Rules and Response Resolution

Rule types currently implemented

Rule type resolver coverage

rule_type	Resolver	When to use
EXACT	ExactRuleTypeResolver	Exact text-driven trigger
REGEX	RegexRuleTypeResolver	Pattern-driven trigger
JSON_PATH	JsonPathRuleTypeResolver	Condition over session/context/payload facts

Rule actions in detail

Action behavior

action	What it changes	Typical real-world use
SET_STATE	session.state	Move flow to AWAITING_ACCOUNT_ID
SET_INTENT	session.intent	Reclassify after deterministic signal
SET_JSON	input param from JSON path extract	Pull ticketId/correlationId from context
GET_CONTEXT	input param snapshot of context	Use context in downstream prompts/tasks
GET_SESSION	input param snapshot of session facts	Expose runtime flags to prompts
GET_SCHEMA_JSON	input param schema-only field subset	Use only extracted form fields
SET_TASK	executes custom task bean methods	Raise incident, call internal orchestrator

Response selection algorithm

1. Find enabled ce_response candidates matching intent/state
2. Score by exact state/intent > ANY/null fallback > priority
3. Choose top candidate
4. Resolve by response type

EXACT vs DERIVED

• EXACT: deterministic payload (exact_text)
• DERIVED: template-rendered + LLM-generated output

Derived response path (concept)

TEXT

ce_response(response_type=DERIVED, output_format=TEXT)
 -> template from ce_prompt_template(response_type=TEXT)
 -> render variables
 -> llm.generateText(...)
 -> optional ResponseTransformer
 -> persisted assistant output

Intervention points

• Use ResponseTransformer after response resolution to apply consumer-specific post-processing.
• Use hook on RulesStep to inspect rule cascades in production debugging.

Enum drift warning

Do not insert unimplemented rule types/actions in DB. Keep ce_rule values aligned with installed resolver classes.