browser-test

name	browser-test
description	Validates browser behavior with agent-browser and sends a browser-check report back to the requester session.

Browser Test

Run browser validation from a browser_check_requested mailbox body and report the result back to the requester session.

Workflow protocol baseline: use the agent-deck-workflow skill.

Input

Provide the mailbox body from browser_check_requested.

Primary Tool

Use agent-browser as the main validation tool.

Preferred commands:

• agent-browser open
• agent-browser wait
• agent-browser snapshot -i
• agent-browser click
• agent-browser fill
• agent-browser get
• agent-browser console
• agent-browser errors
• agent-browser screenshot

Basic agent-browser Help

Keep this basic CLI shape in mind while executing browser checks:

agent-browser - fast browser automation CLI for AI agents

Usage: agent-browser <command> [args] [options]

Core Commands:
  open <url>                 Navigate to URL
  click <sel>                Click element (or @ref)
  type <sel> <text>          Type into element
  fill <sel> <text>          Clear and fill
  press <key>                Press key (Enter, Tab, Control+a)
  wait <sel|ms>              Wait for element or time
  screenshot [path]          Take screenshot
  snapshot                   Accessibility tree with refs (for AI)
  eval <js>                  Run JavaScript
  close                      Close browser

Navigation:
  back                       Go back
  forward                    Go forward
  reload                     Reload page

Get Info:  agent-browser get <what> [selector]
  text, html, value, attr <name>, title, url, count, box, styles, cdp-url

Check State:  agent-browser is <what> <selector>
  visible, enabled, checked

Examples:
  agent-browser open example.com
  agent-browser snapshot -i
  agent-browser click @e2
  agent-browser fill @e3 "[e-mail]"
  agent-browser get text @e1
  agent-browser screenshot --full
  agent-browser wait --load networkidle

For browser-test work, default to snapshot -i to get stable element refs, then interact via @e... refs when possible.

First-Use Environment Check

Before the first browser action in a workflow turn, run a minimal environment check:

1. confirm agent-browser is available with command -v agent-browser
2. use agent_mailbox

Output Format

Use this exact structure as the mailbox body:

Task: <task_id>
Action: browser_check_report
From: browser-tester <browser_tester_session_id>
To: <requester_role> <requester_session_id>
Planner: <planner_session_id_or_N/A>
Round: <round>

## Decision
PASS / FAIL / UNKNOWN

## Coverage
[What batch of scenarios or checks were actually exercised]

## Findings
- [finding or `None`]

## Code Change Summary
- Code changed: [yes/no]
- Branch: [branch name or `N/A`]
- Commit: [short prefix or `N/A`]
- Files changed: [list or `None`]

## Evidence
- Steps executed: [summary]
- Console errors: [summary or `None`]
- Page errors: [summary or `None`]
- Network observations: [summary or `None`]
- Screenshots: [paths or `None`]

## Reproduction
1. [short repro path]

## Residual Risk
[What remains unverified]

Agent Deck Mode

Use the agent-deck-workflow skill for shared protocol.

Skill-specific context resolution:

• task_id: explicit -> mailbox body -> ask
• planner_session_id: explicit -> mailbox body -> ask
• browser_tester_session_id: explicit -> mailbox body To header -> bound mailbox sender context -> ask
• requester_session_id: explicit -> mailbox body From header -> ask
• requester_workspace: explicit -> mailbox body -> ask
• requester_role: explicit -> mailbox body From header -> default requester
• round: explicit -> mailbox body Round header -> default 1

Execution flow:

1. run the first-use environment check
   • if agent-browser is unavailable, stop and report the blocker instead of improvising with another browser tool
2. execute the requested browser steps with agent-browser
   • if the request explicitly allows browser-tester edits, it may modify display-adjacent code on the requested branch before rerunning browser validation
   • if login, auth, environment, or test-data prerequisites are missing, ask the requester first; ask the user directly when requester context is unavailable or user input is clearly required
3. collect runtime evidence
4. produce one browser_check_report
5. use agent_mailbox
6. first call agent_deck_require_session with:
   • session_id = <requester_session_id>
   • workdir = <requester_workspace>
7. send it back to the requester with mailbox_send
   • from_address = agent-deck/<browser_tester_session_id>
   • to_address = agent-deck/<requester_session_id>
   • subject = "browser report: <task_id> r<round>"
   • body = <browser-check report body>

Rules

• validate the full requested browser test batch, not just one narrow sub-step and not unrelated product areas
• prefer the shortest path that still covers the requested scenarios, assertions, and regression checks
• when the request includes multiple related test points, report which ones were covered, which failed, and which remained unverified
• return UNKNOWN when environment, auth, data, or setup blocks a reliable result
• if agent-browser is missing, or required session identity cannot be resolved from explicit metadata plus bound mailbox sender context, state that explicitly in the report or blocker message
• by default, do not change code from this role
• if the request explicitly allows browser-tester edits, limit them to display-adjacent code and keep them on the requested branch
• keep findings factual and tied to observed browser evidence
• prefer requester-provided login/auth/setup context over re-discovering it from scratch
• use the requester workspace from the mailbox body for reply-path session verification; do not substitute the browser-tester's current workspace
• Do not naturally end after writing the report; this workflow turn is complete only after the required mailbox_send back to the requester has succeeded