Imported from kevinpbuckley/VibeUE (Content/Skills/blueprint-graphs/SKILL.md). Install upstream with npx skills add kevinpbuckley/VibeUE --skill blueprint-graphs. Copyright stays with the author.

🧠 Brains complement: IF an unreal-engine-skills-manager tool (external MCP) exists in this session, call it with {action: "load", skill: "blueprint-fundamentals"} for UE domain knowledge on this topic — correct APIs, architecture, best practices — and treat it as the rubric for any review / "best practices" question. If no such tool is available (e.g. running under Claude Code or Codex without that MCP), skip this line entirely and proceed with this skill alone — do NOT attempt the call.

Blueprint Graph Editing Skill

Also load blueprints skill when creating new blueprints, adding variables/components, or overriding functions. This skill covers node-level graph editing — adding nodes, wiring pins, setting values, and layout.

Critical Rules

⚠️ Nothing works during Play-In-Editor — and it fails SILENTLY

Every BlueprintService method resolves its target through LoadBlueprint(), which fails while PIE is running. You get no error. You get an empty array, False, or an empty-string node id — which reads exactly like "that node type doesn't exist" or "that spawner key is wrong". It is very easy to spend a long time debugging a correct call that was only ever blocked by PIE.

les = unreal.get_editor_subsystem(unreal.LevelEditorSubsystem)
if les.is_in_play_in_editor():
    raise RuntimeError("Stop PIE first - BlueprintService returns empty results during play")

Check this FIRST when any call returns nothing unexpectedly, before doubting a spawner key, a search term, or a graph name. The tell: unreal.EditorAssetLibrary.does_asset_exist(bp_path) also returns False for an asset you know exists. Asset writes during PIE are unsafe anyway — stop PIE, then re-run.

⚠️ Method Name Gotchas

⚠️ Engine BlueprintTools call_tool args are {refPath} objects, not strings. Every arguments={"blueprint": path, ...} / {"blueprint": bp_path} below is shorthand — the engine toolset needs {"refPath": "/Game/Dir/BP.BP"} (full object path, .AssetName suffix required) and returns {"returnValue": null} on success. See the blueprints skill for the bp_ref_of() helper. VibeUE's own unreal.BlueprintService.* methods are unaffected — they take plain string paths.

Node creation moved to the engine toolset. The old per-node convenience creators (add_function_call_node, add_event_node, add_get_variable_node, add_set_variable_node, add_member_get_node, add_validated_get_node) and the basics (create_blueprint, compile_blueprint, add_variable, create_function) are gone from unreal.BlueprintService. Create nodes one of three ways:

1. Batch (preferred for 2+ nodes): unreal.BlueprintService.build_graph(...) — the VibeUE delta. Node type values (function_call, event, variable_get, variable_set, branch, print_string, member_get, validated_get, …) live inside build_graph. See build-graph.md.
2. Single node: the engine BlueprintTools.create_node tool via call_tool (args graph, type_id, pos):

   call_tool(
       tool_name="create_node",
       toolset_name="editor_toolset.toolsets.blueprint.BlueprintTools",
       arguments={"blueprint": bp_path, "graph": "EventGraph",
                  "type_id": "Development|PrintString", "pos": [400, 0]})

   type_id examples: 'Development|PrintString', 'AddEvent|EventBeginPlay', 'AddEvent|Custom|MyEvent', a function 'Utilities|...', a macro 'Utilities|FlowControl|ForLoop', a dispatcher handler 'Default|EventDispatcherOnDamaged'.
3. Editor-style spawner key: create_node_by_key() still works for deterministic creation after discover_nodes().

The surviving wiring/inspection methods (connect_nodes, disconnect_pin, get_connections, get_node_pins, set_node_pin_value, add_custom_event_node, add_macro_instance_node, the timeline + custom-event-input CRUD, build_graph, auto_layout_*, etc.) are the VibeUE focus and remain unchanged. Engine basics — BlueprintTools.create, .compile_blueprint, .add_variable, .add_function_graph — also go through call_tool against that same toolset.

⚠️ disconnect_pin() Signature — 4 Args Only

disconnect_pin breaks all connections from a single named pin. Do not call it like connect_nodes (which takes 6 args):

# CORRECT — 4 args: path, graph, node_id, pin_name
unreal.BlueprintService.disconnect_pin(bp_path, "EventGraph", custom_event_id, "then")

# WRONG — 6 args crashes with "takes at most 4 arguments (6 given)"
# unreal.BlueprintService.disconnect_pin(bp_path, graph, src_id, src_pin, tgt_id, tgt_pin)

To remove a specific edge, disconnect the output pin on the source node (e.g. "then"). Because the other end is the only connection on that exec pin, the single-pin disconnect is equivalent to removing the edge.

⚠️ Property Name Gotchas

⚠️ Read graphs cheaply: get_graph_summary() first, filtered get_nodes_in_graph() second

A full get_nodes_in_graph() dump of a big graph is the most token-expensive read in this toolset (every node ships its full pin list twice). Default reading order:

# 1. Fixed-size overview — payload independent of graph size
s = unreal.BlueprintService.get_graph_summary(bp_path, "EventGraph")  # None if bp/graph not found
print(s.node_count, s.connection_count, s.compile_status)
print(s.entry_points)      # ["Event BeginPlay|<node_id>", ...]
print(s.node_type_counts)  # ["K2Node_CallFunction x12", ...]

# 2. Drill in with filters: max nodes, name/type/id substring, and pins off
nodes = unreal.BlueprintService.get_nodes_in_graph(bp_path, "EventGraph", 25, "SpawnActor", False)
# args 3-5 are optional: max_nodes=0 (all), name_filter="" (all), include_pins=True

Only fall back to the unfiltered full dump when you genuinely need every pin of every node.

⚠️ Component/widget bound events: use create_component_bound_event(), never create_node_by_key

A K2Node_ComponentBoundEvent (e.g. On Clicked (MyButton)) built via create_node_by_key + configure_node compiles clean but never fires at runtime — the delegate binding params (including the generated handler function name) can only be initialized atomically:

node_id = unreal.BlueprintService.create_component_bound_event(
    bp_path, "EventGraph", "MyButton", "OnClicked", 100, 100)

Idempotent: if the bound event already exists, the existing node's id is returned. Works for any component delegate (OnComponentBeginOverlap, widget OnClicked/OnValueChanged, ...).

⚠️ Override events are idempotent

create_node_by_key("EVENT Actor::ReceiveBeginPlay") (and EVENT CUSTOM::Name) returns the existing node's id when the blueprint already has that event — a blueprint can hold each override event only once, and a duplicate would break compilation. Don't treat "node id I didn't just place" as an error; wire into it.

⚠️ discover_nodes() vs get_nodes_in_graph() — Different Object Types

get_nodes_in_graph() returns FBlueprintNodeInfo objects — these have node_title, node_id, pos_x, pos_y, node_type.

discover_nodes() returns FBlueprintNodeTypeInfo objects — these have display_name (NOT node_title), spawner_key, category, tooltip, is_pure, is_latent, keywords.

# get_nodes_in_graph — use node_title
nodes = unreal.BlueprintService.get_nodes_in_graph(bp_path, graph)
for n in nodes:
    print(n.node_title, n.node_id)  # node_title is correct here

# discover_nodes — use display_name (NOT node_title)
matches = unreal.BlueprintService.discover_nodes(bp_path, "Broadcast")
for m in matches:
    print(m.display_name, m.spawner_key)  # display_name, NOT node_title

Also: search discover_nodes by function name, not variable type. To find the Broadcast Delegate node for a StateTreeDelegate variable, search "Broadcast" — NOT "Dispatcher" or "StateTreeDelegate".

👀 Reading the User's Live Graph Selection — get_selected_nodes()

get_nodes_in_graph() returns every node in a graph. To act on just the nodes the user has highlighted in the open Blueprint editor, use get_selected_nodes() instead. Returned objects are the same FBlueprintNodeInfo shape as get_nodes_in_graph() — same node_id, node_title, node_type, pos_x, pos_y, pin_names, and pins fields — so you can pass node.node_id straight into get_node_pins(), set_node_position(), delete_node(), etc.

The Blueprint MUST already be open in the editor. Selection state lives on the Slate panel, not on the asset; closing the editor discards it. Calling get_selected_nodes() for a closed asset returns an empty array (and logs a warning).

# Caller knows the asset
selected = unreal.BlueprintService.get_selected_nodes("/Game/BP_Player")

# Caller does NOT know which asset the user is looking at — empty path
# returns the selection from the first open Blueprint editor that has one.
selected = unreal.BlueprintService.get_selected_nodes("")

for n in selected:
    print(f"selected: {n.node_title} ({n.node_type}) @ ({n.pos_x}, {n.pos_y})")
    pins = unreal.BlueprintService.get_node_pins("/Game/BP_Player", "EventGraph", n.node_id)

Use this when the user says "this node", "the selected node(s)", "what I have highlighted", or "the one I'm looking at". An empty result means nothing is selected in any open Blueprint editor — ask the user to click a node in the graph rather than guessing.

🧭 "What am I looking at?" — get_focused_graph_context()

When the user says "this Blueprint", "the graph I have open", "the current material", or "what I'm looking at" without naming the asset, call get_focused_graph_context() first instead of guessing a path or chaining list_graphs → get_selected_nodes. It returns the asset and the focused graph in one shot — VibeUE's analogue of Epic's GetDockedContext(). It anchors on the globally-active editor tab (falling back to the first open supported editor).

It covers both Blueprint-family editors and the Material editor. Fields:

ctx = unreal.BlueprintService.get_focused_graph_context()
if ctx.found and ctx.graph_kind != "Material":
    # asset_path + graph_name drop straight into the other Blueprint tools
    nodes = unreal.BlueprintService.get_nodes_in_graph(ctx.asset_path, ctx.graph_name)
    sel = [n.node_id for n in ctx.selected_nodes]

Materials: when graph_kind == "Material", the Blueprint graph tools (get_nodes_in_graph, build_graph) do not apply — use MaterialNodeService keyed by ctx.asset_path. selected_nodes is still populated (material expression nodes), but graph_name is the internal material-graph name, not a Blueprint graph tab.

found == False means no Blueprint/Material editor is open — ask the user to open the asset rather than guessing a path.

💬 Comment Boxes — add_comment_node() / add_comment_around_nodes()

Comment boxes are the coloured bubbles that visually group nodes (the editor shortcut is C after selecting nodes). Two APIs:

• add_comment_node(blueprint_path, graph_name, comment_text, pos_x, pos_y, width, height, r, g, b, a) — explicit position and size. Use when you already know exactly where you want the box.
• add_comment_around_nodes(blueprint_path, graph_name, comment_text, node_ids, padding, r, g, b, a) — computes the bounding box of the supplied nodes (plus padding, default 40 px) and pads room for the title bar. This is the programmatic equivalent of select-then-press-C.

✍️ Write informative comment text — this is the whole point

Comment boxes are documentation that lives next to the code. A future reader (human or LLM) should be able to glance at the comment and understand what this group of nodes accomplishes and why, without re‑reading every pin. Treat the comment text the same way you'd treat a function header comment.

Good comments explain intent and behaviour:

• "Pick a random patrol point and cache its world location for the Move To task"
• "On Damage > 0: play hit reaction, flash material, decrement Health"
• "Debounce input — ignore re‑presses within 0.25s of the last fire"
• "Early‑out when the owning controller is not locally controlled (server‑auth path)"

Bad comments restate node titles or are filler:

• "Print String" ❌ (just the node name)
• "Logic" / "Stuff" / "TODO" ❌ (says nothing)
• "Branch + Set Variable" ❌ (mechanical, not intent)
• The user's verbatim request like "add a comment around the nodes" ❌ — that's the task, not the meaning

Workflow: before calling add_comment_around_nodes, inspect the selected nodes (titles, pin connections, variables they read/write) and synthesise a one‑line description of what the group does. If the user gave you a hint ("this is the patrol picker"), build on it; if they didn't, derive it from the nodes themselves. Use multiple lines (\n) when the group has a non‑trivial flow worth narrating.

sel = unreal.BlueprintService.get_selected_nodes("/Game/BP_Player")
ids = [n.node_id for n in sel]

# Derive intent from the selected nodes (event, variables read/written, called
# functions) and write a comment that documents WHY this group exists.
comment = (
    "Pick a random patrol point on EnterState\n"
    "Caches the chosen point's world location into PatrolPointLocation\n"
    "so the subsequent Move To task has a stable target."
)

unreal.BlueprintService.add_comment_around_nodes(
    "/Game/BP_Player", "EventGraph", comment, ids)

Both functions return the new comment node's node_id (a GUID string), or "" on failure. Colour parameters are RGBA in 0–1 range; the default is the standard pale yellow. Unknown node_ids passed to add_comment_around_nodes() are skipped with a warning — if all IDs are invalid, the call returns "".

Standard editor behaviour applies: any K2 nodes whose bounds fall inside the comment box get picked up and dragged with it (GroupMovement mode).

⚠️ Branch Node Pin Names

Use then and else, NOT true/false:

# WRONG
connect_nodes(path, func, branch_id, "true", target_id, "execute")

# CORRECT
connect_nodes(path, func, branch_id, "then", target_id, "execute")
connect_nodes(path, func, branch_id, "else", target_id, "execute")

⚠️ UE5.7 Uses Doubles for Math

The editor still displays "Make Literal Float", but its spawner is FUNC KismetSystemLibrary::MakeLiteralDouble. A build_graph function_call node (or engine create_node) targeting "KismetSystemLibrary", "MakeLiteralFloat" resolves to nothing silently — use MakeLiteralDouble, or discover_nodes(bp, "Make Literal Float") and spawn via the returned spawner_key.

⚠️ Compile Before Using Variables in Nodes

The variable must exist on the compiled skeleton class before a getter/setter can resolve it. Add the variable and compile through the engine BlueprintTools toolset, then create the variable_get node (here via build_graph):

# Engine basics go through call_tool / BlueprintTools
call_tool(tool_name="add_variable",
          toolset_name="editor_toolset.toolsets.blueprint.BlueprintTools",
          arguments={"blueprint": path, "name": "Health", "type": "float", "default": "100.0"})
call_tool(tool_name="compile_blueprint",
          toolset_name="editor_toolset.toolsets.blueprint.BlueprintTools",
          arguments={"blueprint": path})  # REQUIRED before adding nodes

# Now the getter resolves — VibeUE build_graph
unreal.BlueprintService.build_graph(path, func,
    [{"ref": "GetHP", "type": "variable_get", "params": {"variable": "Health"}}],
    [], [], True, True)

⚠️ Verify Pins Immediately After Adding Variable Get/Set Nodes

A variable Get/Set node can occasionally be created with zero pins (the variable property failed to resolve on the skeleton class at spawn time). The create call still returns a GUID, so you won't notice until every connect_nodes against it returns False. After creating a variable node, check its pins before wiring anything (create it via build_graph's variable_get; ref_to_node_id maps your ref back to the GUID):

r = unreal.BlueprintService.build_graph(bp_path, graph,
    [{"ref": "Get", "type": "variable_get", "params": {"variable": "Armor"}}],
    [], [], False, False)
gid = r.ref_to_node_id["Get"]
pins = unreal.BlueprintService.get_node_pins(bp_path, graph, gid)
if not pins:  # corrupt node — recover
    unreal.BlueprintService.delete_node(bp_path, graph, gid)
    call_tool(tool_name="compile_blueprint",
              toolset_name="editor_toolset.toolsets.blueprint.BlueprintTools",
              arguments={"blueprint": bp_path})
    r = unreal.BlueprintService.build_graph(bp_path, graph,
        [{"ref": "Get", "type": "variable_get", "params": {"variable": "Armor"}}],
        [], [], False, False)
    gid = r.ref_to_node_id["Get"]
    assert unreal.BlueprintService.get_node_pins(bp_path, graph, gid)

Recovery notes:

• delete_node can return False here — confirm with a fresh get_nodes_in_graph() and don't leave a stray pin-less duplicate behind (two nodes with the same title, one never connected).
• Compiling between delete and re-add refreshes the skeleton class so the retry resolves.
• This is another reason to create one node at a time instead of batching 4–6 creates in one call.

⚠️ The Result Node's execute Pin MUST Be Wired — the Warning Is Not Minor

Compile warning ReturnNode Exec pin has no connections means the execution chain never reaches the Return Node, so the function's output parameters are never set — callers always get default values. The function is broken even though compile.success is True.

Every non-pure function needs an exec path Entry.then → ... → Result.execute. For a "compute-only" function (e.g. read a variable, compare, return bool), wire the entry directly to the result — connecting only the data pin is not enough:

# CheckDeath: IsDead = (Health <= 0)
connect_nodes(bp, "CheckDeath", entry_id, "then", result_id, "execute")        # exec — REQUIRED
connect_nodes(bp, "CheckDeath", lessequal_id, "ReturnValue", result_id, "IsDead")  # data

Do not dismiss this warning in your summary or claim the function "works". Treat any ReturnNode compile warning as a wiring bug and fix it before reporting success. (If the function has no side effects, the cleaner fix is making it pure, but wiring the exec chain is the minimal correct repair.)

⚠️ Node IDs Are GUID Strings, Not Small Integers

Do not assume Blueprint nodes have numeric IDs like 0 or 1. get_nodes_in_graph() returns GUID strings in node.node_id.

nodes = unreal.BlueprintService.get_nodes_in_graph(bp_path, graph)
for node in nodes:
    print(node.node_id, node.node_title)

For function graphs, find nodes by node_type or node_title. For event-style overrides, find nodes by the display title Unreal shows in the graph, not by the raw override function name.

⚠️ Event-Style Overrides Use "EventGraph" — NOT the Function Name

When an override is event-style (void/latent functions like ReceiveTreeStart, ReceiveTick, ReceiveLatentEnterState, ReceiveStateCompleted), the event node lives inside "EventGraph". Do NOT use the function name as the graph name — that graph doesn't exist.

# WRONG — returns 0 nodes, creates nothing
nodes = unreal.BlueprintService.get_nodes_in_graph(bp_path, "ReceiveTreeStart")
# ...and creating into "ReceiveTreeStart" (build_graph or create_node) goes nowhere

# CORRECT — event-style overrides are in EventGraph
nodes = unreal.BlueprintService.get_nodes_in_graph(bp_path, "EventGraph")
unreal.BlueprintService.build_graph(bp_path, "EventGraph",
    [{"ref": "Call", "type": "function_call",
      "params": {"class": "KismetSystemLibrary", "function": "PrintString"}}],
    [], [], True, True)

Common event-style overrides and their graph name → node title:

Only function-style overrides (non-void, like GetStaticDescription) create a separate function graph.

⚠️ Do Not Guess Blueprint Function Node Names

Blueprint display titles and internal UFunction names are often not the same.

Common examples:

If you do not already know the exact callable name, do not guess and do not batch multiple node creations into one shot. Use one of these two patterns:

1. discover_nodes() + create_node_by_key() for deterministic editor-style creation.
2. A build_graph function_call node (or engine BlueprintTools.create_node) only after discovery, or when you already know the exact function name.

import unreal

bp_path = "/Game/StateTree/Tasks/STT_Chase"
graph = "EventGraph"

matches = unreal.BlueprintService.discover_nodes(bp_path, "Get Actor Location")
actor_get_location = next(
    node for node in matches
    if node.spawner_key.startswith("FUNC AActor::") or node.spawner_key.startswith("FUNC Actor::")
)

node_id = unreal.BlueprintService.create_node_by_key(
    bp_path, graph, actor_get_location.spawner_key, 400, 0)
assert node_id, actor_get_location.spawner_key

If a node-create call returns an empty ID, stop immediately. Re-read the graph, inspect the error output, and fix the lookup before creating anything else.

⚠️ Standard Macro nodes (ForEachLoop, etc.) — do NOT use create_node_by_key

Macro instances (K2Node_MacroInstance) have no spawner key, so discover_nodes() won't find them and create_node_by_key() fails silently (returns empty, no error). Use the dedicated method instead:

import unreal
node_id = unreal.BlueprintService.add_macro_instance_node(bp_path, graph, "ForEachLoop", x, y)

Supported shorthands:

A full AssetPath:GraphName string is also accepted for user-defined macro libraries (e.g. /Game/Macros/ML_X.ML_X:MyMacro).

Notes:

• IsValid places one node exposing both Is Valid and Is Not Valid exec outputs — there is no separate IsNotValid macro.
• MultiGate is K2Node_MultiGate (a distinct node type), not a StandardMacro — it is not in this list.
• To create a macro graph on a MacroLibrary Blueprint, use create_macro_graph(lib_path, "MacroName") (the engine BlueprintTools.add_function_graph path asserts on MacroLibrary BPs). To create the library asset itself from Python (issue #449):

  f = unreal.BlueprintMacroFactory()   # NOT BlueprintFactory
  unreal.AssetToolsHelpers.get_asset_tools().create_asset("ML_X", "/Game/Macros", unreal.Blueprint, f)
  unreal.BlueprintService.create_macro_graph("/Game/Macros/ML_X", "MyMacro")
  # reference it from a graph: add_macro_instance_node(bp, graph, "/Game/Macros/ML_X.ML_X:MyMacro", x, y)

🔎 Calling a function (self / parent / library) — discover the spawner key (#449)

build_graph's function_call works for library statics, but the most reliable way to add a call to a self function, a parent-class function (Character/Pawn/Actor…), or any discoverable function is discover_nodes → create_node_by_key:

# Find the function — self functions show category "Self Functions"; parents show "FUNC Parent::Func"
hits = unreal.BlueprintService.discover_nodes(bp, "Jump")        # search_term (NOT a graph name)
key  = hits[0].spawner_key                                       # e.g. "FUNC Character::Jump"
node = unreal.BlueprintService.create_node_by_key(bp, "EventGraph", key, 400, 200)

discover_nodes(blueprint_path, search_term, category="", max_results=20) surfaces the blueprint's OWN functions (your custom functions, callable on self) plus the full parent hierarchy and library functions. The returned spawner_key ("FUNC <Class>::<Func>") feeds straight into create_node_by_key — this is how you add a self-call node deterministically.

⚠️ Leave category empty unless you are echoing a category you saw in a result. It is a substring match against the node's menu category, which is "Self Functions", "Parent: Actor", or the editor's own menu path — not a topic word. Passing something reasonable-looking like "Math" silently filters out everything and returns zero results, which looks identical to "no such node exists". Search first with category="", then read .category off the hits.

⚠️ Search the exact display name, not a fragment. Results fill in order (self → parents → action database) and stop at max_results, so a short common fragment gets crowded out before the node you want is reached. discover_nodes(bp, "Sin", "", 500) returns 500 rows and none of them is Sin (Radians); discover_nodes(bp, "Sin (Radians)", "", 20) finds it immediately as FUNC KismetMathLibrary::Sin. If a search comes back without an obvious node, make the term more specific rather than raising the cap.

Full node coverage — every action-menu node, not just functions/events

discover_nodes enumerates the entire Blueprint action database, so it also returns template/custom K2 nodes and variable get/set on other classes — e.g. Get Subsystem (Get EnhancedInputLocalPlayerSubsystem, Get GameInstanceSubsystem, …), Cast To nodes, etc. These come back with a faithful SPAWN <NodeClass>|<MenuName> key:

# Get Enhanced Input Local Player Subsystem, fully bound (ReturnValue typed to the subsystem)
hits = unreal.BlueprintService.discover_nodes(bp, "EnhancedInputLocalPlayerSubsystem", "", 100)
key  = next(h.spawner_key for h in hits if h.node_class == "K2Node_GetSubsystem")
nid  = unreal.BlueprintService.create_node_by_key(bp, "EventGraph", key, 0, 0)

create_node_by_key resolves a SPAWN key back to the exact action-database spawner and invokes it — the same path the editor's Add-Node menu uses — so the node is created fully bound (Get Subsystem's CustomClass, a variable's member reference, …), not as a blank node.

Search specifically. Node menu names are spaceless (EnhancedInputLocalPlayerSubsystem, not "Enhanced Input …"), and a broad term like "Subsystem" matches hundreds of function nodes that can crowd out the template nodes under max_results. Search the exact spaceless name (or raise max_results) and filter by node_class (K2Node_GetSubsystem, K2Node_GetSubsystemFromPC, K2Node_GetEngineSubsystem, K2Node_GetEditorSubsystem).

⚠️ Complex Graphs: Create and Verify One Node at a Time

For fragile graph work such as STT_* task Blueprints, timers, delegate workflows, or any graph being built from a screenshot:

1. Create one node.
2. Re-read the graph and confirm the returned GUID exists.
3. Read that node's pins.
4. Only then create the next node.
5. After all required nodes exist, connect one edge at a time and verify the new connection appears in get_connections().

Do not create 4-6 nodes in one tool call when some of them have unresolved function names. That makes cleanup harder and leaves duplicate nodes behind on retry.

⚠️ Recovery After a Failed Graph Attempt

If a graph-edit attempt fails partway through:

1. Call get_nodes_in_graph() and get_connections() to see the true post-failure state.
2. Remove orphaned nodes with delete_node().
3. Remove stale links with disconnect_pin().
4. Retry from a clean baseline.

Do not invent helper names like disconnect_all_pins(). Use the real APIs that exist.

Examples for StateTree task blueprints:

⚠️ Delay vs Set Timer by Event

When a user asks for a "timer" or "timed delay" in a Blueprint, use Set Timer by Event, NOT Delay:

The pattern for Set Timer by Event requires a Custom Event:

Important: after add_custom_event_node(...), immediately re-read the graph and re-find that callback by the returned node_id. Do not assume the title will be exactly the event name you passed in. Unreal can display custom event titles as multi-line labels such as OnTimerFinished + Custom Event.

Do not silently substitute a Create Event / Create Delegate node plus a separate function graph when the user asked for a Custom Event node or when the target graph should visibly contain the callback in EventGraph. A Create Event delegate can compile and still be the wrong implementation for the requested graph shape.

import unreal

bp_path = "/Game/MyBlueprint"
graph = "EventGraph"

# 1. Add Set Timer by Event node (build_graph function_call; ref_to_node_id gives the GUID)
r = unreal.BlueprintService.build_graph(bp_path, graph,
    [{"ref": "Timer", "type": "function_call",
      "params": {"class": "KismetSystemLibrary", "function": "K2_SetTimerDelegate"}}],
    [], [], False, False)
timer_id = r.ref_to_node_id["Timer"]

# 2. Set the Time pin
unreal.BlueprintService.set_node_pin_value(bp_path, graph, timer_id, "Time", "1.0")

# 3. Add a Custom Event node — this is the callback
custom_event_id = unreal.BlueprintService.add_custom_event_node(
    bp_path, graph, "OnTimerFinished", 300, 300)

# 4. Connect the Custom Event's delegate output to the timer's Delegate pin
unreal.BlueprintService.connect_nodes(
    bp_path, graph, custom_event_id, "OutputDelegate",
    timer_id, "Delegate")

# 5. Wire: EnterState → Set Timer by Event
unreal.BlueprintService.connect_nodes(
    bp_path, graph, enter_state_id, "then", timer_id, "execute")

# 6. Wire: Custom Event → Finish Task (or whatever follows)
unreal.BlueprintService.connect_nodes(
    bp_path, graph, custom_event_id, "then", finish_id, "execute")

⚠️ StateTree Task Timer Workflow: Always Discover Real Titles and Pins First

For STT_* Blueprint tasks, do not guess event titles or pin names. The stable pattern is:

1. override_function(bp, "ReceiveLatentEnterState")
2. get_nodes_in_graph(bp, "EventGraph") and locate the event by title Event EnterState
3. After add_custom_event_node(...), call get_nodes_in_graph() again and re-find the callback by the returned GUID
4. Treat the callback title as display-only evidence; custom event titles can be multi-line and should not be your primary lookup key
5. get_node_pins() on every newly created node
6. Wire using the actual pin names returned by the graph

Current UE 5.7 / VibeUE graph details for this workflow:

Success for this workflow also requires the callback node in EventGraph to be a real custom event node, typically K2Node_CustomEvent in a fresh node listing. K2Node_CreateDelegate is a different node type and should not be treated as equivalent when the requested outcome is a visible custom event callback in the graph.

import unreal

bp_path = "/Game/StateTree/STT_Rotate"
graph = "EventGraph"

unreal.BlueprintService.override_function(bp_path, "ReceiveLatentEnterState")

# Timer + Finish Task are function_call nodes — create them via build_graph;
# the Custom Event callback uses the surviving add_custom_event_node.
r = unreal.BlueprintService.build_graph(bp_path, graph,
    [{"ref": "Timer", "type": "function_call",
      "params": {"class": "KismetSystemLibrary", "function": "K2_SetTimerDelegate"}},
     {"ref": "Finish", "type": "function_call",
      "params": {"class": "StateTreeTaskBlueprintBase", "function": "FinishTask"}}],
    [], [], False, False)
timer_id = r.ref_to_node_id["Timer"]
finish_id = r.ref_to_node_id["Finish"]
custom_id = unreal.BlueprintService.add_custom_event_node(
    bp_path, graph, "OnTimerFinished", 0, 420)

nodes = unreal.BlueprintService.get_nodes_in_graph(bp_path, graph)
enter_node = next((n for n in nodes if n.node_title == "Event EnterState"), None)
custom_node = next((n for n in nodes if n.node_id == custom_id), None)
assert enter_node, "Event EnterState not found"
assert custom_node, f"Custom event {custom_id} not found after create"

custom_pins = unreal.BlueprintService.get_node_pins(bp_path, graph, custom_node.node_id)
print("CUSTOM EVENT TITLE:", custom_node.node_title)
print("CUSTOM EVENT PINS:", [p.pin_name for p in custom_pins])

unreal.BlueprintService.set_node_pin_value(bp_path, graph, timer_id, "Time", "1.0")
unreal.BlueprintService.set_node_pin_value(bp_path, graph, timer_id, "bLooping", "false")
unreal.BlueprintService.set_node_pin_value(bp_path, graph, finish_id, "bSucceeded", "true")

assert unreal.BlueprintService.connect_nodes(bp_path, graph, enter_node.node_id, "then", timer_id, "execute")
assert unreal.BlueprintService.connect_nodes(bp_path, graph, custom_node.node_id, "OutputDelegate", timer_id, "Delegate")
assert unreal.BlueprintService.connect_nodes(bp_path, graph, custom_node.node_id, "then", finish_id, "execute")

Use add_create_event_node() only when you need a Create Event / Create Delegate node. For Set Timer by Event, a Custom Event node is the simpler callback source and matches the Blueprint editor workflow shown in screenshots.

If a previous attempt already inserted a Create Event node for this timer callback, treat that as the wrong graph shape when the request is for a custom event. Remove the wrong node, remove any stale callback function graph that only existed to support that delegate node, then create the real Custom Event node and verify it by node ID and node type.

⚠️ Verification Is Mandatory Before Claiming Success

After any graph edit, verify all three layers:

1. Connections: call get_connections() and confirm the exact expected wiring.
2. Pins: if a connection fails, call get_node_pins() and use the real pin names.
3. Compile: inspect the engine BlueprintTools.compile_blueprint result's success, num_errors, and errors.

For any node you claim you created, also re-read the graph with get_nodes_in_graph() and confirm that node actually exists in the graph after the edit. A returned node ID from a create call is not enough.

For Custom Event timer callbacks, verify both of these before wiring:

1. The returned GUID appears in a fresh get_nodes_in_graph() result.
2. get_node_pins() on that exact node shows the pins you intend to use, typically OutputDelegate and then for the callback path.

Also verify that the node type is the expected custom event form rather than K2Node_CreateDelegate.

# compile goes through the engine BlueprintTools toolset
result = call_tool(tool_name="compile_blueprint",
                   toolset_name="editor_toolset.toolsets.blueprint.BlueprintTools",
                   arguments={"blueprint": bp_path})
assert result["success"], result.get("errors")

nodes = unreal.BlueprintService.get_nodes_in_graph(bp_path, graph)
for node in nodes:
    print(f"NODE {node.node_id} {node.node_title}")

connections = unreal.BlueprintService.get_connections(bp_path, graph)
for conn in connections:
    print(f"{conn.source_node_title}.{conn.source_pin_name} -> {conn.target_node_title}.{conn.target_pin_name}")

unreal.EditorAssetLibrary.save_asset(bp_path)

Never print MODIFIED or describe the graph as complete until the expected connections are present and compile succeeds. Never describe a node as present until it appears in a fresh get_nodes_in_graph() result. Never treat compile success alone as proof that the graph matches the requested design.

Required Success Gate For Graph Edits

If you edit a graph, your success check must answer all of these from live asset data:

1. Which required node titles are present?
2. Which exact connection lines prove the intended wiring exists?
3. Did compile succeed with zero relevant errors?

If you cannot answer those three questions from tool output, the task is not complete yet.

For complex graph tasks, also answer these two questions before claiming completion:

4. Which node IDs were created successfully, one by one?
5. What cleanup was done after any failed intermediate attempt?

Common mistake: Using Delay when the user says "timer". Delay is a latent action that pauses the execution chain. Set Timer by Event is non-blocking and fires a separate event — this is critical for State Tree tasks, animation blueprints, and any actor that must keep ticking during the wait period.

Sample scripts (run via execute_python_code)

• scripts/build_event_graph.txt — wire Event BeginPlay → Print String, verify connection, compile.
• scripts/timer_with_custom_event.txt — non-blocking Set Timer by Event + Custom Event callback.

Sub-docs Available

This skill index contains the frontmatter, intro, and the Critical Rules / gotchas section. For everything else, load the matching sub-doc by reading its file under this same folder (Plugins/VibeUE/Content/Skills/blueprint-graphs/):

• node-reference.md — Reference for specific node APIs that are part of Critical Rules but read as reference material: the member_get / validated_get build_graph node types (pin tables), Custom Event Input Pins CRUD (*_custom_event_input), and Timelines CRUD (*_timeline*).
• workflows.md — Step-by-step workflows: overriding a parent function (override_function), creating a function with logic, adding an Enhanced Input Action node.
• node-layout.md — Node layout best practices: layout constants, execution flow, data flow above execution, branch layout, repositioning Entry/Result nodes.
• function-classes.md — Quick reference for the common class names you pass to a build_graph function_call node / engine create_node (KismetMathLibrary, KismetSystemLibrary, KismetArrayLibrary, GameplayStatics, etc.).
• array-operations.md — Array operations on wildcard pins: Array_Random, available array functions, K2Node_GetArrayItem, wildcard pin type propagation, common array mistakes.
• build-graph.md — The batch build_graph API: when to use it, node types, connection format, examples (BeginPlay → PrintString, Branch with Math, StateTreeDelegate Broadcast), round-trip export/rebuild, auto-layout, Make Struct / Make Instanced Struct, error handling.

Additional gotchas

• discover_nodes returns clean FUNC <BP>_C::Fn keys now (no SKEL_ prefix to strip). BP custom events are callable cross-BP through such a key; BP functions cannot call custom events — FUNC Self::<Event> in a function graph returns "".
• Read pin defaults from get_node_pins (it now reports non-string defaults too); get_graph_definition's defaults array and a node's default_object are the cross-checks. Bool pins on VariableSet nodes work (the pin name is the variable name).
• Variable GET/SET spawners are scoped to the owning Blueprint now, so their keys are unambiguous — you no longer disambiguate by n.category (Variables|<OwningBlueprint>).
• build_graph that fails on one node spec leaves the OTHER nodes behind; inspect the graph and delete the strays before retrying.
• An interface function spawns as a plain K2Node_CallFunction with an interface-typed self pin — feed self through a Cast To <Interface> node; some query functions spawn PURE (no exec pins).
• Compile errors surface in compile_blueprint's res.error, and get_editor_property("status") reads BS_ERROR; never save on BS_ERROR, and stop PIE before structural edits (service calls against a BP with live PIE instances fail silently or return None).
• Script-baked BP-CDO TArray<TSubclassOf<...>> defaults can be LOST on editor reload — read them back after a restart to confirm they persisted (writing the CDO is allowed; the engine just drops these array defaults across a reload).
• Editor-world edits issued immediately after StopPIE can silently no-op; let the editor settle (poll) before editing.
• list_graphs/get_graph_summary returning empty or None in a long session is a stale-session artifact, not a per-asset fact — restart before diagnosing. list_graphs legitimately returns [] for Anim Blueprints (use AnimGraphService.list_graphs).