Viventium logo
Viventium
Docs
Prompt Workbench

Prompt Workbench

The local control room for Viventium's prompts, evals, drift review, prompt traces, and scheduled prompt work.

The Short Version

Prompt Workbench is Viventium's local control room for prompt quality.

It lets builders and operators inspect what the product is supposed to say, what is live right now, and how proposed changes behave before they are pushed into the running system.

That matters because Viventium is not one static prompt. It has the main assistant, background cortices, memory prompts, scheduler prompts, voice behavior, worker instructions, and safety boundaries. Those instructions need the same kind of review discipline as code.

What You Can See

Prompt Workbench is organized around three states:

StateMeaning
SourceThe tracked prompt source of truth in the repo.
LiveThe prompt or instruction state currently loaded into the running local system.
EvaluatedThe candidate behavior after tests, evals, or draft review.

The app brings those states into practical surfaces:

  • Prompt Atlas - browse the prompt system without hunting through files
  • Prompt Flow Dashboard - see how prompt objects relate to each other
  • Monaco Prompt Detail - inspect prompt text with a real editor
  • Live Drift Board - spot differences between source and live state
  • Draft Review - review proposed prompt changes before they affect the product
  • Eval Designer and Results - test prompts against expected behavior
  • Prompt Traces - understand what prompt material was used in a run
  • Scheduled Prompts - inspect private scheduled prompt work and user-level scheduled tasks
  • Frame Observability - see prompt-frame behavior without guessing
  • LibreChat integration contract panel - confirm the local chat surface is wired to the expected prompt contract

Why It Matters For Users

Even if you never open Prompt Workbench yourself, it changes the reliability story.

It helps Viventium avoid invisible prompt drift, unreviewed instruction changes, and "it probably works" releases. When prompts control memory, voice, background agents, and worker behavior, the product needs a place to inspect, evaluate, and reconcile them.

Scheduled Prompts

Prompt Workbench also exposes scheduled prompt work.

That includes Workbench-private scheduled prompts, user-level Scheduling Cortex tasks, route/config state, and run history. The built-in private nightly template is called Subconscious Deep Thought, but it is not a silent default that rewrites your memory.

How To Open It

Prompt Workbench is a local sidecar. It can be started without stopping the main Viventium stack:

bin/viventium prompt-workbench open
bin/viventium prompt-workbench status --json
bin/viventium prompt-workbench stop

To keep it running with the local runtime, enable the Prompt Workbench sidecar in canonical config:

runtime:
  prompt_workbench:
    enabled: true

Stopping Prompt Workbench stops only the workbench app. It should not stop LibreChat, Mongo, native services, or the rest of Viventium.

Safety Model

Prompt Workbench is intentionally conservative:

  • no new prompt database
  • no generated runtime writes treated as source truth
  • no silent Mongo writes
  • source edits become reviewed drafts
  • live pushes use the guarded prompt sync flow
  • prompt sync dry-runs before reviewed application
  • private workbench state stays under local App Support

The product goal is not "edit a prompt and hope." It is inspect, evaluate, compare, then apply with a review trail.

Keep Reading

Voice, Chat & MessagingSelf-Healing & Feature Requests