paths throughout to serialize as object.

all for clarity! ahh
2026-05-21 02:02:04 +07:00
parent 2fcff9936f
commit e568db3169
19 changed files with 1003 additions and 232 deletions
@@ -0,0 +1,106 @@
+# Structural Refs
+
+Qualified names are display strings. They are not authoritative identifiers.
+
+The platform still accepts old dotted strings at MCP/API boundaries when it can
+parse them, but saved workflow artifacts and deployments should store structural
+refs from now on. A dotted string may be shown to humans and LLM clients, but
+runtime code should not infer source boundaries from it.
+
+## Why
+
+These strings look similar but mean different things:
+
+```text
+context7.default.query-docs
+workflow.echo_wrapper.v1
+demo.foo.bar
+```
+
+`context7.default.query-docs` usually means a concrete external source and a
+capability key. `workflow.echo_wrapper.v1` means a saved workflow artifact and
+artifact version. `demo.foo.bar` is ambiguous: it could be source `demo` with
+capability key `foo.bar`, or source `demo.foo` with capability key `bar`.
+
+No first-dot, last-dot, or regex parser can recover a boundary that was not
+stored.
+
+## Capability Refs
+
+Use a source plus a local capability key:
+
+```json
+{
+  "source": "demo",
+  "capability_key": "foo.bar"
+}
+```
+
+The `capability_key` is local to the known source. It may contain dots. Those
+dots do not carry source meaning.
+
+Old input like `"demo.foo.bar"` may still parse at compatibility boundaries, but
+that parse is best-effort and should not be used for new saves.
+
+## Workflow Artifact Refs
+
+Saved workflow and wrapper capabilities are a separate domain:
+
+```json
+{
+  "artifact_id": "echo_wrapper",
+  "version": 1
+}
+```
+
+The display string `workflow.echo_wrapper.v1` remains useful for lists,
+inspection, and old callers, but it is not the canonical saved shape.
+
+## Deployment Bindings
+
+Deployment bindings map an artifact-local logical source to a concrete source:
+
+```json
+{
+  "logical_source": "demo",
+  "concrete_source": "demo.personal"
+}
+```
+
+Neither field is a capability name. Runtime code uses this source mapping before
+looking up the capability key.
+
+## Graph Paths
+
+Capability refs and graph paths are different domains.
+
+Path strings such as `input.text`, `state.person.name`, and `output.echoed`
+describe graph data movement. Do not reuse capability-ref parsing rules for
+graph paths.
+
+New canonical graph path JSON uses root/parts objects:
+
+```json
+{
+  "root": "input",
+  "parts": ["message"]
+}
+```
+
+```json
+{
+  "root": "state",
+  "parts": ["person.name", "three and four"]
+}
+```
+
+```json
+{
+  "root": "local",
+  "parts": []
+}
+```
+
+Old strings are accepted at parse boundaries for compatibility. Structural
+`parts` are literal field names, so a part may contain dots or spaces without
+being split again.
@@ -0,0 +1,220 @@
+# Structural Graph Paths Implementation Plan
+
+> **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:** Save canonical graph paths as structural objects while keeping old dotted strings as parse-only compatibility input.
+
+**Architecture:** The core already has first-class path types: `GraphSourcePath`, `StatePath`, and `LocalPath`. Update those Pydantic hooks to accept structural dict input and serialize structurally in JSON mode. Keep `str(path)` for display and legacy fields. Do not redesign `NodeUse.input` / `output`; those structs already replaced deprecated `in_map` / `out_map`.
+
+**Tech Stack:** Python 3.14, Pydantic core schema hooks, `wf_core.paths`, `wf_core.models.steps`, pytest.
+
+---
+
+## Current State
+
+Canonical node bindings already exist:
+
+```json
+{
+  "input": [
+    {"path": "input.message", "target": "message"}
+  ],
+  "output": [
+    {"source": "echoed", "target": "state.echoed"}
+  ]
+}
+```
+
+Internally these parse to:
+
+- `GraphSourcePath`
+- `LocalPath`
+- `StatePath`
+
+The remaining problem is serialization. These path objects currently dump as strings, so saved JSON still relies on dot-separated path grammar.
+
+## Canonical Shape
+
+Graph source paths:
+
+```json
+{"root": "state", "parts": ["person", "name"]}
+```
+
+State write paths:
+
+```json
+{"root": "state", "parts": ["person", "name"]}
+```
+
+Local node paths:
+
+```json
+{"root": "local", "parts": ["payload", "text"]}
+```
+
+Local root remains explicit:
+
+```json
+{"root": "local", "parts": []}
+```
+
+Old strings such as `"state.person.name"` and `"."` remain accepted input.
+
+---
+
+## Task 1: Add Structural Serialization for Path Types
+
+**Files:**
+- Modify: `src/wf_core/paths.py`
+- Test: `tests/core/test_path_values.py`
+
+- [ ] **Step 1: Update tests first**
+
+Change `test_pydantic_accepts_path_strings_and_serializes_strings` into structural JSON expectations:
+
+```python
+dumped = payload.model_dump(mode="json")
+assert dumped["source"] == {"root": "input", "parts": ["user"]}
+assert dumped["target"] == {"root": "state", "parts": ["person"]}
+assert dumped["local"] == {"root": "local", "parts": ["user"]}
+```
+
+Keep `model_dump()` expectations if useful for Python-mode compatibility only if the implementation intentionally keeps Python mode as strings. Otherwise assert structural dumps in both modes.
+
+- [ ] **Step 2: Add structural input tests**
+
+Add a test:
+
+```python
+payload = Payload.model_validate({
+    "source": {"root": "input", "parts": ["user.name"]},
+    "target": {"root": "state", "parts": ["person.name"]},
+    "local": {"root": "local", "parts": ["payload.text"]},
+})
+
+assert payload.source == GraphSourcePath.input("user.name")
+assert payload.target == StatePath.of("person.name")
+assert payload.local == LocalPath.of("payload.text")
+```
+
+This documents that structural `parts` are literal field names. Old string inputs still split on dots for compatibility, but structural parts such as `"user.name"` are not split again.
+
+- [ ] **Step 3: Implement path serializers**
+
+In `src/wf_core/paths.py`, update each path type:
+
+- `LocalPath` accepts string, object instance, and dict `{"root": "local", "parts": list[str]}`
+- `GraphSourcePath` accepts string, object instance, and dict `{"root": "input"|"state"|"context", "parts": list[str]}`
+- `StatePath` accepts string, object instance, and dict `{"root": "state", "parts": list[str]}`
+
+Serialize as dicts in JSON mode:
+
+```python
+{"root": "local", "parts": list(value.parts)}
+{"root": value.root, "parts": list(value.parts)}
+{"root": "state", "parts": list(value.parts)}
+```
+
+- [ ] **Step 4: Run focused path tests**
+
+Run:
+
+```bash
+uv run --with pytest pytest tests/core/test_path_values.py -q
+```
+
+Expected: all tests pass.
+
+---
+
+## Task 2: Update Canonical Node Binding Dumps
+
+**Files:**
+- Test: `tests/core/test_canonical_node_bindings.py`
+- Test: `tests/authoring/test_builder.py`
+
+- [ ] **Step 1: Update canonical node dump expectations**
+
+In `tests/core/test_canonical_node_bindings.py`, update JSON-mode expectations:
+
+```python
+assert dumped["input"][1]["path"] == {"root": "input", "parts": ["message"]}
+assert dumped["input"][1]["target"] == {"root": "local", "parts": ["message"]}
+assert dumped["output"][0]["source"] == {"root": "local", "parts": ["echoed"]}
+assert dumped["output"][0]["target"] == {"root": "state", "parts": ["echoed"]}
+```
+
+Deprecated `in_map` / `out_map` inputs should continue parsing, but dumps must omit those old fields and emit structural paths.
+
+- [ ] **Step 2: Update authoring serialization expectations**
+
+In `tests/authoring/test_builder.py`, update any `model_dump(mode="json")` expectations that currently assert path strings.
+
+- [ ] **Step 3: Run focused binding/authoring tests**
+
+Run:
+
+```bash
+uv run --with pytest pytest tests/core/test_canonical_node_bindings.py tests/authoring/test_builder.py -q
+```
+
+Expected: all tests pass.
+
+---
+
+## Task 3: Update Docs
+
+**Files:**
+- Modify: `docs/structural_refs.md`
+- Modify: any path/core docs if directly relevant.
+
+- [ ] **Step 1: Add graph path note**
+
+Extend the path note in `docs/structural_refs.md`:
+
+```text
+New canonical graph path JSON uses root/parts objects. Old strings are accepted
+at parse boundaries for compatibility.
+```
+
+- [ ] **Step 2: Add examples**
+
+Include examples:
+
+```json
+{"root": "input", "parts": ["message"]}
+{"root": "state", "parts": ["echoed"]}
+{"root": "local", "parts": []}
+```
+
+---
+
+## Task 4: Verification
+
+- [ ] **Step 1: Run focused tests**
+
+```bash
+uv run --with pytest pytest tests/core/test_path_values.py tests/core/test_canonical_node_bindings.py tests/authoring/test_builder.py -q
+```
+
+- [ ] **Step 2: Run full tests**
+
+```bash
+uv run --with pytest pytest -q
+```
+
+- [ ] **Step 3: Run checks**
+
+```bash
+uvx ruff check src/wf_core/paths.py tests/core/test_path_values.py tests/core/test_canonical_node_bindings.py tests/authoring/test_builder.py
+uv run basedpyright --level error src/wf_core/paths.py tests/core/test_path_values.py tests/core/test_canonical_node_bindings.py tests/authoring/test_builder.py
+```
+
+---
+
+## Self-Review Notes
+
+- This plan does not revive `in_map` / `out_map`; those remain deprecated parse-only fields.
+- This plan relaxes path segment validation. Structural `parts` preserve literal field names, including dots and spaces. Old dotted string inputs still split on dots for compatibility.
+- This plan changes saved JSON shape for canonical path fields, so broad tests are required.
@@ -6,6 +6,11 @@ things a workflow should usually consume.
 For the short operator/client workflow using these concepts, see
 [`wf_mcp_operator_manual.md`](wf_mcp_operator_manual.md).

+For the identifier rule behind source/capability names, see
+[`structural_refs.md`](structural_refs.md). In short: dotted qualified names
+are display strings and compatibility input only; saved refs should be
+structural.
+
 The short version:

 ```text
@@ -29,7 +29,12 @@ class DiagnosticSeverity(StrEnum):


 class RequiredCapability(BaseModel):
-    """Saved contract for one capability an artifact references."""
+    """Saved contract for one capability an artifact references.
+
+    `ref` is canonical structure. Old dotted strings are accepted as
+    compatibility input, but new saves should preserve the source/capability
+    boundary because capability keys may contain dots.
+    """

    ref: CapabilityRefInput
    kind: Literal["tool", "resource", "prompt", "node_spec", "reducer", "workflow"]
@@ -68,7 +73,12 @@ class RequiredCapability(BaseModel):
            logical_source = data.pop("logical_source", None)
            capability_name = data.pop("capability_name", None)
            if isinstance(logical_source, str) and isinstance(capability_name, str):
-                data["ref"] = f"{logical_source}.{capability_name}"
+                # Preserve the source/capability boundary; capability names may
+                # contain dots, so joining then reparsing would corrupt the ref.
+                data["ref"] = {
+                    "source": logical_source,
+                    "capability_key": capability_name,
+                }
        return data


@@ -90,7 +100,12 @@ class AvailableSource(BaseModel):


 class SourceBinding(BaseModel):
-    """Deployment-time mapping from artifact logical source to concrete source."""
+    """Deployment-time mapping from artifact logical source to concrete source.
+
+    `logical_source` is the artifact-local alias used by saved refs.
+    `concrete_source` is the deployment-selected source id. Neither field is a
+    capability name.
+    """

    logical_source: SourceRefInput
    concrete_source: SourceRefInput
@@ -1,6 +1,9 @@
 from __future__ import annotations

 from dataclasses import dataclass
+from typing import Any
+
+from pydantic_core import core_schema


@dataclass(frozen=True, slots=True)
@@ -31,3 +34,39 @@ class WorkflowCapabilityRef:

    def __str__(self) -> str:
        return f"workflow.{self.artifact_id}.v{self.version}"
+
+    @classmethod
+    def __get_pydantic_core_schema__(
+        cls,
+        _source_type: object,
+        _handler: object,
+    ) -> core_schema.CoreSchema:
+        """Validate legacy display strings but save workflow refs structurally."""
+        return core_schema.no_info_plain_validator_function(
+            cls._validate,
+            serialization=core_schema.plain_serializer_function_ser_schema(
+                cls._serialize,
+                when_used="json",
+            ),
+        )
+
+    @classmethod
+    def _validate(cls, value: Any) -> WorkflowCapabilityRef:
+        if isinstance(value, WorkflowCapabilityRef):
+            return value
+        if isinstance(value, str):
+            return cls.parse(value)
+        if isinstance(value, dict):
+            artifact_id = value.get("artifact_id")
+            version = value.get("version")
+            if isinstance(artifact_id, str) and isinstance(version, int):
+                return cls(artifact_id=artifact_id, version=version)
+        raise TypeError(
+            "workflow capability ref must be a workflow.<artifact>.v<version> "
+            "string or {'artifact_id': str, 'version': int}"
+        )
+
+    @staticmethod
+    def _serialize(value: WorkflowCapabilityRef) -> dict[str, int | str]:
+        """Serialize canonical saved workflow refs without a display-name parser."""
+        return {"artifact_id": value.artifact_id, "version": value.version}
@@ -1,8 +1,7 @@
 from __future__ import annotations

-from dataclasses import dataclass
-import re
 from collections.abc import Mapping, MutableMapping
+from dataclasses import dataclass
 from typing import Any, ClassVar, Literal

 from pydantic_core import core_schema
@@ -12,12 +11,11 @@ class PathResolutionError(ValueError):
    pass


-SEGMENT_RE = re.compile(r"^[A-Za-z_][A-Za-z0-9_]*$")
 GraphRoot = Literal["input", "state", "context"]


 def _validate_segment(segment: str, *, path_kind: str) -> str:
-    if not SEGMENT_RE.fullmatch(segment):
+    if not segment or not segment.strip():
        raise PathResolutionError(f"invalid {path_kind} segment {segment!r}")
    return segment

@@ -36,8 +34,52 @@ def _parse_fragments(*fragments: str, path_kind: str) -> tuple[str, ...]:
    return tuple(parts)


-def _json_schema(pattern: str, description: str) -> dict[str, Any]:
-    return {"type": "string", "pattern": pattern, "description": description}
+def _path_json_schema(root: str | list[str], description: str) -> dict[str, Any]:
+    """Return the canonical structural schema for saved path fields."""
+    root_schema: dict[str, Any]
+    if isinstance(root, str):
+        root_schema = {"const": root}
+    else:
+        root_schema = {"enum": root}
+    return {
+        "type": "object",
+        "description": description,
+        "properties": {
+            "root": root_schema,
+            "parts": {
+                "type": "array",
+                "items": {"type": "string", "minLength": 1},
+            },
+        },
+        "required": ["root", "parts"],
+        "additionalProperties": False,
+    }
+
+
+def _structural_parts(
+    value: Mapping[object, object], *, path_kind: str
+) -> tuple[str, ...]:
+    """Validate literal path segments from canonical structural JSON.
+
+    Old string paths keep dotted parsing for compatibility. Structural `parts`
+    are different: each item is already one field name and may contain dots or
+    spaces, so this helper validates only that segments are non-empty strings.
+    """
+    raw_parts = value.get("parts", [])
+    if not isinstance(raw_parts, list):
+        raise PathResolutionError(f"{path_kind} path parts must be a list")
+    return tuple(
+        _validate_segment(part, path_kind=path_kind)
+        for part in raw_parts
+        if isinstance(part, str)
+    )
+
+
+def _reject_non_string_parts(value: Mapping[object, object], *, path_kind: str) -> None:
+    raw_parts = value.get("parts", [])
+    if isinstance(raw_parts, list) and all(isinstance(part, str) for part in raw_parts):
+        return
+    raise PathResolutionError(f"{path_kind} path parts must be strings")


@dataclass(frozen=True)
@@ -85,22 +127,32 @@ class LocalPath:
                return cls(value.parts)
            if isinstance(value, str):
                return cls.parse(value)
-            raise ValueError("expected local path string")
+            if isinstance(value, Mapping):
+                if value.get("root") != "local":
+                    raise ValueError("expected local path root")
+                _reject_non_string_parts(value, path_kind="local")
+                return cls(_structural_parts(value, path_kind="local"))
+            raise ValueError("expected local path string or structural object")

        return core_schema.no_info_plain_validator_function(
            validate,
            serialization=core_schema.plain_serializer_function_ser_schema(
-                str,
+                cls._serialize,
            ),
        )

+    @staticmethod
+    def _serialize(value: LocalPath) -> dict[str, str | list[str]]:
+        """Serialize canonical path JSON without relying on dotted display text."""
+        return {"root": "local", "parts": list(value.parts)}
+
    @classmethod
    def __get_pydantic_json_schema__(
        cls, _core_schema: core_schema.CoreSchema, _handler: object
    ) -> dict[str, Any]:
-        return _json_schema(
-            cls._JSON_PATTERN,
-            "Node-local dotted path or root marker `.`.",
+        return _path_json_schema(
+            "local",
+            "Node-local path. Use an empty parts list for the whole payload.",
        )


@@ -159,21 +211,32 @@ class GraphSourcePath:
                return cls(value.root, value.parts)
            if isinstance(value, str):
                return cls.parse(value)
-            raise ValueError("expected graph source path string")
+            if isinstance(value, Mapping):
+                root = value.get("root")
+                if root not in cls._ROOTS:
+                    raise ValueError("expected graph source path root")
+                _reject_non_string_parts(value, path_kind="graph source")
+                return cls(root, _structural_parts(value, path_kind="graph source"))  # type: ignore[arg-type]
+            raise ValueError("expected graph source path string or structural object")

        return core_schema.no_info_plain_validator_function(
            validate,
            serialization=core_schema.plain_serializer_function_ser_schema(
-                str,
+                cls._serialize,
            ),
        )

+    @staticmethod
+    def _serialize(value: GraphSourcePath) -> dict[str, str | list[str]]:
+        """Serialize canonical path JSON without relying on dotted display text."""
+        return {"root": value.root, "parts": list(value.parts)}
+
    @classmethod
    def __get_pydantic_json_schema__(
        cls, _core_schema: core_schema.CoreSchema, _handler: object
    ) -> dict[str, Any]:
-        return _json_schema(
-            cls._JSON_PATTERN,
+        return _path_json_schema(
+            sorted(cls._ROOTS),
            "Readable graph path rooted at input, state, or context.",
        )

@@ -220,20 +283,30 @@ class StatePath:
                return cls(value.parts)
            if isinstance(value, str):
                return cls.parse(value)
-            raise ValueError("expected state path string")
+            if isinstance(value, Mapping):
+                if value.get("root") != "state":
+                    raise ValueError("expected state path root")
+                _reject_non_string_parts(value, path_kind="state")
+                return cls(_structural_parts(value, path_kind="state"))
+            raise ValueError("expected state path string or structural object")

        return core_schema.no_info_plain_validator_function(
            validate,
            serialization=core_schema.plain_serializer_function_ser_schema(
-                str,
+                cls._serialize,
            ),
        )

+    @staticmethod
+    def _serialize(value: StatePath) -> dict[str, str | list[str]]:
+        """Serialize canonical path JSON without relying on dotted display text."""
+        return {"root": "state", "parts": list(value.parts)}
+
    @classmethod
    def __get_pydantic_json_schema__(
        cls, _core_schema: core_schema.CoreSchema, _handler: object
    ) -> dict[str, Any]:
-        return _json_schema(cls._JSON_PATTERN, "Writable state path such as state.foo.")
+        return _path_json_schema("state", "Writable workflow state path.")


 def split_graph_path(path: str | GraphSourcePath | StatePath) -> tuple[str, list[str]]:
@@ -1,6 +1,6 @@
 from __future__ import annotations

-from typing import TypeAlias
+from typing import Any, TypeAlias

 from wf_artifacts import WorkflowCapabilityRef
 from wf_platform import CapabilityRef
@@ -8,13 +8,20 @@ from wf_platform import CapabilityRef
 WorkflowSurfaceCapabilityId: TypeAlias = CapabilityRef | WorkflowCapabilityRef


-def parse_workflow_surface_capability_id(value: str) -> WorkflowSurfaceCapabilityId:
+def parse_workflow_surface_capability_id(
+    value: str | dict[str, Any],
+) -> WorkflowSurfaceCapabilityId:
    """Parse a workflow-surface capability name into its real domain ref.

    MCP tools still accept and return plain strings. Internally, workflow-facing
    capability ids are either live source capabilities or saved wrapper
    artifacts, so this parser avoids inventing a third identifier model.
    """
+    if isinstance(value, dict):
+        if "artifact_id" in value and "version" in value:
+            return WorkflowCapabilityRef._validate(value)
+        return CapabilityRef._validate(value)
+
    try:
        return WorkflowCapabilityRef.parse(value)
    except ValueError:
@@ -88,7 +88,7 @@ class CapabilityRef:
        return core_schema.no_info_plain_validator_function(
            cls._validate,
            serialization=core_schema.plain_serializer_function_ser_schema(
-                str,
+                cls._serialize,
                when_used="json",
            ),
        )
@@ -99,4 +99,16 @@ class CapabilityRef:
            return value
        if isinstance(value, str):
            return cls.parse(value)
-        raise TypeError("capability ref must be a string")
+        if isinstance(value, dict):
+            source = value.get("source")
+            name = value.get("capability_key")
+            if isinstance(source, str) and isinstance(name, str):
+                return cls(source=SourceRef.parse(source), name=name)
+        raise TypeError(
+            "capability ref must be a string or {'source': str, 'capability_key': str}"
+        )
+
+    @staticmethod
+    def _serialize(value: CapabilityRef) -> dict[str, str]:
+        """Serialize canonical saved refs structurally; `str(ref)` is display-only."""
+        return {"source": str(value.source), "capability_key": value.name}
@@ -41,7 +41,10 @@ def test_workflow_artifact_serializes_required_capability_contract() -> None:
    assert dumped["version"] == 1
    assert dumped["outcomes"] == ["done", "failed"]
    required = dumped["required_capabilities"][0]
-    assert required["ref"] == "context7.query-docs"
+    assert required["ref"] == {
+        "source": "context7",
+        "capability_key": "query-docs",
+    }
    assert required["input_schema_hash"] == "sha256:input"
    assert artifact.required_capability_map()["context7.query-docs"] == capability

@@ -66,7 +69,8 @@ def test_workflow_artifact_can_be_marked_as_wrapper_intent() -> None:
 def test_workflow_artifact_accepts_legacy_required_capability_map_and_dumps_list() -> (
    None
 ):
-    artifact = WorkflowArtifact.model_validate({
+    artifact = WorkflowArtifact.model_validate(
+        {
            "id": "legacy_capabilities",
            "version": 1,
            "title": "Legacy Capabilities",
@@ -82,19 +86,63 @@ def test_workflow_artifact_accepts_legacy_required_capability_map_and_dumps_list
                    "observed_concrete_source": "demo.personal",
                }
            },
-    })
+        }
+    )

    dumped = artifact.model_dump(mode="json")
    required = dumped["required_capabilities"][0]

    assert isinstance(dumped["required_capabilities"], list)
-    assert required["ref"] == "demo.echo"
+    assert required["ref"] == {"source": "demo", "capability_key": "echo"}
    assert required["observed_concrete_source"] == "demo.personal"
    assert "logical_source" not in required
    assert "capability_name" not in required
    assert artifact.required_capability_map()["demo.echo"].logical_source == "demo"


+def test_required_capability_accepts_legacy_logical_fields_but_dumps_ref_object() -> (
+    None
+):
+    capability = RequiredCapability.model_validate(
+        {
+            "logical_source": "demo",
+            "capability_name": "foo.bar",
+            "kind": "node_spec",
+        }
+    )
+
+    assert capability.logical_source == "demo"
+    assert capability.capability_name == "foo.bar"
+    assert capability.model_dump(mode="json")["ref"] == {
+        "source": "demo",
+        "capability_key": "foo.bar",
+    }
+
+
+def test_workflow_artifact_legacy_required_capability_map_is_best_effort() -> None:
+    artifact = WorkflowArtifact.model_validate(
+        {
+            "id": "legacy_dotted_capability",
+            "version": 1,
+            "title": "Legacy Dotted Capability",
+            "input_schema": {"type": "object", "properties": {}},
+            "output_schema": {"type": "object", "properties": {}},
+            "outcomes": ["done"],
+            "plan": {"name": "legacy_dotted_capability", "nodes": [], "edges": []},
+            "required_capabilities": {
+                "demo.foo.bar": {"kind": "node_spec"},
+            },
+        }
+    )
+
+    required = artifact.model_dump(mode="json")["required_capabilities"][0]
+
+    assert required["ref"] == {
+        "source": "demo.foo",
+        "capability_key": "bar",
+    }
+
+
 def test_workflow_deployment_binds_logical_sources_to_concrete_sources() -> None:
    deployment = WorkflowDeployment(
        id="summarize_docs.personal",
@@ -118,12 +166,14 @@ def test_workflow_deployment_binds_logical_sources_to_concrete_sources() -> None


 def test_workflow_deployment_accepts_legacy_binding_map_and_dumps_list() -> None:
-    deployment = WorkflowDeployment.model_validate({
+    deployment = WorkflowDeployment.model_validate(
+        {
            "id": "legacy_bindings.personal",
            "artifact_id": "legacy_bindings",
            "artifact_version": 1,
            "bindings": {"demo": "demo.personal"},
-    })
+        }
+    )

    dumped = deployment.model_dump(mode="json")
    binding = dumped["bindings"][0]
@@ -73,7 +73,8 @@ def test_file_store_loads_legacy_artifact_and_rewrites_canonical_shape(
    artifact_dir.mkdir(parents=True)
    artifact_path = artifact_dir / "1.json"
    artifact_path.write_text(
-        json.dumps({
+        json.dumps(
+            {
                "id": "legacy_capabilities",
                "version": 1,
                "title": "Legacy Capabilities",
@@ -87,7 +88,8 @@ def test_file_store_loads_legacy_artifact_and_rewrites_canonical_shape(
                        "input_schema_hash": "sha256:input",
                    }
                },
-        }),
+            }
+        ),
        encoding="utf-8",
    )

@@ -96,7 +98,7 @@ def test_file_store_loads_legacy_artifact_and_rewrites_canonical_shape(
    rewritten = json.loads(artifact_path.read_text(encoding="utf-8"))

    required = rewritten["required_capabilities"][0]
-    assert required["ref"] == "demo.echo"
+    assert required["ref"] == {"source": "demo", "capability_key": "echo"}
    assert required["kind"] == "tool"
    assert "logical_source" not in required
    assert "capability_name" not in required
@@ -108,12 +110,14 @@ def test_file_store_loads_legacy_deployment_and_rewrites_canonical_shape(
    store = FileWorkflowArtifactStore(tmp_path)
    deployment_path = store.deployments_dir / "legacy_bindings.personal.json"
    deployment_path.write_text(
-        json.dumps({
+        json.dumps(
+            {
                "id": "legacy_bindings.personal",
                "artifact_id": "legacy_bindings",
                "artifact_version": 1,
                "bindings": {"demo": "demo.personal"},
-        }),
+            }
+        ),
        encoding="utf-8",
    )

@@ -105,10 +105,10 @@ def test_builder_emits_canonical_node_bindings() -> None:

    dumped_node = builder.compile().model_dump(mode="json")["nodes"][0]

-    assert dumped_node["input"][0]["path"] == "input.text"
-    assert dumped_node["input"][0]["target"] == "text"
-    assert dumped_node["output"][0]["source"] == "text"
-    assert dumped_node["output"][0]["target"] == "state.text"
+    assert dumped_node["input"][0]["path"] == {"root": "input", "parts": ["text"]}
+    assert dumped_node["input"][0]["target"] == {"root": "local", "parts": ["text"]}
+    assert dumped_node["output"][0]["source"] == {"root": "local", "parts": ["text"]}
+    assert dumped_node["output"][0]["target"] == {"root": "state", "parts": ["text"]}
    assert "in_map" not in dumped_node
    assert "input_values" not in dumped_node
    assert "out_map" not in dumped_node
@@ -14,9 +14,12 @@ def test_condition_dsl_compiles_to_core_condition() -> None:
    dumped = condition.to_condition().model_dump(mode="json")

    assert dumped["op"] == "and"
-    assert dumped["args"][0]["left"]["path"] == "state.should_email"
+    assert dumped["args"][0]["left"]["path"] == {
+        "root": "state",
+        "parts": ["should_email"],
+    }
    assert dumped["args"][0]["right"]["value"] is True
-    assert dumped["args"][1]["path"] == "state.summary"
+    assert dumped["args"][1]["path"] == {"root": "state", "parts": ["summary"]}


 def test_condition_dsl_compiles_authoring_paths_to_typed_core_paths() -> None:
@@ -28,10 +31,16 @@ def test_condition_dsl_compiles_authoring_paths_to_typed_core_paths() -> None:
    assert isinstance(comparison.right, PathOperand)
    assert comparison.left.path == GraphSourcePath.state("score")
    assert comparison.right.path == GraphSourcePath.state("threshold")
-    assert comparison.model_dump(mode="json")["left"]["path"] == "state.score"
+    assert comparison.model_dump(mode="json")["left"]["path"] == {
+        "root": "state",
+        "parts": ["score"],
+    }
    assert isinstance(existence, ExistsCondition)
    assert existence.path == GraphSourcePath.state("summary")
-    assert existence.model_dump(mode="json")["path"] == "state.summary"
+    assert existence.model_dump(mode="json")["path"] == {
+        "root": "state",
+        "parts": ["summary"],
+    }


 def test_condition_dsl_supports_not_ge_and_ne() -> None:
@@ -6,7 +6,8 @@ from wf_core.paths import GraphSourcePath, LocalPath, StatePath


 def test_node_use_accepts_canonical_input_and_output_bindings():
-    node = NodeUse.model_validate({
+    node = NodeUse.model_validate(
+        {
            "id": "echo",
            "type": "node",
            "node": "echo",
@@ -15,7 +16,8 @@ def test_node_use_accepts_canonical_input_and_output_bindings():
                {"target": "mode", "value": None},
            ],
            "output": [{"source": "echoed", "target": "state.echoed"}],
-    })
+        }
+    )

    path_binding = node.input[0]
    assert isinstance(path_binding, InputPathBinding)
@@ -32,78 +34,112 @@ def test_node_use_accepts_canonical_input_and_output_bindings():


 def test_node_use_converts_old_maps_to_canonical_bindings():
-    node = NodeUse.model_validate({
+    node = NodeUse.model_validate(
+        {
            "id": "echo",
            "type": "node",
            "node": "echo",
            "in_map": {"input.message": "message"},
            "input_values": {"mode": "fast"},
            "out_map": {"echoed": "state.echoed"},
-    })
+        }
+    )

    dumped = node.model_dump(mode="json")
    assert "in_map" not in dumped
    assert "input_values" not in dumped
    assert "out_map" not in dumped
    assert dumped["input"][0]["value"] == "fast"
-    assert dumped["input"][0]["target"] == "mode"
-    assert dumped["input"][1]["path"] == "input.message"
-    assert dumped["input"][1]["target"] == "message"
-    assert dumped["output"][0]["source"] == "echoed"
-    assert dumped["output"][0]["target"] == "state.echoed"
+    assert dumped["input"][0]["target"] == {"root": "local", "parts": ["mode"]}
+    assert dumped["input"][1]["path"] == {"root": "input", "parts": ["message"]}
+    assert dumped["input"][1]["target"] == {"root": "local", "parts": ["message"]}
+    assert dumped["output"][0]["source"] == {"root": "local", "parts": ["echoed"]}
+    assert dumped["output"][0]["target"] == {"root": "state", "parts": ["echoed"]}


-def test_node_use_serializes_canonical_binding_paths_as_strings_in_all_dump_modes():
-    node = NodeUse.model_validate({
+def test_node_use_serializes_canonical_binding_paths_as_structural_json():
+    node = NodeUse.model_validate(
+        {
            "id": "echo",
            "type": "node",
            "node": "echo",
            "input": [{"target": "message", "path": "input.message"}],
            "output": [{"source": "echoed", "target": "state.echoed"}],
-    })
+        }
+    )

    python_dumped = node.model_dump()
    json_dumped = node.model_dump(mode="json")

-    assert python_dumped["input"][0]["target"] == "message"
-    assert python_dumped["input"][0]["path"] == "input.message"
-    assert python_dumped["output"][0]["source"] == "echoed"
-    assert python_dumped["output"][0]["target"] == "state.echoed"
-    assert json_dumped["input"][0]["target"] == "message"
-    assert json_dumped["input"][0]["path"] == "input.message"
-    assert json_dumped["output"][0]["source"] == "echoed"
-    assert json_dumped["output"][0]["target"] == "state.echoed"
+    assert python_dumped["input"][0]["target"] == {
+        "root": "local",
+        "parts": ["message"],
+    }
+    assert python_dumped["input"][0]["path"] == {
+        "root": "input",
+        "parts": ["message"],
+    }
+    assert python_dumped["output"][0]["source"] == {
+        "root": "local",
+        "parts": ["echoed"],
+    }
+    assert python_dumped["output"][0]["target"] == {
+        "root": "state",
+        "parts": ["echoed"],
+    }
+    assert json_dumped["input"][0]["target"] == {
+        "root": "local",
+        "parts": ["message"],
+    }
+    assert json_dumped["input"][0]["path"] == {
+        "root": "input",
+        "parts": ["message"],
+    }
+    assert json_dumped["output"][0]["source"] == {
+        "root": "local",
+        "parts": ["echoed"],
+    }
+    assert json_dumped["output"][0]["target"] == {
+        "root": "state",
+        "parts": ["echoed"],
+    }


 def test_node_use_rejects_mixed_old_and_new_binding_styles():
    with pytest.raises(ValidationError):
-        NodeUse.model_validate({
+        NodeUse.model_validate(
+            {
                "id": "echo",
                "type": "node",
                "node": "echo",
                "input": [{"target": "message", "path": "input.message"}],
                "in_map": {"input.other": "other"},
-        })
+            }
+        )


 def test_input_binding_rejects_path_and_value_together():
    with pytest.raises(ValidationError):
-        NodeUse.model_validate({
+        NodeUse.model_validate(
+            {
                "id": "bad",
                "type": "node",
                "node": "bad",
                "input": [{"target": "message", "path": "input.message", "value": "x"}],
-        })
+            }
+        )


 def test_input_binding_rejects_neither_path_nor_value():
    with pytest.raises(ValidationError):
-        NodeUse.model_validate({
+        NodeUse.model_validate(
+            {
                "id": "bad",
                "type": "node",
                "node": "bad",
                "input": [{"target": "message"}],
-        })
+            }
+        )


@pytest.mark.parametrize(
@@ -115,12 +151,14 @@ def test_input_binding_rejects_neither_path_nor_value():
 )
 def test_bindings_reject_extra_fields(field: str, binding: dict[str, object]):
    with pytest.raises(ValidationError):
-        NodeUse.model_validate({
+        NodeUse.model_validate(
+            {
                "id": "bad",
                "type": "node",
                "node": "bad",
                field: [binding],
-        })
+            }
+        )


@pytest.mark.parametrize(
@@ -133,46 +171,52 @@ def test_bindings_reject_extra_fields(field: str, binding: dict[str, object]):
 )
 def test_deprecated_maps_reject_non_mapping_values(field: str, value: object):
    with pytest.raises(ValidationError):
-        NodeUse.model_validate({
+        NodeUse.model_validate(
+            {
                "id": "bad",
                "type": "node",
                "node": "bad",
                field: value,
-        })
+            }
+        )


 def test_deprecated_conversion_preserves_input_value_then_in_map_order():
-    node = NodeUse.model_validate({
+    node = NodeUse.model_validate(
+        {
            "id": "ordered",
            "type": "node",
            "node": "ordered",
            "input_values": {"first": 1, "second": 2},
            "in_map": {"input.third": "third", "state.fourth": "fourth"},
-    })
+        }
+    )

    dumped_input = node.model_dump(mode="json")["input"]
-    assert dumped_input[0]["target"] == "first"
+    assert dumped_input[0]["target"] == {"root": "local", "parts": ["first"]}
    assert dumped_input[0]["value"] == 1
-    assert dumped_input[1]["target"] == "second"
+    assert dumped_input[1]["target"] == {"root": "local", "parts": ["second"]}
    assert dumped_input[1]["value"] == 2
-    assert dumped_input[2]["target"] == "third"
-    assert dumped_input[2]["path"] == "input.third"
-    assert dumped_input[3]["target"] == "fourth"
-    assert dumped_input[3]["path"] == "state.fourth"
+    assert dumped_input[2]["target"] == {"root": "local", "parts": ["third"]}
+    assert dumped_input[2]["path"] == {"root": "input", "parts": ["third"]}
+    assert dumped_input[3]["target"] == {"root": "local", "parts": ["fourth"]}
+    assert dumped_input[3]["path"] == {"root": "state", "parts": ["fourth"]}


 def test_deprecated_input_value_preserves_explicit_null():
-    node = NodeUse.model_validate({
+    node = NodeUse.model_validate(
+        {
            "id": "null",
            "type": "node",
            "node": "null",
            "input_values": {"maybe": None},
-    })
+        }
+    )

    value_binding = node.input[0]
    assert isinstance(value_binding, InputValueBinding)
    assert value_binding.value is None

    dumped_input = node.model_dump(mode="json")["input"]
-    assert dumped_input[0]["target"] == "maybe"
+    assert dumped_input[0]["target"] == {"root": "local", "parts": ["maybe"]}
    assert dumped_input[0]["value"] is None
@@ -23,6 +23,25 @@ def test_graph_source_path_accepts_root_and_nested_paths() -> None:
    assert str(GraphSourcePath.context("loop_item")) == "context.loop_item"


+def test_structural_path_parts_preserve_literal_field_names() -> None:
+    class Payload(BaseModel):
+        source: GraphSourcePath
+        target: StatePath
+        local: LocalPath
+
+    payload = Payload.model_validate(
+        {
+            "source": {"root": "input", "parts": ["user.name"]},
+            "target": {"root": "state", "parts": ["person name"]},
+            "local": {"root": "local", "parts": ["payload.text"]},
+        }
+    )
+
+    assert payload.source == GraphSourcePath("input", ("user.name",))
+    assert payload.target == StatePath(("person name",))
+    assert payload.local == LocalPath(("payload.text",))
+
+
 def test_state_path_serializes_with_state_prefix() -> None:
    assert str(StatePath.of("person.name")) == "state.person.name"
    assert str(StatePath.parse("state.person.name")) == "state.person.name"
@@ -47,9 +66,6 @@ def test_local_path_supports_root_marker_and_fragments() -> None:
        ".",
        "state.",
        "state..name",
-        "state.items.0",
-        "state.user-name",
-        "state.items[0]",
        "output.foo",
    ],
 )
@@ -71,9 +87,6 @@ def test_graph_source_paths_reject_invalid_segments(raw: str) -> None:
    [
        "state.",
        "state..name",
-        "state.items.0",
-        "state.user-name",
-        "state.items[0]",
    ],
 )
 def test_all_path_types_reject_invalid_segments(factory, raw: str) -> None:
@@ -92,9 +105,9 @@ def test_path_objects_are_immutable_and_hashable() -> None:
@pytest.mark.parametrize(
    ("factory", "args"),
    [
-        (GraphSourcePath, ("output", ("user-name",))),
-        (StatePath, (("0",),)),
-        (LocalPath, (("items[0]",),)),
+        (GraphSourcePath, ("output", ("user",))),
+        (StatePath, (("",),)),
+        (LocalPath, ((" ",),)),
    ],
 )
 def test_direct_constructors_enforce_path_invariants(
@@ -114,64 +127,93 @@ def test_pydantic_revalidates_existing_path_objects() -> None:
    # constructor validation. Pydantic must not blindly trust existing instances.
    source = object.__new__(GraphSourcePath)
    object.__setattr__(source, "root", "output")
-    object.__setattr__(source, "parts", ("user-name",))
+    object.__setattr__(source, "parts", ("user",))
    target = object.__new__(StatePath)
-    object.__setattr__(target, "parts", ("0",))
+    object.__setattr__(target, "parts", ("",))
    local = object.__new__(LocalPath)
-    object.__setattr__(local, "parts", ("items[0]",))
+    object.__setattr__(local, "parts", (" ",))

    with pytest.raises(ValidationError):
-        Payload.model_validate({
+        Payload.model_validate(
+            {
                "source": source,
                "target": StatePath.of("person"),
                "local": LocalPath.root(),
-        })
+            }
+        )
    with pytest.raises(ValidationError):
-        Payload.model_validate({
+        Payload.model_validate(
+            {
                "source": GraphSourcePath.input("user"),
                "target": target,
                "local": LocalPath.root(),
-        })
+            }
+        )
    with pytest.raises(ValidationError):
-        Payload.model_validate({
+        Payload.model_validate(
+            {
                "source": GraphSourcePath.input("user"),
                "target": StatePath.of("person"),
                "local": local,
-        })
+            }
+        )


-def test_pydantic_accepts_path_strings_and_serializes_strings() -> None:
+def test_pydantic_accepts_path_strings_and_serializes_structural_json() -> None:
    class Payload(BaseModel):
        source: GraphSourcePath
        target: StatePath
        local: LocalPath

-    payload = Payload.model_validate({
+    payload = Payload.model_validate(
+        {
            "source": "input.user",
            "target": "state.person",
            "local": "user",
-    })
+        }
+    )

    assert payload.source == GraphSourcePath.input("user")
    assert payload.target == StatePath.of("person")
    assert payload.local == LocalPath.of("user")

    dumped = payload.model_dump(mode="json")
-    assert dumped["source"] == "input.user"
-    assert dumped["target"] == "state.person"
-    assert dumped["local"] == "user"
+    assert dumped["source"] == {"root": "input", "parts": ["user"]}
+    assert dumped["target"] == {"root": "state", "parts": ["person"]}
+    assert dumped["local"] == {"root": "local", "parts": ["user"]}

    python_dumped = payload.model_dump()
-    assert python_dumped["source"] == "input.user"
-    assert python_dumped["target"] == "state.person"
-    assert python_dumped["local"] == "user"
+    assert python_dumped["source"] == {"root": "input", "parts": ["user"]}
+    assert python_dumped["target"] == {"root": "state", "parts": ["person"]}
+    assert python_dumped["local"] == {"root": "local", "parts": ["user"]}


-def test_condition_path_operand_serializes_path_as_string_in_all_dump_modes() -> None:
+def test_path_json_schema_advertises_structural_shape() -> None:
+    class Payload(BaseModel):
+        source: GraphSourcePath
+        target: StatePath
+        local: LocalPath
+
+    schema = Payload.model_json_schema()
+
+    assert schema["properties"]["source"]["type"] == "object"
+    assert schema["properties"]["source"]["properties"]["root"]["enum"] == [
+        "context",
+        "input",
+        "state",
+    ]
+    assert schema["properties"]["target"]["properties"]["root"]["const"] == "state"
+    assert schema["properties"]["local"]["properties"]["root"]["const"] == "local"
+
+
+def test_condition_path_operand_serializes_path_as_structural_json() -> None:
    operand = PathOperand.model_validate({"path": "state.x"})

-    assert operand.model_dump()["path"] == "state.x"
-    assert operand.model_dump(mode="json")["path"] == "state.x"
+    assert operand.model_dump()["path"] == {"root": "state", "parts": ["x"]}
+    assert operand.model_dump(mode="json")["path"] == {
+        "root": "state",
+        "parts": ["x"],
+    }


 def test_pydantic_accepts_existing_path_objects() -> None:
@@ -180,11 +222,13 @@ def test_pydantic_accepts_existing_path_objects() -> None:
        target: StatePath
        local: LocalPath

-    payload = Payload.model_validate({
+    payload = Payload.model_validate(
+        {
            "source": GraphSourcePath.state("person"),
            "target": StatePath.of("person.name"),
            "local": LocalPath.root(),
-    })
+        }
+    )

    assert str(payload.source) == "state.person"
    assert str(payload.target) == "state.person.name"
@@ -205,7 +249,7 @@ def test_existing_source_and_destination_validation_helpers_use_new_parsers() ->
    assert is_valid_source_path("context", set(), set(), allow_context=True) is True
    assert is_valid_source_path("state.person", {"person"}, set()) is True
    assert is_valid_source_path("input.person", set(), {"person"}) is True
-    assert is_valid_source_path("state.person-name", {"person-name"}, set()) is False
+    assert is_valid_source_path("state.person-name", {"person-name"}, set()) is True

    assert is_valid_destination_path("state") is False
    assert is_valid_destination_path("state.person") is True
@@ -29,10 +29,10 @@ def test_raw_canonical_workflow_serializes_new_shape() -> None:
    assert "in_map" not in node
    assert "input_values" not in node
    assert "out_map" not in node
-    assert node["input"][0]["path"] == "input.text"
-    assert node["input"][0]["target"] == "text"
+    assert node["input"][0]["path"] == {"root": "input", "parts": ["text"]}
+    assert node["input"][0]["target"] == {"root": "local", "parts": ["text"]}
    assert node["input"][1]["value"] == "raw:"
-    assert node["output"][0]["source"] == "message"
-    assert node["output"][0]["target"] == "state.message"
+    assert node["output"][0]["source"] == {"root": "local", "parts": ["message"]}
+    assert node["output"][0]["target"] == {"root": "state", "parts": ["message"]}
    assert message_schema["type"] == "string"
    assert message_schema["reducer"] == "wf.std.replace"
@@ -33,16 +33,47 @@ def test_platform_refs_validate_and_serialize_through_pydantic() -> None:
        source: SourceRef
        capability: CapabilityRef

-    payload = Payload.model_validate({
+    payload = Payload.model_validate(
+        {
            "source": "demo.personal",
            "capability": "demo.personal.echo_tool",
-    })
+        }
+    )

    assert payload.source == SourceRef.parse("demo.personal")
    assert payload.capability == CapabilityRef.parse("demo.personal.echo_tool")
    assert payload.model_dump(mode="json") == {
        "source": "demo.personal",
-        "capability": "demo.personal.echo_tool",
+        "capability": {
+            "source": "demo.personal",
+            "capability_key": "echo_tool",
+        },
+    }
+
+
+def test_capability_ref_accepts_structural_input() -> None:
+    class Payload(BaseModel):
+        capability: CapabilityRef
+
+    payload = Payload.model_validate(
+        {"capability": {"source": "demo", "capability_key": "foo.bar"}}
+    )
+
+    assert payload.capability.source == SourceRef.parse("demo")
+    assert payload.capability.name == "foo.bar"
+
+
+def test_capability_ref_serializes_structurally() -> None:
+    class Payload(BaseModel):
+        capability: CapabilityRef
+
+    payload = Payload(
+        capability=CapabilityRef(source=SourceRef.parse("demo"), name="foo.bar")
+    )
+
+    assert payload.model_dump(mode="json")["capability"] == {
+        "source": "demo",
+        "capability_key": "foo.bar",
    }


@@ -1,5 +1,7 @@
 from __future__ import annotations

+from pydantic import BaseModel
+
 from wf_artifacts import WorkflowCapabilityRef


@@ -17,6 +19,40 @@ def test_workflow_capability_ref_preserves_dotted_artifact_ids() -> None:
    assert ref.version == 3


+def test_workflow_capability_ref_validates_legacy_string_input() -> None:
+    class Payload(BaseModel):
+        ref: WorkflowCapabilityRef
+
+    payload = Payload.model_validate({"ref": "workflow.echo_wrapper.v1"})
+
+    assert payload.ref.artifact_id == "echo_wrapper"
+    assert payload.ref.version == 1
+
+
+def test_workflow_capability_ref_validates_structural_input() -> None:
+    class Payload(BaseModel):
+        ref: WorkflowCapabilityRef
+
+    payload = Payload.model_validate(
+        {"ref": {"artifact_id": "echo_wrapper", "version": 1}}
+    )
+
+    assert payload.ref.artifact_id == "echo_wrapper"
+    assert payload.ref.version == 1
+
+
+def test_workflow_capability_ref_serializes_structurally() -> None:
+    class Payload(BaseModel):
+        ref: WorkflowCapabilityRef
+
+    payload = Payload(ref=WorkflowCapabilityRef("echo_wrapper", 1))
+
+    assert payload.model_dump(mode="json")["ref"] == {
+        "artifact_id": "echo_wrapper",
+        "version": 1,
+    }
+
+
 def test_workflow_capability_ref_rejects_other_namespaces() -> None:
    try:
        WorkflowCapabilityRef.parse("demo.echo.v1")
@@ -5,7 +5,7 @@ import shutil
 from typing import Any, cast

 from wf_artifacts import FileDraftWorkspaceStore, WorkflowDeployment
-from wf_authoring import NodeSpec, build_async_registry
+from wf_authoring import NodeSpec, build_async_registry, node
 from wf_core import END, NodeUse, RunStatus, RuntimeContext
 from wf_mcp.broker import WfMcpService
 from wf_mcp.models import AuthRecord, ConnectionConfig, RawWorkflowPlan
@@ -20,6 +20,8 @@ from wf_platform import (
 )

 from .test_support import (
+    EchoInput,
+    EchoOutput,
    FailingDiscoveryAdapter,
    FakeAdapter,
    echo_tool,
@@ -28,6 +30,11 @@ from .test_support import (
 )


+@node(name="foo.bar")
+def pro_dotted_echo_tool(payload: EchoInput) -> EchoOutput:
+    return EchoOutput(echoed=f"pro:{payload.text}")
+
+
 def _single_echo_plan(plan_name: str, node_name: str) -> RawWorkflowPlan:
    return _raw_plan(
        name=plan_name,
@@ -476,6 +483,50 @@ def test_service_runs_logical_source_plan_with_dotted_local_name() -> None:
    assert run.output["echoed"] == "hello"


+def test_service_binds_longest_logical_source_prefix_first() -> None:
+    service = WfMcpService(
+        store=FileStore(local_temp_root() / "longest_logical_source_prefix")
+    )
+    service.register_connection(
+        ConnectionConfig(id="demo.personal", server="demo", account="personal")
+    )
+    service.register_connection(
+        ConnectionConfig(
+            id="demo.pro.personal",
+            server="demo",
+            account="pro.personal",
+        )
+    )
+    service.register_specs("demo.personal", echo_tool)
+    service.register_specs("demo.pro.personal", pro_dotted_echo_tool)
+    plan = _single_echo_plan(
+        "longest_logical_source_prefix_plan",
+        "demo.pro.foo.bar",
+    )
+
+    run = asyncio.run(
+        service.run_workflow_from_plan(
+            plan,
+            {"text": "hello"},
+            deployment=WorkflowDeployment(
+                id="longest_logical_source_prefix.personal",
+                artifact_id="longest_logical_source_prefix",
+                artifact_version=1,
+                bindings=[
+                    {"logical_source": "demo", "concrete_source": "demo.personal"},
+                    {
+                        "logical_source": "demo.pro",
+                        "concrete_source": "demo.pro.personal",
+                    },
+                ],
+            ),
+        )
+    )
+
+    assert run.status == RunStatus.COMPLETED
+    assert run.output["echoed"] == "pro:hello"
+
+
 def test_service_does_not_resolve_specs_hidden_from_planner() -> None:
    service = WfMcpService(store=FileStore(local_temp_root() / "hidden_spec_store"))
    hidden_echo_tool = NodeSpec(
@@ -14,6 +14,19 @@ def test_workflow_surface_capability_id_parses_live_capability_ref() -> None:
    assert capability.name == "echo_tool"


+def test_workflow_surface_capability_id_parses_structural_live_capability_ref() -> None:
+    capability = parse_workflow_surface_capability_id(
+        {
+            "source": "demo",
+            "capability_key": "foo.bar",
+        }
+    )
+
+    assert isinstance(capability, CapabilityRef)
+    assert str(capability.source) == "demo"
+    assert capability.name == "foo.bar"
+
+
 def test_workflow_surface_capability_id_parses_saved_wrapper_ref() -> None:
    capability = parse_workflow_surface_capability_id("workflow.echo_wrapper.v2")

@@ -21,3 +34,15 @@ def test_workflow_surface_capability_id_parses_saved_wrapper_ref() -> None:
    assert str(capability) == "workflow.echo_wrapper.v2"
    assert capability.artifact_id == "echo_wrapper"
    assert capability.version == 2
+
+
+def test_workflow_surface_capability_id_parses_structural_saved_wrapper_ref() -> None:
+    capability = parse_workflow_surface_capability_id(
+        {
+            "artifact_id": "echo_wrapper",
+            "version": 2,
+        }
+    )
+
+    assert isinstance(capability, WorkflowCapabilityRef)
+    assert str(capability) == "workflow.echo_wrapper.v2"