Flows

A flow sends one kind of question down a path you lay out. A skill gives the assistant instructions and leaves the rest to the model. A flow settles the path first: a model sorts the question, then each step runs the assistant with only the instructions, model and routes that step needs. Put one file per flow in your project's flows/ folder:

The file name is the flow's name, here raid-officer. Each entry in nodes is one step: a decide node picks where to go, an act node runs the assistant, and start names the first. Typed as /raid-officer Who is level 60?, the question goes from triage to roster, which can read but never write.

Deciding

questions holds the one question a decide node asks, by name. Here triage learns a loot intent:

• choice always picks one of its keys. The model reads the question's name, each key and the sentence beside it, then judges the message. Add a key such as other: 'Anything else' so an off-topic message has somewhere to go.
• score picks one level of an ordered list, lowest first: { score: ['common', 'rare', 'epic'] }.
• yesNo asks a question answered yes or no: { yesNo: 'Does it name a locale?' }.
• A decide node sees only the words the person typed, never record data or what an earlier node did. It shows the person nothing.

Routing

A decide node's next picks where to go from one answer:

• on names the question, and each cases key is one of its options: a choice key, a score level, or yes and no.
• Every answer comes with a confidence from 0 to 1. A chat model reports how sure it is, and Jev computes it. For yesNo, 0.6 means at least 80% sure either way.
• below catches an unsure model: an answer under 0.6 confidence goes to general, whatever it was.
• An answer without a case ends that path. If nothing else is left to run, the person sees that the assistant found nothing to do, so give every option a case or set below.

Acting

act sets what the assistant does in a node, and within which limits:

• skill follows a skill's instructions, and prompt adds to them or stands alone. With neither, as in general, the node runs the plain assistant.
• tiers offers only the routes of those tiers: ['read'] can never write, and [] proposes nothing. Omitted, every tier is open. A skill is instructions only, so it never narrows the routes. Only tiers does.
• model runs the node on another ai.models entry. Omitted, it runs on the model the person picked.
• The node is done when the assistant answers without proposing more. A request waiting for approval pauses it until the person decides.

Chains and lists

An act node's next runs once it is done:

With loot: ['history', 'loot'] in triage's cases, a loot question runs history, then loot, then recap:

• A list runs in order, never in parallel. A node's own next finishes before the next node in the list starts.
• Each node sees what the nodes before it did, as long as the model stays the same. A node on a different model than the one before it starts over from the typed question.

What a person sees

1. In the palette, type /, the flow's name and the question: /raid-officer Who is level 60?. Without a question, the title is sent in its place.
2. Decide nodes pick the path and show nothing.
3. Each act node streams its answer into the same chat. Its requests wait for approval like any other.
4. The flow ends when nothing is left to run. A follow-up is a plain chat, without the flow.

A person needs the ai.use capability and every skill the flow's act nodes use. One skill out of reach hides the whole flow. The step and token limits count across the whole flow, and can stop it part-way. A flow cannot give the model a route the person lacks: every request still goes through the same approval.

Models for deciding

Deciding needs no big model. A decide node runs on the first of these that is set:

• its own decide.model
• ai.decide
• the model the person picked, else ai.model

Any chat model can decide. Jev is a model built to answer typed questions:

A jev model can only decide. The boot refuses it as an act node's model.

Mistakes it catches

A flow that could not run fails when the file loads or the server boots:

• a node id nodes lacks, or a case for an option its question does not have
• a decide node without a branch, or with a question its branch does not route on
• a node start never reaches
• decide nodes that route in a loop with no act node between them
• an unknown skill or model, or a jev model on an act node

Across layers

Flows are scanned from every layer. A closer layer's flow replaces a further one's with the same name, and disable.flows drops one by name, as skills are.