Concepts

The debugging loop

vite-browser is built around a specific loop:

1. Connect to a running Vite app
2. Inspect current runtime state
3. Compare current failures against recent update activity
4. Narrow likely causes with evidence that can be queried again

This is different from general browser automation. Most debugging pain starts after the browser already rendered something incorrect — vite-browser focuses on that moment.

Evidence types

The tool combines signals that are normally scattered across different panels:

Signal	Source	Typical strength
Current errors	Browser runtime	Strong anchor point
Recent HMR updates	Vite WebSocket	Strong when fresh
Module graph changes	Vite internals	Good for import/dependency issues
Framework component state	Vue/React/Svelte DevTools	Good for state-related bugs
Store updates & changed keys	Framework stores	Timing-sensitive
Console logs & network traces	Browser APIs	Supporting evidence

All output is structured so repeated CLI calls can be compared across a live debugging session.

Confidence levels

vite-browser uses intentional confidence language in its output:

• High confidence — The evidence chain is strong enough to act on. Prioritize this file or repro step.
• Plausible — Useful for narrowing the suspect set, but not enough to call proven.
• Conservative output — The tool preferred to say less rather than construct a weak explanation.

What it does not claim

correlate renders and diagnose propagation are narrowing tools, not proof engines.

They do not guarantee reconstruction of deep chains like:

store update → component A → component B → async side effect → error

across arbitrary frameworks and timing conditions. When evidence is incomplete, the output is intentionally conservative.

Why this works for AI agents

AI agents are weak at visual inspection but strong at repeated structured queries. vite-browser plays to this strength:

vite-browser errors --mapped --inline-source   # what's broken
vite-browser correlate errors --mapped          # what changed recently
vite-browser diagnose hmr --limit 50            # what pattern matches

Each command returns something concrete to compare — no screenshot parsing or DevTools navigation required.