## Context
Following https://github.com/twentyhq/twenty/pull/16399
Now using the new global orm manager everywhere and returning a
GlobalDatasource/WorkspaceDatasource based on a feature flag.
This means we now need to wrap all our ORM calls within
executeInWorkspaceContext callback (at least for now) so the global
datasource can dynamically hydrate its context via the new store (the
global datasource does not store anything related to workspaces as it is
now a unique singleton). If feature flag is off it still uses local data
stored in the workspace datasource.
## Summary
This PR significantly simplifies the AI chat architecture by removing
complex routing/planning mechanisms and introduces clickable record
links in AI responses.
## Changes
### AI Chat Architecture Simplification
- **Removed** the entire `ai-chat-router` module (~850 lines) including:
- Strategy decider service
- Plan generator service
- Complex routing logic
- **Removed** agent execution planning services (~700 lines):
- `agent-execution.service.ts`
- `agent-plan-executor.service.ts`
- `agent-tool-generator.service.ts`
- **Added** centralized `ToolRegistryService` for tool management:
- Builds searchable tool index (database, action, workflow tools)
- Provides tool lookup by name
- Supports agent search for loading expertise
- **Added** `ChatExecutionService` as simple replacement:
- Includes full tool catalog in system prompt
- Pre-loads common tools (find/create/update for company, person,
opportunity, task, note)
- Uses `load_tools` mechanism for dynamic tool activation
- Enables native web search by default
### Record References in AI Responses
- Added `recordReferences` field to tool outputs for create, find, and
update operations
- Implemented `[[record:objectName:recordId:displayName]]` syntax for AI
to reference records
- Created `RecordLink` component that renders clickable chips with
object icons
- Integrated record link parsing into the markdown renderer
- Users can now click directly on created/found records in AI responses
### Workflow Agent Fixes
- Fixed cache invalidation issue when creating agents in workflows
- Added default prompt for workflow-created agents to prevent validation
errors
- Relaxed agent validation to only check properties being updated (not
all required properties)
### Code Quality Improvements
- Extracted `getRecordDisplayName` utility that mirrors frontend's
`getLabelIdentifierFieldValue` logic
- Uses object metadata to determine the correct label identifier field
- Handles `FULL_NAME` composite type for person/workspaceMember objects
- Shared across create, find, and update record services
## Net Impact
- **~1,200 lines deleted** (complex routing/planning code)
- **~500 lines added** (simpler tool registry + record links)
- Significantly reduced code complexity
- Better tool discovery through full catalog in system prompt
- Improved UX with clickable record references
## Testing
- Typecheck passes
- Lint passes
- Manual testing of AI chat with record creation and linking
## Context
Deprecating TwentyORMManager in favor of TwentyORMGlobalManager
(temporarily, as this will simplify the ultimate goal to later replace
all usages with the new TwentyORMGlobalManagerV2 which will have a
similar signature)
This means this PR had to refactor a bit of code to pass down the
workspaceId when not available directly as it is now a requirement,
meaning we also deprecated scopedWorkspaceContextFactory to have a less
obscure way to fetch the workspaceId and have something more
declarative.
Step 3 will be to update TwentyORMGlobalManager to use a featureFlag
toggling and use the new GlobalWorkspaceOrmManager internally using the
new cache service
Step 4 will be to remove the feature flag and pg_pool patch
## Context
Deprecating the old objectMetadataMap type in favour of split flat
entities to match with our new caching.
In the long run, trying to achieve:
- Better performance through caching
- Consistent data access patterns across the codebase
- Reduced database queries
Now that everything is based on flat entities, which are cached, we can
finish the refactoring of workspace context cache which should already
improve performances.
Then the last step will be to consume that new cache in the new global
datasource to get rid of the many workspace datasources stored in the
server
## Summary
This PR introduces a comprehensive agent evaluation system and refactors
the AI module structure for better organization.
## Key Changes
### 🎯 Agent Evaluation System
- Added **Agent Turn Evaluation** entities, DTOs, and database schema
- New GraphQL mutations: `evaluateAgentTurn` and `runEvaluationInput`
- Added `evaluationInputs` field to Agent entity for storing test inputs
- New `AgentTurnGraderService` for automatic turn evaluation
- Added evaluation UI with new **Evals** and **Logs** tabs in agent
detail pages
### 🏗️ Entity & Module Refactoring
- Renamed `AgentChatMessage` → `AgentMessage` for clarity
- Consolidated chat entities: `AgentMessage`, `AgentTurn`, and
`AgentChatThread`
- Reorganized AI modules under `ai/` subdirectory structure
- Updated imports across codebase to reflect new module paths
### 🤖 New Agents & Roles
- Added **Dashboard Builder Agent** for dashboard creation and
management
- Added **Dashboard Manager Role** with appropriate permissions
- Updated role permissions to be more granular (users vs agents vs API
keys)
### 🔐 Permission System Updates
- Added `HTTP_REQUEST_TOOL` permission flag
- Updated Workflow Manager role permissions (restricted tool access)
- Enhanced permission flag types to differentiate between user/agent/API
key contexts
- Added `isRelevantForAgents`, `isRelevantForApiKeys`,
`isRelevantForUsers` to permission flags
### 📨 Message Role Enhancement
- Added `system` role to `AgentMessageRole` enum (alongside
user/assistant)
- Updated message handling to support system prompts
### 🎨 UI/UX Improvements
- New tabs in agent detail: **Evals** and **Logs**
- Added turn detail page: `/ai/agents/:agentId/turns/:turnId`
- Fixed text overflow in `SettingsListItemCardContent`
- Updated role applicability labels ("Assignable to Workspace Members")
### 🛠️ Technical Improvements
- Fixed Zod schema validation for UUID and Date fields (use string
validators)
- Updated `ToolRegistryService` to properly register HTTP tool with
permission flag
- Enhanced error handling in agent execution services
- Updated database migrations for new entity schema
## Database Migrations
- `1764210000000-add-system-role-to-agent-message.ts`
- `1764220000000-add-evaluation-inputs-to-agent.ts`
- `1764200000000-add-agent-turn-evaluation.ts`
- `1764100000000-refactor-agent-chat-entities.ts`
## Testing
- [ ] Agent evaluation flow tested
- [ ] Dashboard Builder agent tested
- [ ] Permission system validated
- [ ] UI tabs and navigation tested
- [ ] Database migrations run successfully
## Breaking Changes
⚠️ **Entity Rename**: `AgentChatMessage` renamed to `AgentMessage` -
GraphQL queries need updating
## Related Issues
<!-- Link any related issues here -->
## Screenshots
<!-- Add screenshots if applicable -->
## Overview
This PR replaces the dynamic agent handoff system with a more
predictable planning-based router that decides upfront how to handle
multi-agent coordination.
## Major Changes
### 🔄 Architecture Shift: Handoffs → Planning
**Removed:**
- `AgentHandoffEntity` and handoff tracking system
- `AgentHandoffService` and `AgentHandoffExecutorService`
- Dynamic agent-to-agent transfers during execution
- Handoff tool generation and description templates
**Added:**
- `AiRouterService` with two strategies: `simple` (single agent) and
`planned` (multi-agent)
- `AgentPlanExecutorService` for executing multi-step plans
- Plan validation (cycle detection, dependency resolution)
- `UnifiedRouterResult` type with discriminated union
### 🤖 New Standard Agents
Added two new specialized agents:
- **Researcher Agent**: Web search, fact-finding, competitive
intelligence
- **Code Agent**: TypeScript function generation for serverless
workflows
### 🏗️ Router Refactoring (Latest)
Split router responsibilities into focused services:
- `AiRouterStrategyDeciderService`: Decides simple vs planned strategy
- `AiRouterPlanGeneratorService`: Generates and validates execution
plans
- `AiRouterService`: Coordinates between services (reduced from 426→275
lines)
### ⚙️ Configuration Improvements
- Added `outputStrategy` to agent definitions (`direct` vs `synthesize`)
- Removed hardcoded special cases for workflow-builder
- Added `plannerModel` field to workspace entity
- Increased `MAX_STEPS` from 10 to 25 for complex workflows
### 📝 Agent Prompt Refinements
Significantly simplified prompts for better clarity:
- Workflow Builder: 51→36 lines
- Helper: 49→28 lines
- Data Manipulator: Enhanced with sorting guidance
### 🔍 Enhanced Debugging
- Plan reasoning and step count in data message parts
- Router debug info with token usage tracking
- Better logging throughout execution pipeline
## Benefits
1. **Simpler Mental Model**: Router decides upfront vs dynamic transfers
2. **Better Predictability**: Users see the plan before execution
3. **Cleaner Architecture**: SRP with focused services
4. **Configuration Over Code**: Agent behavior via config, not hardcoded
logic
5. **Plan Validation**: Catches invalid dependencies and cycles
## Migration Notes
- Database migration removes `agentHandoff` table
- Adds `plannerModel` column to workspace table
- No API breaking changes (agent endpoints unchanged)
## Testing
- Integration tests updated to remove handoff dependencies
- Agent tool test utilities simplified
- Plan validation covered by new logic
## Next Steps (Future PRs)
- Parallel execution of independent plan steps
- Dynamic re-planning based on results
- Plan caching for common routing patterns
- Error recovery strategies in plan executor
## Summary
This PR adds configurable response format support for AI agents,
allowing them to return either plain text or structured JSON data based
on a defined schema.
## Key Features
### 1. Agent Response Format Configuration
- Added `AgentResponseFormat` type supporting:
- `text`: Returns plain text responses (default)
- `json`: Returns structured JSON based on defined schema
- New `AgentResponseSchema` type moved to `twenty-shared/ai` for sharing
between frontend/backend
### 2. Settings UI
- New `SettingsAgentResponseFormat` component for configuring response
format
- Visual schema builder for defining JSON output structure
- Real-time validation and preview
- Integrated into agent settings tab
### 3. Workflow Integration
- AI Agent workflow action automatically uses agent's configured
response format
- Output schema dynamically generated from agent's response format
- Workflow variable picker shows structured fields for JSON responses
- Backward compatible with existing text-only agents
### 4. Backend Implementation
- Added `convertAgentSchemaToZod` utility to validate JSON responses
- Agent executor service handles both text and JSON generation
- Automatic agent creation/cloning when adding AI agent steps to
workflows
- Unique agent naming with conflict resolution
### 5. Database Migration
- Migration `1763622159656-update-agent-response-format.ts`
- Sets default `responseFormat` to `{"type":"text"}` for existing agents
- Updated all standard agents with proper response format
## Changes by Module
### Frontend (`twenty-front`)
- 🆕 `AgentResponseFormat` type
- 🆕 `SettingsAgentResponseFormat` component
- ✏️ Updated `WorkflowEditActionAiAgent` to support response format
configuration
- 🗑️ Removed deprecated `useAiAgentOutputSchema` hook and
`AiAgentOutputSchema` type
### Backend (`twenty-server`)
- 🆕 `AgentResponseFormat` type in agent entity
- 🆕 `convertAgentSchemaToZod` utility for schema validation
- ✏️ Updated `AiAgentExecutorService` to handle both text and JSON
generation
- ✏️ Updated `WorkflowSchemaWorkspaceService` to generate output schema
from agent config
- ✏️ Enhanced `WorkflowVersionStepOperationsWorkspaceService` with agent
creation/cloning
- 🆕 Agent naming constants for conflict resolution
### Shared (`twenty-shared`)
- 🆕 `AgentResponseSchema` type
- 🆕 `ModelConfiguration` type moved to shared package
- Updated exports in `ai/index.ts`
## Code Quality
- Removed useless comments following code style guidelines
- All linter checks passed
- Type-safe implementation with proper TypeScript types
## Testing
- ✅ Database migration tested
- ✅ Agent creation/cloning in workflows verified
- ✅ Response format switching (text ↔ JSON) validated
- ✅ Backward compatibility with existing agents confirmed
## Migration Notes
- Existing agents will have `responseFormat: {type: 'text'}` set
automatically
- No breaking changes - all existing functionality preserved
- Agents can be updated to use JSON format through settings UI
I did an acceptable design for empty node and iterators for release:
- use array field for iterator node. Added an util to stringify arrays
for backward compatibily
- remove icon for empty node
- allow to select a node on empty node selection
https://github.com/user-attachments/assets/b00037a8-aa1d-4784-b973-05973649b46e
Avoid fetching full steps and trigger for versions that are not the
current version. Because those won't be used anyway. Better for
performances.
Only difficulty was for the `createDraft` mutation. I needed to return
the full created version so I can store it in cache and use it as new
`currentVersion`. Otherwise the current version is considered as
incomplete for a short time, since workflow is fetched separately from
the current version.
---------
Co-authored-by: Charles Bochet <charles@twenty.com>
https://github.com/user-attachments/assets/d6c565eb-9a29-4830-9396-5f979c8caa7b
- Added a new component for manual trigger (mostly duplicated from
previous one). Will remove the old one once all data are migrated
- Updated schema output so the current item of the iterator can be typed
Todo left:
- migrate old triggers
- add an util to search iterator output. Today current item fields will
be displayed as not found
- set new manual triggers for workflow runs
This PR should fix optimistic rendering issues on step updates:
- compute a diff for trigger and steps on mutations
- build an util that applies that diff (built a more robust version from
https://github.com/AsyncBanana/micropatch)
- apply diff in cache
Previous refacto was creating output schema V2 which has more specific
schemas based on the step type. Before we were using one common schema,
which was too complex when searching for variable informations.
This PR migrate the deprecated schemas and remove the old code:
- mark previous `BaseOutputSchema` as deprecated
- remove other previous schemas
- use V2 everywhere
- icon should not be stored in schema. Instead it should be generated
based on the fieldmetadata or the item type
- replace save by insert
- if the insert output is needed, cast the generatedMaps
- remove transactions from trigger services. Doing it manually would be
complex and not reliable
Today WorkflowVariablesDropdownAllItems is used for variable picker in
all workflow steps, and for step field picker in filters.
Using that component in filters prevent us from knowing if a full record
or only the id has been selected.
This PR:
- creates a new component WorkflowDropdownStepOutputItems, mostly
copying the logic of WorkflowVariablesDropdownAllItems
- call it WorkflowStepFilterFieldSelect, removing the display logic from
that parent component
- store isFullRecord in filter. So we now know if we should display a
record picker or a uuid picker
Before - selecting id makes picker behaves like when we select an object
https://github.com/user-attachments/assets/bde34dc5-8011-4983-8d0f-d8cb0cb3c045
After - selecting object and id are two different things
https://github.com/user-attachments/assets/49289990-3e6d-4ad7-abc1-e3ade2a821bb
When the target is not in the direct next step ids of the source (step
or trigger), check if there is a filter between both. If yes, delete the
filter and remove it from source next step ids.
Also refactored the existing.
