MCP and Audit Internals

MCP flow

McpToolStep does:

1. skip on pending clarification
2. list enabled tools
3. clear stale context_json.mcp branch
4. planner loop (CALL_TOOL or ANSWER)
5. DB tool execution via McpDbExecutor
6. store observations and final answer in context

MCP-related tables

MCP table usage

Table	Role	Key fields
ce_mcp_tool	Tool registry	tool_code, tool_group, intent_code, state_code, enabled
ce_mcp_db_tool	DB tool SQL template config	sql_template, param_schema, max_rows, safe_mode
ce_config	Planner prompts	McpPlanner.SYSTEM_PROMPT, McpPlanner.USER_PROMPT

Audit stream model

DbAuditService writes normalized payload with _meta block, then:

• applies AuditStageControl (level/include/exclude/rate controls)
• dispatches listeners through AuditEventDispatcher (sync by default, async optional)

Transport outputs:

• SSE stream (/api/v1/conversation/stream/{conversationId})
• optional STOMP topic (/topic/convengine/audit/{conversationId})
• STOMP broker mode: simple in-memory (default) or relay (RabbitMQ/other STOMP relay)

Trace API

GET /api/v1/conversation/audit/{conversationId}/trace

Returns:

• step timeline
• ordered stage list
• source class/file metadata

SSE subscribe example

BASH

curl -N http://localhost:8080/api/v1/conversation/stream/9bf7540a-b129-4685-a120-730e8a0cb94b

Debug strategy

When output is wrong, compare raw audit stages with trace step boundaries. This isolates whether issue came from resolver logic, rule mutation, MCP plan, or response mapping.