Click: cycle status (❓→✅→🔧) | Right-click: notes | Double-click: test

❓ Not Tested | 👁️ Visited | ✅ Verified | 🔧 Needs Fix

🔧 DSL Builder & Validator

Convert Twilio Studio flows, scripts, or descriptions into our DSL format. All prompt sections below are editable - customize them for your specific needs.

📜 System Instructions Core rules for the LLM - edit to adjust strictness

🔀 Node Types Reference DSL node specifications - add/modify as needed

=== SUPPORTED NODE TYPES ===

1. SAY - Play TTS prompt
{ "type": "say", "prompt": "Text to speak", "next": "next_node_id" }

2. CHOICE - Menu with DTMF options
{ "type": "choice", "prompt": "Press 1 for X, press 2 for Y",
  "choices": {
    "option1": { "dtmf": ["1"], "next": "node_for_1", "label": "Option 1" },
    "option2": { "dtmf": ["2"], "next": "node_for_2", "label": "Option 2" }
  },
  "timeout": 10, "fallback": "timeout_node" }

3. BRANCH - Route based on variable value
{ "type": "branch", "input": "{{variable_name}}",
  "branches": { "value1": "node_for_value1", "value2": "node_for_value2", "_default": "fallback_node" } }

4. SET_VAR - Set session variables (MUST use "assignments" object)
{ "type": "set_var", "assignments": { "var_name": "value", "another_var": "{{existing_var}}" }, "next": "next_node" }

5. PROMPT_COLLECT - Collect user input (digits, speech)
{ "type": "prompt_collect", "prompt": "Please enter your order number",
  "input_type": "digits", "max_digits": 8, "store_as": "order_number",
  "next": "next_node", "fallback": "retry_node" }

6. HTTP_DECISION - Make API call and branch on result
{ "type": "http_decision", "endpoint": "endpointName",
  "params": { "order_number": "{{order_number}}" },
  "branches": { "success": "success_node", "not_found": "not_found_node", "_default": "error_node" } }

7. HTTP_ACTION - Make API call without branching
{ "type": "http_action", "endpoint": "endpointName",
  "params": { "data": "{{value}}" }, "next": "next_node", "on_error": "error_node" }

8. TRANSFER - Transfer call to queue/number
{ "type": "transfer", "transfer_type": "queue", "destination": "{{queue_name}}",
  "pre_transfer_prompt": "Please hold while we connect you", "next": "_end" }

9. GOTO - Jump to another flow (with optional return)
{ "type": "goto", "target_flow": "OTHER_FLOW", "target_node": "start", "return_to": "node_to_return_to" }

10. RETURN - Return from a goto
{ "type": "return" }

11. END - End the call
{ "type": "end" }

📋 JSON Structure Templates Expected output format for flows and endpoints

💡 Working Example (SURVEY_FLOW) Real example from our system - helps LLM understand patterns

=== WORKING EXAMPLE: SURVEY_FLOW ===
This flow was converted from an AGI PHP script. Key patterns:

1. ENTRY_PARAMS for required input:
"entry_params": { "survey_type": { "required": true, "description": "1=Sales, 2=CS, 3=Florist", "examples": ["1","2","3"] } }

2. BRANCH on entry param to route:
"survey_router": { "type": "branch", "input": "{{survey_type}}", 
  "branches": { "1": "sales_intro", "2": "cs_intro", "3": "florist_intro", "_default": "invalid_survey" } }

3. CHOICE for collecting 1-5 ratings:
"sales_q1": { "type": "choice", "prompt": "On a scale of 1-5, how satisfied were you?",
  "choices": {
    "1": { "dtmf": ["1"], "next": "set_sales_q1_1" },
    "2": { "dtmf": ["2"], "next": "set_sales_q1_2" },
    ...
  }, "timeout": 15, "max_attempts": 3, "fallback": "sales_q1" }

4. SET_VAR to store responses:
"set_sales_q1_1": { "type": "set_var", "assignments": { "sales_q1": "1" }, "next": "sales_q2" }

5. HTTP_ACTION to submit results:
"submit_survey": { "type": "http_action", "endpoint": "saveSurvey",
  "params": { "lead_id": "{{lead_id}}", "q1": "{{sales_q1}}", "q2": "{{sales_q2}}", "q3": "{{sales_q3}}" },
  "next": "thank_you", "on_error": "thank_you" }

📄 YOUR SOURCE TO CONVERT Paste Twilio JSON, script code, or describe the flow you want

✨ Generated Combined Prompt

Click "Generate Combined Prompt" to assemble all sections above into a single prompt for the LLM.

📥 Import LLM Response

After running the prompt through an LLM (Claude, GPT, etc.), paste the response here. The system will extract FLOWS and ENDPOINTS JSON blocks, validate them, and prepare for import.

📁 Upload JSON File

🕐 Business Hours Manager

Manage named schedules that DSL check_schedule calls reference

📊 Schedule Status

Current status of all named schedules (as referenced by DSL)

📅 Weekly Hours

Select a schedule to edit its weekly hours

🚫 Recurring Holidays

These dates are closed every year (affects schedules with shared holidays enabled)

📆 Special Closures

One-time closures (e.g., Thanksgiving varies each year)

🧪 Test Schedule

Simulate a date/time to test if a schedule would be open

📞 Queue Manager

Map DSL queue destinations to Kazoo extensions and CRM callback IDs

📋 Queue Mappings

DSL Destination	Name	Ext	Callflow ID	Queue ID	CRM ID	Aliases
Loading...

🔍 Test Queue Lookup

Test resolving any identifier to its queue info

📊 Queue Usage in Flows

Shows which flows use each queue

Traffic Router

Route Kazoo pivot calls across multiple IVR servers with weighted traffic splitting

Servers

Shared pool of IVR server base URLs referenced by route profiles

Key	Name	Base URL
Loading...

Primary Server (failover): If routing fails for any reason, calls fall back to this server

Routes

Each route is a Kazoo entry point with switchable traffic profiles

Server Base URL

This URL is used by Kazoo to reach your IVR server. Set it in Settings → Server Settings.

Auto Realtime

📊 Today's Stats

Calls Today

Active Now

Peak Concurrent

Avg Steps

Avg Response: 0ms

Last Call: --

Errors: 0

📞 Order Identification

Auto (Caller ID)

Manual Entry

Avg Steps (all calls): 0

🧠 Prediction Accuracy

Accuracy Rate

Prediction Coverage --

0 of 0 calls

Confirmed

Declined

Auto-routed

Low Confidence

Per-Rule Accuracy

No data yet

🎯 Flow Hits Today

No data yet

🌐 API Endpoint Hits

No data yet

🟢 Active Calls

No active calls

⚙ Feature Flags

Loading…

🚨 Error Log

No errors

Page 1 0 entries

📡

Loading call logs...

🌐 API Endpoints

CRM API integration health and usage dashboard

Total Endpoints

Used in Flows

Orphaned

Local Handlers

Fields Mapped

KB Conditions

Refund Audit Log

Cancel & refund operations with business rule validation

Time	Order #	Caller	Endpoint	Amount	Balance	Delivered	Cancel OK	Status	Flags
Click Refresh to load audit data

Page 1

📁 API Groups

⚠️ Orphaned Endpoints

🌐

Select an endpoint group or endpoint to view details

Bot Hierarchy

Loading hierarchy...

Orphan Nodes

Loading insights...

Loading skills...

| Zoom: 100% |

DSL Nodes

Select a Landbot flow to visualize its graph and DSL nodes.

KB Articles

IVR

Loading articles...

Loading KB overview...

Loading comparison data...

Loading gap analysis...

Loading classification data...

Loading automation paths...

Loading recommendations...

Loading decision trees...

Loading generation readiness...

Article Detail

Click an article to view details.

Loading rules...

Select a rule to edit

🔮

Select a prediction rule from the left panel to edit

Rule Status

No rule selected

Test Predictions

Run a test to see matched rules

Engine Info

Known Fields

🗺 Phase Status / Roadmap

[click to expand]

📋 Agent Context Bundle

The curated payload sent to the LLM at agent_loop entry — order summary, eligibility flags with block reasons, history, frustration signal, and the in-scope tool catalog. Strips PII, internal IDs, and URLs. See src/agent_context.py and Section 11.5 of docs/ARCHITECTURE_DECISIONS.md.

Session ID: or:

// Pick a session above or click a demo to view the bundle.

🧰 Tool Catalog

Auto-generated from config/endpoints.json + manual overlay from config/tool_annotations.json. Filter by flow to see what Claude has access to. See src/tool_schemas.py and Section 11 of docs/ARCHITECTURE_DECISIONS.md.

Flow:

🛡️ Validator Trace

Every LLM tool call passes through validators that gate against the same business rules as the existing flow branches (within_24h, balance_status, network_category, kill switches). This panel shows the audit log of decisions and lets you probe tools against fabricated sessions. See src/validator.py and Section 11.5 of docs/ARCHITECTURE_DECISIONS.md.

🧪 Probe a tool

Tool Args (JSON)

Session overrides (JSON)

Filter:

📝 System Prompts

Per-flow prompts under config/prompts/*.md. Drafts never go live — only "Promote" makes a draft active. Snapshots auto-save to .history/ on each promote. See src/prompt_manager.py.

Flow:

📥 ACTIVE (live)

📝 DRAFT (editable)

🎯 Render preview

📜 Version history

⏱️ Agent Loop — Dry Run + Live Monitor

Run the agent_loop in simulate-tools mode against a fabricated session — real Claude, mocked CRM, no real refunds. See every iteration, tool call, validator decision, and cost. See src/agent_loop.py · Section 9 of docs/ARCHITECTURE_DECISIONS.md.

Flow

Scenario

max_iterations

Customer's first utterance (optional — leave blank to let Claude await one)

⚠️ Hits real Claude API — costs apply (visible in the Usage tab after each run). Set ANTHROPIC_API_KEY in env if not yet configured.

🔬 Order Inspector

Look up any order by number + ZIP. Shows raw CRM fields, computed flags (network_category, balance_status, within_24h, etc.), and a probe of every validator — so you can see in one shot which tools would fire, which would deny, and why. Mirrors python tools/inspect_order.py <order> <zip>.

📄 Prompt Review & Edit

Every customer-facing string in the deterministic flows — DTMF prompts, choice options, transfer announcements, agent_loop filler / silence-goodbye. Three workflows:
• 📥 Download a Markdown doc, share with customer, get marked-up version back, dev applies via JSON edits.
• 👁️ Preview a flow inline (read-only).
• ✏️ Edit directly — every prompt becomes a textarea, save per-prompt, backups + audit log automatic.

💾 Edits persist across deploys. Saves write to the persistent disk at data/flow_overrides/ (Render's mounted volume) — not to the repo's flows/ files. The interpreter reads overrides first, falling back to the repo version. So when we push code updates, your edits stay live. Devs can review the pending overrides below and either integrate them into git (then click ↩ Revert here once committed) or leave them as-is.

Loading…

▸ How do variable placeholders work?

Tokens like {{billing_name}}, {{recipient_name}}, {{delivery_date}} are filled in at runtime from the customer's order data. The editor flags any save that removes a placeholder (a common mistake) but doesn't block — you might intentionally want to remove one.

🎧 Pivot-TAP Live Audio Sessions

Real-time view of every active <Stream> WebSocket from Kazoo. Audio is piped to the configured ASR provider; transcripts (partials + finals) accumulate per call. Auto-refreshes every 2 seconds. Settings → 🎤 ASR Provider picks the backend. See src/pivot_tap.py.

💬 Interactive Chat

Type customer utterances, see live multi-turn agent_loop responses with tool calls, validator decisions, and per-turn cost. Real Claude API + mocked CRM (simulate-tools mode). Best for prompt iteration + stakeholder demos.

Flow: Scenario:

Click "▶️ Start chat" with a flow and scenario to begin.

📚 KB Retrieval (Toggle + Probe)

Pulls top-K relevant chunks from knowledge/articles/*.json into the system prompt's {{kb_chunks}} placeholder per turn. BM25 keyword retrieval scoped by flow category. Off by default — A/B test whether it helps. See src/kb_retrieval.py.

KB retrieval feature flag

Affects new calls only — in-progress calls keep their session-start value.

🧠 Semantic Embeddings (upgrade from BM25)

Loading…

🔍 Probe retrieval

▶️ Transcript Replay

Feeds real call transcripts from har/ through the agent_loop in simulate-tools mode (real Claude, mocked CRM). Compare the bot's first response against the human agent's first response. Useful for vetting prompts and lexicon against real customer language. See src/transcript_replay.py.

🧪 Scenario harness

Run hand-coded scenarios against real Claude. Multi-run mode runs each scenario N times to expose LLM stochasticity.
~$0.05 per single run · ~$0.15 per 3-run.

Scenario: Runs:

🔁 Replay recorded AI calls

Every completed AI agent_loop turn is recorded to disk. Pick a recorded call and replay it through the agent_loop with simulated CRM tools — optionally with a candidate prompt — to see how a change affects past calls.

Days: Flow:

🗣️ Speech-match stats (speech mode only)

Per-CHOICE-node speech recognition outcomes. DTMF callers don't appear here — by design. The lowest match-rate nodes are the highest-leverage place to add to speech_intents.

Days: No-match only

💰 Usage & Billing

Append-only ledger of every metered upstream call (TTS, Claude, Groq, ASR). Filter by tenant or call. Cache hits show up but bill 0 — only generations bill. See src/usage_ledger.py · logs/usage_*.jsonl · Section 11.7 of docs/ARCHITECTURE_DECISIONS.md.

Tenant: Days:

🤖 AI / LLM Mode

Master toggle for the agent_loop on pilot flows (NON_DELIVERY_AI, PRODUCT_QUALITY_AI). When OFF, calls fall back to deterministic CANCEL_ORDER_FLOW / PRODUCT_FLOW (DTMF). Same fallback fires on concurrency-cap rejection or any agent_loop exception. See src/ai_concurrency.py + src/interpreter.py agent_loop branch.

Master Toggle

Max Concurrent AI Sessions

New calls beyond this fall back to DTMF for that call. Existing AI calls finish.

🗣 Groq Utterance Polish (Phase 2)

Routes Claude's draft response through Groq for tone cleanup. Adds ~200-400ms but cheaper Groq tokens. A/B test whether it improves naturalness.

🔇 Silence Hangup (consecutive no-speech)

Mirrors the DTMF gather's retry-then-hangup pattern for AI flows. After N customer turns where nothing is heard (empty SpeechResult AND empty Pivot-TAP transcript), the agent_loop short-circuits BEFORE calling Claude — plays the goodbye message and hangs up. Counter resets the moment the customer actually speaks.

Master toggle kills the feature globally (any value). Threshold + Goodbye are defaults — flows that set max_consecutive_no_speech / silence_goodbye_message in their JSON win. NON_DELIVERY_AI and PRODUCT_QUALITY_AI both set 3 explicitly today.

Threshold (turns)

1–20. Default 3.

Goodbye message (≤500 chars)

Spoken once before hangup. Keep it warm and short.

⏩ VAD-driven endpointing (Phase 3)

Drives turn-end detection from our local VAD + ASR provider fast-path events instead of Kazoo's fixed-window <Gather> timeout (which is DTMF-only, not speech-aware). When ON: the EndpointingCoordinator listens for VAD speech-end (after min_silence_ms debounce) AND Deepgram UtteranceEnd events; whichever fires first triggers a Kazoo metaflow redirect to /pivot/continue after a small grace period. Provider-agnostic — works the same with Whisper-via-Together once that lands.

Grace (ms): how long we wait after the speech-end signal for the ASR to finalize the last word before firing the redirect. 250ms is a balanced default; bump to 400ms if you're losing tail-words, drop to 150ms if turns feel sluggish.

Default OFF — flip ON for a verified test call once you've confirmed Kazoo metaflow is wired (KAZOO_API_URL / ACCOUNT_ID / AUTH_TOKEN env vars set). When OFF, falls back to the legacy 8s gather (existing behavior).

Grace period (ms)

0–2000. Default 250.

Live state

—

▸ Active AI sessions

▸ Recent acquire/reject/release events

Auto-refreshes every 5s while panel is open

📞 Kazoo Integration URLs

Edit the public-facing URLs Kazoo uses to reach this server. Saves to .env and applies live (no restart). The catalog below is auto-derived — copy-paste into your Kazoo Pivot configuration. See src/env_config_manager.py.

Loading…

Paths Kazoo connects to:

Defined in the FastAPI routers — paths themselves are not user-configurable, but the BASE_URL above feeds into all of them. Click any URL to copy.

Loading…

🌐 Voice Gateway

Speech-mode voice transport. Handles VAD / ASR / endpointing on the gateway side (real speech support, unlike the DTMF-only touchtone path). First time setting this up? Follow the checklist below.

Master toggle

—

📋 Setup checklist (follow in order)

Loading…

🔐 Inbound HTTP Basic auth

When you set Basic auth in your voice-gateway application, paste the same credentials here. We'll verify every incoming webhook carries them. Leave both empty to skip verification (less secure but simpler — fine for early testing).

Username

Password

🔗 URLs (click to copy)

Voice URL (initial call — pick flow)

—

Status URL (call lifecycle — optional)

—

Health probe (test the adapter)

—

📡 Recent webhook attempts

No webhook traffic yet. Once the voice gateway hits us, attempts (success + failure) will appear here.

Auto-refreshes every 5s while the Settings tab is open

🔑 LLM API Keys

Anthropic (Claude) for tool-use decisions · Groq for fast utterance generation (Phase 2). Keys are persisted to .env and applied to the running process immediately — no restart required. Values are masked for display. See src/secrets_manager.py.

Loading…

🔊 VAD (Voice Activity Detection) Provider

Detects when the customer is speaking. WebRTC is the default — Google's lightweight C extension, ~0.1ms per frame, no torch/onnx, ~150KB total. Binary speech/silence per 30ms frame. Silero is heavier (~1ms/chunk, ships torch+onnxruntime) but trained on speech formants — better at rejecting noise. SoulX-Duplug is a semantic upgrade (returns idle/nonidle/speak/blank where speak means semantic completeness) — skeleton until hosting is decided. NoOp disables VAD entirely (zero CPU). See src/vad/.

Loading…

🚧 Phase 3 Feature Toggles

Switches for Phase 3 audio-pipeline features. Some are wired to live code paths; others are skeletons until the underlying impl lands. Flipping a flag whose impl isn't ready is a no-op on dead code — safe to toggle ahead of time.

Loading…

🧪 Channel Interrupt Lab — Empirical Testing

Test how Kazoo behaves when you interrupt an in-progress play verb. Documentation is sparse — the only way to know if metaflow / redirect / break actually preempts mid-utterance is to fire the request against a live call and observe.

📚 Kazoo research notes

POST /v2/accounts/{aid}/channels/{call_id}/metaflow — runtime callflow injection. Documented purpose: hot-key callflow features mid-call (e.g. dial a star-code to mute). Whether it preempts an in-progress play verb is unverified.
POST /v2/accounts/{aid}/channels/{call_id} with action=break / queue — older API. Behavior under Pivot's play is unknown.
Pivot redirect: returning a fresh Pivot response with a new play may or may not preempt the previous one — depends on Pivot's playback state machine.
The polite barge-in approach (Section 5.4(a) of arch doc) sidesteps this entirely: let the current play finish a short sentence, capture customer speech in parallel, respond on next turn.

Live call selector

Loading…

Interrupt strategy

📋 Expected observations to record

Did the original play stop immediately? Within how many ms?
Did Kazoo queue the new audio (waited for original to finish)?
Did the call drop or behave erratically?
HTTP response code + body from Kazoo

Append findings to docs/ARCHITECTURE_DECISIONS.md Section 5.4(b).

🎤 ASR (Speech-to-Text) Provider

Selects the streaming-ASR backend used by the voice gateway and the Pivot-TAP path. Deepgram nova-2 is the production-ready default. Together.ai Whisper is the cost target (streaming verification pending). Google Cloud is a future skeleton. NoOp bypasses ASR for tests/dev. See src/asr/.

Loading…

🎙️ TTS Cache Pre-warm + Hash Registry

Pre-generate audio for common phrases at startup so the first call doesn't pay synthesis latency. Hash registry persistence lets dynamic Groq utterances survive restarts. See src/tts_prewarm.py · src/adapters/tts_endpoint.py.

🚨 Frustration Detector

Repeat-call frequency thresholds + keyword signals (BBB, lawyer, chargeback, etc.). Hard signals skip the AI agent entirely and route to the priority queue. Soft signals are exposed to the LLM as context flags. See src/frustration_detector.py · Section 11.6 of docs/ARCHITECTURE_DECISIONS.md.

🧪 Probe a scenario

subject_id (caller phone)

current_transcript

Thresholds are placeholders — see open question #13 (out to customer)

🌐 Server Settings

Configure the public URLs that Kazoo will use to reach this server and the TTS endpoint.

Base URL (Public IP/domain for Pivot)

TTS Base URL (Leave blank to use Base URL)

IVR Base URL (Leave blank to use Base URL)

Host

Port

🔊 TTS Provider

Select and configure the text-to-speech engine for IVR prompts.

Provider

Enable TTS Cache

Max Text Length

🗣️ Pronunciation Mappings

Map words to TTS-friendly pronunciations. Example: "avas" → "ayvas" so TTS says it correctly.

Original Word

→

Pronunciation

☁️ Google Cloud TTS Settings

Configure Google Cloud Text-to-Speech API credentials and voice settings.

Service Account JSON *Required

No file uploaded

Upload your Google Cloud service account JSON key file (e.g., my-project-123456.json)

Or Paste JSON Credentials

Voice

Language

Speaking Rate (0.25 - 4.0)

Pitch (-20 to +20)

Audio Encoding

Sample Rate

Effects Profile

🔊 TTS Testing

Test TTS synthesis directly and via the Kazoo-compatible endpoint.

Current Provider: -- (Change above and Save Settings to switch)

Test Text

📦 TTS Cache & Hash Registry

Monitor TTS caching and hash registry. Hash registry maps clean URLs to text.

🔗 Hash Registry

Registered Hashes

Max Capacity

🎵 Audio Cache

Hit Rate

Entries

Hits

Misses

Memory

Bytes Saved

Evictions

Recent Cache Entries:

                                Click "Refresh Stats" to load entries
                            

🔢 DTMF Settings

Configure timing for DTMF digit collection. Other settings (max digits, timeout) are defined per-node in the DSL.

Inter-digit Timeout (ms) Time between key presses before input is accepted. Increase if callers need more time to type order numbers or zip codes.

📥 Pivot Session Fields

Select which Kazoo Pivot request fields to store as session variables. These become available as {{variable}} in your DSL flows for branching, API calls, etc.
📖 Kazoo Pivot Docs

Standard Fields (every request)

Optional Fields (after certain actions)

Always included:

call_id - unique call identifier (used as session ID)
lead_id - alias for call_id (CRM convenience)
Custom-Application-Vars - unpacked as individual session vars

💡 Usage Example

// In your DSL flow, branch based on caller:
{
"type": "branch",
"input": "{{caller_id}}",
"branches": {
"+15551234567": "vip_greeting",
"_default": "standard_greeting"
}
}

🔀 Relay Proxy

Configure relay proxy for CRM API calls through a remote server.

Enable Relay Proxy

Relay Target URL

Forward requests to IVR after relay

🎤 OSTI Speech Intent Router

Configure the openSpeechToIntent server for voice-based menu navigation. When enabled on a choice node, calls will be redirected to OSTI for speech recognition racing against DTMF.

Enable OSTI Integration (Global)

OSTI Pivot URL (Speech intent server) The OSTI server receives Kazoo pivot redirects, streams audio, races DTMF vs speech, and returns results.

Return URL Template (How OSTI calls back) Template URL OSTI uses to return results. Use {session_id} placeholder.

Default Threshold Confidence score (0-1) required to accept speech intent

Speech Timeout (ms) How long OSTI listens for speech

OSTI WebSocket URL (Audio stream endpoint) WebSocket URL for Kazoo to stream call audio to OSTI

📋 OSTI Redirect Parameters

Parameters sent TO OSTI:
session_id
 - Current call session ID
return_url
 - URL to redirect back with results
prompt_url
 - TTS audio URL for speech prompt
dtmf_prompt_url
 - TTS audio URL for DTMF prompt
intents
 - JSON: {key: {dtmf:["1"], phrases:[...]}}
threshold
 - Confidence threshold
timeout
 - Speech timeout in ms
session_vars
 - Base64 encoded session variables
flow_id
 - Current flow ID
node_id
 - Current node ID
Parameters returned FROM OSTI:
Digits
 - Matched DTMF digit (for DSL compatibility)
source
 - "osti" or "dtmf"
intent
 - Intent key if from speech
confidence
 - Match confidence (0-1)
transcript
 - What caller said
session_id
 - Session ID (echoed back)
                            

📡 Real-time WebSocket

Enable WebSocket event broadcasting for real-time dashboards. When disabled, there is zero performance overhead on the IVR engine.

Enable Real-time Broadcasting

When enabled:

• WebSocket endpoint: ws://host/ws/realtime?channels=stats,calls,all

• REST API: GET /api/realtime/summary

• Events published for: call start/end, flow changes, node execution

Use for external dashboards that need real-time call counts.

Status: Checking...

📱 Twilio SMS

Configure Twilio for sending SMS messages from IVR flows (missing item links, PAN notices, order modification links).

Account SID

Auth Token Keep this secret - stored in settings.json

Messaging Service SID Uses messaging service for automatic number selection

📋 SMS Nodes in Flows

DELIVERY_FLOW / missing_item_message

Sends CS URL for missing item requests

SOP_FLOW / pan_url_text

Sends Product Availability Notice URL + PIN

ORDER_MOD_FLOW / ordermod_send_sms

Sends order modification URL

🔐 Security Settings

Basic login protection for the admin console.

Admin Password

Leave blank to keep current password. Default: avas2026

🔒 Superadmin password

Required to open the Settings tab itself. Stored only in this browser (localStorage). Once unlocked, it stays unlocked for the rest of the session — Logout above clears both the admin and superadmin sessions.

Default: avas-super-2026. Min 8 chars on change.

📁 File Manager

Browse and manage files on the persistent data disk (config, flows, data).

🔍 System Diagnostics

Shows which directories the system is using for config and flows. Helps diagnose persistent disk issues.

Click Refresh to load diagnostics...

📋 Current Configuration

Live Preview

Showing: Last saved configuration

📞 Telephone.org

🔧 DSL Builder & Validator

📥 Import LLM Response

📦 Extracted Items

✅ Validation & Import

🕐 Business Hours Manager

📊 Schedule Status

📅 Weekly Hours

🚫 Recurring Holidays

📆 Special Closures

🧪 Test Schedule

📞 Queue Manager

➕ Add New Queue

📋 Queue Mappings

🔍 Test Queue Lookup

📊 Queue Usage in Flows

Traffic Router

Servers

Routes

Server Base URL

🌐 API Endpoints

Refund Audit Log

🗺 Phase Status / Roadmap

📋 Agent Context Bundle

🧰 Tool Catalog

Per-flow Tool Allowlist Matrix

Tool Annotations Editor

🛡️ Validator Trace

📝 System Prompts

⏱️ Agent Loop — Dry Run + Live Monitor

🔬 Order Inspector

📄 Prompt Review & Edit

🎧 Pivot-TAP Live Audio Sessions

💬 Interactive Chat

📚 KB Retrieval (Toggle + Probe)

▶️ Transcript Replay

🧪 Scenario harness

🔁 Replay recorded AI calls

🗣️ Speech-match stats (speech mode only)

💰 Usage & Billing

Billing Rate Card

🤖 AI / LLM Mode

📞 Kazoo Integration URLs

Paths Kazoo connects to:

🌐 Voice Gateway

📋 Setup checklist (follow in order)

🔐 Inbound HTTP Basic auth

🔗 URLs (click to copy)

📡 Recent webhook attempts

🔑 LLM API Keys

🔊 VAD (Voice Activity Detection) Provider

🚧 Phase 3 Feature Toggles

🧪 Channel Interrupt Lab — Empirical Testing

Live call selector

Interrupt strategy

🎤 ASR (Speech-to-Text) Provider

🎙️ TTS Cache Pre-warm + Hash Registry

🚨 Frustration Detector

🌐 Server Settings

🔊 TTS Provider

🗣️ Pronunciation Mappings

☁️ Google Cloud TTS Settings

🎤 Kokoro TTS Settings

🔗 External TTS Settings

🔊 TTS Testing

📦 TTS Cache & Hash Registry

🔢 DTMF Settings

📥 Pivot Session Fields

🔀 Relay Proxy

🎤 OSTI Speech Intent Router

📡 Real-time WebSocket

📱 Twilio SMS

🔐 Security Settings

🔒 Superadmin password

📁 File Manager

🔍 System Diagnostics

📋 Current Configuration

Node

Endpoint Log