YAML Reference

File Structure

A SirenSpec workflow is a YAML file with a top-level mapping. Three fields are required; the rest are optional.

version: "0.1"         # required
agents: {}             # required
nodes: {}              # required
edges: []              # optional
input: {}              # optional
state: {}              # optional
guardrails: []         # optional
budget: {}             # optional
defaults: {}           # optional
env_file: ""           # optional

`version`

Required. The schema version string. Currently only "0.1" is supported.

version: "0.1"

`agents`

Required. A mapping of agent IDs to agent definitions. Each agent wraps an LLM with a system prompt.

agents:
  assistant:
    model: "openai:gpt-4o-mini"
    system: "You are a helpful assistant."
    guardrails: ["injection", "length"]   # optional

Field	Required	Type	Description
`model`	Yes	string	Provider URI — `provider:model` format.
`system`	Yes	string	System prompt sent to the model.
`guardrails`	No	list of strings or `{name, config}` objects	Agent-level guardrail override. Replaces the workflow-level list. Configurable guardrails (`schema`, `cost_cap`, `pii`) use the `{name, config}` form — see Guardrails.

See Agents for full agent documentation and Providers for valid model URIs.

`nodes`

Required. A mapping of node IDs to node definitions. Six node types are supported: Agent node (default) — binds an LLM agent to an output path:

nodes:
  answer:
    agent: assistant
    writes: output.reply

Tool node — invokes an HTTP endpoint or Python callable:

nodes:
  fetch_diff:
    type: tool
    tool: http
    config:
      url: "https://api.example.com/data"
      method: GET
    output_key: data

Swrm node — runs multiple agents in parallel with optional synthesis:

nodes:
  analyze:
    type: swrm
    agents:
      - id: sentiment
        provider: openai
        model: gpt-4o-mini
        prompt: "Analyze sentiment..."

Factory node — spawns instances (single agents or full swrms) for each item in a list:

nodes:
  process_items:
    type: factory
    agent: worker
    for_each: "{{ inputs.items }}"
    inputs:
      item: "{{ item }}"
    writes: output.results

Or with inline swrm (one swrm per item):

nodes:
  grade_papers:
    type: factory
    swrm:
      agents:
        - id: editor
          provider: openai
          model: gpt-4o-mini
          prompt: "Review: {{ item }}"
        - id: grader
          provider: anthropic
          model: claude-haiku-4-5-20251001
          prompt: "Grade: {{ item }}"
      synthesis:
        provider: anthropic
        prompt: "Final grade for paper {{ index }}: ..."
    for_each: "{{ inputs.papers }}"
    writes: output.grades

Workflow node — executes a sub-workflow inline:

nodes:
  run_child:
    type: workflow
    ref: ./child.yaml
    inputs:
      topic: "{{ extract.output }}"

Human node — pauses execution to collect input from a human operator:

nodes:
  approve_draft:
    type: human
    prompt: |
      {{ draft.output }}

      Approve this draft? (yes/edit/reject)
    writes: working.approval
    timeout: 3600
    on_timeout: use_default
    default_output: "no"

Agent node fields

Field	Required	Type	Description
`agent`	Yes	string	Agent ID from the `agents` map.
`writes`	Yes	string	Dot-notation context path where the agent’s response is stored.
`streaming`	No	boolean	When `true` (default), use token streaming if the provider supports it. Set `false` to opt this node out — `--no-stream` on the CLI is the global override. Guardrails always apply to the fully assembled response, not per-chunk.
`retry`	No	object	Per-node retry policy. Overrides `defaults.retry`. See Retry Policies.
`on_failure`	No	object	Failure action when retries are exhausted. Overrides `defaults.on_failure`. See Retry Policies.
`max_tokens_per_call`	No	integer	Optional ceiling on completion tokens for a single call. Forwarded to the provider so the LLM truncates its own response.

Human node fields

A human node pauses execution to collect a response from a human operator. The node consumes no LLM tokens. The rendered prompt is shown via the configured input source (stdin by default), and the response is written to the workflow context just like an agent’s output. Downstream when: conditions can gate on the response.

Field	Required	Type	Description
`type`	Yes	`"human"`	Discriminator that marks this as a human node.
`writes`	Yes	string	Dot-notation context path where the response is stored.
`prompt`	No	string	Template string shown to the operator. Supports `{{ expr }}` interpolation.
`timeout`	No	number	Wall-clock seconds before `on_timeout` fires. Omitted means wait indefinitely.
`on_timeout`	No	`"abort"` \| `"skip"` \| `"use_default"`	Action when the timeout fires. Default `"abort"` raises `HumanInputError`. `"skip"` writes the empty string. `"use_default"` writes `default_output`.
`default_output`	Required when `on_timeout` is `"use_default"`	string	Fallback response used on timeout.

Factory node fields

A factory node supports three mutually exclusive execution modes: Mode 1: Agent + for_each — one agent call per item in a runtime list (native list, JSON array, or fenced ```json).

nodes:
  execute:
    type: factory
    agent: worker_agent
    for_each: "{{ plan.output }}"
    inputs:
      task: "{{ item }}"
    concurrency: 4
    writes: working.execute.outputs

Mode 2: Agent + swarm_size — N identical agent calls on the same input.

nodes:
  execute:
    type: factory
    agent: worker_agent
    swarm_size: 5
    inputs:
      position: "{{ index }} of {{ total }}"
    concurrency: 5
    writes: working.execute.outputs

Mode 3: Swrm + for_each — one full swrm (parallel specialist agents with optional synthesis) per item in a runtime list (native list, JSON array, or fenced ```json).

nodes:
  grade_papers:
    type: factory
    swrm:
      agents:
        - id: editor
          provider: openai
          model: gpt-4o-mini
          prompt: "Review: {{ item }}"
        - id: grader
          provider: anthropic
          model: claude-haiku-4-5-20251001
          prompt: "Grade: {{ item }}"
      synthesis:
        provider: anthropic
        prompt: "Final grade for paper {{ index }}: ..."
    for_each: "{{ inputs.papers }}"
    concurrency: 3
    writes: working.grades

Field	Required	Type	Description
`agent`	One of `agent` or `swrm`	string	Named agent from the workflow’s top-level agents map. Mutually exclusive with `swrm`.
`swrm`	One of `agent` or `swrm`	object	Inline swrm spec (with `agents`, optional `synthesis`, optional `concurrency`). Executed per item. Mutually exclusive with `agent`.
`for_each`	One of `for_each` or `swarm_size`	string	Template expression resolved to a list at runtime. Accepts a native Python list, a plain JSON array string, or a fenced ```json block — no manual unwrapping of upstream tool or agent output needed.
`swarm_size`	One of `for_each` or `swarm_size`	int or string	Static count or template expression for parallel agent instances. Only valid with `agent` mode.
`inputs`	No	object	Template strings for each input. Supports `{{ item }}`, `{{ index }}`, and `{{ total }}`. Each resolved value is also exposed to the spawned agent’s prompt as `{{ inputs.<key> }}`, in addition to being joined into the user message.
`concurrency`	No	integer	Max parallel worker instances. Default: `1`.
`timeout_per_instance`	No	integer	Per-instance timeout in seconds. Default: `60`.
`on_failure`	No	`"abort"` \| `"continue"`	Failure policy: `abort` (raise) or `continue` (skip). Default: `abort`.
`writes`	Yes	string	Dot-notation context path where the outputs list is stored.

Loop Variables

Inside inputs: templates, agent prompts, and swrm agent/synthesis prompts (when using swrm mode), three special variables are available:

Variable	Availability	Type	Description
`{{ item }}`	`for_each` mode only	any	Current list element.
`{{ index }}`	All modes	int	Zero-based position in the list.
`{{ total }}`	All modes	int	Total number of items in the list.

Each key under inputs: is additionally available in the spawned agent’s system prompt as {{ inputs.<key> }} once resolved. For example, inputs: { task: "{{ item }}" } makes {{ inputs.task }} reference the current item inside the agent’s prompt.

Swrm Within Factory

The swrm field in factory nodes accepts:

Field	Required	Type	Description
`agents`	Yes	list of agents	One or more agents to run in parallel for each factory item.
`synthesis`	No	synthesis object	Optional synthesis step run after all agents complete per item.
`concurrency`	No	integer	Max agents to run concurrently per item. Default: all agents.

Agent prompts and synthesis prompts support all standard interpolation namespaces plus the three loop variables ({{ item }}, {{ index }}, {{ total }}). See the following pages for detailed field reference and examples:

Tool Nodes — HTTP and Python adapters
Swrm — Parallel agents and synthesis
Workflow Nodes — Sub-workflow composition

Context paths

The writes field uses dot-notation to specify where output is written in the workflow context:

Path prefix	Description
`output.*`	Final output — included in the trace’s `output` field.
`working.*`	Intermediate state — readable by downstream nodes but not in final output.

Examples:

output.reply — final response available in the JSON trace.
working.intent — intermediate classification readable by the next node.
working.triage.intent — nested intermediate state.

`edges`

Optional. A list of directed edges connecting nodes. If omitted, all nodes are treated as roots and execute in definition order.

edges:
  - from: classify
    to: reply

  - from: triage
    to: handle_refund
    when: working.triage.intent == "refund"

  - from: triage
    to: handle_general
    when: working.triage.intent == "general"

Field	Required	Type	Description
`from`	Yes	string	Source node ID.
`to`	Yes	string	Destination node ID.
`when`	No	string	Python expression evaluated after the source node completes.

`when` expressions

The when field enables conditional branching. The expression is evaluated after the source node writes its output to the context. Available names in when expressions:

Name	Type	Description
`working`	object	The current `working` context (dot-access).
`output`	object	The current `output` context (dot-access).
`_budget`	object	Budget state from `cost_cap` guardrail (if active). Fields: `total_tokens` (int), `estimated_usd` (float or None).
`true` / `false` / `null`	literals	YAML boolean/null literals.
`len` / `bool` / `str` / `int` / `float` / `abs` / `min` / `max`	builtins	A small set of safe built-ins, e.g. `len(working.items) > 0`.

No imports are available and __builtins__ is otherwise cleared, so arbitrary code execution is not possible — only the names above are in scope.

# Activate handle_refund only if the triage agent classified the intent as "refund"
edges:
  - from: triage
    to: handle_refund
    when: working.triage.intent == "refund"

# Skip expensive node if cost_cap guardrail has exhausted budget
edges:
  - from: initial
    to: expensive_processing
    when: _budget.estimated_usd < 2.0

If a when expression raises an error (missing key, syntax error, type mismatch), it is treated as false and the edge is not traversed. Edges without when are always traversed.

`input`

Optional. A static input message for the first node. Can be overridden at runtime with the --input CLI flag.

input:
  message: "What is the capital of France?"

Field	Required	Type	Description
`message`	No	string	Static user message passed to the first (root) node.

If neither input.message nor --input is provided, the CLI exits with an error.

`state`

Optional. Initial state to seed the workflow context before execution begins.

state:
  working:
    seed: "initial value"
  output:
    default_reply: "No answer yet."

State is merged into the corresponding working and output context buckets at startup. Nodes can read and overwrite these values during execution.

`env_file`

Optional. Path to a .env file (relative to the workflow file) loaded into os.environ before execution. This lets provider clients and {{ env.* }} templates pick up credentials without exporting them manually.

env_file: ".env"

The file is loaded eagerly inside load_workflow() — before the Workflow object is returned — so provider clients that read os.environ at construction time see the values regardless of call order. Existing environment variables are never overwritten: if a key in the .env file is already present in os.environ as an empty string, an EnvFileShadowWarning is emitted (the existing empty value wins, which is usually a misconfiguration). A missing env_file raises FileNotFoundError.

`defaults`

Optional. Workflow-wide defaults for retry and failure handling, inherited by all nodes that do not specify their own.

defaults:
  retry:
    max_attempts: 3
    backoff: exponential
    base_delay: 1.0
    on: [429, network_error]
  on_failure:
    action: abort

See Retry Policies for the full field reference.

`guardrails`

Optional. A list of guardrail names (or guardrail specs with configuration) applied globally to all agents. Defaults to ["injection"] if omitted.

Simple guardrails (no configuration)

guardrails:
  - injection
  - length

Guardrails with configuration

Some guardrails require configuration. Use the name and config fields:

guardrails:
  - name: schema
    config:
      schema:
        type: "object"
        properties:
          intent:
            type: "string"
        required: ["intent"]

Guardrail	Description	Configuration
`injection`	Detects prompt-injection patterns. Applied by default.	None
`length`	Truncates output to 4000 characters.	Optional: `max_chars`, `mode`
`pii`	Detects and redacts (or blocks on) email, phone, SSN, and credit-card data.	Optional: `entities`, `action`, `replacement`
`schema`	Validates JSON output against a JSON Schema Draft 7 dict.	Required: `schema`
`cost_cap`	Enforces token and/or USD budget ceilings.	Required: at least one of `max_usd` or `max_tokens`; optional: `action`

An empty list ([]) disables all guardrails for the entire workflow. Individual agents can override this with their own guardrails field. See Guardrails for full details.

`budget`

Optional. Workflow-level cumulative budget for token, USD, and wall-clock spend. The executor checks the running totals after every node and enforces the ceilings declared here. At least one of max_tokens, max_cost_usd, or max_duration_s must be set — an empty budget: block is rejected at validation time.

budget:
  max_tokens: 50000        # total tokens across all nodes
  max_cost_usd: 5.00       # estimated USD ceiling for the whole run
  max_duration_s: 300      # wall-clock cap for the whole run
  on_exceeded: abort       # abort | warn | skip_remaining

Field	Required	Type	Description
`max_tokens`	One of the three	integer	Maximum total tokens across all nodes in a run.
`max_cost_usd`	One of the three	number	Maximum estimated USD spend across all nodes. Falls back to None for models without pricing entries (e.g. Ollama).
`max_duration_s`	One of the three	number	Maximum wall-clock seconds for the full run.
`on_exceeded`	No	`"abort"` \| `"warn"` \| `"skip_remaining"`	Action when any ceiling is hit. Default `"abort"` raises `BudgetExceededError`. `"warn"` logs and continues. `"skip_remaining"` finishes without new LLM calls.

Budget status in the trace

When a budget: block is configured, the trace summary includes a budget block:

{
  "summary": {
    "total_tokens": 1832,
    "budget": {
      "max_tokens": 50000,
      "max_cost_usd": 5.0,
      "max_duration_s": 300,
      "on_exceeded": "abort",
      "tokens_used": 1832,
      "estimated_usd": 0.0021,
      "duration_s": 3.514,
      "exceeded": false,
      "violations": [],
      "skipped_remaining": false
    }
  }
}

Per-node `max_tokens_per_call`

Agent nodes also support a per-call ceiling:

nodes:
  research:
    agent: researcher
    writes: working.research
    max_tokens_per_call: 500

max_tokens_per_call is forwarded to the provider as the max_tokens API parameter so the LLM truncates its own response. Combined with the workflow budget: block, this gives you two layers of protection: each individual call is bounded and the cumulative spend is bounded.

Complete Example

version: "0.1"

agents:
  triage_agent:
    model: "openai:gpt-4o-mini"
    system: |
      Classify the user's message as either a refund request or a general enquiry.
      Reply with ONLY one word — either "refund" or "general" — and nothing else.

  refund_handler:
    model: "openai:gpt-4o-mini"
    system: |
      You are a customer-support specialist handling refund requests.
      Acknowledge the request warmly and outline the refund process in two or three sentences.

  general_handler:
    model: "openai:gpt-4o-mini"
    system: |
      You are a helpful customer-support agent.
      Answer the user's question clearly and concisely in two or three sentences.

nodes:
  triage:
    agent: triage_agent
    writes: working.triage.intent

  handle_refund:
    agent: refund_handler
    writes: output.reply

  handle_general:
    agent: general_handler
    writes: output.reply

edges:
  - from: triage
    to: handle_refund
    when: working.triage.intent == "refund"

  - from: triage
    to: handle_general
    when: working.triage.intent == "general"

guardrails:
  - injection
  - length

Getting Started

Core Concepts

Reference

Resources

File Structure

`version`

`agents`

`nodes`

Agent node fields

Human node fields

Factory node fields

Loop Variables

Swrm Within Factory

Context paths

`edges`

`when` expressions

`input`

`state`

`env_file`

`defaults`

`guardrails`

Simple guardrails (no configuration)

Guardrails with configuration

`budget`

Budget status in the trace

Per-node `max_tokens_per_call`

Complete Example

​File Structure

​version

​agents

​nodes

​Agent node fields

​Human node fields

​Factory node fields

​Loop Variables

​Swrm Within Factory

​Context paths

​edges

​when expressions

​input

​state

​env_file

​defaults

​guardrails

​Simple guardrails (no configuration)

​Guardrails with configuration

​budget

​Budget status in the trace

​Per-node max_tokens_per_call

​Complete Example

File Structure

`version`

`agents`

`nodes`

Agent node fields

Human node fields

Factory node fields

Loop Variables

Swrm Within Factory

Context paths

`edges`

`when` expressions

`input`

`state`

`env_file`

`defaults`

`guardrails`

Simple guardrails (no configuration)

Guardrails with configuration

`budget`

Budget status in the trace

Per-node `max_tokens_per_call`

Complete Example