Enum NodeKind 

pub enum NodeKind {
    Model {
        model_ref: String,
        prompt_ref: String,
        output_schema: String,
        system_prompt: Option<String>,
        tools: Vec<Value>,
    },
    Tool {
        tool_ref: String,
        input_mapping: HashMap<String, String>,
        output_schema: String,
    },
    PythonFn {
        module: String,
        function: String,
        output_schema: String,
        agent_tool_dispatch: bool,
    },
    JavaFn {
        class_name: String,
        method: String,
        output_schema: String,
        agent_tool_dispatch: bool,
    },
    Condition {
        branches: Vec<ConditionalBranch>,
    },
    Parallel {
        branches: Vec<NodeId>,
    },
    Join {
        wait_for: Vec<NodeId>,
        merge_strategy: MergeStrategy,
    },
    HumanApproval {
        description: String,
        timeout_secs: Option<u64>,
        fallback_node: Option<NodeId>,
    },
    Wait {
        condition: WaitCondition,
        correlation_key: Option<String>,
        timeout_secs: Option<u64>,
    },
    Subgraph {
        workflow_ref: String,
        workflow_version: Option<String>,
        input_mapping: HashMap<String, String>,
        output_mapping: HashMap<String, String>,
    },
    MemoryRetrieval {
        connector_ref: String,
        query_expr: String,
        output_schema: String,
    },
    Policy {
        policy_ref: String,
        on_violation: ViolationAction,
    },
    Finalizer {
        tool_ref: String,
        run_on: FinalizerTrigger,
    },
    Agent {
        agent_ref: String,
        input_mapping: HashMap<String, String>,
        output_schema: String,
    },
    McpTool {
        server: String,
        tool: String,
        input_mapping: HashMap<String, String>,
        output_schema: String,
    },
    A2aTask {
        remote_agent: String,
        skill: String,
        input_mapping: HashMap<String, String>,
        output_schema: String,
        stream: bool,
        on_input_required: Option<NodeId>,
        timeout_secs: Option<u64>,
    },
    AgentDiscovery {
        skill: String,
        protocol: Option<String>,
        output_binding: String,
    },
    Coordinator {
        task: String,
        required_skills: Vec<String>,
        preferred_skills: Vec<String>,
        trust_domain: Option<String>,
        budget: Option<CoordinatorBudget>,
        tiebreaker: Option<TiebreakerConfig>,
        strategy: String,
        weights: DimensionWeights,
        input_mapping: HashMap<String, String>,
        output_key: String,
    },
    AgentTool {
        agent: AgentTarget,
        mode: AgentToolMode,
        input_mapping: HashMap<String, String>,
        output_key: String,
        timeout_ms: Option<u64>,
        budget: Option<AgentToolBudget>,
    },
    Eval {
        scorers: Vec<EvalScorer>,
        on_fail: EvalOnFail,
        max_retries: u32,
        input_expr: Option<String>,
    },
    LimitExceeded,
}

All node kinds supported by the JamJet runtime.

Variants

Model

LLM call with a prompt and structured output.

Fields

model_ref: String

prompt_ref: String

output_schema: String

system_prompt: Option<String>

tools: Vec<Value>

OpenAI-format tool/function schemas offered to the model for this call. Empty means no tools are offered (standard text completion).

Tool

Python function, HTTP endpoint, or gRPC tool.

Fields

tool_ref: String

input_mapping: HashMap<String, String>

output_schema: String

PythonFn

Arbitrary Python function executed by a Python worker.

Fields

module: String

function: String

output_schema: String

agent_tool_dispatch: bool

True when this node is an ADK agent tool-dispatch node, i.e. it runs a whole turn’s model-chosen tool calls in one shot.

Policy for these nodes cannot be derived from the static node kind: the tool names only exist in runtime state. The worker instead evaluates the frozen pending calls carried in the work-item payload. #[serde(default)] so IRs compiled before this field existed deserialize to false and behave exactly as they did before.

JavaFn

Arbitrary Java method executed by an external Java tool-worker.

The Java analog of NodeKind::PythonFn: a class_name + method pair the worker reflects/dispatches. Routes to the java_tool queue, which a net-new Java worker drains over the same HTTP claim/complete API the Python python_tool worker uses — durable, exactly-once.

Fields

class_name: String

method: String

output_schema: String

agent_tool_dispatch: bool

See NodeKind::PythonFn::agent_tool_dispatch. The Java tool-worker receives identical payload enrichment, so it has the identical enforcement gap and needs the identical marker.

Condition

Router — evaluates expressions and branches.

Fields

branches: Vec<ConditionalBranch>

Parallel

Fan-out to multiple branches concurrently.

Fields

branches: Vec<NodeId>

Join

Waits for all parallel branches to complete.

Fields

wait_for: Vec<NodeId>

merge_strategy: MergeStrategy

HumanApproval

Pauses workflow for human decision.

Fields

description: String

timeout_secs: Option<u64>

fallback_node: Option<NodeId>

Wait

Suspends until a timer fires or external event arrives.

Fields

condition: WaitCondition

correlation_key: Option<String>

timeout_secs: Option<u64>

Subgraph

Executes a child workflow.

Fields

workflow_ref: String

workflow_version: Option<String>

input_mapping: HashMap<String, String>

output_mapping: HashMap<String, String>

MemoryRetrieval

Retrieves context from a memory/retrieval connector.

Fields

connector_ref: String

query_expr: String

output_schema: String

Policy

Evaluates policy rules; can block or branch on violation.

Fields

policy_ref: String

on_violation: ViolationAction

Finalizer

Side-effect node (notifications, writes).

Fields

tool_ref: String

run_on: FinalizerTrigger

Agent

Delegates to a local JamJet agent.

Fields

agent_ref: String

input_mapping: HashMap<String, String>

output_schema: String

McpTool

Invokes a tool from an external MCP server.

Fields

server: String

tool: String

input_mapping: HashMap<String, String>

output_schema: String

A2aTask

Delegates a task to an external A2A agent.

Fields

remote_agent: String

skill: String

input_mapping: HashMap<String, String>

output_schema: String

stream: bool

on_input_required: Option<NodeId>

timeout_secs: Option<u64>

AgentDiscovery

👎Deprecated:

Use Coordinator node instead

Dynamically discovers and selects an agent at runtime.

Fields

skill: String

👎Deprecated:

Use Coordinator node instead

protocol: Option<String>

👎Deprecated:

Use Coordinator node instead

output_binding: String

👎Deprecated:

Use Coordinator node instead

Coordinator

Dynamic agent routing with structured scoring + LLM tiebreaker. Supersedes AgentDiscovery.

Fields

task: String

required_skills: Vec<String>

preferred_skills: Vec<String>

trust_domain: Option<String>

budget: Option<CoordinatorBudget>

tiebreaker: Option<TiebreakerConfig>

strategy: String

weights: DimensionWeights

input_mapping: HashMap<String, String>

output_key: String

AgentTool

Invoke a registered agent as a callable tool.

Fields

agent: AgentTarget

mode: AgentToolMode

input_mapping: HashMap<String, String>

output_key: String

timeout_ms: Option<u64>

budget: Option<AgentToolBudget>

Eval

Evaluates the preceding node’s output using configurable scorers.

Supports LLM-judge, deterministic assertions, latency/cost thresholds, and custom Python scorer plugins.

Fields

scorers: Vec<EvalScorer>

Ordered list of scorer configurations.

on_fail: EvalOnFail

Action on overall failure (any scorer below threshold).

max_retries: u32

Maximum retry attempts before propagating failure.

input_expr: Option<String>

Input expression — which state field to evaluate (default: last node output).

LimitExceeded

Terminal node emitted by strategy compilers when an iteration or cost limit is reached. The runtime records the workflow as LimitExceeded and stops further execution. No fields are required — it is purely a marker that carries optional descriptive metadata in the node’s description / labels.

Implementations

impl NodeKind

pub fn queue_type(&self) -> QueueType

Returns the queue type this node should be dispatched to.

pub fn is_durable(&self) -> bool

Returns true if this node requires durable tracking across crashes.

Trait Implementations

impl Clone for NodeKind

fn clone(&self) -> NodeKind

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for NodeKind

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl<'de> Deserialize<'de> for NodeKind

fn deserialize<__D>(__deserializer: __D) -> Result<Self, __D::Error>

where __D: Deserializer<'de>,

Deserialize this value from the given Serde deserializer.

impl Serialize for NodeKind

fn serialize<__S>(&self, __serializer: __S) -> Result<__S::Ok, __S::Error>

where __S: Serializer,

Serialize this value into the given Serde serializer.