Changelog.

A run can no longer come back as Ship when its checks did not run. Any check that failed to complete floors the verdict, half of them unrun returns Inconclusive, and the summary line is built from what actually ran, so "all checks passed" cannot appear above a check that never started.

Link checking has an undetermined state. Probes now carry a real browser User-Agent, anything HTTP calls broken gets a second opinion from a real browser, and a link behind a bot wall or a login screen is reported as "could not be verified" rather than broken. Unverifiable links never count toward the verdict.

Findings are verified against what they claim rather than whether the problem would matter. Quoted text is checked against the rendered page, a finding whose evidence does not reach its claim is rejected, and rulings the judge never reasoned about are no longer recorded as confirmations.

Visual findings are checked against the page they describe, not the screenshot alone, so a wrapped line is not reported as clipped and a selector matching nothing is not reported at all. Copy checks match text a user can actually see, so scripts and hidden elements no longer produce critical false positives.

The performance score is computed on a log-normal curve instead of three buckets, so the headline number reflects how slow a page actually is. Seven sites that previously all scored the same now spread across the range. Which findings fire is unchanged.

Code review no longer asserts external facts it cannot check from the diff, such as whether a package version exists or what a third-party library does at runtime.

Replying "fixed" to a finding now re-checks the other files the pull request changed, not only the file the comment was pinned to, so a fix that lands in a different module is recognised instead of leaving the thread open.

POST /api/runs rejects fields it does not understand instead of silently ignoring them, so a request asking for a specific set of agents either runs exactly that set or returns a 400 naming the field. Agents not implemented yet are rejected at validation rather than accepted and skipped.

Concurrent runs no longer exhaust the browser pool, clearing the "Vision agent unreachable" and "Page returned 0" failures that appeared when several runs overlapped.

Preview URL detection picks up URLs from commit statuses as well as comments, recognises hostnames posted without a scheme, and no longer misses a deployment that finished before the shipguarde label was applied.

The founding pilot programme is open, with a page setting out what is included and how to apply.

PR reviews now diff against the merge base, so a branch that is behind its target is reviewed only for its own changes, matching the GitHub "Files changed" tab.

Committable code suggestions keep the original line's indentation, so "Apply suggestion" produces correct code.

Ask a PR to test the UI with no URL and visual QA now falls back to your project's default URL.

New Features overview page bringing both products together in one place.

Conversational PR commands: mention @shipguarde in plain English and it parses the intent into a run.

Reply "fixed" to any finding and ShipGuarde re-checks the file at the latest commit, verifies the fix, and resolves the thread.

Repository rename and transfer no longer break the project mapping; the webhook remaps automatically.

Pull requests are analyzed in a secret-starved, single-use sandbox. We never run your repository's own scripts, tests, or build.

Opt-in merge gating: turn a block into a required failing check via an in-repo .shipguarde.yml or the GitHub Action.

Inline comments are anchored to real changed lines, with off-diff findings demoted to the summary instead of a random line.

Simplified to USD across every plan.

Run-completion notifications by email and Slack, with per-run recipient lists.