Browser Verification Loop

Browser Verification Loop

Autonomously confirm a list of tasks in a real browser using the Playwright MCP server, then report the result to the linked issue. This is the local half of agentum spec 008a: the agent does the browser work; agentum launches this skill and (for Linear, or strict enforcement) posts the comment.

When To Use

• An agentum Agents & Automation → Browser Verification Loop launch.
• A task list (a GitHub/Linear issue checklist, or an explicit list) that must

be verified against a running web app, unattended.

• You need a pass/fail report posted back to the issue, backed by evidence.

When Not To Use

• There is no Playwright MCP server configured (see Preconditions — fail loudly).
• The work is non-browser (use a normal test runner instead).
• You only need a quick manual look (just drive the browser directly).

Preconditions (check first — fail loudly if unmet)

1. Playwright MCP present. The Playwright MCP tools (e.g. browser_navigate, browser_snapshot, browser_take_screenshot) must be available this session. Confirm via /mcp or by checking the tool list. If they are absent, STOP immediately, emit a failed result naming the reason (e.g. "Playwright MCP not connected — add it to .mcp.json and npx playwright install chromium"), and post nothing green. Never hang waiting for a browser that won't start. 2. A task list. From the linked issue body (checkbox list) or passed in at launch. If empty, report failed with "no tasks to verify". 3. A stop cap. A max-iteration / task budget (passed at launch; default 25). You MUST honor it — never loop unbounded. 4. For direct GitHub posting: gh is authenticated (gh auth status). If not, skip the direct post and rely on the emitted result block (the desktop posts).

The loop (one task per iteration)

For each task, up to the stop cap:

1. Drive it in-browser. Use the Playwright MCP tools to navigate, interact, and observe the running app — actually exercise the behavior the task describes. Do the work in the browser; do not infer pass/fail from code. 2. Capture mandatory evidence. Take a screenshot (browser_take_screenshot) — and/or a browser_snapshot — that shows the verified state. Record its path/ref. This is required for the task to count as passed. 3. Decide pass/fail from what the browser actually showed. 4. Record { task, status: pass|fail, evidence: <screenshot ref>, note }. 5. Stop when every task is done OR the stop cap is reached (note any unrun tasks as skipped: cap reached — do not silently drop them).

Strict evidence (non-negotiable)

A task may be reported passed only if it carries a captured screenshot/snapshot ref. A "pass" with no evidence is invalid — downgrade it to fail with the note "no evidence captured". The desktop also rejects an all-pass result that carries no evidence, so a green report without screenshots will be blocked either way.

Report

Build a markdown summary:

## Browser verification — <passed>/<total> passed
- [x] <task> — evidence: <screenshot ref>
- [ ] <task> — FAILED: <why> (evidence: <ref>)
...
Stopped at: <completed N / cap M>. Tasks unrun: <list or none>.

Then deliver it two ways (do both when possible):

1. Direct post (GitHub, when `gh` is authed): gh issue comment <number> --repo <owner/repo> --body-file <summary.md> (PR conversation: gh pr comment). This is the standalone path. 2. Structured result block so the agentum desktop can enforce strict evidence and post for desktop-managed providers (Linear):

   <<<AGENTUM-VERIFY-RESULT>>>
   { "status": "completed" | "failed",
     "reason": "<only when failed>",
     "summary": "<the markdown above>",
     "tasks": [ { "task": "...", "status": "pass|fail|skipped", "evidence": "<ref>" } ] }
   <<<END>>>

If launched through agentum orchestration, also report via: agentum orchestration task-update <task-id> completed --result '<json above>' (or failed), so the launching desktop pane can post the comment and surface the outcome.

Fail-loud contract

• Playwright MCP missing / browser won't start → failed result with the reason,

no green comment, stop. Never a silent hang.

• A task you cannot drive in the browser → mark it fail with the reason, keep

going, do not fake a pass.

• Cap reached with tasks remaining → report the unrun tasks explicitly.