Skip to main content

Documentation Index

Fetch the complete documentation index at: https://docs.nusomi.com/llms.txt

Use this file to discover all available pages before exploring further.

Nusomi has a small core model. Five primitives, in three layers.

The model in one diagram

                 ┌──────────────┐
   real work ──> │   Session    │ <── you create one per workflow run
                 └──────┬───────┘

        ┌───────────────┼────────────────┐
        ▼               ▼                ▼
  ┌─────────┐    ┌─────────────┐   ┌──────────────┐
  │ Frames  │    │   Events    │   │  Memory graph│
  │ (video) │    │ (actions)   │   │  (cross-run) │
  └────┬────┘    └──────┬──────┘   └──────┬───────┘
       │                │                  │
       └────────┬───────┴──────────────────┘

   ┌─────────────────────────────────┐
   │  Replay · Recovery · Datasets   │
   └─────────────────────────────────┘

Capture layer

PrimitiveWhat it is
SessionA single workflow run. Owns its frames, events, and metadata. Created when recording starts, sealed when it stops.
FramesThe raw video — screen captures at ~30 fps, plus DOM snapshots, browser metadata, and timestamps. The ground truth.
EventsStructured actions extracted from the recording: click_button, input_text, navigate, validation_error, retry, success. Each event is anchored to a specific frame.

Indexing layer

PrimitiveWhat it is
Memory graphA queryable graph across every session in your workspace. Nodes are workflow states, edges are transitions, leaves are outcomes. Lets you find similar runs, prior failures, recovery points.

Output layer

PrimitiveWhat it is
ReplayRe-execute a session. Deterministic, LLM-guided when the UI shifted, or partial-resume from any frame.
RecoveryPick up at the frame just before a workflow broke. Prior state attached.
DatasetsExport frame/action pairs as Parquet, WebDataset, or raw JSONL — for training computer-use models.

Mental model

The capture happens once (a Session). What you do with it splits three ways:
  1. Replay the path the next time the work needs doing.
  2. Recover from a failure with the prior state attached.
  3. Train by exporting frames and actions as a dataset.
If you understand sessions, frames, and events, you understand the whole product. The rest is downstream.