Skip to main content

Issues, chat test, voice test, Copilot

Module goal: find errors before the client does and learn to read the dialog trace.
Four testing tools, from the quickest to the most complete: Issues (instant, no models), the chat test (the real engine, no telephony), the browser voice call (a real voice, real latency) and scenarios (the platform itself runs a batch of test conversations with the agent and gives a verdict).
The Issues panel with one warning

Issues. Errors block publishing, warnings do not. There is one finding here: the node has no outgoing transitions, and the Show on canvas link takes you to it on the canvas.

The filter at the top keeps All, Errors or Warnings, and the findings are grouped by location: node, pipeline or agent. Every finding links to the place where you fix it: Show on canvas pans the canvas to the node, even if it was outside the visible area; pipeline findings lead to Open in Pipeline, agent findings to Open Settings › … or Open Agent variables. In detail — Error checking.
The test panel, the voice test

Test → Voice → Call. A call from the browser against the draft or the published version, with its own call variables and an export of latency metrics.

After a voice test, the Last call block with the conversation recording appears at the bottom of the panel. A test call is recorded even if recording is off for the agent, and the recording is kept for 7 days. In detail — Listening to the voice test.

Starting a test from a selected node

You don’t have to talk your way to a node in the middle of the graph from the start: the conversation follows the edges from Begin through the greeting, data collection and branching, and reaching the right branch on the first try doesn’t always work — one phrase is misunderstood by the classifier, and the conversation goes into the neighbouring branch. Select the node on the canvas, open the ▾ arrow on the green Run button and choose Run from the selected node · <name> (in the voice test — Call from the selected node). The session starts right from it: the edge from Begin is not traversed, and the node runs through the same mechanism as when it is reached by an edge. It works in both the chat and the voice test. In the chat, the restart icon (Restart from <name>) restarts from the same node, so you can run one section as many times in a row as you like. With no selection and on the Begin node, the menu item is disabled and hints Select a node on the canvas. In the session record, a start from a node is marked with a separate Start from node event with the node’s name: whoever reads the record doesn’t have to guess why the conversation started in the middle.
Path variables are not filled — this is the main thing to know.The session gets only the default values from the flow. If the node reads a variable that a node earlier on the path fills, the variable will be empty, and the node will behave differently than in a live conversation. This is not a defect: such values can’t be collected automatically and reliably — global nodes and function results also write to variables, and a static analysis of the graph would give the wrong picture.There is no synthetic conversation history either. The engine does not invent the lines that “would have been” said on the way to the node: the transcript starts right from it.
Limits worth knowing in advance:
  • Test entry points only. The public API, telephony and transferring a call to another agent don’t accept the start node field. A conversation with a real client started from the middle with empty variables is a defect, not a feature.
  • There is one entry point — the Run button menu in the test panel. There is no context menu on the canvas.
  • A component node as a start — only in the draft. When the published version is being tested, the menu item for a component node is disabled and suggests testing the draft. In the draft, a component node resolves to its entry point.

How to read the chat test trace

Every turn in the test panel shows not only the reply but also the engine’s decision: which transition fired and of what type (for example, PROMPT, EQUATION, ALWAYS, DEFAULT, and GLOBAL for global nodes), how long the classifier and generation took, what the function call returned, what was found in the knowledge bases, which objective was closed. This is the main debugging tool: if the agent “went the wrong way”, the answer is visible in the transition chip rather than in guesses. Each agent reply is captioned with its node, and the latency pill under the reply opens a card with the phases of the turn. Session variables are visible in the Dynamic variables (live) drawer under the log. More on the test panel — Test before launch.
The test panel: a live dialog with a trace of transitions, a function call and timings

Test → Chat → Run. A real run: the client asks to reschedule a visit, a prompt transition fires, extract pulls out the date, Function returns a mock CRM response (200), an equation leads to the node with the slots. The settings summary at the top opens the choices: the draft or the published version, rigid or flex, HTTP functions — mock or real calls.

By default, HTTP functions in the chat test answer with mocks and don’t reach your systems; the real value in the test configuration switches on live endpoints, as in production. The voice tab has no such choice. Tip: write test lines in your clients’ language. For an agent in Kazakhstan — in Russian, and for a bilingual agent (kk/ru) in Kazakh too: «Можно перенести визит на пятницу?», «Жұмаға ауыстыруға бола ма?». For an agent in the USA (us_tcpa) — in English, and for a bilingual agent (en/es) in Spanish too: “Can I move my appointment to Friday?”, “¿Puedo cambiar mi cita al viernes?”. Switching languages mid-conversation is a separate test.
The required minimum of tests before publishing (a tip)
  • The whole happy path — all the way to End.
  • Every branch of the branching: “no slots”, “CRM didn’t respond”, “the client cancels”.
  • Both global nodes: an off-topic question (and the return to the script) and a request for a human.
  • Client silence — whether the timer and the reminders fired.
  • By voice — at least one full call: only there do you hear pauses, interruptions and how numbers and dates are pronounced. For an agent in the USA, listen separately to phone numbers, dollar amounts and dates in the US format.
  • Voicemail and call screeners — only with a live outbound call to a real number: the browser voice test doesn’t recognize them, detection works only on outbound phone calls. For the USA, add a call to a phone with unknown-caller screening switched on. In detail — module 15.

Scenarios: testing without you

On the Scenarios tab of the test panel, it’s not you who talks to the agent but the platform’s test callers. You describe the scenarios — who calls, how they behave and what you expect from the agent — or have them generated from the agent settings, 3 to 20 at a time. If there are no enabled scenarios, the check runs with five base callers: calm, in a hurry, confused, drifting off topic and attacking. The Check the agent button runs the conversations in parallel, each up to six turns. After each conversation, the judge decides whether the agent did what was expected of it and marks the line with the error — Error here. The verdict — Ready to launch, Not ready or Not everything was checked — is based on the organization’s thresholds: at least 95% of scenarios passed, 0 breaches. During the check, the agent’s functions answer with test mocks, and the conversations don’t appear in the call history. Test callers speak the agent’s language from its settings, and the judge explains its decision in the interface language. So the check works equally well for an agent speaking Russian or Kazakh and for an English-speaking agent under the US profile. A failure has a Fix with Copilot button: Copilot receives the conversation from the check and proposes a change to the agent. In detail — Testing an agent.

Copilot

The Copilot panel in the editor

Copilot — changes in words: “add a node that asks for a convenient time and lead it to the slot check”. Copilot never changes the flow silently: it shows a proposal, and you accept or reject each change.

The changes of one proposal are checked together, as one set: when a new branch comes in a single proposal, a step inside it can refer to a node that is created by the next step of the same proposal. You can accept changes one at a time or with Accept all. And if something can’t be applied, Copilot names the real cause with the path to the field and marks derived symptoms as “depends on operation …” — so that one typo doesn’t turn into five identical Not applicable cards.