pickValue: Native Framework Support Plan

Problem Summary

CWL v1.2 workflows use pickValue on workflow outputs (and step inputs) to merge multiple sources, selecting non-null values. Galaxy crashes when importing these workflows because parser.py:get_outputs_for_label() hardcodes multiple=False on outputSource, and no runtime logic exists to apply pickValue semantics when collecting workflow outputs.

27 CWL v1.2 conditional tests are RED because of this gap.

pickValue Patterns in CWL Tests

Two distinct patterns exist:

Pattern A: Multiple outputSource (most tests)

outputs:
  out1:
    type: string
    outputSource: [step1/out1, step2/out1]
    pickValue: first_non_null

Steps have when expressions; some produce null. The workflow output gathers from multiple steps and picks among them.

Pattern B: Single outputSource + scatter (cond-wf-009, 010, 011)

outputs:
  out1:
    type: string[]
    outputSource: step1/out1
    pickValue: all_non_null

A single step is scattered with when; some scatter elements produce null. pickValue filters nulls from the scatter result array.

Current Architecture

Workflow Model (model/init.py)

WorkflowOutput lives on a single WorkflowStep:

class WorkflowOutput(Base):
    workflow_step_id  # FK to workflow_step
    output_name       # which output of that step
    label             # the workflow output label

A workflow output is bound to one step and one output_name. There is no mechanism for a workflow output to reference outputs from multiple steps.

WorkflowStepConnection models data flow between steps — it connects an output step to a WorkflowStepInput on the consuming step. Multi-source connections (for step inputs) work via multiple WorkflowStepConnection rows pointing to the same WorkflowStepInput.

CWL Parser (tool_util/cwl/parser.py)

WorkflowProxy.get_outputs_for_label(label) iterates CWL workflow outputs, calls split_step_references(outputSource, multiple=False) which asserts a single reference. It returns a list of {"output_name": ..., "label": ...} dicts that get placed in the step dict’s workflow_outputs list.

WorkflowProxy.to_dict() produces a Galaxy workflow dict where each step has a workflow_outputs key. The problem: a CWL workflow output with outputSource: [step1/out1, step2/out1] references TWO different steps, but Galaxy’s dict format puts workflow_outputs inside each step dict — there’s no place for a cross-step output.

Workflow Import (managers/workflows.py)

_workflow_from_raw_description() walks step dicts. For each step, if workflow_outputs exists, it creates WorkflowOutput model objects bound to that step (line ~1941-1964). There is no mechanism to create a workflow output that spans multiple steps.

Workflow Execution (workflow/run.py)

WorkflowProgress.set_step_outputs() iterates step.workflow_outputs and calls _record_workflow_output() for each. This records the output in the invocation via workflow_invocation.add_output(workflow_output, step, output).

get_replacement_workflow_output() looks up a workflow output by going to its step and finding the output by name in self.outputs[step.id].

Null/Skipped Outputs

When when_values == [False] (step skipped entirely), the tool still executes but produces “empty” datasets. These get hidden (output.visible = False). The outputs dict still has entries — they’re just empty/hidden HDAs, not Python None.

For CWL, a skipped step should produce null for its outputs. Currently Galaxy represents this as an empty HDA, which is not the same thing. The WorkflowInvocationOutputValue table stores JSON values, so it can store None.

Proposed Approach

Strategy: Duplicate-Label WorkflowOutputs + Post-Processing

Rather than fundamentally restructuring WorkflowOutput to span multiple steps (which would be a massive model change touching export, import, editor, API, and every workflow feature), use a simpler approach:

Add pick_value metadata to WorkflowOutput and handle it during output collection in run.py.

The key insight: Galaxy’s workflow model already supports a workflow output being on a specific step. For Pattern A (multiple outputSource), we need multiple WorkflowOutput objects with the same label on different steps. For Pattern B (scatter+pickValue), we need pickValue logic on a single WorkflowOutput.

Currently, the label uniqueness is not enforced at the DB level — it’s just convention. And set_step_outputs() already iterates all WorkflowOutput objects per step. We can:

1. Create multiple WorkflowOutput objects with the same label on different steps
2. Add a pick_value column to WorkflowOutput
3. At the end of execution, post-process outputs with the same label using pickValue semantics

Alternative Considered: Direct Model Restructuring

Adding multi-source workflow outputs to the model would require:

• New join table workflow_output_source (workflow_output_id, step_id, output_name, position)
• Changes to WorkflowOutput to remove step FK or make it nullable
• Changes to every export format (ga, format2, editor dict, instance dict)
• Changes to the workflow editor UI (which we’re explicitly not touching)
• Changes to run.py output collection
• Changes to the API schema for WorkflowOutput

This is a much larger change with much broader impact. The duplicate-label approach is more contained.

Detailed Plan: Native pick_value on WorkflowOutput

Phase 0: Fix noJS 400 Errors (Prerequisite)

File: lib/galaxy/tool_util/cwl/parser.py, cwl_input_to_galaxy_step (~line 909)

CWL workflow inputs with type: int[] AND default: [1,2,3,4,5,6] are imported as data_collection_input steps, which discards the default value. The array-type branch (line 922-924) does NOT enter the USE_STEP_PARAMETERS path that sets tool_state["default"] and tool_state["optional"]. When test job provides no value (relying on defaults), _normalize_inputs finds no input, no default, optional=False → HTTP 400.

Fix: When items is a scalar type (int, string, float, boolean), treat as parameter_input with parameter_type: field instead of data_collection_input. This enters the path that handles default and optional correctly. For File arrays, keep existing data_collection_input behavior.

This unblocks 5 noJS conditional tests that fail before pickValue logic is even reached:

• condifional_scatter_on_nonscattered_true_nojs (should pass with Phase 0 alone — all conditions true)
• condifional_scatter_on_nonscattered_false_nojs
• scatter_on_scattered_conditional_nojs
• conditionals_multi_scatter_nojs
• conditionals_nested_cross_scatter_nojs

Phase 1: Model Changes

Add column to workflow_output table:

# In model/__init__.py, class WorkflowOutput:
pick_value: Mapped[Optional[str]] = mapped_column(String(64), nullable=True)

Valid values: None, "first_non_null", "the_only_non_null", "all_non_null".

Migration:

# New alembic migration
def upgrade():
    add_column("workflow_output", Column("pick_value", String(64), nullable=True))

Update copy():

def copy(self, copied_step):
    copied_output = WorkflowOutput(copied_step)
    copied_output.output_name = self.output_name
    copied_output.label = self.label
    copied_output.pick_value = self.pick_value
    return copied_output

Update _serialize():

def _serialize(self, id_encoder, serialization_options):
    d = dict_for(self, output_name=self.output_name, label=self.label)
    if self.pick_value:
        d["pick_value"] = self.pick_value
    return d

Phase 2: Parser Changes (parser.py)

Fix get_outputs_for_label() to handle multiple outputSource:

def get_outputs_for_label(self, label):
    outputs = []
    for output in self._workflow.tool["outputs"]:
        source = output["outputSource"]
        pick_value = output.get("pickValue")

        # Handle both single and list outputSource
        references = split_step_references(
            source,
            multiple=True,  # Changed from False
            workflow_id=self.cwl_id,
        )

        for step, output_name in references:
            if step == label:
                output_id = output["id"]
                if "#" not in self.cwl_id:
                    _, output_label = output_id.rsplit("#", 1)
                else:
                    _, output_label = output_id.rsplit("/", 1)

                out_dict = {
                    "output_name": output_name,
                    "label": output_label,
                }
                if pick_value:
                    out_dict["pick_value"] = pick_value
                outputs.append(out_dict)
    return outputs

This means if a CWL workflow output has outputSource: [step1/out1, step2/out1], then:

• get_outputs_for_label("step1") returns [{"output_name": "out1", "label": "out1", "pick_value": "first_non_null"}]
• get_outputs_for_label("step2") returns [{"output_name": "out1", "label": "out1", "pick_value": "first_non_null"}]

Both steps get a WorkflowOutput with the same label but pick_value set.

Also handle input steps: The CWL output in cond-wf-003.cwl references both a step output AND a workflow input (def). The cwl_input_to_galaxy_step() method already calls get_outputs_for_label(label), so input steps would also get WorkflowOutput objects if referenced in a multi-source outputSource. This already works.

Phase 3: Import Changes (managers/workflows.py)

Update __module_from_dict to read pick_value:

In the workflow_outputs loop (~line 1944-1964), add:

for workflow_output in workflow_outputs:
    if not isinstance(workflow_output, dict):
        workflow_output = {"output_name": workflow_output}
    output_name = workflow_output["output_name"]
    # ... existing validation ...
    uuid = workflow_output.get("uuid", None)
    label = workflow_output.get("label", None)
    m = step.create_or_update_workflow_output(
        output_name=output_name,
        uuid=uuid,
        label=label,
    )
    # NEW: set pick_value
    pick_value = workflow_output.get("pick_value", None)
    if pick_value:
        m.pick_value = pick_value
    if not dry_run:
        trans.sa_session.add(m)

Relax duplicate label check: Currently found_output_names checks for duplicate output_name within a step, not duplicate labels across steps. This should be fine — duplicate labels across steps is the whole point.

Phase 4: Execution Changes (workflow/run.py)

Add a post-processing step after all steps are scheduled.

Currently set_step_outputs() calls _record_workflow_output() for each WorkflowOutput on each step as it completes. For pickValue, we need to defer final output recording until all source steps have completed, then apply pickValue logic.

Option A: Post-process at invocation completion

After all steps are scheduled in WorkflowInvoker.invoke(), before setting state to SCHEDULED, iterate all workflow outputs with pick_value set and apply the logic:

# In WorkflowProgress or WorkflowInvoker, after all steps scheduled:
def apply_pick_value_outputs(self):
    """Post-process workflow outputs that have pick_value set."""
    # Group WorkflowOutput objects by label
    outputs_by_label = defaultdict(list)
    for step in self.workflow_invocation.workflow.steps:
        for wo in step.workflow_outputs:
            if wo.pick_value:
                outputs_by_label[wo.label].append(wo)

    for label, workflow_outputs in outputs_by_label.items():
        pick_value = workflow_outputs[0].pick_value
        values = []
        for wo in workflow_outputs:
            step_outputs = self.outputs.get(wo.workflow_step.id, {})
            output = step_outputs.get(wo.output_name)
            values.append(output)

        result = apply_pick_value(pick_value, values, label)
        # Record the final aggregated output
        # Use the first workflow_output as the "primary" record
        self.workflow_invocation.add_output(
            workflow_outputs[0], workflow_outputs[0].workflow_step, result
        )

The apply_pick_value function:

def apply_pick_value(pick_value, values, label):
    """Apply CWL pickValue semantics to a list of values."""

    def is_null(v):
        # A value is null if it's None, NO_REPLACEMENT,
        # or a hidden empty HDA from a skipped step
        if v is None or v is NO_REPLACEMENT:
            return True
        if isinstance(v, dict) and v.get("__class__") == "NoReplacement":
            return True
        # For HDA outputs from skipped steps:
        if hasattr(v, "dataset") and not v.producing_job_finished:
            # Skipped step - output is null
            return True
        return False

    non_null = [(i, v) for i, v in enumerate(values) if not is_null(v)]

    if pick_value == "first_non_null":
        if not non_null:
            raise FailWorkflowEvaluation(...)  # "All sources are null"
        return non_null[0][1]

    elif pick_value == "the_only_non_null":
        if len(non_null) != 1:
            raise FailWorkflowEvaluation(...)
        return non_null[0][1]

    elif pick_value == "all_non_null":
        return [v for _, v in non_null]  # Return as list/collection

Option B: Modify _record_workflow_output to defer pick_value outputs

In set_step_outputs(), when encountering a workflow output with pick_value set, don’t record it immediately — instead, accumulate it in a pending_pick_value_outputs dict. Then at the end of scheduling, resolve them.

This is cleaner because it doesn’t double-record outputs.

Phase 5: Null Detection

The hardest part is reliably detecting “this output is null because the step was skipped.”

Currently when a CWL step has when=False:

• The tool still executes (with __when_value__: False)
• Galaxy creates output HDAs that are empty and hidden
• These are not semantically “null” — they’re empty datasets

For pickValue to work correctly, we need to distinguish:

• “Step produced an empty dataset” (not null, just empty)
• “Step was skipped, output is null” (should be null)

Proposed approach: Track skipped-step outputs explicitly.

In set_step_outputs(), when progress.when_values == [False]:

if progress.when_values == [False]:
    for output_name in outputs:
        self._null_outputs[(step.id, output_name)] = True

Then in apply_pick_value, check _null_outputs instead of trying to infer nullness from HDA state.

For Pattern B (scatter+pickValue on single source), the null detection is different — individual scatter elements may be null while others are not. The scatter produces a collection, and the collection elements with when=False are the null ones. Galaxy already has skipped state for collection elements (see migration c39f1de47a04_add_skipped_state_to_collection_job_), so this may already work for detecting null elements within a scatter result.

Phase 6: Pattern B — Scatter + pickValue

For all_non_null on a scattered output, the expected behavior is:

• Scatter produces a list collection
• Elements from skipped iterations are null
• all_non_null filters out null elements, returning a smaller list

Galaxy represents scatter results as HistoryDatasetCollectionAssociation (list collections). The filtered result would be a new collection with only the non-null elements.

This requires:

1. After scatter execution, identify which collection elements came from skipped iterations
2. Create a new filtered collection excluding those elements
3. The filtered collection becomes the workflow output

This is more complex than Pattern A and may warrant a separate implementation phase.

Phase 7: Export Changes

Workflow export (ga format, format2) needs to serialize pick_value:

In _workflow_to_dict_export() (managers/workflows.py), the workflow_outputs serialization already includes output_name and label. Add pick_value:

# In the step dict construction for export
for workflow_output in step.unique_workflow_outputs:
    wo_dict = {
        "output_name": workflow_output.output_name,
        "label": workflow_output.label,
        "uuid": str(workflow_output.uuid),
    }
    if workflow_output.pick_value:
        wo_dict["pick_value"] = workflow_output.pick_value

Phase 8: should_fail Tests

Several CWL v1.2 tests expect workflow execution to FAIL:

• first_non_null_all_null — all sources null, first_non_null should error
• the_only_non_null_multi_true — multiple non-null, the_only_non_null should error
• all_non_null_multi_with_non_array_output — all_non_null on non-array type should error

These currently pass because the import crashes (so the test “succeeds” as a should_fail). After fixing the import, the pickValue runtime logic must produce the correct errors for these to keep passing.

Benefit to Galaxy-Native Workflows

Galaxy-native workflows already support when expressions (added in 23.0). If pickValue were added to the runtime layer:

1. Galaxy workflows could express “take first available output” — e.g., two conditional branches where exactly one runs, merged into a single output via the_only_non_null. Currently Galaxy users must use a “pick value” tool or restructure their workflow.

2. all_non_null for filtered scatter results — Galaxy workflows with conditional scatter could produce filtered output collections.

3. The UI integration could come later — the runtime layer would work, and the Galaxy workflow editor could add pickValue configuration in a future release.

4. Format2 support — Galaxy’s format2 workflow format could natively express pickValue on outputs, making conditional workflow patterns cleaner.

The infrastructure cost is low: one new column, one new post-processing function. The conceptual fit is good since Galaxy already has when, linkMerge/merge_type, and conditional step support.

Size Estimate

Total: Medium-sized change. Pattern A (multi-source pickValue) is the primary blocker for most tests. Pattern B (scatter+pickValue) is more complex and could be a separate phase.

Implementation Order

1. Model + migration (pick_value column on workflow_output)
2. Parser fix (multiple=True in get_outputs_for_label, pass pickValue through)
3. Import fix (read pick_value from workflow dict)
4. Null tracking in execution (when_values==[False] -> mark outputs null)
5. pickValue post-processing in run.py (Pattern A: multi-source)
6. Export serialization
7. Pattern B: scatter + pickValue filtering (separate PR if needed)

Testing Plan

• Red-to-green on CWL conformance tests: The 27 RED tests listed in CWL_CONDITIONALS_STATUS.md are the primary targets.
• Phase 0 first: condifional_scatter_on_nonscattered_true_nojs should go green with just the int[] defaults fix (all conditions true, no pickValue filtering needed).
• Start with Pattern A tests (cond-wf-003 through 007 variants): these are the simplest — two sources, one skipped.
• Then Pattern B tests (cond-wf-009 through 013): scatter+conditional+pickValue.
• Scatter+conditional tests (formerly GAP5, 9 tests total):

• Verify should_fail tests stay green after import no longer crashes.
• Run Galaxy-native workflow tests to confirm no regressions (the new column and execution logic should be no-ops when pick_value is NULL).

linkMerge + pickValue Composition

CWL spec: linkMerge applies BEFORE pickValue. For conditionals_multi_scatter which uses linkMerge: merge_flattened AND pickValue: all_non_null, the runtime must merge sources first, then filter nulls. The apply_pick_value_outputs() post- processing runs after merge operations in replacement_for_input_connections(), so ordering is naturally correct for Pattern A. For Pattern B (scatter), pickValue filters null elements from the scatter result collection after merge.

Review Notes

Reviewed against CWL_CONDITIONALS_STATUS.md and source code.

Factual Corrections

1. Pattern B scope is wrong. Plan says Pattern B is “cond-wf-009, 010, 011” but cond-wf-011 (conditionals_nested_cross_scatter) retains null values in nested arrays — pickValue: all_non_null applies only at the outermost level. cond-wf-013 (conditionals_multi_scatter) is a Pattern A+B hybrid (multiple outputSource + scatter + linkMerge + pickValue). These are distinct patterns the plan doesn’t distinguish.

2. Step input pickValue is deprioritized. Grep of all v1.2 conditional test workflows confirms pickValue appears ONLY on workflow outputs, never on step inputs. Zero conformance tests exercise it. Deprioritize the WorkflowStepInput.pick_value question.

3. condifional_scatter_on_nonscattered_false semantics. This test expects out1: [] when ALL scatter elements are skipped. The entire collection is null, not individual elements — different from Phase 6’s “filtering null elements from a collection.”

Missing Considerations

4. SubworkflowStepProxy when bug (from status doc). SubworkflowStepProxy.to_dict() does NOT extract when. Not a pickValue blocker but a related gap.

5. Editor duplicate-label warning. _workflow_to_dict_editor() tracks output_label_index across steps and flags duplicates as upgrade_message_dict["output_label_duplicate"]. CWL-imported workflows with pickValue will trigger this. May need to suppress when pick_value is set.

6. Import output_name uniqueness guard. workflows.py:1949-1952 raises ObjectAttributeInvalidException for duplicate output_name within a step. Not triggered for Pattern A (different steps), but an implicit constraint.

Approach Correctness

7. Double-recording risk with Option A. add_output() appends without checking for duplicate labels. If set_step_outputs() records per-step AND apply_pick_value_outputs() records aggregated, there’ll be duplicates. Option B (defer recording) is strongly preferred.

8. all_non_null list result type. apply_pick_value returns a Python list. add_output() dispatches on history_content_type — a list has none, so it’d be WorkflowInvocationOutputValue (JSON blob). May work for CWL conformance but for Galaxy-native use should be HDCA.

9. linkMerge + pickValue composition order is answered. CWL spec: linkMerge applies first (merge/flatten), then pickValue filters nulls. Plan should incorporate this.

Unresolved Questions

• How to detect “output is null from skipped step” vs “output is an empty dataset”? The when_values tracking is per-invocation, not per-output. Need reliable null marker.
• For Pattern B, should the filtered collection be a new HDCA or should Galaxy support “sparse” collections with null elements?
• Should pick_value on WorkflowOutput also support step inputs? CWL allows pickValue on step inputs too (not just workflow outputs). Galaxy’s WorkflowStepInput already has merge_type — should we add pick_value there as well?
• The duplicate-label WorkflowOutput approach — will the workflow editor handle two outputs with the same label gracefully, or will it need special-casing?
• For the all_non_null mode returning a list: if the original output type is File but all_non_null returns File[], should this create a list collection? The CWL type system expects this, but Galaxy would need to dynamically produce a collection from scalar outputs.
• Should the first_non_null/the_only_non_null failures produce CWL-spec-compliant error messages? The spec says specific error conditions for each mode.
• cond-with-defaults.cwl uses both linkMerge: merge_flattened AND pickValue: all_non_null on the same output. How do these compose? Does pickValue operate before or after linkMerge?