Intent Resolution and Schema Collection

Intent resolution sources

1. Deterministic classifier rules (ce_intent_classifier)
2. Agent scoring path (AgentIntentResolver) using ce_config prompts
3. Composite arbitration in CompositeIntentResolver

ce_config keys that shape intent behavior

AgentIntentResolver keys

config_key	Purpose	Example
MIN_CONFIDENCE	Minimum confidence to accept agent intent	0.55
COLLISION_GAP_THRESHOLD	Top-score gap threshold for collision	0.10
SYSTEM_PROMPT	System behavior contract	JSON-only intent scoring rules
USER_PROMPT	User template with context/history/allowed intents	{{context}}, {{allowed_intents}}, {{user_input}}

IntentResolutionStep keys

config_key	Purpose	Example
STICKY_INTENT	Keep intent/state stable only during active incomplete schema collection	true

ce_prompt_template for schema extraction

SchemaExtractionStep resolves template by:

• response_type = SCHEMA_JSON
• intent_code
• state_code (or fallback)

Schema extraction prompt pattern

TEXT

Intent lock behavior

When schema is incomplete:

• session.intentLocked = true
• IntentResolutionStep skips re-resolving to prevent intent drift
• lock reason is persisted in context (intent_lock.reason)

When schema becomes complete:

• lock is released
• if STICKY_INTENT=true, resolver can skip and retain existing intent/state only while schema collection remains incomplete
• explicit switch/reset/force signals still allow re-resolution

Why this matters

Without lock behavior, multi-turn collection flows can jump to a new intent after partial input and break deterministic progression.