Assistants.

An assistant in wrxstack is a small, well-scoped program that reads from the work graph, calls a frontier model, and writes back. Every action runs under a permission scope you define. Every step is recorded. Every recorded step can be replayed.

The lifecycle of a run

An assistant doesn't "run" in the open-ended chatbot sense. Each invocation has a beginning and an end. The runtime walks the lifecycle in four phases:

1. Trigger. Something happens in the graph: a task is created, a form is submitted, a webhook fires, or a human says "go." The runtime materializes the trigger as a typed event.
2. Plan. The model is given the event, the workflow definition, and a list of tools it's allowed to call. It returns a sequence of tool calls and the reasoning behind each.
3. Act. The runtime executes the calls one by one. Side effects are batched at the graph layer so a half-completed run never leaves the workspace in a torn state.
4. Audit. The full chain (input, tool calls, side effects, model rationale) is written to the audit log. Replay is one API call away.

Scope is the contract

A wrxstack assistant is defined by what it's allowed to see and do, not by what its prompt says. The prompt is a hint. The scope is the contract.

Scopes are graph-typed. You grant read on crm:opportunities and the assistant can see every opportunity the invoking user can. You don't grant it and no prompt-injection trick gets there. The same is true for write: an assistant with no write:contracts scope cannot draft a contract, even if the user asks nicely.

Default assistants

Three assistants ship in every workspace. They're meant as starting points; copy any of them as a template.

• Triage. Reads a new task, finds related context in the graph, labels and assigns it, posts a one-line summary on the parent project.
• Recap. Reads a meeting transcript, pulls action items, attaches them to the right deal or project, drafts a recap doc in the team's voice.
• Redline. Reads a contract, compares it against the active playbook, flags risky clauses, and writes a redline doc for legal review.

When to write your own

Write a custom assistant when (a) the workflow is specific to your business and (b) the default assistants would need more scope than you want to grant them. Start with a one-tool assistant; expand only when the audit log shows you need the next tool.

The first workflow guide walks the full loop end to end.