concurrent foreach work

docs/adr/0002-concurrent-foreach-policy-and-barrier-commits.md

@@ -95,6 +95,12 @@ Patch creation and commit must extract/reuse the existing node output
 validation, output binding, and reducer logic. Concurrent foreach must not
 create a second write system.
 
+Current sync V1 implements the barrier commit path only for `loop -> one node ->
+END` item bodies. The runtime includes an explicit no-op overlay seam
+(`state_view_for_frame`) so the next slice can add lineage-local reads without
+rewiring node execution. Until that seam becomes real, multi-step concurrent
+item bodies are rejected instead of reading stale parent state.
+
 ## Merge and Reducer Rules
 
 At a barrier, missing reducer means default replace only for single-writer

docs/superpowers/plans/2026-05-22-concurrent-foreach-phase4-roadmap.md

@@ -0,0 +1,166 @@
+# Concurrent Foreach Phase 4 Roadmap
+
+> **For agentic workers:** REQUIRED SUB-SKILL: Use superpowers:subagent-driven-development (recommended) or superpowers:executing-plans to implement this plan task-by-task. Steps use checkbox (`- [ ]`) syntax for tracking.
+
+**Goal:** Split concurrent foreach execution into safe, independently testable runtime slices.
+
+**Architecture:** `foreach(mode="concurrent")` is the workflow mode. Sync runtime should interleave admitted item frames one node call at a time; async runtime may later run admitted async node handlers simultaneously. All item writes must flow through `StatePatch` and commit at a foreach barrier, not directly from child frames into shared state.
+
+**Tech Stack:** Python 3.14, dataclasses, Pydantic v2, pytest, `wf_core.runtime.scheduler`, `wf_core.runtime.foreach_state`, `wf_core.runtime.ops.state`.
+
+---
+
+## Current State
+
+Already implemented:
+
+- `ForeachNode.mode` accepts canonical `"concurrent"`.
+- `ForeachConcurrentPolicy` exists with `max_active`, `max_outstanding`, and `interrupt="quiesce"`.
+- Legacy `mode="parallel"` / `parallel={...}` parse into canonical concurrent shape.
+- `ForeachItemErrorPolicy` exists with `fail`, `skip`, and `collect`.
+- `completed_with_errors` is derived for `skip` and `collect`.
+- `collect_to` is validated as a declared array state field.
+- `StatePatch`, `build_output_patch(...)`, and `commit_state_patch(...)` exist.
+- `ForeachBarrierState`, `PendingItemResult`, and `ItemErrorRecord` exist.
+- Serial foreach progress now uses `ForeachBarrierState`.
+- Runtime still rejects `foreach(mode="concurrent")`.
+
+## Non-Goals For Phase 4
+
+- Do not implement Fork/Gather graph nodes.
+- Do not turn `JoinNode` into a real barrier.
+- Do not run sync node handlers in threads or processes.
+- Do not add OpenTelemetry.
+- Do not add platform-level source/tool semaphores.
+- Do not add persistent run storage.
+
+## Slice 1: Sync Concurrent Foreach, Fail-Only
+
+Implement first because it proves the scheduler and frame admission model without async task orchestration or handled item failures.
+
+Scope:
+
+- `foreach(mode="concurrent", item_error.action="fail")` runs in sync runtime.
+- Parent foreach admits up to `concurrent.max_active` item frames.
+- Scheduler interleaves item frames one step at a time.
+- Child output writes are buffered as per-item `StatePatch` objects.
+- Barrier commits all successful item patches only when every item succeeds.
+- Any item runtime failure fails the whole run.
+- `skip` and `collect` remain runtime-unsupported for concurrent mode.
+
+Plan:
+
+- See [`2026-05-22-concurrent-foreach-v1-sync-fail-only.md`](2026-05-22-concurrent-foreach-v1-sync-fail-only.md).
+
+## Slice 2: Barrier Commit Conflict Semantics
+
+Implement after Slice 1 if V1 keeps conflict behavior too conservative.
+
+Scope:
+
+- Detect sibling lineage writes to the same state path.
+- If exactly one lineage writes a destination path, default replace is allowed.
+- If multiple sibling lineages write the same destination path, a declared reducer is required.
+- Ancestor/descendant writes across sibling lineages are conflicts unless an explicit future merge strategy covers them.
+- Commit order is item index order, never completion order.
+
+Files likely touched:
+
+- `src/wf_core/runtime/foreach_state.py`
+- `src/wf_core/runtime/ops/state.py`
+- `src/wf_core/runtime/ops/foreach.py`
+- `tests/core/test_concurrent_foreach.py`
+
+Key tests:
+
+- `test_concurrent_foreach_rejects_sibling_writes_without_reducer`
+- `test_concurrent_foreach_applies_reducer_in_item_index_order`
+- `test_concurrent_foreach_rejects_ancestor_descendant_write_conflict`
+
+## Slice 3: Item Error Policies
+
+Implement after barrier success commits are correct.
+
+Scope:
+
+- `item_error.action="skip"` continues after item runtime failures.
+- `item_error.action="collect"` continues and writes structured errors to `collect_to`.
+- Both emit `completed_with_errors` if at least one item failed.
+- `collect` writes an empty list and emits `done` if all items succeed.
+- Failed item frames remain `FAILED`; parent foreach decides whether the failure is handled.
+
+Files likely touched:
+
+- `src/wf_core/runtime/foreach_state.py`
+- `src/wf_core/runtime/ops/foreach.py`
+- `tests/core/test_concurrent_foreach_errors.py`
+
+Key tests:
+
+- `test_concurrent_foreach_skip_emits_completed_with_errors`
+- `test_concurrent_foreach_collect_writes_ordered_error_records`
+- `test_concurrent_foreach_collect_writes_empty_list_on_clean_success`
+
+## Slice 4: Async Concurrent Foreach
+
+Implement only after sync semantics are stable.
+
+Scope:
+
+- Async runtime may have multiple async node handler calls in flight.
+- `concurrent.max_active` caps admitted/running item work.
+- Sync handlers are still called normally; no thread/process executor.
+- Trace remains append-only chronological execution history.
+- Barrier commit order remains item index order.
+
+Files likely touched:
+
+- `src/wf_core/runtime/engine.py`
+- `src/wf_core/runtime/step.py`
+- `src/wf_core/runtime/ops/nodes.py`
+- `src/wf_core/runtime/ops/foreach.py`
+- `tests/core/test_concurrent_foreach_async.py`
+
+Key tests:
+
+- `test_async_concurrent_foreach_respects_max_active`
+- `test_async_concurrent_foreach_commits_in_item_index_order`
+
+## Slice 5: Interrupt Quiescence
+
+Implement after async execution exists.
+
+Scope:
+
+- If any concurrent item interrupts, the whole run pauses.
+- No new item frames are admitted after the interrupt.
+- Already-started async node calls drain to pending results.
+- The caller gets control only at a quiescent point.
+- Pending results do not commit until resume/commit policy allows it.
+
+Files likely touched:
+
+- `src/wf_core/runtime/engine.py`
+- `src/wf_core/runtime/preparation.py`
+- `src/wf_core/runtime/ops/interrupts.py`
+- `src/wf_core/runtime/ops/foreach.py`
+- `tests/core/test_concurrent_foreach_interrupts.py`
+
+Key tests:
+
+- `test_concurrent_foreach_interrupt_returns_after_quiescence`
+- `test_resume_prioritizes_interrupted_item_frame_before_siblings`
+
+## Execution Order
+
+1. Sync concurrent foreach, fail-only.
+2. Barrier conflict semantics, if not fully covered by slice 1.
+3. `skip` / `collect` item error policies.
+4. Async concurrent foreach.
+5. Interrupt quiescence.
+
+## Self-Review
+
+- Spec coverage: the roadmap covers scheduler admission, barrier commits, reducer conflicts, handled item failures, async handler execution, and interrupt quiescence.
+- Placeholder scan: each slice has scope, likely files, and named tests; concrete code lives in the slice-specific plan.
+- Type consistency: the roadmap uses canonical `concurrent`, `ForeachConcurrentPolicy`, `ForeachBarrierState`, `PendingItemResult`, and `StatePatch`.

docs/superpowers/plans/2026-05-22-concurrent-foreach-roadmap.md

@@ -22,6 +22,10 @@
 - Phase 4 is not implemented: `foreach(mode="concurrent")` still validates as a
   model shape but runtime execution rejects it until concurrent scheduling,
   barrier commits, and item failure handling are implemented.
+- Phase 4 is expanded into a dedicated roadmap:
+  [`2026-05-22-concurrent-foreach-phase4-roadmap.md`](2026-05-22-concurrent-foreach-phase4-roadmap.md).
+  Start with
+  [`2026-05-22-concurrent-foreach-v1-sync-fail-only.md`](2026-05-22-concurrent-foreach-v1-sync-fail-only.md).
 
 ---
 

src/wf_core/runtime/foreach_state.py

@@ -6,6 +6,7 @@ from typing import Any, Literal
 from wf_core.errors import WorkflowExecutionError
 from wf_core.run_state import ExecutionFrame
 from wf_core.runtime.ops.state import StatePatch
+from wf_core.runtime.scheduler import ForeachIterationMetadata
 
 _BARRIER_METADATA_KEY = "foreach_barriers"
 
@@ -120,6 +121,7 @@ class ForeachBarrierState:
     """Resumable state owned by one foreach parent frame."""
 
     next_index: int = 0
+    mode: Literal["serial", "concurrent"] = "serial"
     active_frame_ids: tuple[str, ...] = ()
     outstanding_frame_ids: tuple[str, ...] = ()
     pending_results: dict[int, PendingItemResult] = field(default_factory=dict)
@@ -156,11 +158,14 @@ class ForeachBarrierState:
         if not isinstance(raw, dict):
             raise WorkflowExecutionError("malformed foreach barrier state")
         next_index = raw.get("next_index")
+        mode = raw.get("mode", "serial")
         active_frame_ids = _string_tuple(raw.get("active_frame_ids", ()))
         outstanding_frame_ids = _string_tuple(raw.get("outstanding_frame_ids", ()))
         pending_results = raw.get("pending_results", {})
         if not isinstance(next_index, int):
             raise WorkflowExecutionError("malformed foreach barrier next_index")
+        if mode not in {"serial", "concurrent"}:
+            raise WorkflowExecutionError("malformed foreach barrier mode")
         if not isinstance(pending_results, dict):
             raise WorkflowExecutionError("malformed foreach barrier pending results")
         parsed_results: dict[int, PendingItemResult] = {}
@@ -174,6 +179,7 @@ class ForeachBarrierState:
             parsed_results[index] = PendingItemResult.from_metadata(raw_result)
         return cls(
             next_index=next_index,
+            mode=mode,
             active_frame_ids=active_frame_ids,
             outstanding_frame_ids=outstanding_frame_ids,
             pending_results=parsed_results,
@@ -196,6 +202,7 @@ class ForeachBarrierState:
     def to_metadata(self) -> dict[str, Any]:
         return {
             "next_index": self.next_index,
+            "mode": self.mode,
             "active_frame_ids": list(self.active_frame_ids),
             "outstanding_frame_ids": list(self.outstanding_frame_ids),
             "pending_results": {
@@ -204,6 +211,45 @@ class ForeachBarrierState:
             },
         }
 
+    def start_child(self, frame_id: str) -> None:
+        """Record one admitted child frame as active and outstanding."""
+        if frame_id in self.active_frame_ids or frame_id in self.outstanding_frame_ids:
+            raise WorkflowExecutionError(
+                f"foreach child frame {frame_id!r} already active"
+            )
+        self.active_frame_ids = (*self.active_frame_ids, frame_id)
+        self.outstanding_frame_ids = (*self.outstanding_frame_ids, frame_id)
+
+    def finish_child(self, frame_id: str) -> None:
+        """Record one child frame as no longer active or outstanding."""
+        self.active_frame_ids = tuple(
+            item for item in self.active_frame_ids if item != frame_id
+        )
+        self.outstanding_frame_ids = tuple(
+            item for item in self.outstanding_frame_ids if item != frame_id
+        )
+
+    def add_success_patch(
+        self, *, index: int, frame_id: str, patch: StatePatch
+    ) -> None:
+        """Buffer one successful item patch by item index."""
+        self.pending_results[index] = PendingItemResult(
+            index=index,
+            frame_id=frame_id,
+            status="succeeded",
+            patch=patch,
+        )
+
+
+def item_frame_owner(frame: ExecutionFrame) -> tuple[str, str, int] | None:
+    """Return parent frame id, foreach node id, and item index for item frames."""
+    if frame.kind != "foreach_iteration" or frame.parent_frame_id is None:
+        return None
+    metadata = ForeachIterationMetadata.from_frame(frame)
+    if metadata is None:
+        return None
+    return frame.parent_frame_id, metadata.foreach_node_id, metadata.loop_index
+
 
 def _string_tuple(raw: object) -> tuple[str, ...]:
     if isinstance(raw, tuple) and all(isinstance(item, str) for item in raw):

src/wf_core/runtime/ops/flow.py

@@ -15,7 +15,7 @@ from wf_core.runtime.ops.schemas import validate_payload_against_schema
 from wf_core.runtime.ops.state import project_output
 from wf_core.runtime.scheduler import (
     mark_frame_pending,
-    wake_parent_if_children_complete,
+    wake_parent_for_child_progress,
 )
 from wf_core.tokens import END
 
@@ -81,7 +81,7 @@ def advance_frame(
     if next_node_id == END:
         frame.status = FrameStatus.COMPLETED
         frame.finished_at_node_id = END
-        wake_parent_if_children_complete(run, frame.id)
+        wake_parent_for_child_progress(run, frame.id)
     else:
         frame.finished_at_node_id = None
         mark_frame_pending(run, frame.id)

src/wf_core/runtime/ops/foreach.py

@@ -1,19 +1,24 @@
 from __future__ import annotations
 
+from collections.abc import Mapping
+
 from wf_core.conditions import safe_resolve_path
 from wf_core.errors import WorkflowExecutionError
-from wf_core.models.steps import ForeachNode
+from wf_core.models.steps import ForeachNode, NodeUse
 from wf_core.models.workflow import Workflow
 from wf_core.run_state import ExecutionFrame, FrameStatus, RunState, StepExecutionResult
 from wf_core.runtime.foreach_state import ForeachBarrierState
 from wf_core.runtime.ops.flow import advance_frame, append_step_result_trace
 from wf_core.runtime.ops.frames import frame_context_values
 from wf_core.runtime.ops.index import WorkflowIndex
+from wf_core.runtime.ops.merges import ReducerDefinition
+from wf_core.runtime.ops.state import build_barrier_patch, commit_state_patch
 from wf_core.runtime.scheduler import (
     ForeachIterationMetadata,
     add_frame,
     block_frame_on_children,
 )
+from wf_core.tokens import END
 
 
 def step_foreach(
@@ -21,25 +26,32 @@ def step_foreach(
     run: RunState,
     step: ForeachNode,
     index: WorkflowIndex,
+    *,
+    reducers: Mapping[str, ReducerDefinition] | None = None,
+) -> RunState:
+    if step.mode == "serial":
+        return _step_foreach_serial(workflow, run, step, index)
+    return _step_foreach_concurrent(
+        workflow,
+        run,
+        step,
+        index,
+        reducers=reducers,
+    )
+
+
+def _step_foreach_serial(
+    workflow: Workflow,
+    run: RunState,
+    step: ForeachNode,
+    index: WorkflowIndex,
 ) -> RunState:
     if step.mode != "serial":
-        raise WorkflowExecutionError(
-            "concurrent foreach execution is not implemented yet"
-        )
+        raise WorkflowExecutionError("serial foreach helper received non-serial mode")
 
     frame = run.current_frame()
     barrier = ForeachBarrierState.from_frame(frame, step.id) or ForeachBarrierState()
-
-    iterable = safe_resolve_path(
-        str(step.over),
-        state=run.state,
-        workflow_input=run.workflow_input,
-        context=frame_context_values(frame),
-    )
-    if not isinstance(iterable, list):
-        raise WorkflowExecutionError(
-            f"foreach source {str(step.over)!r} must resolve to a list"
-        )
+    iterable = _resolve_foreach_iterable(run, frame, step)
 
     loop_index = barrier.next_index
     if loop_index >= len(iterable):
@@ -62,17 +74,10 @@ def step_foreach(
         return run
 
     loop_start = index.next_node_id(frame.node_id, "loop")
-
     item = iterable[loop_index]
     barrier.next_index = loop_index + 1
     barrier.save_to_frame(frame, step.id)
     child_id = f"{frame.id}:{step.id}:{loop_index}"
-    child_metadata = ForeachIterationMetadata(
-        foreach_node_id=step.id,
-        loop_index=loop_index,
-        loop_item=item,
-        loop_alias=step.as_,
-    )
     add_frame(
         run,
         ExecutionFrame(
@@ -81,7 +86,12 @@ def step_foreach(
             node_id=loop_start,
             status=FrameStatus.PENDING,
             parent_frame_id=frame.id,
-            metadata=child_metadata.to_metadata(),
+            metadata=ForeachIterationMetadata(
+                foreach_node_id=step.id,
+                loop_index=loop_index,
+                loop_item=item,
+                loop_alias=step.as_,
+            ).to_metadata(),
         ),
         ready=True,
     )
@@ -101,3 +111,210 @@ def step_foreach(
     )
     run.sync_from_current_frame()
     return run
+
+
+def _step_foreach_concurrent(
+    workflow: Workflow,
+    run: RunState,
+    step: ForeachNode,
+    index: WorkflowIndex,
+    *,
+    reducers: Mapping[str, ReducerDefinition] | None = None,
+) -> RunState:
+    if step.item_error.action != "fail":
+        raise WorkflowExecutionError(
+            "concurrent foreach v1 only supports item_error.action='fail'"
+        )
+    if step.concurrent is None:
+        raise WorkflowExecutionError("concurrent foreach requires concurrent policy")
+    _validate_single_node_loop_body(index, step)
+
+    frame = run.current_frame()
+    barrier = ForeachBarrierState.from_frame(frame, step.id)
+    if barrier is None:
+        barrier = ForeachBarrierState(mode="concurrent")
+    elif barrier.mode != "concurrent":
+        raise WorkflowExecutionError("malformed concurrent foreach barrier mode")
+
+    _finish_completed_children(run, barrier)
+    iterable = _resolve_foreach_iterable(run, frame, step)
+    _admit_concurrent_children(
+        run=run,
+        frame=frame,
+        step=step,
+        index=index,
+        barrier=barrier,
+        iterable=iterable,
+    )
+
+    if barrier.next_index >= len(iterable) and not barrier.outstanding_frame_ids:
+        return _finish_concurrent_foreach(
+            workflow=workflow,
+            run=run,
+            frame=frame,
+            step=step,
+            index=index,
+            barrier=barrier,
+            reducers=reducers,
+        )
+
+    barrier.save_to_frame(frame, step.id)
+    block_frame_on_children(run, frame.id, barrier.outstanding_frame_ids)
+    run.sync_from_current_frame()
+    return run
+
+
+def _resolve_foreach_iterable(
+    run: RunState,
+    frame: ExecutionFrame,
+    step: ForeachNode,
+) -> list[object]:
+    iterable = safe_resolve_path(
+        str(step.over),
+        state=run.state,
+        workflow_input=run.workflow_input,
+        context=frame_context_values(frame),
+    )
+    if not isinstance(iterable, list):
+        raise WorkflowExecutionError(
+            f"foreach source {str(step.over)!r} must resolve to a list"
+        )
+    return iterable
+
+
+def _validate_single_node_loop_body(index: WorkflowIndex, step: ForeachNode) -> None:
+    """Reject multi-step concurrent item bodies until item overlays are real.
+
+    The current slice has a no-op item-state overlay seam. Without a real overlay,
+    multi-node item bodies would read stale parent state after earlier item-local
+    writes, so V1 only allows loop -> one node -> END.
+    """
+    loop_start = index.next_node_id(step.id, "loop")
+    loop_step = index.nodes_by_id.get(loop_start)
+    if not isinstance(loop_step, NodeUse):
+        raise WorkflowExecutionError(
+            "concurrent foreach v1 only supports loop bodies with one node"
+        )
+    node_def = index.node_defs[loop_step.node]
+    for outcome in node_def.outcomes:
+        if index.next_node_id(loop_step.id, outcome) != END:
+            raise WorkflowExecutionError(
+                "concurrent foreach v1 only supports loop bodies with one node"
+            )
+
+
+def _finish_completed_children(run: RunState, barrier: ForeachBarrierState) -> None:
+    for child_id in tuple(barrier.outstanding_frame_ids):
+        child = run.frames[child_id]
+        if child.status == FrameStatus.COMPLETED:
+            barrier.finish_child(child_id)
+        elif child.status == FrameStatus.FAILED:
+            message = child.metadata.get("error", "unknown item failure")
+            raise WorkflowExecutionError(
+                f"concurrent foreach item frame {child_id!r} failed: {message}"
+            )
+
+
+def _admit_concurrent_children(
+    *,
+    run: RunState,
+    frame: ExecutionFrame,
+    step: ForeachNode,
+    index: WorkflowIndex,
+    barrier: ForeachBarrierState,
+    iterable: list[object],
+) -> int:
+    if step.concurrent is None:
+        raise WorkflowExecutionError("concurrent foreach requires concurrent policy")
+
+    admitted = 0
+    loop_start = index.next_node_id(frame.node_id, "loop")
+    while (
+        barrier.next_index < len(iterable)
+        and len(barrier.active_frame_ids) < step.concurrent.max_active
+        and len(barrier.outstanding_frame_ids) < step.concurrent.max_outstanding
+    ):
+        loop_index = barrier.next_index
+        item = iterable[loop_index]
+        child_id = f"{frame.id}:{step.id}:{loop_index}"
+        active_count = len(barrier.active_frame_ids)
+        barrier.next_index = loop_index + 1
+        barrier.start_child(child_id)
+        add_frame(
+            run,
+            ExecutionFrame(
+                id=child_id,
+                kind="foreach_iteration",
+                node_id=loop_start,
+                status=FrameStatus.PENDING,
+                parent_frame_id=frame.id,
+                metadata=ForeachIterationMetadata(
+                    foreach_node_id=step.id,
+                    loop_index=loop_index,
+                    loop_item=item,
+                    loop_alias=step.as_,
+                ).to_metadata(),
+            ),
+            ready=True,
+        )
+        append_step_result_trace(
+            run,
+            frame_id=frame.id,
+            node_id=frame.node_id,
+            step_type=step.type,
+            next_node_id=loop_start,
+            result=StepExecutionResult(
+                outcome="loop",
+                resolved_input={
+                    "item": item,
+                    "index": loop_index,
+                    "active_count": active_count,
+                },
+                output={},
+                state_changes={},
+            ),
+        )
+        admitted += 1
+    return admitted
+
+
+def _finish_concurrent_foreach(
+    *,
+    workflow: Workflow,
+    run: RunState,
+    frame: ExecutionFrame,
+    step: ForeachNode,
+    index: WorkflowIndex,
+    barrier: ForeachBarrierState,
+    reducers: Mapping[str, ReducerDefinition] | None = None,
+) -> RunState:
+    next_node_id = index.next_node_id(frame.node_id, "done")
+    combined = build_barrier_patch(
+        workflow,
+        [
+            barrier.pending_results[item_index].patch
+            for item_index in sorted(barrier.pending_results)
+        ],
+        run.state,
+        reducers=reducers,
+    )
+    state_changes = commit_state_patch(run.state, combined)
+    append_step_result_trace(
+        run,
+        frame_id=frame.id,
+        node_id=frame.node_id,
+        step_type=step.type,
+        next_node_id=next_node_id,
+        result=StepExecutionResult(
+            outcome="done",
+            resolved_input={
+                "count": barrier.next_index,
+                "index": barrier.next_index,
+                "committed_items": len(barrier.pending_results),
+            },
+            output={},
+            state_changes=state_changes,
+        ),
+    )
+    advance_frame(run, frame, outcome="done", next_node_id=next_node_id)
+    return run

src/wf_core/runtime/ops/nodes.py

@@ -11,10 +11,12 @@ from wf_core.models.schemas import NodeDef
 from wf_core.models.steps import InputPathBinding, InputValueBinding, NodeUse
 from wf_core.models.workflow import Workflow
 from wf_core.run_state import RunState, RuntimeContext, StepExecutionResult
+from wf_core.runtime.foreach_state import ForeachBarrierState, item_frame_owner
 from wf_core.runtime.ops.frames import frame_context_values
 from wf_core.runtime.ops.merges import ReducerDefinition
+from wf_core.runtime.ops.overlays import state_view_for_frame
 from wf_core.runtime.ops.schemas import validate_payload_against_schema
-from wf_core.runtime.ops.state import apply_output_bindings
+from wf_core.runtime.ops.state import build_output_patch, commit_state_patch
 
 NodeHandler = Callable[[dict[str, Any], RuntimeContext], NodeResult | dict[str, Any]]
 AsyncNodeHandler = Callable[
@@ -32,6 +34,7 @@ def _resolve_node_execution(
 ) -> tuple[dict[str, Any], RuntimeContext]:
     frame = run.current_frame()
     context_values = frame_context_values(frame)
+    state_view = state_view_for_frame(run, frame)
     resolved_input: dict[str, Any] = {}
     for binding in node.input:
         if isinstance(binding, InputValueBinding):
@@ -39,7 +42,7 @@ def _resolve_node_execution(
         elif isinstance(binding, InputPathBinding):
             value = safe_resolve_path(
                 str(binding.path),
-                state=run.state,
+                state=state_view,
                 workflow_input=run.workflow_input,
                 context=context_values,
             )
@@ -85,13 +88,30 @@ def _finalize_node_execution(
     validate_payload_against_schema(
         node_def.output_schema, result.output, f"node output for {node.id}"
     )
-    state_changes = apply_output_bindings(
+    patch = build_output_patch(
         workflow,
         node.output,
         result.output,
         run.state,
         reducers=reducers,
     )
+    owner = item_frame_owner(run.current_frame())
+    if owner is None:
+        state_changes = commit_state_patch(run.state, patch)
+    else:
+        parent_frame_id, foreach_node_id, item_index = owner
+        parent_frame = run.frames[parent_frame_id]
+        barrier = ForeachBarrierState.from_frame(parent_frame, foreach_node_id)
+        if barrier is not None and barrier.mode == "concurrent":
+            barrier.add_success_patch(
+                index=item_index,
+                frame_id=run.current_frame().id,
+                patch=patch,
+            )
+            barrier.save_to_frame(parent_frame, foreach_node_id)
+            state_changes = {}
+        else:
+            state_changes = commit_state_patch(run.state, patch)
     return StepExecutionResult(
         outcome=result.outcome,
         resolved_input=resolved_input,

src/wf_core/runtime/ops/overlays.py

@@ -0,0 +1,16 @@
+from __future__ import annotations
+
+from typing import Any
+
+from wf_core.run_state import ExecutionFrame, RunState
+
+
+def state_view_for_frame(run: RunState, frame: ExecutionFrame) -> dict[str, Any]:
+    """Return the state view visible to one execution frame.
+
+    This is intentionally a no-op seam for concurrent foreach V1. The first
+    sync-concurrent slice only supports single-node item bodies, so item frames
+    do not need to read their own prior buffered writes yet. The overlay slice
+    should replace this with parent-state plus item-local staged writes.
+    """
+    return run.state

src/wf_core/runtime/ops/state.py

@@ -145,6 +145,46 @@ def commit_state_patch(state: dict[str, Any], patch: StatePatch) -> dict[str, An
     return dict(patch.changes)
 
 
+def build_barrier_patch(
+    workflow: Workflow,
+    item_patches: Sequence[StatePatch],
+    state: dict[str, Any],
+    *,
+    reducers: Mapping[str, ReducerDefinition] | None = None,
+) -> StatePatch:
+    """Build one committed barrier patch by replaying item writes in order.
+
+    Item patches are built against the parent-visible state. Their prepared
+    writes cannot be blindly merged because reducers must see the value produced
+    by earlier item patches. The barrier therefore replays trace-facing incoming
+    changes against a single staged state in deterministic item order.
+    """
+    state_fields = workflow.state_schema.field_index()
+    staged_state = deepcopy(state)
+    prepared_patch: dict[StatePath, tuple[list[str], Any]] = {}
+    committed_changes: dict[str, Any] = {}
+    for item_patch in item_patches:
+        for destination, incoming_value in item_patch.changes.items():
+            destination_path = StatePath.parse(destination)
+            key_path, merged_value = prepare_state_value(
+                workflow,
+                staged_state,
+                destination_path,
+                incoming_value,
+                reducers=reducers,
+                state_fields=state_fields,
+            )
+            safe_set_nested_value(staged_state, key_path, merged_value)
+            prepared_patch[destination_path] = (key_path, merged_value)
+            committed_changes[destination] = merged_value
+    validate_staged_state_patch(staged_state, prepared_patch, state_fields)
+    return StatePatch(
+        changes=committed_changes,
+        _prepared_writes=prepared_patch,
+        _staged_state=staged_state,
+    )
+
+
 def apply_mapped_state(
     workflow: Workflow,
     source_data: dict[str, Any],

src/wf_core/runtime/scheduler.py

@@ -177,6 +177,21 @@ def wake_parent_if_children_complete(run: RunState, child_frame_id: str) -> None
         wake_frame(run, parent_id)
 
 
+def wake_parent_for_child_progress(run: RunState, child_frame_id: str) -> None:
+    """Wake a blocked parent after one child finishes so it can refill slots."""
+    child = _frame(run, child_frame_id)
+    parent_id = child.parent_frame_id
+    if parent_id is None:
+        return
+    parent = _frame(run, parent_id)
+    if parent.status != FrameStatus.BLOCKED:
+        return
+    block = BlockedOnChildren.from_frame(parent)
+    if block is None or child_frame_id not in block.child_frame_ids:
+        return
+    wake_frame(run, parent_id)
+
+
 def resolve_no_ready_frames(run: RunState) -> RunStatus:
     """Classify an empty ready queue into terminal, paused, or deadlocked state."""
     if run.status == RunStatus.INTERRUPTED:

src/wf_core/runtime/step.py

@@ -99,7 +99,7 @@ def step_workflow(
     elif isinstance(step, InterruptNode):
         return handle_interrupt_step(run, step)
     elif isinstance(step, ForeachNode):
-        return step_foreach(workflow, run, step, index)
+        return step_foreach(workflow, run, step, index, reducers=reducers)
     else:
         raise WorkflowExecutionError(
             f"unsupported step type {getattr(step, 'type', type(step).__name__)!r}"
@@ -152,7 +152,7 @@ async def step_workflow_async(
     elif isinstance(step, InterruptNode):
         return handle_interrupt_step(run, step)
     elif isinstance(step, ForeachNode):
-        return step_foreach(workflow, run, step, index)
+        return step_foreach(workflow, run, step, index, reducers=reducers)
     else:
         raise WorkflowExecutionError(
             f"unsupported step type {getattr(step, 'type', type(step).__name__)!r}"

tests/core/test_concurrent_foreach.py

@@ -0,0 +1,279 @@
+from __future__ import annotations
+
+from typing import Any
+
+import pytest
+
+from wf_core import (
+    END,
+    Edge,
+    ForeachNode,
+    NodeDef,
+    NodeUse,
+    ReducerRef,
+    SchemaRef,
+    StateField,
+    StateSchema,
+    Workflow,
+    WorkflowExecutionError,
+    execute_workflow,
+)
+
+
+def test_sync_concurrent_foreach_interleaves_items_and_commits_at_barrier() -> None:
+    workflow = _workflow(
+        state_schema=StateSchema.from_field_map(
+            {
+                "items": StateField(type="array"),
+                "seen": StateField(
+                    type="array",
+                    reducer=ReducerRef(name="wf.std.append"),
+                ),
+            }
+        ),
+        foreach=ForeachNode.model_validate(
+            {
+                "id": "each",
+                "type": "foreach",
+                "over": "state.items",
+                "as": "item",
+                "mode": "concurrent",
+                "concurrent": {"max_active": 2, "max_outstanding": 2},
+            }
+        ),
+    )
+
+    run = execute_workflow(
+        workflow,
+        {"items": ["a", "b", "c"]},
+        {"record": lambda payload, _ctx: {"outcome": "ok", "output": payload}},
+    )
+
+    assert run.output["seen"] == ["a", "b", "c"]
+    assert run.state["seen"] == ["a", "b", "c"]
+    foreach_entries = [entry for entry in run.trace if entry.step_type == "foreach"]
+    assert foreach_entries[-1].outcome == "done"
+    assert foreach_entries[-1].state_changes["state.seen"] == ["a", "b", "c"]
+
+
+def test_sync_concurrent_foreach_respects_max_active_by_refill_trace() -> None:
+    workflow = _workflow(
+        state_schema=StateSchema.from_field_map(
+            {
+                "items": StateField(type="array"),
+                "seen": StateField(
+                    type="array",
+                    reducer=ReducerRef(name="wf.std.append"),
+                ),
+            }
+        ),
+        foreach=ForeachNode.model_validate(
+            {
+                "id": "each",
+                "type": "foreach",
+                "over": "state.items",
+                "as": "item",
+                "mode": "concurrent",
+                "concurrent": {"max_active": 2, "max_outstanding": 2},
+            }
+        ),
+    )
+
+    run = execute_workflow(
+        workflow,
+        {"items": ["a", "b", "c", "d"]},
+        {"record": lambda payload, _ctx: {"outcome": "ok", "output": payload}},
+    )
+
+    loop_entries = [
+        entry
+        for entry in run.trace
+        if entry.step_type == "foreach" and entry.outcome == "loop"
+    ]
+    assert loop_entries[0].resolved_input["active_count"] == 0
+    assert loop_entries[1].resolved_input["active_count"] == 1
+    assert all(entry.resolved_input["active_count"] < 2 for entry in loop_entries)
+
+
+def test_sync_concurrent_foreach_rejects_non_fail_item_policy_for_now() -> None:
+    workflow = _workflow(
+        state_schema=StateSchema.from_field_map(
+            {
+                "items": StateField(type="array"),
+                "seen": StateField(
+                    type="array",
+                    reducer=ReducerRef(name="wf.std.append"),
+                ),
+                "errors": StateField(type="array"),
+            }
+        ),
+        foreach=ForeachNode.model_validate(
+            {
+                "id": "each",
+                "type": "foreach",
+                "over": "state.items",
+                "as": "item",
+                "mode": "concurrent",
+                "concurrent": {"max_active": 2, "max_outstanding": 2},
+                "item_error": {"action": "collect", "collect_to": "state.errors"},
+            }
+        ),
+        include_completed_with_errors=True,
+    )
+
+    with pytest.raises(
+        WorkflowExecutionError,
+        match="only supports item_error.action='fail'",
+    ):
+        execute_workflow(
+            workflow,
+            {"items": ["a"]},
+            {"record": lambda payload, _ctx: {"outcome": "ok", "output": payload}},
+        )
+
+
+def test_sync_concurrent_foreach_fails_run_on_item_runtime_error() -> None:
+    workflow = _workflow(
+        state_schema=StateSchema.from_field_map(
+            {
+                "items": StateField(type="array"),
+                "seen": StateField(
+                    type="array",
+                    reducer=ReducerRef(name="wf.std.append"),
+                ),
+            }
+        ),
+        foreach=ForeachNode.model_validate(
+            {
+                "id": "each",
+                "type": "foreach",
+                "over": "state.items",
+                "as": "item",
+                "mode": "concurrent",
+                "concurrent": {"max_active": 2, "max_outstanding": 2},
+            }
+        ),
+    )
+
+    def fail_on_b(payload: dict[str, Any], _ctx: object) -> dict[str, Any]:
+        if payload["value"] == "b":
+            raise ValueError("bad item")
+        return {"outcome": "ok", "output": payload}
+
+    with pytest.raises(ValueError, match="bad item"):
+        execute_workflow(workflow, {"items": ["a", "b", "c"]}, {"record": fail_on_b})
+
+
+def test_sync_concurrent_foreach_rejects_multi_step_item_body_for_now() -> None:
+    workflow = _workflow(
+        state_schema=StateSchema.from_field_map(
+            {
+                "items": StateField(type="array"),
+                "seen": StateField(
+                    type="array",
+                    reducer=ReducerRef(name="wf.std.append"),
+                ),
+            }
+        ),
+        foreach=ForeachNode.model_validate(
+            {
+                "id": "each",
+                "type": "foreach",
+                "over": "state.items",
+                "as": "item",
+                "mode": "concurrent",
+                "concurrent": {"max_active": 2, "max_outstanding": 2},
+            }
+        ),
+    )
+    workflow.nodes.append(
+        NodeUse.model_validate(
+            {
+                "id": "after_record",
+                "type": "node",
+                "node": "record",
+                "input": [{"target": "value", "path": "state.seen"}],
+                "output": [{"source": "seen", "target": "state.seen"}],
+            }
+        )
+    )
+    workflow.edges = [
+        Edge.model_validate({"from": "each", "outcome": "loop", "to": "record"}),
+        Edge.model_validate({"from": "record", "outcome": "ok", "to": "after_record"}),
+        Edge.model_validate({"from": "after_record", "outcome": "ok", "to": END}),
+        Edge.model_validate({"from": "each", "outcome": "done", "to": END}),
+    ]
+
+    with pytest.raises(
+        WorkflowExecutionError,
+        match="only supports loop bodies with one node",
+    ):
+        execute_workflow(
+            workflow,
+            {"items": ["a"]},
+            {"record": lambda payload, _ctx: {"outcome": "ok", "output": payload}},
+        )
+
+
+def _workflow(
+    *,
+    state_schema: StateSchema,
+    foreach: ForeachNode,
+    include_completed_with_errors: bool = False,
+) -> Workflow:
+    edges = [
+        Edge.model_validate({"from": "each", "outcome": "loop", "to": "record"}),
+        Edge.model_validate({"from": "record", "outcome": "ok", "to": END}),
+        Edge.model_validate({"from": "each", "outcome": "done", "to": END}),
+    ]
+    if include_completed_with_errors:
+        edges.append(
+            Edge.model_validate(
+                {"from": "each", "outcome": "completed_with_errors", "to": END}
+            )
+        )
+    return Workflow(
+        name="concurrent_foreach_v1",
+        input_schema=SchemaRef(
+            type="object",
+            properties={"items": {"type": "array"}},
+        ),
+        state_schema=state_schema,
+        output_schema=SchemaRef(
+            type="object",
+            properties={"seen": {"type": "array"}},
+        ),
+        node_defs=[
+            NodeDef(
+                name="record",
+                input_schema=SchemaRef(
+                    type="object",
+                    properties={"value": {}, "seen": {}},
+                    required=["value", "seen"],
+                ),
+                output_schema=SchemaRef(
+                    type="object",
+                    properties={"value": {}, "seen": {}},
+                    required=["seen"],
+                ),
+                outcomes=["ok"],
+            )
+        ],
+        start="each",
+        nodes=[
+            foreach,
+            NodeUse.model_validate(
+                {
+                    "id": "record",
+                    "type": "node",
+                    "node": "record",
+                    "input": [
+                        {"target": "value", "path": "context.item"},
+                        {"target": "seen", "path": "context.item"},
+                    ],
+                    "output": [{"source": "seen", "target": "state.seen"}],
+                }
+            ),
+        ],
+        edges=edges,
+    )

tests/core/test_foreach_barrier_state.py

@@ -66,6 +66,25 @@ def test_foreach_barrier_state_rejects_malformed_metadata() -> None:
         ForeachBarrierState.from_frame(frame, "each")
 
 
+def test_foreach_barrier_tracks_active_and_outstanding_children() -> None:
+    barrier = ForeachBarrierState()
+
+    barrier.start_child("child-0")
+    barrier.start_child("child-1")
+    barrier.finish_child("child-0")
+
+    assert barrier.active_frame_ids == ("child-1",)
+    assert barrier.outstanding_frame_ids == ("child-1",)
+
+
+def test_foreach_barrier_rejects_duplicate_child_start() -> None:
+    barrier = ForeachBarrierState()
+    barrier.start_child("child-0")
+
+    with pytest.raises(WorkflowExecutionError, match="already active"):
+        barrier.start_child("child-0")
+
+
 def test_item_error_record_rejects_negative_index() -> None:
     with pytest.raises(WorkflowExecutionError, match="index"):
         ItemErrorRecord.from_metadata(

tests/core/test_scheduler.py

@@ -15,6 +15,7 @@ from wf_core.runtime.scheduler import (
     select_next_frame,
     resolve_no_ready_frames,
     wake_frame,
+    wake_parent_for_child_progress,
     wake_parent_if_children_complete,
 )
 
@@ -160,6 +161,27 @@ def test_child_completion_wakes_blocked_parent() -> None:
     assert run.ready_frame_ids == ["parent"]
 
 
+def test_wake_parent_when_child_finishes_for_refill() -> None:
+    run = _run()
+    add_frame(run, ExecutionFrame(id="parent", kind="root", node_id="foreach"))
+    add_frame(
+        run,
+        ExecutionFrame(
+            id="child",
+            kind="foreach_iteration",
+            node_id="__end__",
+            parent_frame_id="parent",
+        ),
+    )
+    block_frame_on_children(run, "parent", ("child", "other"))
+    run.frames["child"].status = FrameStatus.COMPLETED
+
+    wake_parent_for_child_progress(run, "child")
+
+    assert run.frames["parent"].status == FrameStatus.PENDING
+    assert run.ready_frame_ids == ["parent"]
+
+
 def test_resume_wakes_interrupted_frame_at_front() -> None:
     run = _run()
     add_frame(run, ExecutionFrame(id="waiting", kind="root", node_id="ask"))