Real World Examples

This page is intentionally dense. Each flow is shown as:

1. Chat transcript
2. Per-turn step trace (what executed and why)
3. Required control-plane rows (ce_*)
4. Audit timeline with payload-level checkpoints

Example 1: FAQ (informational path)

Chat Transcript

FAQ conversation

User

How do I transfer my electricity connection?

Info

• Informational FAQ request.
• No pending action path required.

ConvEngine

Share customerId, phone, email, source city, and target city. Then confirm to submit transfer.

Tables

• ce_intent
• ce_intent_classifier
• ce_response

Info

• Intent=FAQ.
• FAQ response returned from configured mapping.

Intent and state outcome

Field	Expected value	Why
intent	FAQ	Classifier matches informational query language
state	IDLE	No slot collection / no pending action required
responseType	DERIVED (TEXT)	FAQ answer generated from prompt/context

Per-turn Step Trace

Turn T1 runtime trace

Order	Step	Input facts	Output facts
1	LoadOrCreateConversationStep	conversationId present or new	session opened/hydrated
1a	CacheInspectAuditStep	cache-inspector enabled	CACHE_INSPECTION audit row (conditional)
2	AuditUserInputStep	raw user text	USER_INPUT audit row
3	DialogueActStep	question sentence	dialogue_act=QUESTION
4	InteractionPolicyStep	no pending_action, no pending_slot	policy_decision=RECLASSIFY_INTENT
5	IntentResolutionStep	classifiers + agent hints	intent=FAQ, state=IDLE
6	SchemaExtractionStep	FAQ schema (light)	schema facts available
7	RulesStep	no FAQ rule action	RULE_NO_MATCH or no mutation
8	ResponseResolutionStep	FAQ response row + prompt	assistant output prepared
9	MemoryStep	recent turn context	memory_session_summary
10	PersistConversationStep	final session	conversation/audit persisted

Why no pending-action path here

InteractionPolicyStep sees no pending_action signal and does not produce EXECUTE_PENDING_ACTION. The turn continues as a regular intent-resolution flow.

Required ce_* Rows

Minimum rows required for FAQ flow

Table	Required row shape	Used by
ce_intent	intent_code=FAQ	IntentResolutionStep
ce_intent_classifier	intent_code=FAQ, rule_type=REGEX	IntentResolutionStep
ce_prompt_template	intent_code=FAQ, response_type=TEXT	ResponseResolutionStep (DERIVED)
ce_response	intent_code=FAQ, state_code=IDLE	ResponseResolutionStep
ce_container_config	intent_code=FAQ (optional context source)	AddContainerDataStep
ce_config	AgentIntentResolver + SchemaExtractionStep prompt keys	Intent/Schema steps

FAQ seed snippet

SQL

INSERT INTO ce_intent (intent_code, description, priority, enabled)
VALUES ('FAQ', 'Answer informational questions from FAQ knowledge base', 10, true);

INSERT INTO ce_intent_classifier (intent_code, rule_type, pattern, priority, enabled)
VALUES ('FAQ', 'REGEX', '(?i)\b(what|how|help|faq|information|details|explain)\b', 10, true);

INSERT INTO ce_prompt_template (intent_code, state_code, response_type, system_prompt, user_prompt, temperature, enabled)
VALUES ('FAQ', 'IDLE', 'TEXT',
      'You are a concise FAQ assistant. Answer directly and clearly.',
      'User question: {{user_input}}\nFAQ context: {{container_data}}\nReturn short helpful answer.',
      0.10, true);

INSERT INTO ce_response (intent_code, state_code, output_format, response_type, derivation_hint, priority, enabled)
VALUES ('FAQ', 'IDLE', 'TEXT', 'DERIVED',
      'Answer FAQ using available context and user question.', 10, true);

Audit + Payload Postmortem

Expected audit stages (ordered)

Order	Stage	Must contain
1	USER_INPUT	original user message
2	DIALOGUE_ACT_CLASSIFIED	dialogueAct=QUESTION
3	INTERACTION_POLICY_DECIDED	policyDecision=RECLASSIFY_INTENT
4	INTENT_RESOLVED_BY_*	intent=FAQ
5	RESOLVE_RESPONSE_SELECTED	responseType=DERIVED
6	ASSISTANT_OUTPUT	output text or json payload
7	MEMORY_UPDATED	summaryChars/recalled flags

Audit payload checkpoint: DIALOGUE_ACT_CLASSIFIED

JSON

{
  "_meta": {
    "stage": "DIALOGUE_ACT_CLASSIFIED"
  },
  "userText": "How do I transfer my electricity connection?",
  "dialogueAct": "QUESTION",
  "dialogueActConfidence": 0.7,
  "dialogueActSource": "REGEX"
}

Audit payload checkpoint: INTERACTION_POLICY_DECIDED

JSON

{
  "_meta": {
    "stage": "INTERACTION_POLICY_DECIDED"
  },
  "dialogueAct": "QUESTION",
  "policyDecision": "RECLASSIFY_INTENT",
  "skipIntentResolution": false,
  "hasPendingAction": false,
  "hasPendingSlot": false
}

Example 2: CONNECTION_TRANSFER (multi-turn + confirmation + pending action)

Chat Transcript

Connection transfer conversation

User

Move my electricity connection from City1 to City2.

Info

• Initial transfer request.
• State expected to move to collection mode.

ConvEngine

Please provide customerId, phone, email, sourceCity, and targetCity.

Tables

• ce_output_schema

Info

• Required fields are missing.
• Assistant prompts for slots.

User

customerId 9912, phone 9988776655, email user@zapper.com, sourceCity City1, targetCity City2

Info

• All required fields provided.
• Schema should become complete.

ConvEngine

Do you want to move this connection right away?

Tables

• ce_rule
• ce_pending_action

Info

• Transition to AWAITING_CONFIRMATION.
• Pending action staged.

User

yes please

Info

• AFFIRM dialogue act.
• Policy should execute pending action.

ConvEngine

Connection transfer request submitted successfully.

Tables

• ce_audit
• ce_response

Info

• Pending action executed.
• Final completion response emitted.

Per-turn Step Trace

Turn T1 (initial request)

Step	Key output	Notes
IntentResolutionStep	intent=CONNECTION_TRANSFER	Classifier picks transfer intent
SchemaExtractionStep	missing fields computed	Required slots are incomplete
RulesStep	state -> COLLECT_INPUTS	Bootstrap collection state
ResponseResolutionStep	ask for required fields	Prompt user for missing slots

Turn T2 (slot fill)

Step	Key output	Notes
SchemaExtractionStep	customerId/phone/email/sourceCity/targetCity extracted	Schema complete
RulesStep	COLLECT_INPUTS -> AWAITING_CONFIRMATION	User must confirm execution
ResponseResolutionStep	confirmation prompt

Turn T3 (affirm)

Order	Step	Key output	Why it matters
1	DialogueActStep	AFFIRM	Converts 'yes please' into execution signal
2	InteractionPolicyStep	EXECUTE_PENDING_ACTION	Policy routes to pending action
3	ActionLifecycleStep	OPEN -> IN_PROGRESS	Runtime status tracked in context
4	DisambiguationStep	not required	Single eligible action
5	GuardrailStep	ALLOW	No sensitive/approval block
6	PendingActionStep	EXECUTED	Task bean/method invoked from ce_pending_action
7	RulesStep	AWAITING_CONFIRMATION -> COMPLETED	Final state transition
8	ResponseResolutionStep	success response	User receives completion

Required ce_* Rows

Control-plane requirements

Table	Must have	Consumed by
ce_intent	CONNECTION_TRANSFER	IntentResolutionStep
ce_intent_classifier	transfer/move regex	IntentResolutionStep
ce_output_schema	required slots for transfer	SchemaExtractionStep
ce_rule	state bootstrap + completeness + completion rules	RulesStep
ce_pending_action	action_key + bean_name + method_names	ActionLifecycleStep/PendingActionStep
ce_response	COLLECT_INPUTS, AWAITING_CONFIRMATION, COMPLETED responses	ResponseResolutionStep
ce_config	dialogue/prompt keys	DialogueActStep/Schema steps
ce_policy	optional policy blocks	PolicyEnforcementStep

Seed snippet: pending action + rules

SQL

INSERT INTO ce_pending_action
(intent_code, state_code, action_key, bean_name, method_names, priority, enabled, description)
VALUES
('CONNECTION_TRANSFER', 'AWAITING_CONFIRMATION', 'FINALIZE_CONNECTION_TRANSFER',
'faqRuleTask', 'injectContainerData', 1, true,
'Demo finalize action for connection transfer');

INSERT INTO ce_rule (phase, intent_code, state_code, rule_type, match_pattern, "action", action_value, priority, enabled)
VALUES
('PRE_RESPONSE_RESOLUTION', 'CONNECTION_TRANSFER', 'IDLE', 'REGEX', '.*', 'SET_STATE', 'COLLECT_INPUTS', 10, true),
('PRE_RESPONSE_RESOLUTION', 'CONNECTION_TRANSFER', 'COLLECT_INPUTS', 'JSON_PATH',
'$[?(@.state == ''COLLECT_INPUTS'' && @.customerId && @.phone && @.email && @.sourceCity && @.targetCity)]',
'SET_STATE', 'AWAITING_CONFIRMATION', 100, true),
('PRE_RESPONSE_RESOLUTION', 'CONNECTION_TRANSFER', 'AWAITING_CONFIRMATION', 'JSON_PATH',
'$[?(@.state == ''AWAITING_CONFIRMATION'' && (@.pending_action_result == ''EXECUTED'' || @.inputParams.pending_action_result == ''EXECUTED''))]',
'SET_STATE', 'COMPLETED', 200, true);

Audit + Payload Postmortem

Confirmation turn audit chain (must appear in order)

Order	Stage	Payload keys to verify
1	DIALOGUE_ACT_CLASSIFIED	dialogueAct=AFFIRM, confidence, source
2	INTERACTION_POLICY_DECIDED	policyDecision=EXECUTE_PENDING_ACTION
3	PENDING_ACTION_LIFECYCLE	status=IN_PROGRESS
4	GUARDRAIL_ALLOW	result=ALLOW
5	PENDING_ACTION_EXECUTED	pendingActionResult=EXECUTED
6	RULE_APPLIED	state transition to COMPLETED
7	ASSISTANT_OUTPUT	final completion response
8	MEMORY_UPDATED	summary updated

Audit payload checkpoint: PENDING_ACTION_EXECUTED

JSON

{
  "_meta": {
    "stage": "PENDING_ACTION_EXECUTED"
  },
  "policyDecision": "EXECUTE_PENDING_ACTION",
  "pendingActionRef": "faqRuleTask:injectContainerData",
  "pendingActionResult": "EXECUTED",
  "intent": "CONNECTION_TRANSFER",
  "state": "AWAITING_CONFIRMATION"
}

Audit payload checkpoint: RULE_APPLIED after execution

JSON

{
  "_meta": {
    "stage": "RULE_APPLIED"
  },
  "rulePhase": "PRE_RESPONSE_RESOLUTION",
  "intent": "CONNECTION_TRANSFER",
  "ruleStateCode": "AWAITING_CONFIRMATION",
  "action": "SET_STATE",
  "actionValue": "COMPLETED"
}

Fast failure triage

If user says "yes" but flow loops back to confirmation, check in order:

1. DIALOGUE_ACT_CLASSIFIED must be AFFIRM
2. INTERACTION_POLICY_DECIDED must be EXECUTE_PENDING_ACTION
3. PENDING_ACTION_EXECUTED must exist
4. RULE_APPLIED must transition state from AWAITING_CONFIRMATION to terminal state.

Example 3: ORDER_CANCELLATION (multi-turn with interaction policy)

Chat Transcript

Protected Order Cancellation

User

Cancel my shoe order.

Info

• Cancellation intent request.
• Order number not yet present.

ConvEngine

I can help you cancel your order. What is your order number?

Tables

• ce_output_schema
• ce_response

Info

• Missing field follow-up.
• Collect orderNumber before confirm step.

User

It's 48392.

Info

• Provides required orderNumber.
• Unlocks confirmation stage.

ConvEngine

Are you sure you want to cancel order 48392? This action is irreversible.

Tables

• ce_rule
• ce_pending_action

Info

• Awaiting explicit confirmation.
• Pending action in OPEN/IN_PROGRESS path.

User

Yes please do that.

Info

• AFFIRM turn.
• Interaction policy decides EXECUTE_PENDING_ACTION.

ConvEngine

Order 48392 has been completely cancelled.

Tables

• ce_audit
• ce_response

Info

• Action executed and state transitioned.
• Cancellation response returned.

Per-turn Step Trace

Action Policy Trace

Turn	Pipeline Step	Internal Engine Action
Turn 1	SchemaExtractionStep	Detects `CANCEL_ORDER`. Fails to extract `orderNumber`. State remains `IDLE`.
Turn 1	ResponseResolutionStep	Outputs IDLE message asking for order number.
Turn 2	SchemaExtractionStep	Engine extracts `orderNumber: 48392`. Flags `schemaComplete=true`.
Turn 2	RulesStep	Auto-advance rule fires, transitioning state to `CONFIRM_CANCEL`.
Turn 2	ActionLifecycleStep	Stages `stripeCancelTask` into `pending_action_runtime` context with `OPEN` status.
Turn 2	ResponseResolutionStep	Generates derived string parsing the memory context `orderNumber`.
Turn 3	DialogueActStep	User says 'Yes'. Act classifier marks input as `AFFIRM`.
Turn 3	InteractionPolicyStep	Matrix outputs `EXECUTE_PENDING_ACTION`. Locks classifier logic.
Turn 3	GuardrailStep	Checks patterns for malicious prompt. Trace output: `GUARDRAIL_ALLOW`.
Turn 3	PendingActionStep	Execution triggered. External API drops order. Status updated to `EXECUTED`.
Turn 3	RulesStep	Rule sees `EXECUTED` task status and triggers state transition to `CANCELLED`.
Turn 3	ResponseResolutionStep	Resolves final derived string using historical `orderNumber`.

Required ce_* Rows

Dynamic Slot & Task DML

SQL

INSERT INTO ce_intent (intent_code, enabled) VALUES ('CANCEL_ORDER', true);

INSERT INTO ce_output_schema (intent_code, format_type, schema_json, enabled)
VALUES ('CANCEL_ORDER', 'JSON', '{"type":"object","properties":{"orderNumber":{"type":"string"}},"required":["orderNumber"]}', true);

INSERT INTO ce_prompt_template (intent_code, response_type, system_prompt, enabled)
VALUES ('CANCEL_ORDER', 'SCHEMA_COLLECTION', 'Extract JSON schema fields.', true);

INSERT INTO ce_rule (intent_code, state_code, rule_type, rule_value, action, action_value, phase, priority, enabled)
VALUES ('CANCEL_ORDER', 'IDLE', 'JSON_PATH', '$[?(@.context.orderNumber != null && @.context.orderNumber != '''')]', 'SET_STATE', 'CONFIRM_CANCEL', 'PRE_RESPONSE_RESOLUTION', 10, true);

INSERT INTO ce_response (intent_code, state_code, response_type, exact_text, enabled)
VALUES ('CANCEL_ORDER', 'IDLE', 'EXACT', 'I can help you cancel your order. What is your order number?', true);

INSERT INTO ce_prompt_template (intent_code, response_type, system_prompt, user_prompt, template_id, enabled)
VALUES ('CANCEL_ORDER', 'DERIVED', 'Format order number cleanly.', 'Are you sure you want to cancel order {{context.orderNumber}}? This action is irreversible.', 'tmpl_confirm_cancel', true);

INSERT INTO ce_response (intent_code, state_code, response_type, template_id, enabled)
VALUES ('CANCEL_ORDER', 'CONFIRM_CANCEL', 'DERIVED', 'tmpl_confirm_cancel', true);

INSERT INTO ce_pending_action (intent_code, state_code, action_key, bean_name, method_names, priority, enabled, description)
VALUES ('CANCEL_ORDER', 'CONFIRM_CANCEL', 'stripe_cancel', 'stripeCancelTask', 'execute', 10, true, 'Cancel the order in the provider system');

INSERT INTO ce_rule (intent_code, state_code, rule_type, rule_value, action, action_value, phase, priority, enabled)
VALUES ('CANCEL_ORDER', 'CONFIRM_CANCEL', 'EXACT', 'ANY', 'SET_STATE', 'CANCELLED', 'PRE_RESPONSE_RESOLUTION', 20, true);

INSERT INTO ce_prompt_template (intent_code, response_type, system_prompt, user_prompt, template_id, enabled)
VALUES ('CANCEL_ORDER', 'DERIVED', 'Format cleanly.', 'Order {{context.orderNumber}} has been completely cancelled.', 'tmpl_cancelled', true);

INSERT INTO ce_response (intent_code, state_code, response_type, template_id, enabled)
VALUES ('CANCEL_ORDER', 'CANCELLED', 'DERIVED', 'tmpl_cancelled', true);

Audit + Payload Postmortem

Confirmation turn audit chain

Order	Stage	Must contain
1	DIALOGUE_ACT_CLASSIFIED	dialogueAct=AFFIRM
2	INTERACTION_POLICY_DECIDED	policyDecision=EXECUTE_PENDING_ACTION
3	GUARDRAIL_ALLOW	result=ALLOW
4	PENDING_ACTION_EXECUTED	pendingActionResult=EXECUTED
5	RULE_APPLIED	state transition to CANCELLED
6	ASSISTANT_OUTPUT	final response

Audit payload checkpoint: PENDING_ACTION_EXECUTED

JSON

{
  "_meta": {
    "stage": "PENDING_ACTION_EXECUTED"
  },
  "policyDecision": "EXECUTE_PENDING_ACTION",
  "pendingActionRef": "stripeCancelTask",
  "pendingActionResult": "EXECUTED",
  "intent": "CANCEL_ORDER",
  "state": "CONFIRM_CANCEL"
}

Example 4: INVENTORY_LOOKUP (Tool Orchestration)

Chat Transcript

DB Orchestration Flow

User

How many running shoes are left in warehouse B?

Info

• Inventory question needing tool/data call.
• Intent resolves to DATABASE_QUERY flow.

ConvEngine

Checking the inventory database...

Tables

• ce_mcp_tool

Info

• Tool dispatch started.
• Waiting for adapter/tool result.

User

[System delay executing Tool]

Info

• Represents in-flight tool call latency.
• No state mutation by user here.

ConvEngine

We currently have 421 units of running shoes allocated to warehouse B.

Tables

• ce_audit
• ce_response

Info

• tool_result consumed by response derivation.
• Final count rendered for user.

Per-turn Step Trace

Tool Dispatch Trace

Pipeline Step	Internal Engine Action
IntentResolutionStep	Detects `DATABASE_QUERY` intent. State=`IDLE`.
ToolOrchestrationStep	Identifies `INVENTORY_DB` mapping. Routes payload to connected Java component adapting `McpToolExecutor`.
Tool Execution (Java Adapter)	Adapter processes input, runs `SELECT count FROM inventory`. Returns JSON payload `{'count': 421}`.
ToolOrchestrationStep	Writes returned JSON into engine session variable `tool_result.dbList`. Traces `TOOL_ORCHESTRATION_RESULT`.
RulesStep	Evaluates optional `POST_TOOL_EXECUTION` rules.
ResponseResolutionStep	Injects the `421` count into the derived template context and calls LLM.

Required ce_* Rows

Tool Orchestration DML

SQL

INSERT INTO ce_intent (intent_code, enabled) VALUES ('DATABASE_QUERY', true);

INSERT INTO ce_mcp_tool (intent_code, tool_group, required_params_json, execution_timeout_ms, enabled)
VALUES ('DATABASE_QUERY', 'INVENTORY_DB', '["itemType", "warehouseId"]', 5000, true);

INSERT INTO ce_prompt_template (intent_code, response_type, system_prompt, user_prompt, template_id, enabled)
VALUES ('DATABASE_QUERY', 'DERIVED', 'You are an inventory assistant.', 'Answer from these database results: {{tool_result.dbList}}', 'tmpl_inventory_gen', true);

INSERT INTO ce_response (intent_code, state_code, response_type, template_id, priority, enabled)
VALUES ('DATABASE_QUERY', 'IDLE', 'DERIVED', 'tmpl_inventory_gen', 10, true);

Audit + Payload Postmortem

Tool Orchestration Checkpoints

Order	Stage	Payload details
1	TOOL_ORCHESTRATION_DISPATCH	toolGroup=INVENTORY_DB, toolParams
2	TOOL_ORCHESTRATION_RESULT	toolResult={'count': 421}, executionMs
3	ASSISTANT_OUTPUT	final response interpolation

Example 5: SUPPORT_DIAGNOSTIC (Memory Summarization)

Chat Transcript

Long Context Support Thread

User

My router is broken.

Info

• Initial problem statement.
• Starts long troubleshooting flow.

ConvEngine

Let's check the lights. Are they green?

Info

• Assistant begins diagnostic questions.
• No tool/action execution yet.

User

[12 more dialogue turns testing different cables and lights]

Info

• Represents many intermediate turns.
• Conversation memory grows significantly.

ConvEngine

I suggest performing a factory reset via http://router.local/admin_reset with PIN 1234.

Tables

• ce_conversation_history

Info

• Assistant gives reset URL and PIN.
• Critical details should be retained in memory summary.

User

[User is quiet for minutes]

Info

• Time gap before follow-up.
• Memory recall scenario.

User

[Turn 15]: Wait, what link did you tell me to use for the firmware reset again?

Info

• Late-turn recall question.
• Requires memory-assisted response.

ConvEngine

Based on our previous troubleshooting, you should use http://router.local/admin_reset with the default PIN 1234.

Tables

• ce_audit
• ce_conversation_history

Info

• Memory summary injected.
• Assistant recalls prior URL and PIN correctly.

Per-turn Step Trace

Memory Compression Trace (Turn 15)

Pipeline Step	Internal Engine Action
MemoryStep (Background)	Fetches raw history and asks LLM to re-summarize it into a tight timeline snapshot. Persists to `ce_memory`.
DialogueActStep	User asks 'what link did you tell me to use?'. Abstracted act=`QUESTION`.
MemoryStep (Injection)	Injects the compressed `memory.session_summary` variable directly into current Turn 15 context.
IntentResolutionStep	Resolves exact `SUPPORT_DIAGNOSTIC` intent leveraging memory context.
ResponseResolutionStep	Template retrieves the exact `http://router.local` string generated 10 turns ago.

Required Configuration

Memory Auto-Summarization Config

YAML

convengine:
  flow:
  memory:
    enabled: true
    # Max token size for the compressed summary string
    summary-max-chars: 1200
    # Number of turns to trigger a new summary compression cycle
    recent-turns-for-summary: 3

Audit + Payload Postmortem

Memory Checkpoints

Order	Stage	Payload details
1	MEMORY_EVALUATED	evaluatedTurns, previousSummaryLen
2	MEMORY_UPDATED	newSummary generated and injected
3	INTENT_RESOLVED	context aware resolution