Debug Diagnosis — Execution Protocol

Internal template. Entry point: agent-browser skill.

NEVER

NEVER jump to DOM inspection before checking `errors` — JS errors explain 80% of broken UIs
NEVER assume auth is fine — check `cookies` + `storage local` first when login suspected
NEVER think `inspect` kills CLI — DevTools and CLI work simultaneously

Triage Sequence (run in order, stop at first finding)

agent-browser --headed open <url>
agent-browser wait --load networkidle

# Step 1: JS errors (explains 80% of broken UIs)
agent-browser errors

# Step 2: Console warnings
agent-browser console --pattern "error|warn"

# Step 3: Failed network requests — filter to relevant calls
agent-browser network requests --filter "api"
agent-browser network requests --type xhr,fetch --status 4xx   # 4xx only
agent-browser network requests --method POST --status 5xx      # failed writes

# Step 4: Inspect full request/response body of a failed request
agent-browser network request <requestId>   # get requestId from step 3

# Step 5: Auth state (check first if login/session suspected)
agent-browser cookies         # token cookie present?
agent-browser storage local   # JWT in localStorage?

# Step 6: DOM state
agent-browser snapshot -i
agent-browser is enabled @e5
agent-browser is visible @e3

# Step 7: Form validity (if button appears enabled but won't submit)
agent-browser eval "document.querySelector('form').checkValidity()"

# Step 8: Deep investigation
agent-browser inspect         # DevTools (CLI stays live)

Network HAR for Persistent Evidence

# Record full network trace (start BEFORE navigation)
agent-browser network har start ./debug-trace.har
agent-browser open <url>
agent-browser wait --load networkidle
# ... reproduce the issue ...
agent-browser network har stop
# Open debug-trace.har in browser DevTools → Network → Import

Recovery Patterns

Judgment Gates

Pause when:

• JS errors found (show to user, ask: bug or known?)
• Auth token missing (hard to auto-recover without credentials)
• API returning unexpected errors (user needs to decide next step)
• network request <id> reveals sensitive data in response (don't log, ask user)

Pitfalls