feat(standalone_driver): single-node NetCDF/UGRID output via the common io module by kotsaloscv · Pull Request #1307 · C2SM/icon4py

kotsaloscv · 2026-06-11T07:41:51Z

What

Wires icon4py.model.common.io into the standalone driver for single-node
NetCDF/UGRID output of the full model state (phase 0; parallel output is a follow-up).
Built on the config-driven driver (after merging main).

Output

Opt-in via --enable-output (or DriverConfig.enable_output; default off). Writes the
initial state and then on a schedule, into one file per segment
<output_path>/icon4py_output_NNNN.nc (+ a *_ugrid.nc mesh) — all 10 fields together,
by CF name:

prognostic: air_density, exner_function, virtual_potential_temperature,
upward_air_velocity, normal_velocity
diagnostic (computed at output time): eastward_wind, northward_wind,
temperature, virtual_temperature, pressure

Schedule: every N model steps (output_interval_steps, default 1) or a
simulation-time interval (output_interval + start_time, e.g. "10 HOURS").

How

driver_io.py — state → CF/UGRID DataArrays (host-side, GPU-safe); metadata driven
from states/data.py (CF names + icon_var_name + is_on_half_levels);
DiagnosticsComputer (scratch/output buffers allocated once, reused per step).
standalone_driver.py — monitor built in initialize_driver (single-node guard),
per-step store in the time loop, close() in finally.
common/io — step- or time-based scheduling; grid_id strictly uuid.UUID;
to_data_array(to_host=...).
pyproject.toml — icon4py-common[io] is a required driver dependency.

Testing

Unit tests (test_driver_io.py, test_io.py) + JW integration datatest
(test_driver_io_output.py). pre-commit, unit suites, and the datatest are green.

Out of scope (follow-ups)

Single-node only — parallel output (gather-to-rank-0 / Zarr) is the next PR.

…prognostics + diagnostics) via common io module

…together in the same output

Common io module: - remove the 1 ms output-time tolerance: the model clock advances by exact datetime arithmetic, so no drift is possible (keep >= so a step landing past a scheduled output time is not skipped) - drop redundant mode=0o777 from mkdir calls (it is the default, umask-masked) - add GridIdentifier type alias (uuid.UUID | str) for uuidOfHGrid - IOMonitor accepts pathlib.Path for the grid file - to_data_array copies the attrs dict internally instead of mutating the caller's (fixes latent pollution of the shared CF tables); add to_host_data_array, moved from the driver, which forces buffers to host - add is_on_interface to OptionalMetaData Standalone driver: - replace the ad-hoc prognostic spec tuple with an explicit _PrognosticFieldSpec (state attribute name + FieldMetaData) - output is on by default, so import the io stack unconditionally: lazy imports, TYPE_CHECKING guards and pytest.importorskip removed (broken installs now fail loudly); uv.lock now actually pins icon4py-common[io] - rename the output subfolder from io/ to output/ - create_io_monitor rejects sub-second dtime instead of silently truncating the interval to 0 seconds; drop the unused process_properties parameter - tighten annotations (gtx.Field, gtx.typing.Backend | None, wpfloat) - main() takes all arguments explicitly; defaults live only in cli() - tests: dims-based expected layouts, backend-parameterized host-transfer test, remove "Phase 0" comments

- driver_io: drop the _PrognosticFieldSpec dataclass and getattr; assemble prognostics via direct, type-checked attribute access keyed by output name - io: type IOMonitor.grid_file_name strictly as pathlib.Path (callers convert) - driver_io: re-add the process_props parameter to create_io_monitor (named per convention, reserved for the distributed IO path) - standalone_driver main: replace the click() console-script shim with a module-level typer.Typer() app; point the pyproject entry point at main:app - tests: express expected prognostic output layout with FieldMetaData

Revert the GridIdentifier (uuid.UUID | str) union and type the grid id strictly as uuid.UUID throughout the IO layer: - writers: uuidOfHGrid back to uuid.UUID; drop the GridIdentifier alias - io: IOMonitor / FieldGroupMonitor take grid_id: uuid.UUID - driver: convert at the boundary with uuid.UUID(grid.id) (Grid.id holds the file's uuidOfHGrid as a string; the synthetic simple grid is not a UUID) - tests: stamp a UUID onto the simple grid for the monitor-construction tests Verified: ruff/mypy/tach, driver + common io unit tests, and the JW output integration datatest (gtfn_cpu) all pass.

common/io: - schedule output by step count: FieldGroupIOConfig.output_interval_steps (write every N calls to store); remove the time-based path (to_delta, OutputInterval, start_time, next_output_time). Output now lands on exact step boundaries regardless of the (possibly non-integer) time step length - rename FieldMetaData.is_on_interface -> is_on_half_levels - merge to_host_data_array into to_data_array(..., to_host=...) - update the module docstring examples to the step-based config standalone_driver: - drive prognostic output from states.data.PROGNOSTIC_CF_ATTRIBUTES directly (CF names, icon_var_name lookup, is_on_half_levels); drop the parallel metadata table - rename DiagnosticInputs -> DiagnosticFields; derive them from the field factories instead of passing them into the driver constructor - add opt-in initial-conditions output (output_initial_state) - nc_title/nc_comment tidy-up; de-"boxed" docstrings - revert the main:app entry-point change back to the click() shim (the entry point fix comes with #1304) Verified: pre-commit, common io + driver unit tests, and the JW output integration datatest (gtfn_cpu) all pass.

- to_data_array: document that the result usually references the field buffer (a view), and that to_host only copies on GPU - _store_output: note the assembled DataArrays alias the live state and must be written before the next step mutates it

Copilot

Pull request overview

This PR wires icon4py.model.common.io into the standalone driver to produce single-node NetCDF/UGRID output for the full model state, and updates the common IO layer to be more robust and step-scheduled.

Changes:

Add a standalone-driver IO bridge (driver_io.py) that assembles prognostic state + computes diagnostics at output time, and creates an IOMonitor.
Enable (default-on) output in the standalone driver loop with CLI toggles, and disable output automatically for MPI runs.
Update common.io to use step-based output scheduling and to support GPU-safe host-buffer extraction for NetCDF writing; add/adjust unit and integration tests.

Reviewed changes

Copilot reviewed 19 out of 20 changed files in this pull request and generated 4 comments.

Show a summary per file

File	Description
uv.lock	Update lockfile to include `icon4py-common[io]` for the standalone driver.
model/standalone_driver/pyproject.toml	Make `icon4py-common[io]` a required dependency (output default-on).
model/standalone_driver/src/icon4py/model/standalone_driver/main.py	Split CLI from programmatic `main()`, add `--enable-output/--no-enable-output`, add `click()` entry point.
model/standalone_driver/src/icon4py/model/standalone_driver/standalone_driver.py	Initialize IO monitor, store during time loop, close at end; disable output in MPI runs.
model/standalone_driver/src/icon4py/model/standalone_driver/driver_io.py	New driver-side bridge: DataArray assembly, diagnostic computation, and monitor factory.
model/standalone_driver/tests/standalone_driver/unit_tests/test_driver_io.py	Unit tests for DataArray assembly and IO monitor construction (no disk IO).
model/standalone_driver/tests/standalone_driver/integration_tests/test_driver_io_output.py	Datatest that runs one JW step and verifies NetCDF/UGRID output contents/shapes.
model/standalone_driver/tests/standalone_driver//test_.py	Update existing tests to pass new required args and disable output where not under test.
model/common/src/icon4py/model/common/io/io.py	Switch IO scheduling from time-based strings to step-based intervals; output-dir handling change.
model/common/src/icon4py/model/common/io/utils.py	Add `to_host` option and rename `is_on_interface` → `is_on_half_levels`; copy attrs defensively.
model/common/src/icon4py/model/common/io/ugrid.py	Rename vertical dim mapping flag to `is_on_half_levels`.
model/common/src/icon4py/model/common/io/init.py	Update module docs to reflect step-based IO config and fix links/examples.
model/common/src/icon4py/model/common/states/data.py	Mark `upward_air_velocity` as being on half-levels in metadata.
model/common/src/icon4py/model/common/states/model.py	Extend metadata TypedDict with optional `is_on_half_levels`.
model/common/tests/common/io/utils.py	Update test helper for renamed `to_data_array` arguments.
model/common/tests/common/io/unit_tests/test_io.py	Update tests for step-based scheduling and new config validation.

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

kotsaloscv · 2026-06-19T12:28:20Z

+    def _zero_full() -> gtx.Field:
+        return data_alloc.zero_field(
+            grid, dims.CellDim, dims.KDim, dtype=ta.wpfloat, allocator=backend
+        )
+
+    def _zero_interface() -> gtx.Field:
+        return data_alloc.zero_field(
+            grid,
+            dims.CellDim,
+            dims.KDim,
+            dtype=ta.wpfloat,
+            extend={dims.KDim: 1},
+            allocator=backend,
+        )
+
+    # dry air: all hydrometeors are zero
+    qv, qc, qi, qr, qs, qg = (_zero_full() for _ in range(6))
+    temperature = _zero_full()
+    virtual_temperature = _zero_full()
+    u = _zero_full()
+    v = _zero_full()
+    pressure = _zero_full()
+    # Typed as Any: gt4py's NDArrayObject protocol does not expose __setitem__, so the
+    # in-place buffer fill below would not type-check against the precise Field type.
+    pressure_ifc: Any = _zero_interface()
+    surface_pressure_k = _zero_interface()
+


Done — compute_diagnostics became a DiagnosticsComputer class that allocates the ~14 scratch/output buffers once and reuses them on every compute() call. The driver holds one instance (via a cached property), so the buffers are allocated only when output is actually enabled, then reused each step.

This is already great, but possibly should be synced with however we decide to group states in the near/mid future (and IO should follow what the driver does and reuse from there).
can you add a todo as a reminder also mentioning physics (which heavily uses diagnostic vars)?

or open an issue documenting this at https://github.com/orgs/C2SM/projects/27

# Conflicts: # model/standalone_driver/src/icon4py/model/standalone_driver/main.py # model/standalone_driver/src/icon4py/model/standalone_driver/standalone_driver.py # model/standalone_driver/tests/standalone_driver/integration_tests/test_initial_condition.py # model/standalone_driver/tests/standalone_driver/integration_tests/test_standalone_driver.py # model/standalone_driver/tests/standalone_driver/mpi_tests/test_parallel_initial_conditions.py # model/standalone_driver/tests/standalone_driver/mpi_tests/test_parallel_standalone_driver.py

…o the run dir - main: --enable-output/--no-enable-output toggles DriverConfig.enable_output (default off), so output is selectable from the console script - driver_io: drop the redundant OUTPUT_SUBDIR nesting; the monitor writes directly into the run output directory instead of <output_path>/output/

- common/io: FieldGroupIOConfig accepts exactly one of output_interval_steps (every N steps) or output_interval + start_time (time string, e.g. "10 HOURS"); restore OutputInterval/to_delta and a time-based capture path (>= so a step past the scheduled time still writes). Driver keeps using step-based. - driver: compute_diagnostics -> DiagnosticsComputer, which allocates its ~14 scratch/output buffers once and reuses them per step (avoids per-step allocation, notably on GPU); driver holds it via a cached_property. - tests: restore to_delta tests, add dual-mode validation + a write-free time-based capture test; document both schedule modes in the io docstring.

…s enabled Drop the output_initial_state flag from DriverConfig; the driver now always writes the initial state (before the first step) whenever output is enabled. Integration test asserts the two resulting time slices (initial state + one step).

Output is opt-in (--enable-output, default off), so the "enabled by default" comment was inaccurate. icon4py-common[io] stays a required dependency for now; making it an optional extra is left to a follow-up.

# Conflicts: # model/standalone_driver/src/icon4py/model/standalone_driver/standalone_driver.py

msimberg · 2026-06-19T13:09:37Z

        return refinement_control_fields

+    @property
+    def file_path(self) -> str:


Argh, not new... but why do we store this as a str inside GridManager 😞

If you assume at the grid_manager.file_path call site that you can convert it to a Path, I'd force it already here to be a Path, or even better in GridManager.init (I'm not insisting since that touches parts way outside this PR).

I agree that this is weird. However, let's leave it like this for now and deal with it in a follow up PR. It is a bit out of scope currently ;)

msimberg · 2026-06-19T13:13:18Z

+        # Exactly one scheduling mode is set (enforced by the config validation).
+        self._step_based = config.output_interval_steps is not None
+        if config.output_interval_steps is not None:
+            self._output_interval_steps = config.output_interval_steps
+            self._step_counter = 0
+        else:
+            assert config.output_interval is not None and config.start_time is not None
+            self._next_output_time = dt.datetime.fromisoformat(config.start_time)
+            self._time_delta = to_delta(config.output_interval)


Same comment here about time or step intervals: I think it's nicer if it's normalized to steps only at this point, and then all the logic afterwards can use steps?

Exactly what it does now: the interval is normalized to steps once (in _interval_in_steps, called from FieldGroupMonitor.__init__), and from there on the monitor only ever counts steps — _at_capture_time is a plain step_counter % N. The model-clock path is gone. (c95a7de)

To do for later: use the type aliases from

icon4py/model/standalone_driver/src/icon4py/model/standalone_driver/config.py

Lines 37 to 40 in a9a63fd

RelativeTime: TypeAlias = datetime.timedelta

AbsoluteTime: TypeAlias = datetime.datetime

NumTimeSteps: TypeAlias = int

EndOfSimulation: TypeAlias = RelativeTime | AbsoluteTime | NumTimeSteps

.

…write output - io: replace the dual output_interval_steps/output_interval(str)+start_time schedule with a single output_interval (int steps or datetime.timedelta), normalized to steps once at monitor construction via dtime; drop the model-clock path and the to_delta string parser. Add DeltaT/NumTimeSteps/OutputInterval type aliases. - io: FieldGroupMonitor refuses to overwrite an existing data file (the per-run file counter restarts at 0), so a rerun into a populated directory fails loudly. - driver: thread dtime through create_io_monitor -> IOMonitor -> FieldGroupMonitor. - driver_io: add TODO on sharing diagnostic buffers with future state-grouping/physics. - tests: stamp the fake grid uuid in the grid fixture instead of per-test; cover the timedelta->steps normalization, too-short-interval error, and overwrite refusal.

The IOMonitor source grid file is used solely to regenerate the UGRID topology (`_write_ugrid`); the grid identity comes from `grid_id` (the `Grid` object), not the file. Remove the unused `_read_grid_attrs` so the file path has one clear purpose, and leave a TODO to eventually build the UGRID topology from `Grid`/`GridGeometry` and drop the file dependency entirely.

…ticsComputer

Use the OutputInterval union alias directly in the isinstance check and render it into the error message, so both update automatically if the alias changes.

jcanton

lots of minor comments, it's looking good!

jcanton · 2026-06-22T08:28:08Z

-                self.close()
+        if not self._at_capture_time():
+            return
+        # TODO(halungge): this should do a deep copy of the data


what about this TODO?

Kept it on purpose. It's not needed today because we write synchronously: store() reads the field buffers and appends them to the NetCDF file immediately, before the model takes the next step and mutates them. The deep copy only becomes relevant once IO is asynchronous/buffered (queueing captures or accumulating slices in memory) — likely with the parallel-IO work — so the TODO is a useful marker for that.

jcanton · 2026-06-22T08:30:38Z

-        return self._next_output_time == model_time
+    def _at_capture_time(self) -> bool:
+        # fire every N model steps (one call to ``store`` == one model step)
+        self._step_counter += 1


this counter update should probably not live in this method

Agreed — moved it out. _at_capture_time() is now a pure predicate (step_counter % N == 0); the increment happens in store() instead, where "one call == one model step". (878bb18)

jcanton · 2026-06-22T08:40:36Z

    ],
 )
-def test_fieldgroup_config_validate_filename(start_time, filename, interval, variables, message):
+def test_fieldgroup_config_validate_filename(filename, output_interval, variables, message):


not you but missing type annotations
can we add some parameterized with the DeltaT output_interval?

also: when parameterizing by output_interval: OutputInterval specify the parameterization as NumTimeSteps(1) / DeltaT(whatever)
safer for (disabled) type checking

The whole file is type-checked now ;)

jcanton · 2026-06-22T08:40:56Z

not you but missing type annotations in several (all?) test defs

can you enable type checking for this file?

The NETCDFWriter assigned gt4py/cupy device arrays directly into netCDF4 variables. On a GPU backend this raised `TypeError: Implicit conversion to a NumPy array is not allowed` from cupy, because netCDF4/HDF5 write from host memory and cupy forbids implicit device->host conversion. Route the three write sites (interface heights, new-variable init, and time-slice append) through `data_alloc.as_numpy(...)`, which performs an explicit D2H copy for cupy/gt4py fields and is a no-op for numpy (so CPU backends are unaffected). The CPU-only unit `test_writers.py` did not exercise this; the GPU integration test (`test_driver_io_output.py`) now passes on both gtfn_gpu and dace_gpu. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

- Rename the output dimension `interface_level` -> `half_level` everywhere (writers/ugrid/cf_utils + tests) to match the `is_on_half_levels` metadata flag; CF standard_name values are left untouched. (no UGRID/CF standard mandates the name) - simple_grid() now carries a fixed synthetic UUID id, so tests no longer patch it. - Use the OutputInterval/NumTimeSteps/DeltaT aliases explicitly; add a ruff extend-immutable-calls allowlist so they are valid as defaults (one noqa for RUF009). - Move the per-step counter increment out of the _at_capture_time predicate into store(). - Add a restart TODO at the refuse-to-overwrite check and a state-grouping/physics refactor TODO on DiagnosticsComputer. - standalone_driver: _store_output takes simulation_current_datetime: AbsoluteTime; annotate the try/finally; shorten the --enable-output help text. - Enable mypy type-checking for test_io.py: annotate all test defs, wrap grid_id with uuid.UUID, fix two errors in the io test utils helper.

jcanton · 2026-06-22T14:05:36Z

+    #: Output schedule: either a number of model steps (``int``) or a simulation-time
+    #: delta (``datetime.timedelta``); a delta is normalized to steps using the model time
+    #: step. Defaults to every step.


not necessary now, but the types in the comment could be updated

jcanton

very nice

github-actions · 2026-06-22T14:18:46Z

Mandatory Tests

Please make sure you run these tests via comment before you merge!

cscs-ci run default
cscs-ci run distributed

Optional Tests

To run benchmarks you can use:

cscs-ci run benchmark-bencher

To run tests and benchmarks with the DaCe backend you can use:

cscs-ci run dace

To run test levels ignored by the default test suite (mostly simple datatest for static fields computations) you can use:

cscs-ci run extra

For more detailed information please look at CI in the EXCLAIM universe.

kotsaloscv · 2026-06-22T14:19:04Z

cscs-ci run default

kotsaloscv · 2026-06-22T14:19:11Z

cscs-ci run distributed

kotsaloscv · 2026-06-22T14:19:26Z

cscs-ci run dace

kotsaloscv · 2026-06-22T14:19:36Z

cscs-ci run extra

feat(standalone_driver): add opt-in single-node NetCDF/UGRID output (…

96e25bb

…prognostics + diagnostics) via common io module

msimberg self-requested a review June 11, 2026 07:45

kotsaloscv added 2 commits June 11, 2026 13:06

Output always on by default + prognostic & diagnostic fields written …

e92106b

…together in the same output

Merge branch 'main' into io_module

d72e4f5

kotsaloscv marked this pull request as ready for review June 11, 2026 11:12

kotsaloscv requested a review from jcanton June 11, 2026 11:13

jcanton reviewed Jun 11, 2026

View reviewed changes

Comment thread model/standalone_driver/src/icon4py/model/standalone_driver/standalone_driver.py

msimberg requested changes Jun 12, 2026

View reviewed changes

kotsaloscv added 2 commits June 12, 2026 11:41

Merge branch 'main' into io_module

34246a9

kotsaloscv changed the title ~~feat(standalone_driver): add opt-in single-node NetCDF/UGRID output~~ feat(standalone_driver): default-on single-node NetCDF/UGRID output via the common io module Jun 12, 2026

kotsaloscv added 2 commits June 15, 2026 13:03

jcanton requested changes Jun 15, 2026

View reviewed changes

kotsaloscv added 3 commits June 18, 2026 12:53

refactor(io): rename DiagnosticFields -> DiagnosticInputFields

1519191

jcanton reviewed Jun 18, 2026

View reviewed changes

Comment thread model/standalone_driver/src/icon4py/model/standalone_driver/driver_io.py Outdated

jcanton reviewed Jun 18, 2026

View reviewed changes

Comment thread model/standalone_driver/src/icon4py/model/standalone_driver/standalone_driver.py Outdated

msimberg requested review from Copilot and msimberg June 19, 2026 07:51

Copilot started reviewing on behalf of msimberg June 19, 2026 07:52 View session

Copilot AI reviewed Jun 19, 2026

View reviewed changes

msimberg reviewed Jun 19, 2026

View reviewed changes

Comment thread model/common/tests/common/io/unit_tests/test_io.py Outdated

Comment thread model/standalone_driver/src/icon4py/model/standalone_driver/driver_io.py Outdated

kotsaloscv added 3 commits June 19, 2026 12:35

kotsaloscv changed the title ~~feat(standalone_driver): default-on single-node NetCDF/UGRID output via the common io module~~ feat(standalone_driver): single-node NetCDF/UGRID output via the common io module Jun 19, 2026

kotsaloscv added 2 commits June 19, 2026 14:05

chore(standalone_driver): drop stale io-dependency comment

f5f8d71

Output is opt-in (--enable-output, default off), so the "enabled by default" comment was inaccurate. icon4py-common[io] stays a required dependency for now; making it an optional extra is left to a follow-up.

Merge remote-tracking branch 'origin/main' into io_module

7a8fea4

# Conflicts: # model/standalone_driver/src/icon4py/model/standalone_driver/standalone_driver.py

msimberg reviewed Jun 19, 2026

View reviewed changes

kotsaloscv added 3 commits June 19, 2026 16:47

docs(driver_io): reword 'static inputs' -> 'static fields' in Diagnos…

ed1453d

…ticsComputer

jcanton reviewed Jun 21, 2026

View reviewed changes

Comment thread model/common/src/icon4py/model/common/io/io.py Outdated

jcanton reviewed Jun 21, 2026

View reviewed changes

Comment thread model/common/src/icon4py/model/common/io/io.py Outdated

jcanton reviewed Jun 21, 2026

View reviewed changes

Comment thread model/common/src/icon4py/model/common/io/io.py

kotsaloscv added 2 commits June 22, 2026 09:56

Merge branch 'main' into io_module

be41f90

refactor(io): validate output_interval against the OutputInterval alias

a72c4f0

Use the OutputInterval union alias directly in the isinstance check and render it into the error message, so both update automatically if the alias changes.

jcanton requested changes Jun 22, 2026

View reviewed changes

kotsaloscv and others added 2 commits June 22, 2026 12:27

jcanton reviewed Jun 22, 2026

View reviewed changes

jcanton approved these changes Jun 22, 2026

View reviewed changes

Merge branch 'main' into io_module

cf860db

kotsaloscv enabled auto-merge (squash) June 22, 2026 14:23

kotsaloscv disabled auto-merge June 23, 2026 06:33

kotsaloscv merged commit 678e45c into main Jun 23, 2026
70 of 71 checks passed

kotsaloscv deleted the io_module branch June 23, 2026 06:39

	RelativeTime: TypeAlias = datetime.timedelta
	AbsoluteTime: TypeAlias = datetime.datetime
	NumTimeSteps: TypeAlias = int
	EndOfSimulation: TypeAlias = RelativeTime \| AbsoluteTime \| NumTimeSteps

Uh oh!

Conversation

kotsaloscv commented Jun 11, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What

Output

How

Testing

Out of scope (follow-ups)

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

jcanton left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

kotsaloscv commented Jun 11, 2026 •

edited

Loading