Dfcx Framework Ingestor

Dialogflow CX Ingestor Skill

This skill standardizes the ingestion, parsing, and behavior extraction of conversational agents exported in the official, public Dialogflow CX (DFCX) format.

1. DFCX Export Package Layout

An official Dialogflow CX agent export is packaged as a .zip archive. When extracted, it conforms to the following public directory layout:

<agent_zip_root>/
├── agent.json                       # Global agent metadata (default language, timezone, display name)
├── intents/                         # Declared intents and training phrases
│   └── <intent_name>/
│       └── <intent_name>.json       # Intent configuration
├── entityTypes/                     # Custom entity declarations
│   └── <entity_name>/
│       └── <entity_name>.json       # Entity configuration and synonyms
├── transitionRouteGroups/           # Shared flow-level transition groups
├── webhooks/                        # Declared webhooks and backend REST integrations
│   └── <webhook_name>.json          # Webhook configuration (defining target URI and credentials)
└── flows/                           # Subfolders representing individual Flows
    └── <flow_name>/                 # Flow directory (e.g., Default Start Flow, Billing)
        ├── <flow_name>.json         # Flow configuration listing flow-level transition routes and entry points
        ├── transitionRouteGroups/   # Shared transition route groups specific to this flow
        └── pages/                   # Directory containing individual page configurations
            └── <page_name>.json     # Page definition containing entry prompts, form parameters, and transitions

2. Ingestion & Flow Tracing Directives

The Ingestor Agent MUST parse the extracted file structures using standard directory traversal and JSON parsing, mapping them to the primary coverage checklist:

A. CUJ Mapping

• CUJ Discovery: Treat each Flow folder under flows/ (e.g., flows/Default Start Flow/ or flows/Billing Flow/) as a high-level Parent CUJ. The folder's directory name represents the CUJ title.

B. Sub-intent & Trigger Extraction

For each Flow, traverse <flow_name>.json and all .json files inside its pages/ subdirectory:

1. Flow-Level Transition Routes:
   • Parse the "transitionRoutes" array inside <flow_name>.json.
   • Identify each transition triggered by an intent (where "intent" is declared, e.g., projects/.../intents/<intent_id>).
   • Map each unique triggered transition as a Sub-intent / Scenario (using the target page/flow name to synthesize its name).
2. Page-Level Transition Routes:
   • For each page .json inside pages/, parse the "transitionRoutes" array to map conversational branches and trigger conditions.
3. Form Parameter nodes:
   • Parse the "form.parameters" array inside a page JSON.
   • Each parameter collected (e.g., email_address, postal_code) represents a sub-intent step in the conversational walkthrough.
4. Fulfillment Webhooks:
   • Identify "triggerFulfillment.webhook" inside transition routes or page configs.
   • Resolve the webhook's target endpoint destination URI by looking up its matching configuration file inside the global webhooks/ directory (e.g. webhooks/<webhook_name>.json's "genericWebService.uri").
   • Note: Webhooks represent standard REST integrations defined as toolsets (indexed via OpenAPI specifications open_api_schema.yaml under the toolsets/ directory).

C. Alignment with Canonical cxas-agent-migration Standards

When parsing Dialogflow CX layouts, the Ingestor Agent SHOULD align the extracted behaviors with the open-source cxas-agent-migration topology (N->M agents consolidation):

1. Playbook/Flow to Agent Persona: Each extracted Flow directory represents a distinct sub-agent boundary. Treat the flow's transitions and event handlers as the sub-agent's private prompt limits and routing callbacks.
2. Deduplicate & Group Variables: Consolidate intent parameters and page forms variables using cxas-agent-migration Stage 1 rules (variable deduplication) to avoid parameter bloat in transcripts.
3. Topological Rewiring: Track the source dependency graph (parent flows -> children flows) using cxas-agent-migration Stage 3 rules, ensuring that transitions from steering accurately reflect the legacy DFCX intent routing.

3. Dialogue Simulation & Transcript Guidelines

When simulating conversation transcripts based on parsed DFCX folder packages, enforce the following rules:

A. Verbatim Prompts Extraction

To guarantee the simulated agent speaks the exact developer-defined prompts, parse the fulfillment blocks:

1. Page Entry Prompts:
   • Read "entryFulfillment.messages" inside pages/<page_name>.json.
   • Extract speech variants under "text": {"text": [...]}. The simulated Agent turn MUST output this string verbatim.
2. Form Slot-Filling Prompts:
   • Read "initialPromptFulfillment.messages" inside "form.parameters[].fillBehavior" in the page JSON.
   • The simulated Agent turn asking for the parameter MUST output these prompt strings verbatim.
3. Transition Prompts:
   • Read "triggerFulfillment.messages" inside transition routes.

B. Programmatic Flow Traversal & Intent Discovery

To discover conversational paths and map execution sequences:

1. Flow Start Point: Access flows/<flow_name>/<flow_name>.json and locate the "transitionRoutes" inside the root flow config. These define the entry intents (e.g. "intent": "support_request").
2. State Traversal:
   • For each route, trace the target node: if the target is a sub-flow (e.g. "targetFlow": "flows/Billing/"), pivot to that sub-flow config.
   • If the target is a page (e.g. "targetPage": "Collect Email"), load and parse the page configuration file pages/Collect Email.json.
3. Fulfillment Sequence Tracking:
   • Trace visited pages in chronological order along the path to extract the sequential webhook invocations (triggerFulfillment.webhook) and parameter mappings, mapping each valid end-to-end path as a scenario sequence.