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
INLINECODE0	CES agent definition (importable via Console or REST API)
INLINECODE1

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. 1. GCP Auth: INLINECODE4
2. Project access: Dialogflow CX API enabled on the project
3. Python: 3.10+
4. Packages: INLINECODE5

Usage

Run Migration

CODEBLOCK0

Dry Run (fetch + preview, no files written)

CODEBLOCK1

Custom location (non-global agents)

CODEBLOCK2

Example (carconnect agent)

CODEBLOCK3

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. 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

1. 3. Upload Golden Evals:

- Evaluate tab → + Add test case → Golden → Upload file → select golden_evals.csv

1. 4. Re-create Entity Types from entity_types.json (CES uses them as tool parameters)
2. Update webhook endpoints in the Tools section of your CES agent
3. 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. 1. EVALFLOWS: All non-default flows appear as sub-agents in ces_agent.json?
2. EVALTOOLS: All webhooks appear as tools with OpenAPI schemas in ces_agent.json?
3. EVALENTITIES: All entity types exported to entity_types.json?
4. EVALEVALSCSV: golden_evals.csv has correct header + at least one golden eval row?
5. EVALINSTRUCTIONS: Each sub-agent has non-empty instructions?
6. EVAL_REPORT: migration_report.md exists and contains a stats table?