YAML Workflows

YAML workflows let you define Fibonacci workflows in a declarative format, separating workflow structure from application code. This approach is ideal for version control, code review, and non-developer workflow authoring.

Basic Structure

A YAML workflow file defines the complete workflow configuration:

name: text-analyzer
description: Analyze text sentiment and extract key themes
version: "1.0"

nodes:
  - id: sentiment
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Analyze the sentiment of this text:
      {{input.text}}
      
      Return: positive, negative, or neutral

  - id: themes
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Extract key themes from this text:
      {{input.text}}
      
      Return as a comma-separated list.
    dependencies:
      - sentiment

outputs:
  sentiment: "{{sentiment}}"
  themes: "{{themes}}"

Loading YAML Workflows

From File

from fibonacci import Workflow

# Load from file
workflow = Workflow.from_yaml("workflows/text-analyzer.yaml")

# Execute
result = workflow.execute(inputs={"text": "I love this product!"})
print(result)

From String

from fibonacci import Workflow

yaml_content = """
name: quick-summary
nodes:
  - id: summarize
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: "Summarize: {{input.text}}"
"""

workflow = Workflow.from_yaml_string(yaml_content)

From URL

from fibonacci import Workflow

# Load from remote URL
workflow = Workflow.from_yaml_url(
    "https://example.com/workflows/analyzer.yaml"
)

Node Types

LLM Nodes

nodes:
  - id: writer
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Write a {{input.style}} paragraph about {{input.topic}}.
    temperature: 0.7
    max_tokens: 500
    
  - id: structured_output
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: "Extract entities from: {{input.text}}"
    output_format: json
    schema:
      type: object
      properties:
        people:
          type: array
          items:
            type: string
        places:
          type: array
          items:
            type: string

Tool Nodes

nodes:
  - id: fetch_data
    type: tool
    tool: google_sheets.read_range
    inputs:
      spreadsheet_id: "{{input.sheet_id}}"
      range: "A1:D100"
      
  - id: send_notification
    type: tool
    tool: slack.post_message
    inputs:
      channel: "#alerts"
      message: "Analysis complete: {{analyzer}}"
    dependencies:
      - analyzer

Conditional Nodes

nodes:
  - id: route_by_sentiment
    type: conditional
    conditions:
      - if:
          field: "{{sentiment}}"
          operator: equals
          value: "positive"
        then: positive_handler
      - if:
          field: "{{sentiment}}"
          operator: equals
          value: "negative"
        then: negative_handler
    default: neutral_handler
    dependencies:
      - sentiment

Critic Nodes

nodes:
  - id: quality_check
    type: critic
    model: claude-sonnet-4-5-20250929
    target_node: writer
    criteria:
      - "Content is factually accurate"
      - "Tone matches requested style"
      - "Length is appropriate"
    min_score: 7
    max_iterations: 3

Advanced Configuration

Variables and Secrets

Use environment variables and secrets in YAML:

name: secure-workflow
version: "1.0"

# Reference environment variables
env:
  API_ENDPOINT: ${API_ENDPOINT}
  DEFAULT_MODEL: ${FIBONACCI_MODEL:-claude-sonnet-4-5-20250929}

# Reference secrets (from secure storage)
secrets:
  - OPENAI_API_KEY
  - DATABASE_URL

nodes:
  - id: fetch
    type: tool
    tool: http.request
    inputs:
      url: "{{env.API_ENDPOINT}}/data"
      headers:
        Authorization: "Bearer {{secrets.API_KEY}}"

Retry Configuration

nodes:
  - id: unreliable_api
    type: tool
    tool: http.request
    inputs:
      url: "https://api.example.com/data"
    retry:
      max_attempts: 3
      delay: 1.0
      backoff: exponential
      max_delay: 30.0
      retry_on:
        - timeout
        - 5xx

Memory Configuration

name: stateful-workflow

memory:
  backend: redis
  connection: ${REDIS_URL}
  key_prefix: "myapp:"

nodes:
  - id: chatbot
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Conversation history:
      {{memory.history}}
      
      User: {{input.message}}
    memory_read:
      - history
    memory_write:
      key: last_response
      scope: user

Timeouts

nodes:
  - id: long_analysis
    type: llm
    model: claude-opus-4-5-20251101
    prompt: "Deep analysis of: {{input.data}}"
    timeout: 120  # seconds
    
  - id: quick_check
    type: tool
    tool: http.request
    inputs:
      url: "https://api.example.com/health"
    timeout: 5

Workflow Composition

Including Other Workflows

name: master-pipeline
version: "1.0"

includes:
  - path: ./preprocessing.yaml
    as: preprocess
  - path: ./analysis.yaml
    as: analyze
  - path: ./reporting.yaml
    as: report

nodes:
  - id: run_preprocess
    type: workflow
    workflow: preprocess
    inputs:
      data: "{{input.raw_data}}"
      
  - id: run_analysis
    type: workflow
    workflow: analyze
    inputs:
      clean_data: "{{run_preprocess.output}}"
    dependencies:
      - run_preprocess
      
  - id: run_report
    type: workflow
    workflow: report
    inputs:
      results: "{{run_analysis}}"
    dependencies:
      - run_analysis

Parallel Execution

name: parallel-analysis
version: "1.0"

nodes:
  # These run in parallel (no dependencies between them)
  - id: sentiment_analysis
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: "Analyze sentiment: {{input.text}}"
    
  - id: entity_extraction
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: "Extract entities: {{input.text}}"
    
  - id: topic_classification
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: "Classify topics: {{input.text}}"
    
  # This waits for all parallel nodes
  - id: combine_results
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Combine these analysis results:
      Sentiment: {{sentiment_analysis}}
      Entities: {{entity_extraction}}
      Topics: {{topic_classification}}
    dependencies:
      - sentiment_analysis
      - entity_extraction
      - topic_classification

Validation

Schema Validation

Fibonacci validates YAML workflows against a JSON Schema:

from fibonacci import Workflow, WorkflowValidationError

try:
    workflow = Workflow.from_yaml("workflow.yaml")
except WorkflowValidationError as e:
    print(f"Validation errors: {e.errors}")

CLI Validation

# Validate a workflow file
fibonacci validate workflow.yaml

# Validate with verbose output
fibonacci validate workflow.yaml --verbose

# Validate all workflows in a directory
fibonacci validate workflows/

Custom Validation Rules

name: validated-workflow
version: "1.0"

validation:
  require_descriptions: true
  max_nodes: 20
  allowed_models:
    - claude-sonnet-4-5-20250929
    - claude-haiku-4-5-20251001
  forbidden_tools:
    - http.request  # Require specific API tools instead

nodes:
  - id: analyzer
    type: llm
    model: claude-sonnet-4-5-20250929
    description: "Analyzes input text for sentiment"  # Required
    prompt: "Analyze: {{input.text}}"

Complete Example

Here’s a full production workflow example:

name: customer-support-router
description: Routes customer inquiries to appropriate handlers
version: "2.1"
author: [email protected]

env:
  SLACK_CHANNEL: ${SUPPORT_SLACK_CHANNEL:-#support}
  ESCALATION_THRESHOLD: ${ESCALATION_THRESHOLD:-0.8}

memory:
  backend: redis
  connection: ${REDIS_URL}

nodes:
  # Classify the inquiry
  - id: classify
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Classify this customer inquiry:
      {{input.message}}
      
      Categories: billing, technical, general, complaint
      Urgency: low, medium, high, critical
      
      Return JSON: {"category": "...", "urgency": "..."}
    output_format: json
    timeout: 30

  # Check for escalation triggers
  - id: check_escalation
    type: conditional
    conditions:
      - if:
          field: "{{classify.urgency}}"
          operator: equals
          value: "critical"
        then: escalate_immediately
      - if:
          field: "{{classify.category}}"
          operator: equals
          value: "complaint"
        then: complaint_handler
    default: standard_response
    dependencies:
      - classify

  # Generate standard response
  - id: standard_response
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Previous interactions:
      {{memory.customer_history}}
      
      Customer inquiry ({{classify.category}}):
      {{input.message}}
      
      Provide a helpful response.
    memory_read:
      - customer_history
    dependencies:
      - classify

  # Handle complaints specially
  - id: complaint_handler
    type: llm
    model: claude-sonnet-4-5-20250929
    prompt: |
      Handle this complaint with empathy:
      {{input.message}}
      
      Previous issues:
      {{memory.customer_history}}
      
      Acknowledge concerns and offer resolution.
    memory_read:
      - customer_history
    temperature: 0.3  # More consistent tone
    dependencies:
      - classify

  # Immediate escalation
  - id: escalate_immediately
    type: tool
    tool: slack.post_message
    inputs:
      channel: "{{env.SLACK_CHANNEL}}"
      message: |
        🚨 CRITICAL ESCALATION
        Customer: {{input.customer_id}}
        Category: {{classify.category}}
        Message: {{input.message}}
    dependencies:
      - classify

  # Log interaction
  - id: log_interaction
    type: tool
    tool: database.insert
    inputs:
      table: support_interactions
      data:
        customer_id: "{{input.customer_id}}"
        category: "{{classify.category}}"
        urgency: "{{classify.urgency}}"
        timestamp: "{{now}}"
    dependencies:
      - classify

outputs:
  response: "{{standard_response || complaint_handler}}"
  category: "{{classify.category}}"
  urgency: "{{classify.urgency}}"
  escalated: "{{escalate_immediately != null}}"

Best Practices

Use version control

Store YAML workflows in Git for change tracking, code review, and rollback capabilities.

Separate environments

Use environment variables for configuration that differs between dev/staging/production:

env:
  API_URL: ${API_URL}
  MODEL: ${MODEL:-claude-sonnet-4-5-20250929}

Document your workflows

Add descriptions to workflows and nodes:

name: my-workflow
description: |
  This workflow processes customer feedback.
  Author: [email protected]
  Last updated: 2025-01-15

Validate before deployment

Always run fibonacci validate in CI/CD pipelines before deploying workflow changes.

Getting Started

Core Concepts

Advanced Guides

Basic Structure

Loading YAML Workflows

From File

From String

From URL

Node Types

LLM Nodes

Tool Nodes

Conditional Nodes

Critic Nodes

Advanced Configuration

Variables and Secrets

Retry Configuration

Memory Configuration

Timeouts

Workflow Composition

Including Other Workflows

Parallel Execution

Validation

Schema Validation

CLI Validation

Custom Validation Rules

Complete Example

Best Practices

Next Steps

CLI Guide

Best Practices

Getting Started

Core Concepts

Advanced Guides

​Basic Structure

​Loading YAML Workflows

​From File

​From String

​From URL

​Node Types

​LLM Nodes

​Tool Nodes

​Conditional Nodes

​Critic Nodes

​Advanced Configuration

​Variables and Secrets

​Retry Configuration

​Memory Configuration

​Timeouts

​Workflow Composition

​Including Other Workflows

​Parallel Execution

​Validation

​Schema Validation

​CLI Validation

​Custom Validation Rules

​Complete Example

​Best Practices

​Next Steps

CLI Guide

Best Practices

Basic Structure

Loading YAML Workflows

From File

From String

From URL

Node Types

LLM Nodes

Tool Nodes

Conditional Nodes

Critic Nodes

Advanced Configuration

Variables and Secrets

Retry Configuration

Memory Configuration

Timeouts

Workflow Composition

Including Other Workflows

Parallel Execution

Validation

Schema Validation

CLI Validation

Custom Validation Rules

Complete Example

Best Practices

Next Steps