Skip to content
QGIS SDK

0.9.0 release gate (2026-10-31)

What was run on this checkout at the branch head built on main f466d86 (the PR #157 merge) on Linux x86-64, CPython 3.11.2 in a throwaway venv with the runtime and the tools installed from their src trees, Ruff 0.17.3, Alint 0.20.1. No QGIS and no Qt are present in this sandbox, so every native layer - the offscreen Desktop harness, the in-process QgsServer, the QgsTaskManager job tests - is recorded here as not run, and the Linux checks, Linux coverage, QGIS 3.44 LTS, macOS and Windows CI jobs are the rows that carry them. The tree is unstamped: the manifests still say 0.8.0 / 0.6.0, so the artifact names below carry 0.8.0. Stamping, tagging and publishing wait for an explicit authorization.

  • Required work (AC 1). TASK-115, 82, 116, 117, 80 and 81 are Done, and the milestone’s last two surfaces - the reference example’s server variant (examples/server, PR #157) and doc-12’s 0.9.0 notes - landed before this gate. TASK-83 is the gate itself.
  • Pure Python suite. pytest -c packages/qgis-sdk-python/pyproject.toml packages/qgis-sdk-python/tests 2411 passed, 93 skipped, 7 deselected in 2:00. The 93 skips are the native layers, each naming what it lacks. -m "pure and not smoke" is the coverage lane: 2411 passed, 32 skipped, 68 deselected.
  • Developer scripts and policy. pytest -c tests/scripts/pyproject.toml tests/scripts 289 passed (the two example-manifest tests added with the server variant included), alint check . 9/9 rules, ruff check . clean, ruff format --check . 470 files.
  • The example projects. qgis-sdk test --project examples/simple 20 passed, 7 deselected (the qgis target needs a QGIS, absent here), qgis-sdk test --project examples/server 8 passed, qgis-sdk check examples/simple checked 11 files: no problems, qgis-sdk check examples/server checked 3 files: no problems. qgis-sdk build on examples/server writes metadata.txt with server=True, keeps the source’s serverClassFactory and adds no classFactory.
  • Clean installation (AC 3). pip wheel --no-deps builds the three distributions (qgis_sdk-0.8.0, qgis_sdk_runtime-0.8.0, qgis_sdk_tools-0.8.0); a fresh venv installs qgis-sdk from those wheels with --no-index, getting the kit, the runtime and the tools plus typer==0.27.3, and its qgis-sdk new --host server writes the server starter (__init__.py with serverClassFactory, plugin.py, tests/test_plugin.py).
  • A project from the previous minor (AC 3). A 0.8-shaped project - async def job, ctx: JobContext, await ctx.workers.offload(...), and ctx.jobs.bind(count).delay(...) in an action - is read by qgis-sdk check as legacy08/plugin.py:17:18: JOB002 job countreadsctx.workers in its body, which a WorkerContext refuses, checked 2 files: 1 problem, exit 1: a finding with its remediation, not a crash. The shapes themselves are refused when the plugin binds, which is the message doc-12’s 0.9.0 notes carry.
  • Version policy, rehearsed (AC 2). release.py plan --scope python --trigger dispatch --bump minor --dry-run true reads python-v0.8.0 and plans python-v0.9.0; classify --tag python-v0.9.0 gives 0.9.0 (not a prerelease) and --tag python-v0.9.0-rc.1 gives 0.9.0rc1 (prerelease). A 0.9.0 stamp rehearsed on a scratch copy of the manifests (stamp --scope python) moves the kit, the runtime and the tools to 0.9.0, the tools’ runtime range to >=0.9.0,<0.10, the recipe’s runtime_version / tools_version / tools_runtime_requirement, and both reference examples’ windows to >=0.9,<0.10; verify-versions --scope python --version 0.9.0 then agrees on all eight entries, and --scope typescript --version 0.6.0 still agrees on its four. The real manifests are untouched.
  • Zero bridge delta (AC 5). git diff origin/main --stat -- packages/qgis-sdk-typescript is empty: no file under the bridge changed on this branch, no npm version is consumed, and verify-versions reports plugin description contract version 1 agrees across the Python declarations and the TypeScript bridge. HOST_FINGERPRINT, HOST_SCHEMA_VERSION and BRIDGE_PROTOCOL_VERSION are unchanged from 0.8.0, so doc-13 needs no entry for this release.
  • A lifetime defect the gate had to fix (AC 1). The autorelease quality gate went red on main after PR #157 with exit code 139: QgsTaskWorkers released its last Python reference inside _JobTask.finished(), while QgsTaskManager - which owns the task - still had processSubTasksForCompletion and the completion progress queued for it, so the process died in QgsTaskManager::progressChanged. It reproduced in whole-suite runs only (3 of 4 runs in one job, 1 of 2 in another; never in isolation), which is why CI’s split lanes had stayed green. The fix lets Qt delete the task (deleteLater()), retires the wrapper and releases it from the adapter’s own stack after destroyed; a native regression test pins it. After the fix twelve consecutive whole-suite runs (6 + 6, ~2609 passed each) were green. TASK-82’s notes carry the defect.
  • The LTS row itself was flaky, not failing. QGIS 3.44 LTS came back red twice on the gate PR with apt’s exit 100 and no test output: qgis.org terminated TLS mid-download for two of the ~30 packages (Could not handshake: The TLS connection was non-properly terminated). The row now retries with Acquire::Retries=5 and one step-level retry, then fails as before; it went green on the retry with all six rows passing.
  • The autorelease quality gate is green again (the defect’s last evidence). PR #158 merged as 63b6d108, and the autorelease run on that push (41222400860) passed every step including Run quality gate - the step that had been failing with exit 139 on f466d86, 7319794 and bc3fb1f - plus the dry-run packaging check. The whole-suite crash does not reproduce on main.
  • Not run here, and where it is run. The native layers: tests/test_jobs_tasks_native.py and the rest of -m native in Linux checks, the pure-layer coverage job in Linux coverage, and the server host plus the Desktop harnesses in QGIS 3.44 LTS. The windows/QGIS and macOS/QGIS rows stay unqualified per TASK-33.