Four products, four verification questions

The protocol standards reinforce the separation. The Chrome DevTools Protocol defines JSON commands and events across Chromium domains, while its tip-of-tree form changes frequently and carries no backward-compatibility guarantee [1]. The W3C WebDriver BiDi Working Draft defines bidirectional remote control of user agents and is the native protocol boundary for the Firefox adapter [3].

What the local evidence establishes

The focused suites were executed through their real repository entry paths and their outputs frozen before manuscript generation. The recorded outcomes are 85 passed and one skipped for Chrome CDP, 19 passed for firefox-cdp, and 98 passed for File Blast Radius. The skip remains a skip. Each total describes the named suite at the recorded source state; none measures product size, adoption, or comparative quality.

The Chrome repository also reports a first-party synthetic-fixture benchmark: 0.33 ms p50 for an evaluate round trip after 20 warm-ups across 200 measured samples on Chrome 151 [4]. That fixture is reproducible, but the timing is environment-specific. The companion accessibility-tree field study shows why even a raw node count can move with rendering conditions while page content remains byte-identical [12].

Remote Skynet PC: progressive evidence, not manufactured parity

A bounded run on DESKTOP-OJ9K9QR tested what was actually available. One localhost scan found no usable Chrome CDP endpoint across ports 9222–9230. Firefox was not found. Three exact public low-tier Blast Radius gate tests passed, but installed/public source drift prevented a broader product pass. Skynet Desktop was not installed in the probed registry or default paths, and the run observed an older release state rather than v0.1.3.

The run did not install software, kill or relaunch a browser, or repeat the earlier 58-probe sequence. Those limits are not defects to hide; they define the measured boundary.

This is valuable negative evidence. It shows that a progressive harness can continue across independent products without converting an absent capability into a pass or declaring permanent impossibility from one machine state.

The harness upgrade derived from DeepSeek

DeepSeek Harness is a developer-preview, plugin-oriented agent harness [9]. Its testing policy separates unit, per-file coverage, real-API end-to-end, snapshot, and browser-snapshot tiers. It states that coverage is necessary but insufficient and emphasizes real entry-path tests that “check the world” because mocks can leave a green suite around a broken product [10].

Skynet adopted that principle as behavior, not branding. The installed tools/skynet_audit_harness.py classifies refusal, policy, rate limit, capability absence, transport failure, empty response, worker error, and timeout. Every class maps to a bounded recovery route; none maps to approval. Thirty-two focused tests lock installed/candidate byte identity and recovery behavior.

The smallest reusable capability is an evidence-contract record per field:

{
  "subject": "chrome_cdp_endpoint",
  "environment": "DESKTOP-OJ9K9QR",
  "entry_path": "127.0.0.1 ports 9222-9230",
  "expected": "reachable CDP endpoint",
  "observed": "none found in one bounded scan",
  "state": "capability_absence",
  "next_route": "rerun exact benchmark after safe endpoint setup",
  "receipt": "sha256:…"
}

That record is falsifiable: a later run can challenge the same subject under declared conditions. It is also operational: the next route is executable. The design avoids two failure modes at once—stopping the entire mission because one field is unavailable, and pretending the unavailable field passed so the mission can continue.

Security and trust boundaries

Attaching to a signed-in browser reaches real cookies, sessions, and extensions. Attach-first operation and exact tab ownership are therefore constraints, not proof that remote debugging is generally safe. A routable debugging endpoint can expose an entire profile; loopback binding, per-run secrets, owned-target checks, and cleanup evidence remain mandatory.

The Firefox adapter reduces collision risk by launching an isolated visible profile and limiting cleanup to processes carrying its unique profile path. Skynet Desktop reuses existing CLI sign-ins and requires no API keys, but the Windows installer is unsigned. Electron’s own security guidance treats remote content and powerful desktop capabilities as a trust boundary [8]. Users should verify the published checksum and treat artifact provenance, publisher identity, runtime permissions, and installation success as separate questions.

What this study does not establish

The local suites were run on one source checkout and one host. Repository statements and timing figures remain first-party until independently replicated. The Remote-PC run evaluated useful exclusions, an older Desktop state, and a source-drifted Blast Radius installation; it is not a four-product success run.

The study is authored by the product builder, which creates selection and interpretation risk. Exact hashes, raw receipts, public entry paths, explicit failed/unavailable states, and the frozen evidence appendix reduce that risk without eliminating it.

The safe thesis is narrow: verification must follow the product layer; unlike measurements cannot be averaged, and a typed exclusion is a measured boundary with a next route—not a pass, approval, or permanent impossibility. It would overreach to say these four products are “fully validated,” “machine-comparable,” or remotely installed from the evidence reported here.