## Problem
On a self-hosted instance, an AI Agent workflow action fails at run time
with an opaque model error:
```
The model returned the following errors: tools.0.custom.input_schema.properties:
Property keys should match pattern '^[a-zA-Z0-9_.-]{1,64}$'
```
This is Anthropic's validation on tool `input_schema` **property keys**.
An AI Agent's structured **Output** fields are turned into a JSON schema
and passed to the model as a tool; each output **variable name** becomes
a property key. Anthropic rejects any key that does not match
`^[a-zA-Z0-9_.-]{1,64}$` — most commonly a name containing a **space**
(e.g. `meetings brief`), but also names over 64 characters or with other
symbols.
Until now nothing validated this: `fieldsToSchema` writes
`properties[field.name]` verbatim, so a bad name only failed once the
workflow executed, with an error that gives the user no idea what to
fix. It doesn't reproduce on every instance — it depends purely on how
the workflow's output variables happen to be named.
## Fix
Introduce a single shared check,
`isValidAgentResponseSchemaPropertyKey`, and enforce it in two places:
- **Backend** — `validateAgentResponseFormat` now rejects invalid output
field names at agent **save time** with a clear `userFriendlyMessage`,
instead of letting the broken schema reach the model. This also gates
agents created via the API and re-saves of existing bad data.
- **Frontend** — the output schema builder shows an inline error on the
Variable Name field as soon as an invalid name is entered.
## Tests
- Unit test for the shared validity check (valid + invalid cases:
spaces, leading space, empty, > 64 chars, symbols, unicode).
- Unit test for `validateAgentResponseFormat` covering text/json
formats, valid names, a space in a name, an over-length name, and
reporting multiple invalid names at once.
## Notes for the reporter
The immediate unblock for an affected workflow is to rename the output
variable to remove the space (e.g. `meetings brief` → `meetings_brief`)
and retry the run. With this change the bad name is caught up front with
an explanation rather than failing mid-run.
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/21834?utm_source=github"
target="_blank" rel="noopener noreferrer"
data-no-image-dialog="true"><picture><source
media="(prefers-color-scheme: dark)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"><source
media="(prefers-color-scheme: light)"
srcset="https://www.cubic.dev/buttons/review-in-cubic-light.svg"><img
alt="Review in cubic"
src="https://www.cubic.dev/buttons/review-in-cubic-dark.svg"></picture></a>
<!-- End of auto-generated description by cubic. -->
<img width="948" height="593" alt="image"
src="https://github.com/user-attachments/assets/d990fa98-3cfd-469d-ab7f-0b2d4ccf3afc"
/>
<img width="1361" height="802" alt="image"
src="https://github.com/user-attachments/assets/1091f598-49f3-4c16-92ea-1e1c200181e2"
/>
## Add `runAgent()` to the Logic Function SDK
Lets an app's logic function run one of its own AI agents server-side
and get the result back synchronously — reusing the existing agent
executor instead of a new bespoke transport.
### Backend
- New **`runAgent` GraphQL mutation** (metadata schema) in
`ai-agent-execution`, wrapping the existing
`AgentAsyncExecutorService.executeAgent`. Scopes the agent lookup to the
calling
application and runs it under an application auth context.
- New `@AuthApplication()` param decorator (mirrors `@AuthWorkspace()`)
— first GraphQL resolver authenticated by an **application access
token**.
- Guarded by `WorkspaceAuthGuard` +
`SettingsPermissionGuard(PermissionFlagType.AI)`: the app's role must
grant the `AI` permission flag.
### SDK
- `runAgent({ agentUniversalIdentifier, prompt })` posts the mutation to
`/metadata` with the app token via a new runtime GraphQL transport.
Returns `{ result, hasNoMoreAvailableCredits
}`.
- Refactored the connections helpers onto a shared `postAppEndpoint`
util (removes duplicated transport logic).
### Frontend
- App install permission modal now shows an explicit consent line —
_"Run AI agents and bill AI credits to your workspace"_ — when the app's
role requests the `AI` flag.
### Docs
- Documented `runAgent` and its `AI` permission-flag requirement in
_Skills & Agents_.
- Fixed outdated role-permission examples in _Roles & Permissions_
(`permissionFlags` → `permissionFlagUniversalIdentifiers`,
`PermissionFlag` → `SystemPermissionFlag`).
### Test plan
- [x] SDK unit tests (`run-agent.spec.ts`) — request shape, GraphQL/HTTP
error handling, missing env vars
- [x] `twenty-server`, `twenty-front`, `twenty-shared` typecheck + lint
- [ ] Manual: install an app granting the `AI` flag, call `runAgent()`
from a logic function, confirm the agent runs and credits are billed
---------
Co-authored-by: cubic-dev-ai[bot] <191113872+cubic-dev-ai[bot]@users.noreply.github.com>
## Summary
Exposes two Twenty primitives to the AI chat that it could not
previously manage:
- **Navigation menu items** — workspace nav and personal favorites
(favorites are just nav items with `scope: 'user'`).
- **Webhooks** — full CRUD with a structured operations input (record +
metadata events).
Page layouts and workflow runs were originally in this PR but have been
split out — they touch heavier surfaces (21 widget configurations and
the workflow runner cycle, respectively) and deserve their own focused
PRs.
### Tool inventory (8 new tools across 2 providers)
| Provider | Tools |
|---|---|
| NavigationMenuItem | `list_`, `create_`, `update_`,
`delete_navigation_menu_item` |
| Webhook | `list_`, `create_`, `update_`, `delete_webhook` |
### Design notes
- Both providers follow the established **view-style pattern**: tool
workspace service lives in the entity module's `tools/` folder, is
provided + exported by the entity module, and `ToolProviderModule`
imports the entity module. No `@Global()` modules or injection tokens
introduced.
- `create_navigation_menu_item` uses a Zod `discriminatedUnion` on
`type` (`FOLDER` / `LINK` / `OBJECT` / `VIEW` / `RECORD` /
`PAGE_LAYOUT`). `scope: 'workspace' | 'user'` switches between shared
nav and personal favorites — the underlying
`NavigationMenuItemAccessService` enforces LAYOUTS for workspace writes.
- Webhook operations accept both record events (`{kind:'record', object,
event}` → `<object>.<event>`) and metadata events (`{kind:'metadata',
metadataName, operation}` → `metadata.<metadataName>.<operation>`).
- Permissions reuse existing flags (`LAYOUTS`, `API_KEYS_AND_WEBHOOKS`).
No new permission flags, no migrations.
### Category cleanup
- New: `ToolCategory.NAVIGATION_MENU_ITEM`, `ToolCategory.WEBHOOK`.
- `ToolCategory.VIEW_FIELD` → folded into `VIEW`. Same permission gate,
same domain — separate category was organizational drift.
- `navigate_app` action stays in `ToolCategory.ACTION` where it belongs.
### System prompt addition
[chat-system-prompts.const.ts](packages/twenty-server/src/engine/metadata-modules/ai/ai-chat/constants/chat-system-prompts.const.ts)
now teaches the AI:
- Favorites are nav items with `scope: 'user'`.
- A default OBJECT nav item is auto-created with
`create_object_metadata` — don't double-create.
### One file = one export
Every new schema / type / util file has exactly one top-level export.
## Test plan
- [ ] `npx nx typecheck twenty-server` — passes
- [ ] Spin up locally and exercise via AI chat:
- [ ] "Pin the Companies view to my favorites in a folder called
Important." → `create_navigation_menu_item` (FOLDER, user) then (VIEW,
user, folderId)
- [ ] "Register a webhook to https://example.com firing when any person
is created or updated." → `create_webhook` with discriminated operations
- [ ] Verify workspace-scoped nav writes are denied for a user without
LAYOUTS permission
- [ ] Verify user-scoped nav writes work without LAYOUTS permission
## Follow-ups (separate PRs)
- Page layout tools (record-page, record-index, standalone) — needs
widget-config strategy.
- Workflow run tools (list, get, run, stop) — uses the workflow-runner
cycle path.
- Dashboard / page-layout tool unification —
`DashboardToolWorkspaceService` and a future
`PageLayoutToolWorkspaceService` both inject the same trio
(PageLayout/Tab/Widget services).
- Webhook Settings page reads from raw Apollo query — switch to the
metadata store so it refreshes when the AI mutates webhooks.
Billing is now decremented per-step, not per-turn. The onStepFinish
callback in chat-execution.service.ts calls a new
decrementAndCheckAvailableCredits method on each model step, so Redis is
debited incrementally as the agent runs rather than all at once at the
end.
Credit exhaustion stops the stream mid-run. When a step depletes the
remaining credits, a hasNoMoreAvailableCredits flag is set and passed
into the stopWhen predicate of streamText, causing the agent to halt
before starting the next step.
A new credits-exhausted event is introduced. After the stream drains and
the response is persisted, if credits ran out the job publishes a
dedicated credits-exhausted event to the frontend instead of the normal
message-persisted event.
The frontend handles this new event. useAgentChatSubscription has a new
credits-exhausted case that sets a BILLING_CREDITS_EXHAUSTED-coded error
on the atom, closes the writer, and stops the streaming state —
triggering the existing AiChatCreditsExhaustedMessage UI.
## Intent
This is a small foundation cleanup for the tool architecture.
The main decision is: registry tools and native SDK/model tools are
different things.
- Registry tools have descriptors, schemas, catalog entries, and execute
through `ToolExecutorService`
- Native model tools are opaque AI SDK objects, bound directly into the
model `ToolSet`
- Surfaces still own their policy: chat, MCP, and workflow agents decide
what they expose
## What changed
- Removed `NATIVE_MODEL` from `ToolCategory`
- Kept `ToolRegistryService` focused on registry-backed tools only
- Moved native model tool binding through `NativeToolBinderService`
- Reused native binding from chat instead of duplicating
provider-specific web-search logic
- Kept MCP local execution exclusions in a dedicated constant
- Moved surface-specific constants into dedicated constant files
## What comes next
- Move hardcoded chat app preloads, like Exa web search, into
app/manifest metadata
- Decide a clearer policy for local runtime tools like code interpreter
and HTTP request
- Gradually document the three tool shapes: registry tools, native model
tools, and local runtime tools
---------
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: Félix Malfait <FelixMalfait@users.noreply.github.com>
Co-authored-by: Félix Malfait <felix.malfait@gmail.com>
## Summary
The "AI" acronym was rendered inconsistently across the codebase. The
backend AI module had settled on PascalCase `Ai` (`AiAgentModule`,
`AiBillingService`, `AiChatModule`, `AiModelRegistryService`, etc.),
while frontend components, several DTOs, a few types, and shared
identifiers still used all-caps `AI` (`AIChatTab`,
`AISystemPromptPreviewDTO`, `SettingsPath.AIPrompts`, ...). CLAUDE.md
specifies PascalCase for classes; this PR normalizes everything internal
to `Ai`.
**This is a pure internal rename.** The GraphQL schema is untouched —
`@ObjectType` decorator string arguments, resolver method names (which
become Query/Mutation field names), gql template contents, and the
`generated-metadata/graphql.ts` file are preserved verbatim. The only
visible change is TypeScript identifiers and file names.
## Also folded in (adjacent cleanups)
- **`AgentModelConfigService` → `AiModelConfigService`**. Lives in
`ai-models/` and is used by multiple AI code paths, not just the Agent
entity. The "Agent" prefix was misleading.
- **`generate-text-input.dto.ts` → `generate-text.input.ts`**. The
`ai-agent/dtos/` folder already uses `<entity>.input.ts` convention for
Input classes (`create-agent.input.ts` etc.); the old path mixed
`.dto.ts` file extension with a class that has no DTO suffix. File
rename only; class stays `GenerateTextInput`.
- **Removed stale TODO** in `ai-model-config.type.ts` that asked for the
`AiModelConfig` rename that this PR performs.
## Rename methodology
Bulk rename via perl with anchored regex
`(?<!['"])(?<![A-Z.])AI([A-Z])(?=[a-z])/Ai$1/g`:
- **Lookbehind for non-uppercase** skips adjacent acronyms (`MOSAIC`,
`OIDCSSO`) and leaves `AIRBNB_ID` alone.
- **Lookbehind for non-quote** protects most string literals.
- **Lookahead for lowercase** restricts matches to PascalCase
identifiers (`AIChatTab`), leaving SCREAMING_SNAKE constants untouched.
Strict file-scope exclusions: `generated-metadata/**`, `generated/**`,
`locales/**`, `migrations/**`, `illustrations/**`, `halftone/**`, and
the two gql template files (`queries/getAISystemPromptPreview.ts`,
`mutations/uploadAIChatFile.ts`).
Post-rename reverts for identifiers where the regex was too eager:
- Backend resolver method names kept: `getAISystemPromptPreview`,
`uploadAIChatFile` (they are GraphQL field names).
- `@ObjectType('AdminAIModels')` / `('AISystemPromptPreview')` /
`('AISystemPromptSection')` kept as-is.
- Backend classes `ClientAIModelConfig` / `AdminAIModelConfig` kept
as-is (they use `@ObjectType()` with no argument, so the class name IS
the schema name).
- External-library symbols restored: `OpenAIProvider`,
`createOpenAICompatible`, `vercelAIIntegration`.
File renames use a two-step rename to work on macOS case-insensitive
filesystems: `git mv X.tsx X.tsx.tmp && git mv X.tsx.tmp renamed.tsx`.
## Diff audit
- 0 changes to migrations
- 0 changes to locale `.po` / `.ts` files
- 0 changes to `generated-metadata/graphql.ts`
- 0 changes to website illustration files (base64 blobs preserved)
- 0 renames inside user-facing translation strings (`t\`…\``,
`msg\`…\``, `<Trans>…</Trans>`)
## Test plan
- [x] `npx nx typecheck twenty-server` — PASS
- [x] `npx nx typecheck twenty-front` — PASS
- [x] `npx jest ai-model admin agent-role` — 79/79 PASS
- [x] `npx oxlint --type-aware` on 118 changed files — 0 errors
- [x] `npx prettier --check` on 118 changed files — clean
- [ ] CI
## Summary
- **Queue messages while streaming**: Messages sent during active AI
streaming are queued server-side and auto-flushed when the current
stream completes. Frontend renders queued messages optimistically in a
dedicated queue UI.
- **Drop `@ai-sdk/react` + `resumable-stream`**: Replace the dual HTTP
SSE + AI SDK client architecture with a single GraphQL SSE subscription
per thread. All events (token chunks, message persistence, queue
updates, errors) flow through Redis PubSub → GraphQL subscription.
- **Server-driven architecture**: The server decides whether to queue or
stream (via `POST /:threadId/message`). The frontend mirrors this
decision for optimistic rendering but defers to the server response.
- **Reuse AI SDK accumulation logic**: `readUIMessageStream` from the
`ai` package handles chunk-to-message accumulation on the frontend,
avoiding a custom 780-line accumulator.
## Key files
**Backend:**
- `agent-chat-event-publisher.service.ts` — publishes events to Redis
PubSub
- `agent-chat-subscription.resolver.ts` — GraphQL subscription resolver
- `stream-agent-chat.job.ts` — publishes chunks via PubSub instead of
resumable-stream
- `agent-chat.controller.ts` — unified `POST /:threadId/message`
endpoint
**Frontend:**
- `useAgentChatSubscription.ts` — subscribes to `onAgentChatEvent`,
bridges to `readUIMessageStream`
- `useAgentChat.ts` — send/stop/optimistic rendering (no more AI SDK)
- `AgentChatStreamSubscriptionEffect.tsx` — replaces
`AgentChatAiSdkStreamEffect.tsx`
## Test plan
- [ ] Send message on new thread → optimistic render, streaming response
appears
- [ ] Send message while streaming → queued instantly (no flash in main
thread)
- [ ] Queued message auto-flushes after current stream completes
- [ ] Remove queued message via queue UI
- [ ] Stop streaming mid-response
- [ ] Leave chat idle for several minutes → streaming still works after
(SSE client recycling)
- [ ] Token refresh during session → requests succeed (authenticated
fetch)
- [ ] Switch threads while streaming → clean subscription handoff
Made with [Cursor](https://cursor.com)
---------
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>
## Summary
- Replaces per-provider TypeScript constant files
(`openai-models.const.ts`, `anthropic-models.const.ts`, etc.) with a
single `ai-providers.json` catalog as the source of truth
- Adds runtime model discovery via AI SDK for self-hosted providers,
with `models.dev` enrichment for pricing/capabilities
- Introduces composite model IDs (`provider/modelId`) for canonical,
conflict-free identification
- Simplifies provider configuration: API keys are injected from
environment variables (e.g., `OPENAI_API_KEY`)
- Adds admin panel UI for provider management (add/remove/test), model
discovery, recommended model configuration, and default fast/smart model
selection per workspace
- Removes deprecated config variables (`AI_DISABLED_MODEL_IDS`,
`AUTO_ENABLE_NEW_AI_MODELS`, etc.)
- Adds database migration for composite model ID format
## Test plan
- [ ] Server typecheck passes
- [ ] Frontend typecheck passes
- [ ] Server unit tests pass
- [ ] Frontend unit tests pass
- [ ] CI pipeline green
- [ ] Admin panel AI tab loads correctly
- [ ] Provider discovery works for configured providers
- [ ] Model recommendation toggles persist
- [ ] Default fast/smart model selection works
Made with [Cursor](https://cursor.com)
This PR fixes what allows to have a working demo workspace skill.
- Skill updated many times into something that works
- Fixed infinite loop in AI chat by memoizing ai-sdk output
- Finished navigateToView implementation
- Increased MAX_STEPS to 300 so the chat don't quit in the middle of a
long running skill
- Added CreateManyRelationFields
This PR adds the necessary tool to create a demo workspace with :
relevant custom objects and fields, mock data and a real dashboard with
graph widgets.
It is still a bit under-optimized and slow but it works.
This PR also adds an AI tool that allows to see what happens in real
time, it navigates the app and waits when necessary.
---------
Co-authored-by: Félix Malfait <felix.malfait@gmail.com>
- Remove feature flag
- Remove legacy methods in file-upload and file-service
- Migrate AI Chat to new file management
---------
Co-authored-by: Charles Bochet <charles@twenty.com>
## Summary
Replaces the static "Ask AI" header in the command menu with the
conversation’s auto-generated title once it’s set after the first
message.
## Changes
- **Backend:** Title is generated after the first user message (existing
behavior).
- **Frontend:** After the first stream completes, we fetch the thread
title and sync it to:
- `currentAIChatThreadTitleState` (persists across command menu
close/reopen)
- Command menu page info and navigation stack (so the title survives
back navigation)
- **Entry points:** Opening Ask AI from the left nav or command center
uses the same title resolution (explicit `pageTitle` → current thread
title → "Ask AI" fallback).
- **Race fix:** Title sync only runs when the thread that finished
streaming is still the active thread, so switching threads mid-stream
doesn’t overwrite the current thread’s title.
---------
Co-authored-by: Félix Malfait <felix@twenty.com>
Co-authored-by: Cursor <cursoragent@cursor.com>
⚠️ **AI-generated PR — not ready for review** ⚠️
cc @FelixMalfait
---
## Changes
### System prompt improvements
- Explicit skill-before-tools workflow to prevent the model from calling
tools without loading the matching skill first
- Data efficiency guidance (default small limits, use filters)
- Pluralized `load_skill` → `load_skills` for consistency with
`load_tools`
### Token usage reduction
- Output serialization layer: strips null/undefined/empty values from
tool results
- Lowered default `find_*` limit from 100 → 10, max from 1000 → 100
### System object tool generation
- System objects (calendar events, messages, etc.) now generate AI tools
- Only workflow-related and favorite-related objects are excluded
### Context window display fix
- **Bug**: UI compared cumulative tokens (sum of all turns) against
single-request context window → showed 100% after a few turns
- **Fix**: Track `conversationSize` (last step's `inputTokens`) which
represents the actual conversation history size sent to the model
- New `conversationSize` column on thread entity with migration
### Workspace AI instructions
- Support for custom workspace-level AI instructions
---------
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
## Summary
- Add code interpreter tool that enables AI to execute Python code for
data analysis, CSV processing, and chart generation
- Support for both local (development) and E2B (sandboxed production)
execution drivers
- Real-time streaming of stdout/stderr and generated files
- Frontend components for displaying code execution results with
expandable sections
## Code Quality Improvements
- Extract `getMimeType` to shared utility to reduce code duplication
between drivers
- Fix security issue: escape single quotes/backslashes in E2B driver env
variable injection
- Add `buildExecutionState` helper to reduce duplicated state object
construction
- Add `DEFAULT_CODE_INTERPRETER_TIMEOUT_MS` constant for consistency
- Fix lingui linting warning and TypeScript theme errors in frontend
## Test Plan
- [ ] Test code interpreter with local driver in development
- [ ] Test code interpreter with E2B driver in production environment
- [ ] Verify streaming output displays correctly in chat UI
- [ ] Verify generated files (charts, CSVs) are uploaded and
downloadable
- [ ] Test file upload flow (CSV, Excel) triggers code interpreter
<!-- CURSOR_SUMMARY -->
---
> [!NOTE]
> Updates generated i18n catalogs for Polish and pseudo-English, adding
strings for code execution/output (code interpreter) and various UI
messages, with minor text adjustments.
>
> - **Localization**:
> - **Generated catalogs**: Refresh `locales/generated/pl-PL.ts` and
`locales/generated/pseudo-en.ts`.
> - Add strings for code execution/output (e.g., code, copy code/output,
running/waiting states, download files, generated files, Python code
execution).
> - Include new UI texts (errors, prompts, menus) and minor text
corrections.
> - No changes to `pt-BR`; other files unchanged functionally.
>
> <sup>Written by [Cursor
Bugbot](https://cursor.com/dashboard?tab=bugbot) for commit
befc13d02c21e5a6647bc1aa6daa2a89f60b7ef8. This will update automatically
on new commits. Configure
[here](https://cursor.com/dashboard?tab=bugbot).</sup>
<!-- /CURSOR_SUMMARY -->
## Summary
- Add a context usage indicator to the AI chat interface inspired by
Vercel's AI SDK Context component
- Display token consumption, context window utilization percentage, and
estimated cost in credits
- Show a circular progress ring with percentage, revealing detailed
breakdown on hover
## Changes
### Backend
- Stream usage metadata (tokens, model config) via `messageMetadata`
callback in `agent-chat-streaming.service.ts`
- Return model config from `chat-execution.service.ts`
- Add usage and model types to `ExtendedUIMessage` metadata
### Frontend
- New `ContextUsageProgressRing` component - circular SVG progress
indicator
- New `AIChatContextUsageButton` component with hover card showing:
- Progress bar with used/total tokens
- Input/output token counts with credit costs
- Total credits consumed
- Track cumulative usage in Recoil state (`agentChatUsageState`)
- Reset usage when creating new chat thread
- Integrate button into `AIChatTab`
## Test plan
- [ ] Open AI chat and send a message
- [ ] Verify the context usage button appears with percentage
- [ ] Hover over the button to see detailed breakdown
- [ ] Verify credits are calculated correctly
- [ ] Create a new chat thread and verify usage resets to 0
## 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
Adds intelligent routing system that automatically selects the best
agent for user queries based on conversation context.
### Changes:
- Added `routerModel` column to workspace table for configurable router
LLM selection
- Implemented `RouterService` with conversation history analysis and
agent matching logic
- Created router settings UI in AI Settings page with model dropdown
- Removed agent-specific thread associations - threads are now
agent-agnostic
- Added real-time routing status notification in chat UI with shimmer
effect
- Removed automatic default assistant agent creation
- Renamed GraphQL operations from agent-specific to generic (e.g.,
`agentChatThreads` → `chatThreads`)
---------
Co-authored-by: Félix Malfait <felix.malfait@gmail.com>
Co-authored-by: Félix Malfait <felix@twenty.com>
Charles Bochet
