Agentic QA

FORM PRIMITIVES

Puppeteer 24.43 finally teaches its locators about checkboxes and radios

via GitHub

For years, automating a checkbox click in Puppeteer has been the QA equivalent of parallel parking: doable, but rarely on the first try, and somebody always ends up writing a custom helper.

The new puppeteer-core 24.43.0 rolls to Chrome 148.0.7778.56 and extends the locator API to operate on checkboxes and radios directly. That closes a long-standing ergonomic gap that pushed test authors back to raw page.click calls plus ad-hoc waits, which is exactly how flakes get born.

The release also lands an allowlist implementation for locator ops and Firefox 150.0 updates, while the sibling browsers package 2.13.1 patches the WebUIReloadButton experiment.

• Auto-waiting locators now cover the most common form primitives end-to-end.
• Allowlist support lets you constrain what locator operations can touch.
• Firefox 150.0 is supported alongside the Chrome roll.

Why it matters: One of the top reasons teams fall back to raw clicks and brittle waits just got removed from the menu. read more.

HOTFIX HUSTLE

Claude Code shipped twice in six hours to unbreak Windows VS Code

via GitHub

Anthropic spent its Wednesday morning doing the dev-tools equivalent of a fire drill: two Claude Code releases in roughly six hours, the second one specifically to undo the first.

v2.1.129 went out at 01:40 UTC. v2.1.131 followed at 07:47 UTC, patching a createRequire polyfill bug whose hardcoded build path in the bundled SDK had silently bricked the VS Code extension on Windows. If your Windows agents stopped activating Claude Code overnight, this is your fix.

The earlier release wasn't just a stub either. It added --plugin-url for fetching plugin zips, a CLAUDE_CODE_FORCE_SYNC_OUTPUT env var for terminal output sync, gated gateway /v1/models discovery behind CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1, and stopped the Mantle endpoint from dropping the x-api-key header on auth.

The bits that matter for test harnesses

• Windows CI agents running the VS Code extension are unbroken.
• CLAUDE_CODE_FORCE_SYNC_OUTPUT kills interleaved-stdout flakes when capturing Claude Code from a test runner.
• Gateway model discovery is now opt-in, so existing pipelines won't surprise-call new endpoints.

Why it matters: Anyone running Claude Code inside a CI or test loop just had a silent Windows regression and a stdout-flake foot-gun fixed in the same morning. read more.

FALSE NEGATIVES

Your LLM code-translation eval is over-counting failures

via Unsplash

Turns out a meaningful chunk of the red on your AI code-translation dashboard isn't the model whiffing. It's your test harness crying wolf.

A new arXiv paper argues that a meaningful share of "failures" in LLM-based code-translation evals are false failures: tests fail not because the translated code is wrong, but because the harness mis-judges semantically equivalent output (different floats, different ordering, different exception text, same behavior).

The contribution is a methodology for separating genuine translation defects from harness-induced false negatives, which is exactly the diagnostic step most internal eval pipelines skip.

• If you run LLM-as-judge or differential testing on AI-generated code, your failure counts are probably overstated.
• Useful framework for sanitizing internal eval pipelines before you ship a "regression" report up the chain.
• Pairs naturally with mutation-style sanity checks on the harness itself.

Why it matters: Before you tell the team an AI tool got worse this week, prove the harness didn't. read more.

VERSION VERTIGO

Vitest 5.0 beta is the breaking-change release you should read before pinning

via GitHub

A week after the agent reporter landed in 4.1, Vitest's v5 beta is out, and it's the kind of release where letting Renovate auto-bump on a Friday is a self-own.

The headline breakages: the attachments directory has been restructured, the sequential test option is gone in favor of concurrent, and expect is now inlined. There are good additions too: merge-reports for multi-environment runs and expanded browser-mode capability.

• sequential removal will turn perfectly-good test files into red CI overnight if you don't update.
• expect inlining changes how custom matchers and shared assertion utilities resolve.
• Merge-reports finally make multi-environment runs reportable as one artifact.

Why it matters: Read the changelog now, on your terms, instead of at 9am Monday when the bot has already opened the PR. read more.