web: expose screen recording on `--platform web` (agent-browser supports `.webm`, not surfaced by agent-device)

[kacper-mikolajczak]

Summary

On --platform web, agent-device record start/stop is unsupported - record is absent from WEB_SUPPORTED_COMMANDS, so web sessions cannot capture a screen recording. The managed backend (agent-browser) already supports native video recording (.webm), so this is about surfacing an existing backend capability through the agent-device web layer, not new recording plumbing. It mirrors how full-page screenshot + viewport were surfaced for web in #865.

Environment

• agent-device: 0.17.10 (and main as of 2026-06-26)
• agent-browser (managed backend): 0.27.x
• Node: 24 (web backend)
• Reproduced on macOS (local) and Ubuntu headless CI.

Current behavior

record is not part of the web capability surface:

• src/core/capabilities.ts - WEB_SUPPORTED_COMMANDS = open, close, <query commands>, click, fill, focus, press, scroll, type, viewport; record is not included, so isCommandSupportedOnDevice is false for web.
• The daemon recording handler has start branches only for ios-device / macos / ios-sim / android - no web branch - and returns UNSUPPORTED_OPERATION.
• WebProvider exposes no record method.

So agent-device record start ./clip.mp4 --platform web is unsupported, while record works on Android (adb screenrecord) and iOS.

Proposed solution

Surface agent-browser's native recording through the web provider:

• add record to WEB_SUPPORTED_COMMANDS;
• implement start/stop recording on WebProvider (agent-browser-provider) delegating to agent-browser's native record start ./clip.webm / record stop;
• add a web branch to the daemon recording handler.

Optionally normalize the output to MP4 for parity with the device record output.

Why / use case

We drive agent-device --platform web to verify and capture evidence for web app behavior. The mobile QA path records a video of each reproduction; the web path cannot, so an "always capture a clip" evidence convention can't hold on web - we currently fall back to screenshots only. Exposing the web video that agent-browser already produces would give web verify-runs the same temporal evidence as mobile, without scroll-and-stitch screenshot workarounds.

Notes

• Related: fix: add web viewport control and screenshot aliases #865 surfaced full-page screenshot + viewport for --platform web using the same "expose an existing backend capability through the web layer" pattern.
• Verification correctness is not blocked (we rely on snapshot / is / get); this is specifically about video evidence on web.

[thymikee]

This was generated by AI during triage.

Triage Notes

Outcome: enhancement + ready-for-agent.

What is established:

• The request is real on main: WEB_SUPPORTED_COMMANDS includes web query/interaction/viewport support but not record, and the record capability only covers Apple/Android device targets.
• The current web interactor/provider path exposes screenshot and viewport, but no recording method.
• The existing record daemon handler is platform-specific for Apple/Android recordings, so web currently falls through as unsupported.
• fix: add web viewport control and screenshot aliases #865 is a useful precedent for surfacing an existing agent-browser backend capability through the --platform web layer; it is not a duplicate.

Agent brief:

Expose web screen recording through the agent-device web layer by adding record support for --platform web, delegating start/stop to the managed agent-browser backend, and preserving the existing record command/session semantics. Keep output agent-friendly: return artifact paths and explicit format information, likely .webm unless MP4 normalization is intentionally added.

Implementation should update the web command capability surface, the web provider/interactor contract, daemon record start/stop routing, CLI/client output expectations where affected, docs/help if the command becomes user-visible for web, and focused tests for unsupported-before/supported-after behavior. Add live web backend verification when practical; otherwise state the exact blocker.