Developers add Playwright browser demos to AI-authored pull request reviews

• Review packages for AI-authored changes now combine readable diffs, pseudocode, Mermaid diagrams, and TDD evidence, as catalogued in mattpocockuk's list of examples.
• A Playwright-recorded browser walkthrough with text-to-speech narration has entered the proposed review stack for frontend changes, according to mattpocockuk's example.
• AI-generated review comments are becoming another output to inspect, after GergelyOrosz's follow-up called out teams that measure useful findings against noise.
• Agent-produced optimization work still ends at a maintainer decision: LearnOpenCV's update says its OpenCV pull request was submitted for the core team to accept or reject.

The Playwright video documentation says recordings are saved only when the browser context closes, a lifecycle detail for publishing test footage. A request captured in LearnOpenCV's first task simply asked an agent to make OpenCV's imread faster; the resulting change later entered core-team review.

A developer behind an autonomous GitHub setup wrote that 26 agent roles had merged 1,135 pull requests only after code review, security review, and acceptance in a September 4 account. In mattpocockuk's proposed packet, the reviewer gets five different ways to inspect an AI-authored change.

1. A simplified diff through /show-me.
2. A pseudocode description of the change.
3. A Mermaid diagram.
4. Evidence of test-driven development.
5. A Playwright recording of the working frontend feature, narrated with text-to-speech.

dexhorthy said they had already experimented with Playwright recordings where the model adds captions and circles page elements in an overlay, but had not built the speech-to-text version.

Playwright Test exposes four video modes in its official guide:

• off, which records nothing.
• on, which records every test.
• retain-on-failure, which removes footage from successful runs.
• on-first-retry, which records the first retry only.

The guide says video files usually land in test-results after browser-context closure.

GergelyOrosz opened the discussion by asking teams that still do human review how they keep up with the increase in code and pull requests. His follow-up singled out a separate practice, reviewing the AI review itself and looking for noise.

A reply from GergelyOrosz added that schema checks alone do not cover business logic, which had traditionally been a core target of code review.

DeepLearningAI's summary of an Andrew Ng letter places reviewing work inside a five-part coding-agent skill map.

1. Directing the workflow.
2. Enabling agent autonomy.
3. Reviewing the work.
4. Customizing the agent and its environment.
5. Understanding coding-agent foundations.

The resulting OpenCV pull request changes PNG header parsing so PngDecoder::readHeader() can skip materializing large IDAT data that would otherwise be copied and then read again. Its tests cover memory and file inputs, corrupt and partial IDAT cases, APNG behavior, and multiple PNG formats.

LearnOpenCV reported roughly 9% faster loading on macOS and 17% on Linux for a large uncompressed PNG in the final result, then said the pull request had been submitted for core-team review.

GitHub's status-check documentation defines checks as pull-request readiness signals and says required checks must pass before a protected branch can be merged.

daniel_mac8 describes astra-advisor as an open-source orchestration plug-in for using OpenAI models at different reasoning efforts inside Codex.

Its completed-subagent view lists:

• Product ship review.
• Fitting engine.
• Product final review.
• Configuration ship review.
• Configuration final review.
• Configuration observer.
• Fitting app.
• Configuration review.

A separate agent session reported a validated backend build and tests, an integration commit, and a drafted Unity pull request, while read-only GitHub Actions blocked publication.