Intent and Schema Internals

Dialogue act resolution

DialogueActStep handles linguistic classification before intent resolution even begins. It supports three configuration modes via convengine.flow.dialogue-act.resolute:

• REGEX_ONLY
• REGEX_THEN_LLM
• LLM_ONLY

Core Dialogue Acts

Act	Meaning	Regex Sample
AFFIRM	User agrees or confirms	(?i)^\\s*(yes|yep|yeah|sure)\\s*$
NEGATE	User disagrees or cancels	(?i)^\\s*(no|nope|cancel)\\s*$
ANSWER	User answers an active clarification or slot question	short direct answer in active flow
EDIT	User requests a correction	(?i)^\\s*(edit|change|modify)\\s*$
RESET	User requests a fresh start	(?i)^\\s*(reset|start over)\\s*$
QUESTION	User asks a clarifying question	text ending in '?'
NEW_REQUEST	(Fallback) User begins a new topic	default fallback

When the LLM path is engaged (confidence below llm-threshold), it must return a strict JSON payload containing the act and confidence to be parsed by the engine. It can also derive standaloneQuery, which becomes resolved_user_input when present.

Interaction policy decision

InteractionPolicyStep derives the routing policy for the current turn based on existing context.

Engine Routing Policies

Policy Decision	Trigger Context	Pipeline Impact
EXECUTE_PENDING_ACTION	DialogueAct=AFFIRM + hasPendingAction	Triggers PendingActionStep; Skips Intent Resolution
REJECT_PENDING_ACTION	DialogueAct=NEGATE + hasPendingAction	Marks Pending task REJECTED; Skips Intent Resolution
FILL_PENDING_SLOT	DialogueAct!=NEW_REQUEST + hasPendingSlot	Passes input to SchemaExtractionStep; Skips Intent Resolution
RECLASSIFY_INTENT	DialogueAct=NEW_REQUEST (or no pending state)	Forces standard Intent Resolution

Confirmation routing and correction

CorrectionStep runs after InteractionPolicyStep and before intent/schema rework. It handles confirmation-state turns without forcing full re-extraction:

• AFFIRM in confirmation state can set routing_decision=PROCEED_CONFIRMED
• EDIT in confirmation state can patch a single field in-place and keep the flow in CONFIRMATION
• downstream rules can use POST_SCHEMA_EXTRACTION and PRE_AGENT_MCP to route the next state deterministically

Intent resolver behavior

The pipeline resolves intent in order of priority via CompositeIntentResolver.

Lock and Skip Behaviors

The engine prevents intent drift during active schema collection automatically. You will see these audit signals in the trace when intent resolution is intentionally bypassed:

• INTENT_RESOLVE_SKIPPED_POLICY
• INTENT_RESOLVE_SKIPPED_SCHEMA_COLLECTION
• INTENT_RESOLVE_SKIPPED_STICKY_INTENT

Schema extraction

SchemaExtractionStep drives missing-field follow-up flows and automatically formats prompt templates utilizing ce_prompt_template.

Schema LLM Injection Template

JSON

Runtime facts are evaluated and written into the EngineSession context by AutoAdvanceStep, making them available for ce_rule evaluation:

• missing_fields
• missing_field_options
• schema_json
• schema_complete

Once extraction is complete, POST_SCHEMA_EXTRACTION rules can move the flow into confirmation without waiting for MCP execution.

Practical design rule

Use schema extraction for required slot collection, and use dialogue acts + interaction policies for short confirmations (like replying "yes" to "Are you sure?"). This ensures intent reclassification does not disrupt your active multi-turn flows.