docs: reorganize testing docs for copy/paste

[abitrolly]

Description

Make it easier to scan and copy/paste the needed command. I forgot how to run tests, and my reading skills were all time low today.

Changelog

• Added changelog fragment: docs/changelog/<pr_number>.<type>.rst
  • Types: feature, bugfix, doc, removal, misc
  • Example: 123.feature.rst containing Add custom backend support - by :user:`yourname`

Checklist

• Tests pass locally (tox)
• Code follows project style (tox -e fix)
• Type checks pass (tox -e type)
• Documentation builds (tox -e docs)

[gaborbernat]

I disagree with removing existing information. I understand that I would prefer if we extended it, not clean it up. Your addition here should be strictly additive.

[abitrolly]

@gaborbernat I think I was careful enough to compress info. but not remove. I removed explanation like ``--only-integration` means only integration tests are run", because it is logical if not obvious. And merged examples directly into text that previously explained it. Actually I needed only copy/pasteable examples on one page to quickly fix tests while I have time. So feel free to format how do you see fit as long as the code blocks are here. The PR allows edits for maintainers, so we should not spend too much time here.

By the way, did you use any AI to make #988 ? It is pretty awesome piece of change that I would personally find trouble doing by hand. But I still haven't figure out how to run agents securely.

[henryiii]

What do you mean by "securely? If you mean access to things on your computer, then you can run in a container (I haven't done that yet, just been manually okaying operations), or GitHub's got various mechanisms to run agents (see copilot-instructions.yml, or codespaces).

For private code, then I think you'd need to run locally; I don't really do much that's not open source, so not something I've had to worry about yet. I've been playing with Qwen 3.6 locally and have had issue with it mis-formatting tool calls and paths. (It's pretty smart, it just randomly loses characters; it usually iterates and fixes it, but it's really annoying and makes it take 3x longer). I'm thinking it might be an issue with a small model + OpenCode's 10K token preamble, I'm tempted to try Pi instead, which has a 1K preamble. I also have access to Kimi-K2.5 on nrp.ai, which is really nice (open weight!), but not small enough to run locally. I'm planning on trying Gemma 4 soon. Kimi-K2.5 feels pretty close to GitHub Copilot + claude-haiku-4.5, maybe even a bit above it.

(I think that particular PR looked like it was assisted by Claude Code, at a guess)

[henryiii]

To me, it looks pretty similar in content, with just some lost text for the introduction that should be pretty easy to simply put back in.

[abitrolly]

I've heard some scary stories about agents following rogue instructions while doing lookups on the internet, and sending out private keys and other sensitive data. Also masking instructions with Unicode symbols to make them invisible. Also the statements like that Aider code is 80% written by AI, and also projects like beads + GasTown + Wasteland that run completely on autopilot using repository credentials make me feel really uncomfortable. Especially when code reviews are also done by agents. llama.cpp that exposes API through a web-server that is, well, probably written in C++? I already run all things in a container, but I've also heard that containers also have attack vectors, and with every layer of security I just fall into a trap of analysis paralysis.

[henryiii]

I've used Kimi-K2.6 to restore missing information while keeping the original structure intact (that was basically my prompt!). Maybe that's better? Might be okay to trim a bit.

[layday]

I agree with the general sentiment that the docs are too verbose and visually indistinct. A few suggestions below.