Skip to content

Repository files navigation

boundary-inspector

Inspect deterministic boundary execution over probabilistic systems.

boundary-inspector is an open source UI for exploring how boundary runtimes execute, fail, repair, and succeed.

It turns raw boundary event streams into an interactive view of:

  • runs — top-level contract executions
  • attempts — retry cycles within a run
  • events — verification, repair, retry, and terminal outcomes

The goal is simple: make probabilistic systems understandable when they are embedded inside deterministic software.


The problem

Systems built on probabilistic components (LLMs, agents, automation systems) behave differently from traditional software.

They don't fail once.
They iterate, repair, retry, and converge.

Raw logs make this behavior difficult to understand:

  • Where exactly did the failure occur?
  • What triggered a repair?
  • How did the system evolve between attempts?
  • Why did the run ultimately succeed or fail?

boundary-inspector exists to make these execution boundaries visible.


The model

Boundary runtimes enforce deterministic guarantees around probabilistic outputs.

Execution typically follows a pattern like:


input
↓
probabilistic system (LLM, agent, automation)
↓
verification
↓
repair
↓
retry
↓
terminal outcome

boundary-inspector visualizes this lifecycle as structured execution traces.


Architecture

The inspector is designed to sit on top of boundary event streams.


Producer (llm-contract or other boundary runtime)
↓
BoundaryEvent stream
↓
boundary-inspector

The UI itself is runtime-agnostic. Any system that emits compatible boundary events can be inspected.


Supported connection modes

Currently supported:

  • import or paste JSON event arrays
  • live WebSocket event stream

Future versions will support additional transport adapters.


Running locally


npm install
npm run dev


Event contract

The inspector expects events shaped like:

type BoundaryEvent = {
  system: string
  runId: string
  ts: string
  type: string
  level?: "info" | "warn" | "error"
  attempt?: number
  data: Record<string, unknown>
}

Events are grouped into runs and attempts automatically by the UI.


Interface overview

The UI is organized into three main views.

Runs

Top-level executions.

Shows:

  • run id
  • terminal status
  • attempt count
  • duration

Run inspector

Visualizes the retry lifecycle for a selected run.

Each attempt contains a sequence of events such as:

  • attempt started
  • output observed
  • verification failed
  • repair generated
  • retry scheduled
  • verification passed

Event details

Displays the raw payload for the selected event, including failure explanations and reproduction payloads.

This allows developers to inspect exactly how a boundary execution evolved.


Screenshot anchors

Stable DOM anchors are included to support documentation screenshots and automated capture.

Available targets:

  • header: #shot-header
  • controls: #shot-controls
  • runs list: #shot-runs-list
  • run inspector: #shot-run-inspector
  • event details: #shot-event-details
  • full layout: #shot-layout

Roadmap

Planned improvements include:

  • richer attempt timeline visualization
  • event filtering and search
  • attempt-to-attempt diffing
  • replay debugger mode
  • OTEL and JSONL adapters
  • multi-run comparison

Project status

boundary-inspector is an early stage project.

Event schemas and UI behavior may evolve as boundary runtimes mature.


Contributing

Issues and PRs are welcome.

If proposing schema changes, please include:

  • sample events
  • migration guidance
  • expected UI behavior impact

License

MIT

About

UI for exploring boundary traces.

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages