Julia 1.11 aborts at runtime and no Julia code declares a minimum version

[egpbos]

Problem description

Julia 1.11.9 aborts on Kapteyn (eddington) whenever Julia runs its event loop, which kills install.sh at phase 5 and import proteus. Julia 1.12.7 on the same machine is fine. None of our Julia codes declares a minimum Julia version, so nothing catches this: AGNI has no [compat] at all, Obliqua has [compat] but no julia entry, and the SOCRATES Julia wrappers are not a package.

Expected behavior

An unsupported Julia should give a version error, not a core dump.

Evidence

On eddington with Julia 1.11.9, bash install.sh dies in phase 5 at get_agni.sh line 46:

julia: src/unix/linux.c:1369: uv__io_poll: Assertion `(*__errno_location ()) == 17' failed.
[3221611] signal 6 (-6): Aborted
src/get_agni.sh: line 46: 3221611 Aborted (core dumped) julia -e 'using Pkg; Pkg.activate("."); Pkg.instantiate()'

After juliaup default 1.12, the same clean clone installs all 8 phases, proteus doctor passes, and the unit and AGNI test suites are green.

Additional notes

Proposal: julia = "1.12" compat in AGNI and Obliqua, and a VERSION check in the SOCRATES wrappers since they have no Project.toml.

Two things would then need to agree with that: install.sh still accepts 1.11 (ACCEPTED_JULIA_MINORS="11 12") and deliberately falls back to it when Python links OpenSSL below 3.5, and Obliqua's troubleshooting docs still tell users to pin 1.11 for that same OpenSSL reason. A 1.12 floor implies requiring OpenSSL 3.5 or newer.

I have only shown 1.11.9 failing on this one host, not that 1.11 is broken everywhere.

My computer

eddington, Rocky Linux 9.8, x86_64, Python 3.12.14, Julia 1.11.9 and 1.12.7 via juliaup.

Relevant people

@nichollsh @timlichtenberg @MarijnJ0

[nichollsh]

We should probably just enforce Julia=1.13, no? This is the latest release version and comes with a bunch of nice improvements/changes compared to 1.12 and 1.11

Adding a [compat] section to the AGNI/Obliqua Project.toml files seems like a good idea.

[egpbos]

1.13 also fine for me! I thought because it's just out perhaps it may still be a bit unstable. At least with Python, usually people try to support a range of recent (actively supported) versions, but perhaps this is different in the Julia community.

[egpbos]

If we do this, we can also remove the OpenSSL self-heal block in verify_proteus_import and all its helpers.

[egpbos]

If we do this,

... and also pin the Python side, see #887.

[MarijnJ0]

I just tested Obliqua with Julia 1.13 and this seems to work fine, the changes will be added shortly (FormingWorlds/Obliqua#33). Can we treat this seperate from the existing Obliqua branch in PROTEUS? or do we need to add this check before merging #859 ?

[nichollsh]

I've made similar changes to the AGNI repo. I think this can be closed when #859 updates the pinned versions.

[MarijnJ0]

Ok cool, I have updated the pinned versions now.

[egpbos]

Follow-up after #859: a fresh bash install.sh --no-data on main (82c0f3e) on eddington now fails in phase 5 (AGNI + FastChem):

ERROR: julia version requirement for package at `.../PROTEUS/AGNI` not satisfied: compat entry "julia = ~1.11, ~1.13" does not include Julia version 1.12.7

The AGNI pin and install.sh now only overlap at 1.11:

• pyproject.toml [tool.proteus.modules.agni] pins nichollsh/AGNI@1c4baad ("Specify Julia 1.11 and 1.13"), which sets julia = "~1.11, ~1.13" in Project.toml. That means [1.11.0, 1.12.0) plus [1.13.0, 1.14.0), so every 1.12.x is excluded.
• install.sh still has REQUIRED_JULIA_MINOR=12 and ACCEPTED_JULIA_MINORS="11 12". Fresh installs pin 1.12, and a 1.13 default is rejected in phase 2 and repinned to 1.12 via juliaup.
• 1.11 is the version that aborts on eddington (this issue), so there is no Julia version that installs there on current main. The OpenSSL < 3.5 fallback (fix_julia_111_for_openssl) also lands on 1.11.

CI does not hit this because it installs through .github/actions/setup-proteus, not install.sh, and #859 also moved that action's default julia-version from 1.12.6 to 1.13.0. The 2026-09-23 nightly (https://github.com/FormingWorlds/PROTEUS/actions/runs/35836548378) ran on 1.13.0 and passed on ubuntu-latest and macos-latest, including the 4 AGNI integration tests below.

Workaround: nichollsh/AGNI@414b56c (version 1.12.5) differs from the pin only in .github/workflows/* and the [compat] lines (git diff --stat 414b56cc 1c4baad0). With AGNI_GIT_REF=414b56cc6558f62b2c211731d4e88aa05c6829f8 under Julia 1.12.7, all 8 install phases completed on eddington and the 4 AGNI coupling tests passed (PROTEUS_CI_NIGHTLY=1 pytest tests/integration/test_integration_agni_transparent_limit.py tests/integration/test_integration_agni_greygas_interior.py). proteus doctor then only warns about the pin mismatch.

Two ways to make them agree: move install.sh to 1.13 (REQUIRED_JULIA_MINOR=13, ACCEPTED_JULIA_MINORS matching AGNI's compat, and the "matches what CI tests" comment is then true again), or add 1.12 to AGNI's compat. @nichollsh was leaving out 1.12 intended? The thread above points at 1.13 as the target, but I found no stated reason to drop 1.12 while keeping 1.11.

Not verified:

• Other hosts. The compat check does not depend on the host, so I expect any fresh install.sh run that ends up on 1.12 to fail the same way, but I have only seen it on eddington.
• AGNI and juliacall under 1.13 on eddington or any machine other than the GitHub runners.

[nichollsh]

I have updated AGNI to allow support for Julia 1.12, which I believe should resolve this issue. It now has compat for 1.11/1.12/1.13

I strongly recommend we use 1.13 as the main target. We should use this as the default for the pixi installs with PROTEUS and elsewhere.

@egpbos does this resolve the issue?

[egpbos]

I will make a PR that will close this from the PROTEUS side: 1. pin to updated AGNI, 2. update install.sh's default Julia version to 1.13 as well, 3. add a deprecation warning for 1.11, nudging people towards updating so we can drop support later. The latter is for discussion, let me know if you disagree.