[Feat] Add Critique visual review to frontend tasks

[roomote-roomote]

​Created by Roomote. Follow up by mentioning @roomote-roomote, in the web UI, or in Slack.

Related issue

No linked issue. This is internal Roomote product work requested through a Roomote task.

Why this PR exists

• A maintainer explicitly invited this PR in the linked issue or discussion
• I am a maintainer / this is internal Roomote work

Critique needs access to the exact authenticated UI state a coding task has already prepared, without asking an external service to reload the page or introducing a second browser runtime. The external credential must also remain outside task sandboxes because each review is billable.

What changed

• adds an opt-in critique_visual_review task tool for viewport screenshots and sanitized rendered-DOM snapshots from the active task-scoped agent-browser session
• accepts only bounded capture/review/comparison intent from task code; the API uses a short-lived purpose-specific sandbox-control token to invoke a worker-owned capture procedure and keeps capture bytes in worker memory
• revalidates worker captures on the control plane, re-encodes PNGs, sanitizes the connected DOM tree, excludes task-supplied review context, and constructs the external multipart without accepting task-supplied pixels or DOM
• enforces the two-call paid review/comparison limit with a dedicated Redis quota that fails closed when Redis is unavailable
• bounds task-controlled JSON intent to 64 KB before parsing, including chunked requests without Content-Length
• preserves partial advisory findings and ships a packaged skill that limits autonomous repair to one fix/comparison pass
• documents CRITIQUE_BASE_URL and CRITIQUE_API_TOKEN across supported deployment templates

How it was tested

• focused worker tests cover capture, DOM correctness/sanitization, multipart construction, and task-facing review/comparison behavior
• focused API tests cover trusted sandbox-control capture routing, rejection of task-supplied image/DOM/context, fail-closed quota behavior, two-call enforcement, and declared/streamed oversized intent rejection
• the full worker run-task.test.ts suite passes all 83 tests
• pnpm lint:fast, pnpm check-types:fast, pnpm knip, focused suites, commit hooks, and pre-push gates pass
• current-head GitHub Test, Roomote code review, Lint, Type Check, Knip, docs, build, upgrade, backup, and security checks pass
• visual proof timed out under the shared five-minute budget; no artifacts were retained and no retry was attempted

Checklist

• The PR title follows the repo convention
• This PR is scoped to one feature
• pnpm lint and pnpm check-types pass locally
• I added tests or included a clear manual validation note above
• I removed secrets, tokens, private keys, and customer data from code, logs, and screenshots
• If this change should appear in the changelog, I ran pnpm changeset

[roomote-community]

No code issues found. See task

• The Critique proxy accepts arbitrary multipart task data, bypassing capture sanitization (apps/api/src/handlers/critique/index.ts:76).
• The Critique submission capability remains recoverable by arbitrary task code (apps/worker/src/commands/setup/setup-mcps.ts:365).
• The proxy accepts fabricated task-originated image and DOM data as a Critique capture (apps/api/src/handlers/critique/index.ts:480).
• The paid Critique-call cap fails open while Redis is unavailable (apps/api/src/route-policies.ts:373).
• Task-controlled review context is forwarded to Critique without a trusted capture source (apps/api/src/handlers/critique/index.ts:632).
• The Critique JSON endpoint parses unbounded task-controlled request bodies before enforcing its size cap (apps/api/src/handlers/critique/index.ts:553).

_{Reviewed 33fe41b}