Intents

Intents

An intent is what the orchestrator thinks a message wants to do: search, add something, open a page, pay, or just chat. The planner turns the intent into a goal and picks tools that can reach it.

Current behaviour

Today there is one fixed taxonomy, written for an online shop (packages/core/src/goap/intent-taxonomy.ts). It is the shop-domain layer of the planner; everything else (tool names, schemas, descriptions) comes from the site.

Intent	Meaning
chat	Small talk; no tool is needed.
chooseStore	Go to or switch to a particular store.
navigate	Open a page of the site.
search	Look for items or ask what is available.
filter	Narrow results.
select	Open or pick one item from the results.
addToCart	Buy something or put it in the cart.
removeFromCart	Take something out.
checkout	Pay and place the order.
paginate	More results, next page.
compare	Compare items.
other	Anything else.

(readState also exists, but only as a tool role: a tool that reads what the site holds, such as a cart or bookings. A message is never classified as readState.)

Each intent has fixed effects and preconditions (for example, addToCart needs a store open, a search done and a parsed request). The planner chains tools through them. A tool is assigned an intent from its name (see Tools ).

How a message is classified

In this order, first match wins:

1. Words. Lexical rules decide checkout and removeFromCart only. They are checked against the original text and its English translation, in English and Russian.
2. Site memory. Per channel, a small classifier trained on examples an administrator approved (or that were auto-approved at high confidence). It stays silent below 0.85 confidence and is never used for checkout or removeFromCart.
3. Laya, the decision model, on the English translation of the message. Its answer for checkout and removeFromCart is discarded and treated as chat.
4. Fallback to chat if Laya fails or returns nothing valid (the words are tried once more first).

checkout and removeFromCart have side effects that cannot be undone by replanning, so they come only from explicit words in the message. A model guess alone never pays or empties a cart.

Laya's answers are stored per channel as examples. Each one is pending until an administrator approves, rejects or corrects it on the admin panel's Intents page, or until its calibrated confidence reaches LAYA_AUTO_APPROVE_CONFIDENCE (default 0.85) and it is approved automatically. An approved example changes what the channel's memory answers next time.

What this means off-shop

The taxonomy fits sites where the visitor searches, adds and pays. On other sites some actions have no matching intent. The pizza-size example: "make the pizza bigger" may be read as addToCart, and a tool like set_pizza_size, which is classified other and has no effects, cannot become the goal of a plan. This is the known limit that the planned work below removes.

Planned: intent packs per channel

Planned, not implemented. Tracked as GitHub issue PO-41. Nothing below works yet.

The idea is that the intent set belongs to the channel, not to the code:

• Built-in "Online shop" pack. The current taxonomy above, with its Laya descriptions, effects, preconditions, the "only by explicit words" rule for checkout and removeFromCart, and its step texts. Read-only; copy and edit, like agent profiles.
• Channel without a pack (planned default for new channels). A message picks the site's tool directly: Laya chooses by tool name and description (plus chat), and the goal is that tool's result. No taxonomy, so it works on any site.
• Existing channels are migrated to the "Online shop" pack, so their behaviour does not change.
• Admin panel. Choose the pack in channel settings; a screen for the pack.