ChatDev/docs/user_guide/en/workflow_authoring.md

Workflow Authoring Guide

This guide covers YAML structure, node types, provider configuration, edge conditions, and design template export so you can build and debug DevAll DAGs efficiently. Content mirrors docs/user_guide/workflow_authoring.md with English copy for global contributors.

1. Prerequisites

• Know the layout of yaml_instance/ and yaml_template/.
• Understand the core node types (model, python, agent, human, subgraph, passthrough, literal).
• Review FIELD_SPECS (see field_specs.md) and the Schema API contract (config_schema_contract.md) if you rely on dynamic forms in the frontend/IDE.

2. YAML Top-level Structure

Every workflow file follows the DesignConfig root with only three keys: version, vars, and graph. The snippet below is adapted from yaml_instance/net_example.yaml and can run as-is:

version: 0.4.0
vars:
  BASE_URL: https://api.example.com/v1
  API_KEY: ${API_KEY}
graph:
  id: paper_gen
  description: Article generation and refinement
  log_level: INFO
  is_majority_voting: false
  initial_instruction: |
    Provide a word or short phrase and the workflow will draft and polish an article.
  start:
    - Article Writer
  end:
    - Article Writer
  nodes:
    - id: Article Writer
      type: agent
      config:
        provider: openai
        base_url: ${BASE_URL}
        api_key: ${API_KEY}
        name: gpt-4o
        params:
          temperature: 0.1
    - id: Human Reviewer
      type: human
      config:
        description: Review the article. Type ACCEPT to finish; otherwise provide revision notes.
  edges:
    - from: Article Writer
      to: Human Reviewer
    - from: Human Reviewer
      to: Article Writer
      condition:
        type: keyword
        config:
          none:
            - ACCEPT
          case_sensitive: false

• version: optional configuration version (defaults to 0.0.0). Increment it whenever schema changes in entity/configs/graph.py require template or migration updates.
• vars: root-level key-value map. You can reference ${VAR} anywhere in the file; fallback is the same-name environment variable. GraphDefinition.from_dict rejects nested vars, so keep this block at the top only.

Environment Variables & .env File

The system supports referencing variables in YAML configurations using ${VAR} syntax. These variables can be used in any string field within the configuration. Common use cases include:

• API keys: api_key: ${API_KEY}
• Service URLs: base_url: ${BASE_URL}
• Model names: name: ${MODEL_NAME}

The system automatically loads the .env file from the project root (if present) when parsing configurations. Variable resolution follows this priority order:

Priority	Source	Description
1 (highest)	Values defined in vars	Key-value pairs declared directly in the YAML file
2	System/shell environment variables	Values set via export or system config
3 (lowest)	Values from .env file	Only applied if the variable doesn't already exist

Tip

The .env file does not override existing environment variables. This allows you to define defaults in .env while overriding them via export or deployment platform configurations.

Warning

If a placeholder references a variable that is not defined in any of the three sources above, a ConfigError will be raised during configuration parsing with the exact path indicated.

• graph: required block that maps to the GraphDefinition dataclass:
  • Metadata: id (required), description, log_level (default DEBUG), is_majority_voting, initial_instruction, and optional organization.
  • Execution controls: start/end entry lists (the system executes nodes listed in start at the beginning), plus nodes and edges. Provider/model/tooling settings now live inside each node.config; the legacy top-level providers table is deprecated. In the example the keyword condition on Human Reviewer -> Article Writer keeps looping unless the reviewer types ACCEPT.
  • Shared resources: memory defines stores available to node.config.memories. The validator ensures every attachment points to a declared store.
  • Schema references: yaml_template/design.yaml mirrors the latest GraphDefinition shape. After editing configs run python -m tools.export_design_template or hit the Schema API to validate.

Further reading: docs/user_guide/en/field_specs.md (field catalog), docs/user_guide/en/runtime_ops.md (runtime observability), and yaml_template/design.yaml (generated baseline template).

3. Node Type Cheatsheet

Type	Description	Key fields	Detailed Docs
agent	Runs an LLM-backed agent with optional tools, memories, and thinking phases.	provider, model, prompt_template, tooling, thinking, memories	agent.md
python	Executes Python scripts/commands sharing the code_workspace/.	entry_script, inline_code, timeout, env	python.md
human	Pauses in the Web UI awaiting human input.	prompt, timeout, attachments	human.md
subgraph	Embeds a child DAG to reuse complex flows.	graph_path or inline graph	subgraph.md
passthrough	Pass-through node that forwards only the last message by default and can be configured to forward all messages; used for context filtering and graph structure optimization.	only_last_message	passthrough.md
literal	Emits a fixed text payload whenever triggered and discards inputs.	content, role (user/assistant)	literal.md
loop_counter	Guard node that limits loop iterations before releasing downstream edges.	max_iterations, reset_on_emit, message	loop_counter.md

Fetch the full schema via POST /api/config/schema or inspect the dataclasses inside entity/configs/.

4. Providers & Agent Settings

• When a node omits provider, the engine uses globals.default_provider (e.g., openai).
• Fields such as model, api_key, and base_url accept ${VAR} placeholders for environment portability.
• When combining multiple providers, define globals in the workflow root ({ default_provider: ..., retry: {...} }) if supported by the dataclass.

4.1 Gemini Provider Config Example

model:
  provider: gemini
  base_url: https://generativelanguage.googleapis.com
  api_key: ${GEMINI_API_KEY}
  name: gemini-2.0-flash-001
  input_mode: messages
  params:
    response_modalities: ["text", "image"]
    safety_settings:
      - category: HARM_CATEGORY_SEXUAL
        threshold: BLOCK_LOWER

The Gemini Provider supports multi-modal input (images/video/audio are automatically converted to Parts) and supports function_calling_config to control tool execution behavior.

5. Edges & Conditions

• Basic edge:

  - source: plan
    target: execute
  

• Conditional edge:

  edges:
    - source: router
      target: analyze
      condition: should_analyze   # functions/edge/should_analyze.py
  

• If a condition function raises, the scheduler marks the branch as failed and stops downstream execution.

5.1 Edge Payload Processors

• Add process to an edge when you want to transform or filter the payload after the condition is met (e.g., extract a verdict, keep only structured fields, or rewrite text).
• Structure mirrors condition (type + config). Built-ins include:
  • regex_extract: Python regex with optional group, mode (replace_content, metadata, data_block), multiple, and on_no_match (pass, default, drop).
  • function: calls helpers under functions/edge_processor/. The handler signature is def foo(payload: Message, **kwargs) -> Message | None. Note: The Processor interface is now standardized, and kwargs includes context: ExecutionContext, allowing access to the current execution context.
• Example:

  - from: reviewer
    to: qa
    process:
      type: regex_extract
      config:
        pattern: "Score\\s*:\\s*(?P<score>\\d+)"
        group: score
        mode: metadata
        metadata_key: quality_score
        case_sensitive: false
        on_no_match: default
        default_value: "0"
  

6. Agent Node Advanced Features

• Tooling: Configure AgentConfig.tooling; see the Tooling module (Chinese for now).
• Thinking: Enable staged reasoning via AgentConfig.thinking (e.g., chain-of-thought, reflection). Reference entity/configs/thinking.py for parameters.
• Memories: Attach MemoryAttachmentConfig through AgentConfig.memories; details live in the Memory module.

7. Dynamic Execution (Map-Reduce/Tree)

Nodes support a sibling field dynamic to enable parallel processing or Map-Reduce patterns.

7.1 Core Concepts

• Map Mode (type: map): Fan-out. Splits list inputs into multiple units for parallel execution, outputting List[Message] (flattened results).
• Tree Mode (type: tree): Fan-out & Reduce. Splits inputs for parallel execution, then recursively reduces results in groups of group_size until a single result remains (e.g., "summary of summaries").
• Split Strategy: Defines how to partition the output of the previous node or current input into parallel units.

7.2 Configuration Structure

nodes:
  - id: Research Agents
    type: agent
    # Standard config (behaves as template for parallel units)
    config:
      provider: openai
      model: gpt-4o
      prompt_template: "Research this topic: {{content}}"
    # Dynamic execution config
    dynamic:
      type: map
      # Split strategy (first layer only)
      split:
        type: message             # Options: message, regex, json_path
        # pattern: "..."          # Required for regex mode
        # json_path: "$.items[*]" # Required for json_path mode
      # Mode-specific config
      config:
        max_parallel: 5           # Concurrency limit

7.3 Tree Mode Example

Ideal for chunked summarization of long texts:

dynamic:
  type: tree
  split:
    type: regex
    pattern: "(?s).{1,2000}(?:\\s|$)"  # Split every ~2000 chars
  config:
    group_size: 3   # Reduce every 3 results into 1
    max_parallel: 10

This mode automatically builds a multi-level execution tree until the result count is reduced to 1. Split config is the same as map mode.

8. Design Template Export

After editing configs or FIELD_SPECS, regenerate templates:

python -m tools.export_design_template \
  --output yaml_template/design.yaml \
  --mirror frontend/public/design_0.4.0.yaml

• The script scans registered nodes, memories, tooling, and FIELD_SPECS to emit YAML templates plus the frontend mirror file.
• Commit the generated files and notify frontend owners to refresh static assets.

9. CLI / API Execution Paths

• Web UI: Choose a YAML file, fill run parameters, start execution, and monitor in the dashboard. Recommended path.
• HTTP: POST /api/workflow/execute with session_name, graph_path or graph_content, task_prompt, optional attachments, and log_level (defaults to INFO, supports INFO or DEBUG).
• CLI: python run.py --path yaml_instance/demo.yaml --name test_run. Provide TASK_PROMPT via env var or respond to the CLI prompt.

10. Debugging Tips

• Use the Web UI context snapshots or WareHouse/<session>/context.json to inspect node I/O. Note that all node outputs are now standardized as List[Message].
• Leverage the Schema API breadcrumbs (config_schema_contract.md) or run python run.py --inspect-schema to view field specs quickly.
• Missing YAML placeholders trigger ConfigError during parsing with a precise path surfaced in both UI and CLI logs.