[rudycelekli]

Three existing positive validation controls launch actual search CLIs without a response fixture: two in LinkedIn and one in Freehire. They assert only that the error is not BAD_ARG, so they remain green when the child attempts an external request and subsequently fails for a network reason. The Freehire comment incorrectly claimed that checking the absence of an error made the case offline.

Serve the LinkedIn controls a saved two-card response over loopback HTTP, and point Freehire's control at its existing loopback API override. Assert exit 0, empty stderr, actual freshness/pagination parameters and output; LinkedIn also checks IDs and client-side result limits. The four Danish validation suites already stop valid-value controls at required-argument guards, so they need no change.

Credit: portapulito captured the real LinkedIn guest response in #153, closed as an accidental personalized-fork submission. The JSON-encoded excerpt preserves its HTML bytes; source commit, original/excerpt hashes and capture date are recorded beside it. Freehire uses an explicitly synthetic empty JSON protocol response because this test concerns option/request construction, not portal content parsing.

Validation on Bun 1.4.3 / TypeScript 5.9.3:

• Before: unchanged six flag-validation suites pass, but a deny-and-record observer in their actual CLI children records 3 nonloopback request attempts (LinkedIn 2, Freehire 1). Every attempt is denied before transport; no live portal request was allowed.
• After: the complete six reference CLI suites pass (341 tests, 0 failures) under an OS loopback-only sandbox, and the helper-child observer records 0 nonloopback attempts. All six typechecks and repository guards pass.
• Evaluator controls: deleting the actual freshness-parameter construction makes both LinkedIn controls fail and the Freehire control fail; the source is restored afterward.

This isolates these three controls. A JavaScript preload is not whole-CI OS egress enforcement; our complete local qualification separately uses an OS network sandbox. Existing public CI runs of unchanged upstream tests should not be described as intrinsically network-free on this basis. No production code, runtime dependencies, credentials or profile data changed.

At exact head 86bdf759, the complete 579-test Python suite and all three repository guards also pass.

[MadsLorentzen]

Verified on the real path and merge-ready; one ask before it lands: add a line under [Unreleased] > Fixed in CHANGELOG.md (one sentence in the file's house style, PR number in parentheses). Merge on push.

[rudycelekli]

Addressed in 931e0d88393b: added the requested one-sentence entry under [Unreleased] → Fixed, with the PR number in parentheses.

The follow-up changes only CHANGELOG.md; changelog structure, skill lint, framework-version and security guards passed locally.

[MadsLorentzen]

Merged. The three master controls really did attempt linkedin.com and freehire.me requests and stayed green with the parameter construction deleted; the new controls fail on that mutation and record zero non-loopback attempts. Fixture verified verbatim against the #153 capture.