Christos, thanks for the quick turnaround. Your 0.1.41 patch addresses all four review points: the bash wrapper and its RUN_DEPENDS are gone, the "# machine-learning" leftover is gone, find-new-version does real work now, and the jail test in comment #4 is exactly what I was asking for. I have taken your patch, updated it and tested it here. The result is attached, and that is the version I would commit. Four changes on top of yours. 1) 0.1.41 to 0.1.43 (built 2026-08-31), distinfo regenerated. 2) find-new-version now reads the upstream manifest instead of comparing sizes: https://downloads.claude.ai/claude-science/latest/manifest.json It carries the version and the per-platform SHA256, and its linux-x64 value matches our distinfo exactly. So the check no longer has to download 143 MB to compare checksums, it can name the new version instead of sending the maintainer into the binary to read a string, and the "upstream rebuilt the same version" case you added still works, now for free. I corrected the comment above the target as well: upstream has no directory index, but it does publish a version index. 3) The licence block, and files/LICENSE is dropped: LICENSE= Proprietary LICENSE_NAME= Anthropic terms of service LICENSE_TEXT= Full terms can be found at https://www.anthropic.com/legal LICENSE_PERMS= auto-accept There is no document called "Anthropic Proprietary License". What exists is the Commercial and the Consumer terms of service, and which of the two applies depends on the user's plan, not on the port. files/LICENSE was not a licence text either but a hand written one line summary, and the framework installs it as share/licenses//LICENSE, so we would be shipping our own summary as if it were the vendor's licence. The substantive part is LICENSE_PERMS. dist-mirror plus pkg-mirror is a claim about redistribution rights, and I could not find a basis for it in Anthropic's terms. With plain auto-accept the framework's default deny sets RESTRICTED_FILES and the clean-restricted targets by itself, so the binary is not mirrored and no package is distributed from it. That matches the other proprietary vendor binaries in the tree, www/linux-widevine-cdm, benchmarks/linux-unigine-* and games/nwndata, which all decline redistribution explicitly. The practical cost is that users build the port themselves, which for a NO_BUILD port means fetching the binary from the vendor and installing two files. 4) The wrapper. DISABLE_AUTOUPDATER and DISABLE_UPDATES do not exist in this binary; a raw grep over the 143 MB finds zero occurrences of either name. What the binary does read is --no-auto-update, taken straight out of argv and stored as update.auto_update, and it is listed in the accepted flag set for "serve". There is also an "update" subcommand ("Self-update", per its own help text) plus a self-update route in the daemon, and a self-update would overwrite the pkg-managed binary in libexec. So the wrapper appends --no-auto-update to "serve" only, and refuses "update" unless it is "update --check", which upstream documents as report-only. pkg-message explains that, and it now also carries upstream's own wording about --dangerously-no-sandbox (full $HOME read/write and unrestricted network access) and the telemetry opt-out, since telemetry defaults to on. Verified here: poudriere testport on 151amd64 is green, no stage-qa warnings and the plist is complete; portlint only reports the PORTNAME hyphen; distinfo matches the upstream manifest; and the wrapper was exercised against a stub binary, where "serve --port 8000" gains --no-auto-update, "url" is passed through unchanged, "update" exits 1 and "update --check" passes through. You stay MAINTAINER, so please look the four changes over and say whether you agree, in particular the licence block, since that is a change of substance and not of style. If you would rather keep dist-mirror and pkg-mirror, say so and I will put that question to portmgr before committing instead of deciding it here. Mikhail, on the native node approach: I cannot commit it today, because archivers/unhare (bug 298005) is not in the tree and your patch needs it at extract time. Beyond that, replacing the vendor binary with an extracted and repackaged JS tree is a decision for this port's maintainer rather than for me as the committer, and licence-wise it is the larger question, not the smaller one. My plan is to commit the Linuxulator version now, and to leave the choice of approach to you and Christos once unhare has landed.