CLI Reference

Installation

The sirenspec CLI is installed automatically when you add the package:

uv add sirenspec
# or
pip install sirenspec

After installation, the sirenspec command is available on your PATH. You can also check the installed version:

sirenspec --version

`sirenspec init`

Scaffold a new workflow interactively. Prompts for a template and provider, then writes a workflow.yaml and .env.example in the target directory. The generated workflow passes sirenspec validate immediately.

sirenspec init [OPTIONS]

Options:

Option	Short	Description
`--output PATH`	`-o`	Directory for the generated files. Defaults to the current working directory.

Exit codes:

Code	Meaning
`0`	Files were created successfully.
`1`	The generated workflow failed validation (rare — indicates a template bug).

Example session:

$ sirenspec init

Welcome to SirenSpec! Let's scaffold a new workflow.

Choose a template:
  1. simple-agent          — Single agent that answers a question
  2. sequential-pipeline   — Two agents in sequence
  ...

Template [1]: 1

Choose a provider:
  1. openai
  2. anthropic

Provider [1]: 1

Enable guardrails? (injection + length) [Y/n]: Y

✓ workflow.yaml created
✓ .env.example created

Next steps:
  Copy .env.example → .env and add your API key
  sirenspec validate workflow.yaml
  sirenspec run workflow.yaml

`sirenspec run`

Execute a workflow and display the results. By default, node execution is rendered with Rich formatting to the terminal (streaming per-node view). Use --trace or --output json to output the full JSON execution trace instead.

sirenspec run <workflow-file> [OPTIONS]

Arguments:

Argument	Description
`workflow-file`	Path to the workflow YAML file.

Options:

Option	Short	Description
`--input TEXT`	`-i`	User input message. Overrides `input.message` in the YAML.
`--trace`		Print full JSON trace to stdout (disables per-node panels).
`--output TEXT`		Output format. Use `"json"` for raw JSON trace (equivalent to `--trace`).
`--trace-file TEXT`		Write full JSON trace to this file path (in addition to streaming output).
`--quiet`		Suppress node panels; print only the summary (useful for non-TTY environments).
`--no-stream`		Disable per-token streaming inside node panels. Panels and summary still render.

Exit codes:

Code	Meaning
`0`	Workflow completed successfully.
`1`	File not found, validation error, execution error, or guardrail violation.

Examples:

# Run with streaming output (default)
sirenspec run workflow.yaml --input "What is the capital of France?"

# Print full JSON trace
sirenspec run workflow.yaml --input "Hello" --trace | jq '.output'
sirenspec run workflow.yaml --input "Hello" --output json | jq '.output'

# Suppress node panels but keep summary
sirenspec run workflow.yaml --input "Hello" --quiet

# Write trace to a file and see streaming output
sirenspec run workflow.yaml --input "Hello" --trace-file trace.json

Streaming Output Format: When run in a TTY without --trace or --output json, the CLI renders each completed node as a Rich panel:

╭─ answer ─────────────────────────────╮
│ Paris is the capital of France.      │
╰───────────────────────────────────────╯
  ↓ output.reply

Run complete  │  1 node  │  24 tokens  │  cost unavailable (pricing not configured)

For swarm nodes (type: swrm), the output shows a dedicated parallel execution block:

─── Swarm: analyze [3 agents] ─────────────────────────
╭─── sentiment ─────────────────────────────────────╮
│ The market sentiment is cautiously optimistic...   │
╰───────────────────────────────────────────────────╯
╭─── risk ─────────────────────────────────────────╮
│ 1. Rising interest rates                          │
│ 2. Geopolitical uncertainty                       │
│ 3. Earnings weakness...                           │
╰───────────────────────────────────────────────────╯
╭─── opportunity ───────────────────────────────────╮
│ - Beaten-down tech sector                         │
│ - Infrastructure momentum...                      │
╰───────────────────────────────────────────────────╯
─── Swarm complete: 3/3 succeeded  (2.8s total) ───
╭─── analyze ───────────────────────────────────────╮
│ Based on the three perspectives, this stock...    │
╰───────────────────────────────────────────────────╯
  ↓ output.analyze

When synthesis is configured, its output is shown in a final panel after the swarm footer. If any agent fails, the footer shows the failure count (e.g., 2/3 succeeded, 1 failed). Failed agents display their error message in a red-bordered panel. JSON Trace Format: When --trace, --output json, or --trace-file is used, the full execution trace is written as JSON:

{
  "workflow": { "version": "0.1" },
  "input": { "message": "What is the capital of France?" },
  "nodes": [
    {
      "id": "answer",
      "type": "agent",
      "output": "Paris is the capital of France.",
      "writes": "output.reply",
      "status": "success",
      "error": null,
      "tokens": 24,
      "usage": {
        "prompt_tokens": 18,
        "completion_tokens": 6,
        "estimated_usd": null
      },
      "guardrails_passed": [
        "InjectionGuardrail.check_input",
        "InjectionGuardrail.check_output"
      ],
      "duration_ms": 312.5
    }
  ],
  "output": {
    "reply": "Paris is the capital of France."
  },
  "summary": {
    "total_tokens": 24,
    "total_usage": {
      "prompt_tokens": 18,
      "completion_tokens": 6,
      "total_tokens": 24,
      "estimated_usd": null
    },
    "total_duration_ms": 312.5,
    "status": "success"
  }
}

Trace fields:

Field	Description
`workflow.version`	Schema version from the YAML file.
`input.message`	The user input that was used.
`nodes`	Ordered list of executed nodes.
`nodes[].id`	Node identifier.
`nodes[].type`	Node type: `"agent"`, `"tool"`, `"swrm"`, `"factory"`, or `"workflow"`.
`nodes[].output`	The node’s output value (string, dict, list, or null).
`nodes[].writes`	Context path where the output was written.
`nodes[].status`	`"success"`, `"skipped"`, or `"failed"`.
`nodes[].skip_reason`	Why a node was skipped: `"branch_not_taken"` (no incoming edge’s `when:` matched) or `"budget_exceeded"` (workflow budget exhausted). `null` for executed nodes. Branch-not-taken nodes are suppressed from the terminal panels but still appear in the JSON trace.
`nodes[].error`	Error message if the node failed, otherwise `null`.
`nodes[].tokens`	Total tokens used by this node.
`nodes[].usage`	Detailed token usage with `prompt_tokens`, `completion_tokens`, `total_tokens`, and `estimated_usd`.
`nodes[].guardrails_passed`	List of guardrail check methods that succeeded.
`nodes[].duration_ms`	Wall-clock time for this node in milliseconds.
`output`	Final output context (all `output.*` paths).
`summary.total_tokens`	Sum of tokens across all nodes.
`summary.total_usage`	Aggregated token usage with detailed breakdown.
`summary.total_duration_ms`	Sum of durations across all nodes.
`summary.status`	`"success"` or `"failed"`.
`summary.budget`	Present only when a `budget:` block is configured. Reports the configured ceilings plus `tokens_used`, `estimated_usd`, `duration_s`, `exceeded`, `violations`, and `skipped_remaining`. See YAML Reference → budget.

`sirenspec explain`

Print a human-readable execution plan for a workflow without making any LLM calls.

sirenspec explain <workflow-file> [OPTIONS]

Arguments:

Argument	Description
`workflow-file`	Path to the workflow YAML file.

Options:

Option	Short	Description
`--format TEXT`	`-f`	Output format: `"text"` (default) for human-readable or `"json"` for machine-readable.

Exit codes:

Code	Meaning
`0`	Plan generated successfully.
`1`	File not found, validation error, or structural errors in the workflow graph.

Examples:

# Print text-based execution plan
sirenspec explain workflow.yaml

# Output as JSON for programmatic use
sirenspec explain workflow.yaml --format json
sirenspec explain workflow.yaml -f json

Text Output Format: The text format shows the execution order, node types, guardrails, and edges in a human-readable table:

Workflow: router  │  3 nodes  │  2 agents
  
1. triage            agent=router (openai:gpt-4o-mini)  guardrails=[injection, length]
   writes → working.triage.result
   ↳ [working.triage.intent == "refund"]        → handle_refund
   ↳ [working.triage.intent == "general"]       → handle_general

2. handle_refund     agent=processor (anthropic:claude-3-5-sonnet-20241022)
   writes → output.refund_response
   ↳ [default]                                  → end

3. handle_general    agent=processor (anthropic:claude-3-5-sonnet-20241022)
   writes → output.general_response
   ↳ [default]                                  → end

Guardrails (workflow-level): injection, length

JSON Output Format: The JSON format is useful for integration into CI/CD systems or analysis tools:

{
  "workflow_name": "router",
  "node_count": 3,
  "agent_count": 2,
  "execution_order": ["triage", "handle_refund", "handle_general", "end"],
  "nodes": [
    {
      "id": "triage",
      "type": "agent",
      "type_info": "agent=router (openai:gpt-4o-mini)",
      "agent": "router",
      "model": "openai:gpt-4o-mini",
      "writes": "working.triage.result",
      "guardrails": ["injection", "length"],
      "outgoing_edges": [
        {
          "to": "handle_refund",
          "when": "working.triage.intent == \"refund\""
        },
        {
          "to": "handle_general",
          "when": "working.triage.intent == \"general\""
        }
      ]
    }
  ],
  "workflow_guardrails": ["injection", "length"],
  "warnings": [],
  "errors": []
}

`sirenspec test`

Discover and run YAML test fixtures. Each fixture file (*.test.yaml) contains a workflow definition plus a set of assertions that are validated during execution.

sirenspec test <test-path> [OPTIONS]

Arguments:

Argument	Description
`test-path`	Path to a single test fixture file or a directory of `*.test.yaml` files.

Options:

Option	Description
`--mock`	Replay LLM responses from a cassette file (no live API calls). Requires `--cassette`.
`--record`	Run live and record LLM responses into a cassette file. Requires `--cassette`. Mutually exclusive with `--mock`.
`--cassette PATH`	Path to the cassette file for `--mock` or `--record`. Required when either is used.

Exit codes:

Code	Meaning
`0`	All fixtures passed assertions.
`1`	One or more fixtures failed assertions or encountered execution errors.
`2`	Configuration error (e.g., missing `--cassette` when using `--mock`).

Examples:

# Run a single test fixture with live LLM calls
sirenspec test tests/answer.test.yaml

# Run all fixtures in a directory with live calls
sirenspec test tests/

# Replay recorded LLM responses from a cassette
sirenspec test tests/ --mock --cassette cassettes/responses.yaml

# Record live LLM responses into a cassette
sirenspec test tests/ --record --cassette cassettes/responses.yaml

Output format: For each fixture, a per-fixture result is printed:

PASS tests/answer.test.yaml
  ✓ assertion[0]: output matches expected
  ✓ assertion[1]: tokens < 100

FAIL tests/edge_case.test.yaml
  ✗ assertion[0]: status is success

After all fixtures, a summary line is printed:

3/3 fixtures passed

`sirenspec validate`

Validate a workflow YAML file without executing it or making any API calls.

sirenspec validate <workflow-file>

Arguments:

Argument	Description
`workflow-file`	Path to the workflow YAML file.

Exit codes:

Code	Meaning
`0`	File is valid.
`1`	File not found or validation failed.

Examples:

sirenspec validate workflow.yaml
# ✓ workflow.yaml is valid (2 agents, 3 nodes)

sirenspec validate broken.yaml
# ✗ Validation failed: agents.assistant.model: field required

What is validated:

YAML syntax (including duplicate key detection).
Required fields (version, agents, nodes).
Agent fields (model, system).
Node fields (agent, writes).
Edge references — all from and to values must refer to existing nodes.
Node agent references — all agent values must refer to existing agents.

After model validation, a static template linter runs over every template-bearing field (agent/swrm/synthesis prompts, factory/workflow inputs, for_each, swarm_size):

Rule	Level	Trigger
`working_dot_node_id`	error (blocks load)	`{{ working.<node_id>.* }}` where `<node_id>` is a known node — the canonical form is `{{ <node_id>.output }}`.
`unknown_namespace`	warning	A top-level template name that is neither reserved (`inputs`, `env`, `item`, `index`, `total`) nor a known node ID — usually a typo.

Errors raise WorkflowLintError and fail validation; warnings are printed without blocking. The same linter runs inside load_workflow() before any LLM call is made. Provider credentials and when: expressions are not evaluated during validation.

`sirenspec render`

Render a workflow as a diagram without executing it or making any API calls.

sirenspec render <workflow-file> --target <format> [OPTIONS]

Arguments:

Argument	Description
`workflow-file`	Path to the workflow YAML file.

Options:

Option	Short	Description
`--target TEXT`		Render target format. Currently only `mermaid` is supported.
`--output TEXT`	`-o`	Write output to a file instead of stdout.

Exit codes:

Code	Meaning
`0`	Rendered successfully.
`1`	File not found, validation error, or unsupported target.

Examples:

# Print a Mermaid flowchart to stdout
sirenspec render workflow.yaml --target mermaid

# Save to a file
sirenspec render workflow.yaml --target mermaid --output diagram.md
sirenspec render workflow.yaml --target mermaid -o diagram.md

Output format: The mermaid target produces a graph TD flowchart. Unconditional edges are plain arrows; conditional edges are labeled with their when: expression.

graph TD
    triage[triage]
    handle_refund[handle_refund]
    handle_general[handle_general]
    triage -->|"working.triage.intent == 'refund'"| handle_refund
    triage -->|"working.triage.intent == 'general'"| handle_general

Paste this into any Mermaid renderer (GitHub markdown, Mermaid Live Editor, etc.) to visualise the workflow graph.

Getting Started

Core Concepts

Reference

Resources

Installation

`sirenspec init`

`sirenspec run`

`sirenspec explain`

`sirenspec test`

`sirenspec validate`

`sirenspec render`

​Installation

​sirenspec init

​sirenspec run

​sirenspec explain

​sirenspec test

​sirenspec validate

​sirenspec render

Installation

`sirenspec init`

`sirenspec run`

`sirenspec explain`

`sirenspec test`

`sirenspec validate`

`sirenspec render`