PR reviews page, chat agent upgrades, and mobile network logs

Previously, triggering a change review via the API would fail if the repository wasn't connected to a project through a GitHub or GitLab integration. Now, requests that include application overrides (specifying which environment to test against) skip the integration check entirely and run against the provided URL. Existing API calls that don't supply an environment continue to use the integration mapping as before.

Browser-based tests have always shown full network activity – requests, headers, payloads, and responses. Mobile tests via Appetize used to show an empty Network tab. That data is now captured and displayed in the same Network tab. Binary responses and payloads over 100 KB are filtered, consistent with browser test behaviour.

Previously, these agent actions would pause in a "Waiting for approval" state – the change was proposed but nothing happened until you approved it. That step is removed. The agent now applies test case edits and device preset changes as soon as it decides to, and shows you what changed.

Previously, if you wanted to discuss a specific PR with the chat agent, you had to paste the relevant context yourself. The agent can now fetch PR information and review diffs directly – ask it to summarise a PR, explain what changed, or dig into a diff.

When the chat agent had prepared a PR review and the activity appeared in the in-progress banner, clicking the dismiss button looked like it worked but the activity stayed stuck. Dismissing a pending PR review now correctly cancels the activity.

The test agent used to treat any step it couldn't directly verify as exercised as an automatic failure – even when the step completed on its own (a cookie banner dismissing on page load, credentials submitted via Enter rather than a button click). Those cases now pass correctly rather than overruling the agent's own observed result.

Some older chat sessions contained tool outputs saved in a format that no longer matched the current schema. Opening those sessions caused the chat to fail to render entirely. The parser now degrades gracefully when it encounters an unrecognised shape – the affected tool result renders as empty rather than crashing the whole conversation view.