Skip to content

Ean Schuessler / mo-mcp

Go to a project
Toggle navigation
Toggle navigation pinning
Switch branch/tag
  1. 05 Feb, 2026 7 commits
    • Refactor Agent Runtime: General-purpose LLM Request Service · ab967d1c 
      BREAKING: Introduces new LLM Request/Response pattern for CSR agents

Major Changes:
- Add new SystemMessage types: SmtyLlmRequest, SmtyLlmResponse
- Extend SystemMessage entity with callback and audit fields:
  * parentSystemMessageId - Links Response → Request
  * callbackServiceName - Service to call when LLM completes
  * callbackParameters - JSON params for callback
  * sourceTypeEnumId/sourceId - Audit trail (Order, CommEvent, etc.)
  * llmResponse - Raw response stored for debugging

New Services:
- AgentServices.process#LLMRequest: General-purpose async LLM service
  * Any trigger (SECA, UI, Order, etc.) can call this
  * Creates SmtyLlmRequest SystemMessage
  * Triggers async processing via poller
  * Validates callback service exists before creating task

Refactored Services:
- AgentServices.run#AgentTaskTurn: Universal agent processor
  * Supports both SmtyAgentTask (legacy) and SmtyLlmRequest (new)
  * ALWAYS provides MCP tools to LLM (for CSR agent pattern)
  * Creates SmtyLlmResponse and calls callback for new pattern
  * Maintains backward compatibility with SmtyAgentTask

- AgentServices.callback#CommunicationEvent: Saves LLM responses to conversation
  * Callback service for CommunicationEvent-triggered requests
  * Maintains conversation thread via rootCommEventId

Updated SECA:
- AgentTriggerOnCommunication now calls AgentServices.process#LLMRequest
  * Uses callback pattern instead of direct SystemMessage creation
  * Enables full audit trail via SystemMessage Request :left_right_arrow: Response

Benefits:
- General-purpose: Any trigger can request LLM processing (orders, inventory, support, etc.)
- Traceability: Full audit trail via linked SystemMessages
- RBAC: Agent impersonates users, respects permissions
- Same UX: Agent uses same screens humans use (via MCP tools)
- Flexible: Different callbacks handle responses differently
      Ean Schuessler committed
    • Enhance agent reliability and debug logging · 5c6a9826 
      - AgentServices.xml: Added authenticate='false' to OpenAiChatCompletion, added logging to AgentQueue poller
- Agent.secas.xml: Added extensive logging to trigger SECA
- AgentData.xml: Fixed SmtyAgentTask to use run#AgentTaskTurn as consumeServiceName
      Ean Schuessler committed
    • Implement Reliable Agent Architecture using CommunicationEvent thread logic · 39daa815 
      - Update AgentServices.xml: Single-turn state machine (process one turn, re-queue if tool called)
- Add Agent.secas.xml: Trigger Agent Turn on new CommunicationEvent to Agent Party
- Update AgentEntities.xml: Add rootCommEventId to SystemMessage for thread tracking
- Update AgentData.xml: Define Agent Party, Role, and default VLLM config
      Ean Schuessler committed
    • Add empty receiveServiceName to SmtyAgentTask · 85eedd26 
      Prevent receive#IncomingSystemMessage from auto-consuming messages
by setting receiveServiceName to empty string instead of null.
This allows AgentQueuePoller to pick up SmtyAgentTask messages before
they are consumed by the receive service.
      Ean Schuessler committed
    • Restore consumeServiceName to SmtyAgentTask · a8843c73 
      The data file in runtime/component directory didn't have consumeServiceName,
so adding it back. Needs rebuild to load properly.
      Ean Schuessler committed
    • Remove consumeServiceName from SmtyAgentTask · 6b4c8b0a 
      This fixes race condition where receive#IncomingSystemMessage
was immediately consuming the message instead of letting
AgentQueuePoller pick it up. Now:
- Message created with SmsgReceived status
- AgentQueuePoller (calls poll#AgentQueue) finds and processes it
      Ean Schuessler committed
    • Fix duplicate XML content in AgentServices · 5db18bb1 
      Remove duplicate/corrupted content at end of file
- Lines were orphaned from previous edit
- File now has proper structure with 3 services
      Ean Schuessler committed
  3. 04 Feb, 2026 17 commits
  5. 30 Jan, 2026 1 commit
  7. 28 Jan, 2026 2 commits
    • Refactor MCP servlet to adapter architecture · 3a39f97b 
      Introduce clean adapter layer between Moqui infrastructure and MCP protocol:

- transport/MoquiMcpTransport: Interface abstracting transport concerns
- transport/SseTransport: SSE implementation with session management
- adapter/McpSessionAdapter: Maps Moqui Visit to MCP sessions
- adapter/McpToolAdapter: Maps MCP tools/methods to Moqui services
- adapter/MoquiNotificationMcpBridge: Bridges Moqui notifications to MCP

Simplify EnhancedMcpServlet to orchestrator role, removing inline session
management, SSE logic, and tool dispatch. Remove redundant session
validation in Initialize service (MoquiAuthFilter handles auth).

Delete obsolete files:
- VisitBasedMcpSession.groovy (replaced by McpSessionAdapter)
- JsonRpcMessage.groovy (using plain Maps)
- MoquiMcpTransport.groovy (replaced by new interface)
      Ean Schuessler committed
    • Add workflow indexing and improve documentation · 9aefba5e 
      - Index business process workflows from BUSINESS_PROCESSES wiki space
- Add workflow support to search and help tools
- Improve documentation with better organized help content
- Add type attributes to seed data files
- Enhance error handling for workflow loading
      Ean Schuessler committed
  9. 23 Jan, 2026 1 commit
    • Improve ARIA grid rendering and action result feedback · d6d8f38c 
      - Fix column extraction to include fields with both search and display widgets
- Add all ID fields to row ref as pipe-delimited key=value pairs for unambiguous LLM parsing
- Show NULL values as '(none)' in grid cells for consistency
- Add ARIA-compliant action result feedback (role=alert for errors, role=status for success)
- Pass actionResult through BrowseScreens ARIA mode
- Hide batch_operations tool until fixed
- Default renderMode to 'aria' instead of 'compact'
      Ean Schuessler committed
  11. 21 Jan, 2026 9 commits
    • Extract actual entity IDs from row links for service calls · 33352aca 
      When pseudoId differs from the internal productId (e.g., DEMO_VAR_GR_SM_COPY
vs 100204), models were using the display ID for service calls causing FK errors.

Now extracts the actual entity ID from the row's link URL and includes it in
the row data:
- compact: {id: 'DEMO_VAR_GR_SM_COPY', productId: '100204', link: '...'}
- aria: {ref: 'DEMO_VAR_GR_SM_COPY', productId: '100204', ...}

This is a generic solution that works for any entity type by parsing the
primary key parameter from the Edit link URL.

Fixes QA issue where models used pseudoId for applyProductFeatures causing
'record does not exist' referential integrity errors.
      Ean Schuessler committed
    • Add nextCursor pagination to grid responses · 7f95b312 
      - FTL: Capture Moqui pagination vars (pageIndex, pageSize, pageMaxIndex, listCount)
- Compact mode: Add nextCursor object with pageIndex, pageSize, hasMore
- ARIA mode: Add nextCursor to grid nodes
- Add totalCount to show full result size

Usage: Pass nextCursor values as parameters to get next page:
  parameters: {pageIndex: 1, pageSize: 20}

Response format:
  nextCursor: {pageIndex: 1, pageSize: 20, hasMore: true}
  totalCount: 29
      Ean Schuessler committed
    • Add search form discovery for form-lists · 0153ca18 
      - FTL: Detect header-field search widgets (text-find, date-find, etc.)
- FTL: Add 'searchable' flag and search-specific types to field metadata
- ARIA: Add role='search' landmark before grids with searchable fields
- Compact: Add searchParams array listing filterable parameters
- Filter out 'get*' transitions (autocomplete helpers) from transitions list

This helps LLM clients understand how to filter form-list results by
passing parameters, rather than trying to call autocomplete transitions.
      Ean Schuessler committed
    • Remove xml render mode from MCP options · 98e52cc1 
      XML mode renders nearly empty screen hierarchy (just path structure, no content
or data), which isn't useful for MCP clients.

Remaining render modes:
- compact: structured actionable summary (default)
- aria: accessibility tree for field discovery
- text: plain text rendering
- html: full HTML rendering
      Ean Schuessler committed
    • Remove vuet/qvt render modes from MCP options · f18695f7 
      These modes render Vue.js wrapper templates for web frontends, not actual
screen content. They only show outer container HTML with subscreen placeholders,
which isn't useful for MCP clients.

Remaining render modes:
- compact: structured actionable summary (default)
- aria: accessibility tree for field discovery
- text: plain text rendering
- html: full HTML rendering
- xml: XML rendering
      Ean Schuessler committed
    • Remove 'mcp' render mode from user-facing options · 691089f7 
      The mcp render mode exposed too much internal detail (service names, full
metadata) which confused smaller models into trying to call services directly.

Changes:
- Remove 'mcp' from renderMode descriptions in tool definitions
- Map renderMode='mcp' to 'compact' if explicitly requested
- Update AGENTS.md to remove mcp mode documentation
- Internal rendering still uses mcp mode, just not exposed to users

Users should use:
- compact (default): actionable summary
- aria: accessibility tree for field discovery
- text/html: rendered output
      Ean Schuessler committed
    • Update wiki docs with guidance on using MCP tools · 932ea3f4 
      Adds clear instructions in root wiki doc explaining:
- How to execute actions (use action parameter, NOT service calls)
- How to get dropdown options (use moqui_get_screen_details)
- How to get action documentation (use moqui_get_help with help URI)
- Updated screen paths to use apps/PopcAdmin format
- Added global navigation paths (Messages, Tasks, Calendar)

This helps models understand how to use the system after we removed
service names and dropdown examples from responses.
      Ean Schuessler committed
    • Reduce noise in MCP responses - remove unactionable context · 1b58daf3 
      ARIA and compact modes now cleaner for smaller models:

- Remove service names from actions (confuses models into trying to call services)
- Rename 'describedby' to 'help' for clarity (points to moqui_get_help tool)
- Remove wiki:screen references (not useful, wiki content already in response)
- Fix grid row refs to use specific IDs (productFeatureId, toProductId) instead
  of repeating generic productId for all rows
- Filter navigation links: remove cross-app links, remove action URLs with
  encoded timestamps (fromDate/thruDate parameters)
- Remove dropdown options counts and random examples - just indicate type,
  use moqui_get_screen_details for actual options

This significantly reduces context volume and eliminates patterns that
caused smaller models to hallucinate service calls.
      Ean Schuessler committed
    • Remove truncation from ARIA mode - show all rows, links, actions · b52c9ca0 
      ARIA mode was truncating grid rows to 3, links to 10, and actions to 15.
This caused models to miss important data (e.g., only seeing 3 of 8 features
or 3 of 16 variant associations).

Now ARIA mode shows complete data like text mode does.
      Ean Schuessler committed
  13. 20 Jan, 2026 3 commits
    • Add global navigation to MCP browse_screens response · c4610d5e 
      Adds globalNav array to every browse_screens response with links to:
- My Notifications (apps/my/User/Notifications)
- My Messages (apps/my/User/Messages/FindMessage)
- My Calendar (apps/my/User/Calendar/MyCalendar)
- My Tasks (apps/my/User/Task/MyTasks)

These mirror the MyAccountNav component in the web UI and are always
available regardless of which app the user is currently browsing.

Also adds a hint in uiNarrative.notes about globalNav and search.
      Ean Schuessler committed
    • Fix: Remove userId/username injection from screen render params · c36293a2 
      The username parameter was being injected into screen render params, which
conflicted with FindCustomer's 'username' search field. This caused the
screen to filter results to only show customers with username matching
the current user (john.sales*), returning 0 results.

User context is already available via ec.user for authorization purposes,
so injecting these params was redundant and harmful.
      Ean Schuessler committed
    • Fix validation error reporting and clarify feature service docs · 72108ddb 
      Critical fix: Errors now correctly captured and reported
- Capture errors immediately after render() in CustomScreenTestImpl
- Added saveMessagesToSession() to WebFacadeStub for framework compatibility
- Actions that fail validation now report status:'error' instead of false success

Documentation fix: Clarify two different feature services
- applyProductFeatures: Apply EXISTING features to products (most common)
- createProductFeature: Create NEW feature definitions (rare)
- Updated EditProduct and ProductFeature wiki docs with correct examples
      Ean Schuessler committed