since-cutoff 0.5.0: dependency switches (httpx -> httpx2), published benchmark results, and an explainer video

[MohammadHijjawi97]

since-cutoff 0.5.0 is on PyPI (uvx since-cutoff@latest scan). It adds three things: a new kind of change, published benchmark results, and a video.

1. Dependency switches: the break a diff of the library's own API misses

openai 3.x, anthropic 1.8, huggingface-hub 2.0 and mcp 2.2 now require httpx2 and no longer httpx, and their public signatures take httpx2 objects. Code written from memory passes an httpx.Client or an httpx.Timeout, or catches httpx.HTTPError, and fails at run time. In our benchmark this switch was behind 5 of the 17 post-cutoff tasks, and 0.4.1 could not see it.

0.5 reads both releases' declared requirements and checks where the public API names the dropped library's types. It then writes one short note:

openai 3.19.2 requires httpx2 instead of httpx: OpenAI(http_client=...) takes httpx2.Client (httpx2.AsyncClient for AsyncOpenAI), OpenAI(timeout=...) takes httpx2.Timeout and OpenAI(base_url=...) takes httpx2.URL. Use httpx2 there, not httpx. [diff + metadata]

It also handles a switch inside an extra (fastmcp-slim), a copy the package ships (typer's own typer._click), and a library the old release already required (gradio). Code is only flagged as the old form where it actually hands the package an old object; importing httpx next to openai for your own use is fine.

Tested on 1,081 unit tests and 13 network tests against the real releases, after two independent reviews on more than 25 packages.

2. The benchmark is published

360 Claude Code sessions (Opus 5.5) on 24 tasks with hidden tests, with the protocol frozen before the main run.

• Pass rate: the agent alone already passed 94% of the post-cutoff tasks, so we make no pass-rate claim.
• With since-cutoff 0.4.1's notes against the agent alone: cost ratio 0.80 (95% CI 0.70-0.90), 17% fewer turns, 13% less time.
• Stale API code in the first edit fell from 11.8% to 2.0%, and no control task got worse.
• Context7 ran without a key and hit its monthly quota partway through, so the Context7 comparisons are weaker. This is disclosed in full.

Write-up: https://mohammadhijjawi97.github.io/since-cutoff/benchmark.html. Everything, including all 360 transcripts, is at https://github.com/MohammadHijjawi97/since-cutoff-benchmark.

3. A 2½-minute explainer video

The video is attached to this release (since-cutoff-explainer.mp4, with captions).

Also:

• The agent skill now follows the Agent Skills spec (skills-ref validate passes).
• unapply is allowed in the skill.

Want to help?

There are about 60 open issues with context and acceptance criteria, labelled good first issue and help wanted. Two were picked up today.

Full changelog: https://github.com/MohammadHijjawi97/since-cutoff/blob/main/CHANGELOG.md