Dialogflow CX to CES Migration

Full production-grade migration of a Dialogflow CX agent (v3beta1) to Google Customer Engagement Suite (CES) Conversational Agents. Migrates flows→sub-agents, pages→instructions, intents→routing hints, entity types, webhooks→tools, and test cases→golden evaluations CSV. Includes retry logic, dry-run mode, and a complete migration report.

Dialogflow CX → CES Migration Skill

What This Skill Does

Migrates a Dialogflow CX v3beta1 agent to Google Customer Engagement Suite (CES) Conversational Agents format, producing:

Output File	Contents
ces_agent.json	CES agent definition (importable via Console or REST API)
golden_evals.csv	CES batch evaluation CSV (golden test cases)
entity_types.json	Entity type definitions for manual re-creation
migration_report.md	Full migration summary with next steps

Migration Mapping

Dialogflow CX	→	CES Conversational Agents
Agent	→	Agent Application + Root Agent
Flow (non-default)	→	Sub-Agent
Pages + Routes	→	Agent Instructions (natural language)
Intent training phrases	→	Root agent routing hints
Entity Types	→	Exported JSON (manual import)
Webhook	→	Tool (OpenAPI schema)
Form Parameters	→	Instruction slot-filling steps
Test Cases	→	Golden Evaluations CSV

Prerequisites

1. GCP Auth: gcloud auth application-default login
2. Project access: Dialogflow CX API enabled on the project
3. Python: 3.10+
4. Packages: pip install google-cloud-dialogflow-cx google-auth

Usage

Run Migration

bash

python migrate.py \
  --project YOUR_PROJECT_ID \
  --agent-id YOUR_AGENT_UUID \
  --output ./migration_output

Dry Run (fetch + preview, no files written)

bash

python migrate.py \
  --project YOUR_PROJECT_ID \
  --agent-id YOUR_AGENT_UUID \
  --dry-run

Custom location (non-global agents)

bash

python migrate.py \
  --project YOUR_PROJECT_ID \
  --agent-id YOUR_AGENT_UUID \
  --location us-central1 \
  --output ./migration_output

Example (carconnect agent)

bash

python migrate.py \
  --project genaiguruyoutube \
  --agent-id 3736c564-5b3b-4f93-bbb2-367e7f04e4e8 \
  --output ./carconnect_ces

Expected output:

• 14 flows → 13 sub-agents + root agent enrichment
• 31 intents → root routing hints
• 5 entity types → exported JSON
• 2 webhooks → 2 OpenAPI tools
• Test cases → golden_evals.csv

Post-Migration Steps

1. Review ces_agent.json — check sub-agent instructions make sense, update tool endpoints
2. Import into CES Console:
   • Go to ces.cloud.google.com
   • Select project → Import Agent → Upload ces_agent.json
3. Upload Golden Evals:
   • Evaluate tab → + Add test case → Golden → Upload file → select golden_evals.csv
4. Re-create Entity Types from entity_types.json (CES uses them as tool parameters)
5. Update webhook endpoints in the Tools section of your CES agent
6. Run evaluation suite → review pass rates → iterate on instructions

Retry Logic

All Google API calls use exponential backoff (up to 4 attempts, base delay 1.5s × 2ⁿ). If the API is rate-limited or temporarily unavailable, the tool retries automatically.

Limitations & Known Gaps

• Rich response types (carousels, chips, images) are converted to text messages. Update manually in CES.
• Conditional routes using session parameter syntax ($session.params.X) are preserved as-is in instructions but may need updating for CES parameter syntax.
• DTMF / telephony settings are not migrated (CES has different telephony config).
• Entity type import: CES does not have a direct batch import API for entity types — use the exported JSON as reference to create them manually or via REST.
• Webhook auth: OAuth and mTLS configs are noted but credentials must be re-configured in CES.

Autoresearch Evals (binary pass/fail)

When running autoresearch on this skill, use these evals:

1. EVAL_FLOWS: All non-default flows appear as sub-agents in ces_agent.json?
2. EVAL_TOOLS: All webhooks appear as tools with OpenAPI schemas in ces_agent.json?
3. EVAL_ENTITIES: All entity types exported to entity_types.json?
4. EVAL_EVALS_CSV: golden_evals.csv has correct header + at least one golden eval row?
5. EVAL_INSTRUCTIONS: Each sub-agent has non-empty instructions?
6. EVAL_REPORT: migration_report.md exists and contains a stats table?