Skip to main content
v2

ConvEngine Overview

ConvEngine is a deterministic conversation workflow engine with explicit turn-control steps for dialogue act, policy, pending-action lifecycle, guardrails, tool orchestration, state-graph validation, and memory.

Core Features​

  • Turn understanding before intent re-resolution:
    • DialogueActStep
    • InteractionPolicyStep
    • CorrectionStep
  • Pending action runtime control:
    • ActionLifecycleStep
    • DisambiguationStep
    • PendingActionStep
  • Safety controls:
    • GuardrailStep
    • StateGraphStep (validate-only)
  • Tool runtime unification:
    • ToolOrchestrationStep
    • McpToolStep with executor-per-tool-group
  • Memory and replay:
    • MemoryStep
    • ConversationReplayService
  • Rule phases expanded with schema/MCP hooks:
    • POST_SCHEMA_EXTRACTION
    • PRE_AGENT_MCP
    • POST_AGENT_MCP
    • POST_TOOL_EXECUTION
  • New control tables:
    • ce_pending_action
    • ce_verbose (runtime progress/error messaging)

Runtime model​

ConvEngine remains config-first:

  • DB control plane: ce_intent*, ce_rule, ce_response, ce_prompt_template, ce_pending_action, ce_verbose, ce_mcp_*
  • Runtime policy config: convengine.flow.*
  • Deterministic pipeline with audited stage output

Key outcomes​

  • Handles yes / ok / do that against pending context, not standalone intent.
  • Avoids hidden state mutation by validating transitions (StateGraphStep) and auditing violations.
  • Supports multiple tool groups (DB, HTTP_API, WORKFLOW_ACTION, DOCUMENT_RETRIEVAL, CALCULATOR_TRANSFORM, NOTIFICATION, FILES).
  • Keeps every control decision visible in audit timeline.

Next Reads​