documents: async? mcp? schema type validation?

authoring_sketch.md

@@ -124,6 +124,101 @@ Preferred design:
 If sync runtime encounters an async node, fail clearly rather than faking a sync
 bridge.
 
+## Async runtime seam
+
+The intended async work should be additive, not a rewrite.
+
+Recommended shape:
+
+- keep current sync runtime as the stable baseline
+- add async siblings instead of mutating the sync path into a mixed mode
+- avoid hidden sync-to-async or async-to-sync bridges in core execution
+
+Likely async entry points:
+
+- `execute_workflow_async(...)`
+- `resume_workflow_async(...)`
+- `step_workflow_async(...)`
+- `execute_node_use_async(...)`
+
+Likely registry split:
+
+- sync registry: `dict[str, SyncNodeHandler]`
+- async registry: `dict[str, AsyncNodeHandler]`
+
+`NodeSpec` should support both export paths:
+
+- `to_registry_handler()` for sync callables only
+- `to_async_registry_handler()` for sync or async callables
+- `build_registry(...)` for sync specs
+- `build_async_registry(...)` for mixed or async specs
+
+This keeps the rules simple:
+
+- sync runtime executes sync handlers only
+- async runtime can execute both sync and async handlers
+- async runtime is the natural home for future MCP-backed tool nodes
+
+## MCP proxy layer
+
+MCP integration should sit above `wf_core`, not inside it.
+
+Recommended layering:
+
+1. MCP client or proxy code discovers tools
+2. each MCP tool is wrapped as a `NodeSpec`
+3. wrapped specs enter the same `NodeCatalog` as handwritten nodes
+4. the client LLM builds graphs against one unified catalog
+5. workflows compile to the existing core `Workflow`
+6. runtime executes registry handlers without caring whether the backing tool is local Python or MCP
+
+This means MCP tools should look like ordinary nodes at the authoring boundary:
+
+- declared input model
+- declared output model
+- declared outcomes
+- description/docs for LLM consumption
+- sync or async execution capability
+
+The proxy/MCP layer should be responsible for:
+
+- tool discovery
+- schema translation
+- auth/session concerns
+- wrapping tool calls into `NodeSpec`s
+
+The core runtime should remain responsible only for:
+
+- mapping input/state/context into node payloads
+- validating payloads
+- routing outcomes
+- writing mapped output into workflow state
+- interrupts, frames, trace, and foreach semantics
+
+## Type validation stance
+
+Current core validation is intentionally shallow. Today it mainly enforces:
+
+- object payloads are dict-like
+- required keys exist
+
+It does not yet fully enforce:
+
+- scalar field types
+- nested object structure
+- array item types
+- enums/literals
+
+Near-term recommendation:
+
+- keep `SchemaRef` as the portable contract/export shape
+- keep using `pydantic.BaseModel` as the strongest validation layer for authored nodes
+- gradually strengthen core schema validation where it pays off
+
+This is especially useful for MCP wrapping, because the proxy layer can often
+normalize a tool contract into Pydantic models before the workflow runtime sees
+it.
+
 ## Future extension
 
 ### `Workflow -> NodeSpec`

pyproject.toml

@@ -5,14 +5,14 @@ description = "Add your description here"
 readme = "README.md"
 requires-python = ">=3.14"
 dependencies = [
-    "mcp[cli,rich]>=1.27.0",
+    "mcp[cli,rich]>=1",
-    "pydantic>=2.13.3",
+    "pydantic>=2",
 ]
 
 [dependency-groups]
 dev = [
-    "pytest>=8.4.0",
+    "pytest>=8",
 ]
 
-[tool.pytest.ini_options]
+# [tool.pytest.ini_options]
-addopts = "-p no:cacheprovider"
+# addopts = "-p no:cacheprovider"

scratchpad.md

@@ -1,88 +1,17 @@
 # Graph as struct? workflow as struct
 
-One graph is one input output
+This file is now mostly a working design note.
 
-i decide to make edge dumb as shit. Nodes do all the work
+The original top section was rough brainstorming around a few ideas that did end
+up surviving:
 
-struct needs to define in, out and ALL states. any output from a node is either writing to state or writing to a key of it.
+- edges should stay dumb
+- nodes should own business outcomes
+- graph use-sites should explicitly map input and output
+- workflow state needs declared merge behavior
+- a workflow/subgraph should eventually be reusable like a function
 
-metadata for states to hide in tracing
+The cleaned spec starts below.
-
-whatever language works
-
-## plan
-
-dumbass stupid plan fuhh ahh plan
-
-```text
-.in -> pydantic.create_model
-.out -> pydantic.create_model
-.state // langgraph partial update style
-.steps[] // either nodes only or all everything (all everything seems cooler I LIKE ENUMS but it could be stupih)
-.steps[]:
-  .type IN node, edge
-  node:
-    .tool // ours, so do ts well
-    .args
-    .desc?
-    .bind? // either update .state or .state[.bind]
-  edge:
-    .from
-    .to
-```
-
-### next node
-
-GENUINELY HOW do i convey next node?
-do i pass next node in arg? args.next.{this cond, that cond}
-
-function do need to know about where its run. Who called. So it can decide where it goes next. how? idk maybe a context struct that has the full next list from .edges[]?
-
-or can it just reference any next node? but how? who tells it what node?
-
-function 100% can select what node it wants to go next if supported. Since this is MY stuff, a function I make can do that.
-
-### registry
-
-past graph can turn into a function. so a graph is lowk same thing as a fn.
-
-### input
-
-how in the FUCK do i convey input!
-
-this is how:
-
-since input is state, this function kinda doesnt care about args. Args comes from state.
-
-a function cares about certain keys (args) looking this shape, the least i can do is a remap on node {"specialized graph state key": "generic fn input key"}
-
-same as output {"generic fn answer key": "specialized graph state key"}
-
-### this is lowk constricted
-
-But the counterpart is freeform code...
-
-### more problems
-
-batch? foreach?
-
-control flow.
-
-maybe special nodes that does control flow. Like ensure: something that only after all nodes connected to it ran does it run the nodes it connects to?
-
-if bool = exist, maybe truth table node... if .state.key1 and .state.key2
-
-## Alternative plan! god damn
-
-that was looking like langgraphs Graph API, we could have Functional API being a taddddd simpler: functions imported and run in a sealed ahh box.
-
-functional api is actually peam. its like async.
-
-## inspo
-
-my little use of langgraph
-my nonexistent knowledge of openai fn schema
-claude
 
 ---
 

uv.lock

@@ -264,12 +264,12 @@ dev = [
 
 [package.metadata]
 requires-dist = [
-    { name = "mcp", extras = ["cli", "rich"], specifier = ">=1.27.0" },
+    { name = "mcp", extras = ["cli", "rich"], specifier = ">=1" },
-    { name = "pydantic", specifier = ">=2.13.3" },
+    { name = "pydantic", specifier = ">=2" },
 ]
 
 [package.metadata.requires-dev]
-dev = [{ name = "pytest", specifier = ">=8.4.0" }]
+dev = [{ name = "pytest", specifier = ">=8" }]
 
 [[package]]
 name = "markdown-it-py"

wf_authoring/conditions.py

@@ -75,10 +75,10 @@ class PathExpr:
     def lt(self, other: object) -> Expr:
         return self._binary("lt", other)
 
-    def __eq__(self, other: object) -> Expr:  # ty: ignore[invalid-method-override]
+    def __eq__(self, other: object) -> Expr:  # type: ignore[override] # ty: ignore[invalid-method-override]
         return self._binary("eq", other)
 
-    def __ne__(self, other: object) -> Expr:  # ty: ignore[invalid-method-override]
+    def __ne__(self, other: object) -> Expr:  # type: ignore[override] # ty: ignore[invalid-method-override]
         return self._binary("ne", other)
 
     def __gt__(self, other: object) -> Expr: