[MaxGhenis]

Four engine-free (3.14) jobs that started on 2026-11-10 between 22:37 and 23:48 UTC, one each on #779, #1567, #1572 and #1575, each failed one test on an ubuntu-26.04 runner. uv took that image's Python 3.14.4, and test_spawned_workers_score_identically_to_sequential_cli failed, because the head-to-head scorer refuses --worker-max-slices below Python 3.14.7. The same job on an ubuntu-24.04 runner downloaded Python 3.14.10 and passed.

What changes

• Every job in .github/workflows/test.yml runs on ubuntu-24.04. Each said ubuntu-latest, which names whichever image GitHub assigns.
• The workflow sets UV_MANAGED_PYTHON=1. uv's help says uv uses a system Python when it holds no managed one, and that this option disables system Pythons. With it, uv takes only a Python it manages, and it downloads one when it holds none.
• docs/agent-guide.md gains two sentences on the image and the setting, and changelog.d gains one fragment.

Invariant

Every job's runs-on reads ubuntu-24.04, and the workflow's env holds UV_MANAGED_PYTHON: "1". A new test in test_ci_integration_job.py reads both from the workflow with PyYAML. The test fails when I put one job back on ubuntu-latest, and it fails when I delete the variable.

Limits

• The pin defers the move to ubuntu-26.04. A later pull request moves every job at once and updates the test.
• No job has run on ubuntu-26.04 with UV_MANAGED_PYTHON=1, so I haven't watched uv download its own Python on that image.
• select-countries, verify and the ci job's require-results run on the image's own python3, which this pull request leaves alone.

Tests

• The three CI-plan test files (test_ci_test_plan.py, test_ci_integration_job.py, test_run_engine_test_categories.py): 279 tests pass locally.
• python3 tools/ci_test_plan.py verify prints verification=ok, verify-workflow prints workflow=ok, and ruff check and ruff format --check pass on the test file.

axiom: n/a: CI infrastructure

🤖 Generated with Claude Code

[MaxGhenis]

I merged this at 897e021. All 12 checks passed on that head, every job on ubuntu-24.04, and two independent Opus 5.5 reviews agreed at it after one correction to the description.