> ## Documentation Index
> Fetch the complete documentation index at: https://docs.hubtalk.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# 05. Transitions

> Five kinds of transitions, their evaluation order, and why Default on a speaking node is dead.

# Transitions: five kinds and their priority

> [!NOTE]
>
> **Module goal:** stop guessing why the dialog went the wrong way — and learn to read the transition evaluation order.

The evaluation order is fixed, and it is described in the tooltip next to the **Transitions** heading in the node inspector: **Always → equation → prompt → Default**. Equations are evaluated top to bottom, and the first true one wins — so the order of transitions in the list matters; you can drag them by ⋮⋮. The only thing checked before all of a node's transitions is a global node with **Check on every client turn** enabled.

| Kind | Who decides | What it looks like and where it is used |
| - | - | - |
| **Always** | nobody | Unconditional exit. A node can have only one, and no other transitions next to it: the node's other conditions are ignored, and the editor does not let you draw new connections from such a node. A silent node (Function, Code, etc.) leaves by Always immediately; a speaking node leaves after the client's answer. The editor sets Always by itself in one case: the **Skip response** checkbox turns the node's single transition into Always — the reply is spoken and the dialog moves on at once. |
| **Equation** (`equation`) | code, no LLM | `{{slots_count}} > 0`, `{{status}} == "ok"`. Cheap and deterministic. Advice: if the decision can be expressed as a comparison of variables, use an equation, not a prompt. |
| **Prompt** (`prompt`) | classifier | The meaning of what was said: "the client wants to reschedule the appointment". Advice: phrase it from the client's side and keep conditions distinct — two similar conditions on one node almost guarantee confusion. |
| **Default** | fallback | Fires when nothing else matched. At most one per node. It makes sense on silent nodes and on a node with **Skip response**; on a speaking node — only together with a timer. |
| **Timer** (`timeout`) | runtime timer | "The client has been silent for N seconds" — a deterministic transition without a model. The countdown starts when the agent finishes speaking the node's reply. Conversation nodes only, one timer per node. |

> [!WARNING]
>
> **Pitfall: Default on a speaking node is dead**
>
> The classifier of a speaking node sees only prompt conditions. If the client says something unexpected, the node simply replies on its own and stays put — a Default without a timer never fires. Validation warns about this: "default edge of a speaking node never fires in dialogue (the fallback is STAY — the node generates its own reply)". The behavior you need is set with a prompt condition or an equation, not with Default.

> [!NOTE]
>
> **Three forms of the timer**
>
> The seconds live in the condition field (**Waiting time, seconds**), so the timer comes in three forms:
>
> * **Timer** — a separate transition "the client has been silent for N seconds".
> * **Prompt + timer** — the transition fires on two inputs: "the client said this" *or* "the client stayed silent". This is how you draw the scenario "pick a time… or just stay silent and I'll call you back" with a single transition.
> * **Default + timer** — a fallback transition of a speaking node that fires when the client stays silent.
>
> A timer cannot be set on Always or on an equation. If end-call-on-silence is enabled in the call settings, the node's timer must be **shorter** than its threshold, otherwise publishing will fail.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.