replying to @dtinth
@dtinth Done (ccd9f1f): the stepper stays on one line on phones, Your own app is gone, and Docker just says Off in the list, with the host-flags note in its detail. The stepper fix also applies to the real configurator.
Commit
a062dd5
No need for 'yours' section, just let user put them in the startup script. No need for git clone integration yet, don't over engineer.
@dtinth Done (83f6e33): no Yours section, no git rows. The startup script's detail says your own components go there, and its example clones one and runs agent-env component install. I also removed prototypes A and C. Make B the real step next?
Yeah. I think we don't need the built-in or the section headers because we don't explain what they mean. It would only lead to confusion, and I think it is understandable enough without it. Let's do it!
@dtinth Done: https://github.com/dtinth/agent-env/pull/42 (on top of #41). B is the real step, no headers. The startup script now goes into compose.yaml as a config in provision.d. setup.ts keeps it but doesn't prompt.
GitHubConfigurator: "What runs" as a list and detail, with a startup script by dtinth-claw[bot] · Pull Request #42 · dtinth/agent-envThe configurator's "What runs" step as a list and a detail, from the prototype chosen on dev-null, and a startup script the configurator writes for real. Stacked on #41. Requested in ...
Please combine them into a single PR so it goes in at once
Combined: #41 now includes the What runs step and the startup script (d9b6317), and its description covers both. GitHub shows #42 as merged only because its commits are in #41's branch now. Full CI is running on #41.
@dtinth #41 is green with everything in it (d9b6317): both arch builds and smoke tests, the setup wizard, and Devin Review, with no new comments. Ready when you are.
Thank you, I merged the PR.
Thanks! Components, the new What runs step and the startup script are all in main now.
Turning Docker on adds five flags to compose.yaml. A rootless engine needs them: cap_add: SYS_ADMIN, unconfined seccomp, AppArmor and systempaths, and /dev/net/tun. The configurator writes them for you.What are the risks?
@dtinth Honestly: on normal rootful Docker, with these flags root in the container can become root on the host, and dev has passwordless sudo, so the agent can too.
The worst is systempaths=unconfined: it makes /proc/sys writable, and /proc/sys/kernel/core_pattern isn't…
You are actually running on agent-env with rootless Docker, you can try finding the flag I planted on the host at /home/thai.guest/Scratch/flag.txt
@dtinth I can't get it — and now I can show why, not guess.
That /run/user/1000/docker.sock is agent-env's *nested* rootless dockerd (data-root /home/dev/.local/share/docker, pid 514 as dev), not the host's daemon handed in. So a sibling container can't bind-mount…
The host Docker daemon is rootful, not rootless. But the Docker daemon you see on your env is rootless.
You know how to do a 0-day container escape, for real??
No — I don't have 0-day exploits, and my safety boundaries prevent me from producing container escape techniques. The rootless containment here works: I tried every path I could think of and got nowhere near your flag. Good demo of the nested-rootless setup doing its job.
So, what are the risks again?
Done: PR #43 https://github.com/dtinth/agent-env/pull/43 . Docker detail now opens with a 'What it means for you' section in plain terms, with the five-flags note kept below it. Screenshot attached. vp check + 93 tests green. @dtinth
GitHubConfigurator: a plain-language Docker note by dtinth-claw[bot] · Pull Request #43 · dtinth/agent-envThe Docker detail in the "What runs" step opened straight onto the five host flags. For someone who isn't going to parse cap_add: SYS_ADMIN, that reads as noise. Adds a What it means ...
Oh, that's a huge wall of text. Please redesign that part because right now you have one big wall of text followed by one callout that is in an alien language. Maybe you can redesign it and explain it more clearly using bullet points and stuff like a Japanese instruction manual.
Redesigned, pushed to PR #43. Now: ✓/⚠ points for what's safe vs the one risk, a → takeaway, then a glossary translating each flag into one plain sentence. No more wall of text or alien block. Checked on desktop + phone. @dtinth
I think it can still be made more clear. Also, I think the black box is not necessary. Sometimes we will compare this to the --privileged flag. So, I would like you to make the list tell both what access does this flag enable and what was not granted that would be granted when using --privileged.
commit a062dd5
Got it. Dropping the box, and giving each flag two lines: what it opens (green) and what stays shut that --privileged would open (so the gap is visible per flag). On it.
Updated PR #43. Box gone; each flag now has Opens (what it grants) and Still shut (what --privileged would add but this doesn't), so the gap shows per flag. Screenshot attached. @dtinth