Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
Move the browsing context out of the system prompt and injecting it
directly into the last user message instead.
Previously (before this PR), browsing context change, update system
prompt then break whole conversation history ... and caching. Now,
browsing context is sent with last message only if changed.
"Benchmark" this PR vs main :
- same conv with 3-4 turns - 60% -> 85% cache ratio || 0.31 credits ->
0.13
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
## Summary
`PermissionsException` thrown by `SettingsPermissionGuard` (and other
permission code paths) was bubbling up through every typed REST
exception filter and landing in the global `UnhandledExceptionFilter`,
which falls back to **500** for anything that isn't an `HttpException`.
So a forbidden user (e.g. an API key whose role doesn't have
`DATA_MODEL`) calling `GET /rest/metadata/objects` got:
```
HTTP/1.1 500 Internal Server Error "Entity performing the request does not have permission"
```
GraphQL already had the right plumbing via
`permissionGraphqlApiExceptionHandler` (`ForbiddenError` → 403,
`UserInputError` → 400, `NotFoundError` → 404). This PR mirrors it on
the REST side.
## What
- New util `permissionRestApiExceptionCodeToHttpStatus` mapping every
`PermissionsExceptionCode` → HTTP status, with `assertUnreachable` to
force explicit handling of future codes.
- New filter `PermissionsRestApiExceptionFilter`
(`@Catch(PermissionsException)`) that delegates to
`HttpExceptionHandlerService.handleError(...)` with the resolved status.
- Wired `PermissionsRestApiExceptionFilter` (placed first, so the typed
filter wins over any sibling catch-all) into `@UseFilters(...)` of every
REST controller that uses `SettingsPermissionGuard` or whose service can
throw `PermissionsException`:
- `object-metadata`, `field-metadata`, `webhook`, `api-key`
- `view`, `view-sort`, `view-group`, `view-filter`, `view-filter-group`,
`view-field`
- `page-layout`, `page-layout-widget`, `page-layout-tab`
- `front-component`, `ai-generate-text`
- Unit tests covering 403 / 400 / 404 / 500 mappings.
## Mapping
| Code | Status |
|------|--------|
| `PERMISSION_DENIED`, `NO_AUTHENTICATION_CONTEXT`,
`ROLE_LABEL_ALREADY_EXISTS`, `CANNOT_UNASSIGN_LAST_ADMIN`,
`CANNOT_UPDATE_SELF_ROLE`, `CANNOT_DELETE_LAST_ADMIN_USER`,
`ROLE_NOT_EDITABLE`, `CANNOT_ADD_OBJECT_PERMISSION_ON_SYSTEM_OBJECT`,
`CANNOT_ADD_FIELD_PERMISSION_ON_SYSTEM_OBJECT` | **403** |
| `INVALID_ARG`, `INVALID_SETTING`,
`CANNOT_GIVE_WRITING_PERMISSION_ON_NON_READABLE_OBJECT`,
`CANNOT_GIVE_WRITING_PERMISSION_WITHOUT_READING_PERMISSION`,
`ONLY_FIELD_RESTRICTION_ALLOWED`,
`FIELD_RESTRICTION_ONLY_ALLOWED_ON_READABLE_OBJECT`,
`FIELD_RESTRICTION_ON_UPDATE_ONLY_ALLOWED_ON_UPDATABLE_OBJECT`,
`EMPTY_FIELD_PERMISSION_NOT_ALLOWED`,
`ROLE_MUST_HAVE_AT_LEAST_ONE_TARGET`, `ROLE_CANNOT_BE_ASSIGNED_TO_USERS`
| **400** |
| `ROLE_NOT_FOUND`, `OBJECT_METADATA_NOT_FOUND`,
`FIELD_METADATA_NOT_FOUND`, `FIELD_PERMISSION_NOT_FOUND`,
`PERMISSION_NOT_FOUND` | **404** |
| All remaining "internal" codes (rethrown as-is in GraphQL) | **500** |
## Before
<img width="507" height="216" alt="Screenshot 2026-05-19 at 19 26 07"
src="https://github.com/user-attachments/assets/21d633aa-7ee8-4923-94e4-7ad57258a29e"
/>
## After
<img width="610" height="385" alt="Screenshot 2026-05-19 at 19 26 01"
src="https://github.com/user-attachments/assets/0103b7ee-7df7-4aef-999a-73c22901afd2"
/>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
**Problem**
AI_MODEL_PREFERENCES, JSON env var is not supported +
IS_CONFIG_VARIABLES_IN_DB_ENABLED=false in twenty cloud server
-> No option to set AI_MODEL_PREFERENCES
**Solution**
AI_MODEL_PREFERENCES supports three override sources beyond the
hardcoded code defaults, in priority order:
- DB (IS_CONFIG_VARIABLES_IN_DB_ENABLED=true), the only writable source;
admin-panel mutations persist here
- ENV not usable in Twenty Cloud, which does not handle JSON-format env
vars
- **Introduced in this PR** --> File
(AI_MODEL_PREFERENCES_STORAGE_PATH), a read-only startup fallback, the
only viable override in Cloud/self-managed deployments where DB config
is disabled and JSON env vars are unsupported.
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
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.
## Summary
Extends the asymmetric signing work from #20467 to cover **every
remaining `JwtTokenTypeEnum` value**: `LOGIN`, `WORKSPACE_AGNOSTIC`,
`FILE`, `API_KEY`, `APPLICATION_ACCESS`, `APPLICATION_REFRESH`,
`APP_OAUTH_STATE`, plus the ACCESS-shaped session token issued by the
code interpreter tool.
After this PR, every JWT the server signs is ES256 with a `kid` pointing
at the current `core."signingKey"` row, while legacy HS256 tokens (no
`kid` header) remain verifiable indefinitely through the existing
fallback in `JwtWrapperService.resolveVerificationKey`. No new entity /
migration / config: this is a pure routing change on top of the
infrastructure that already shipped.
## Why
`#20467` only flipped `ACCESS` and `REFRESH` to ES256. Every other JWT
type was still HS256-signed against the global `APP_SECRET`, which kept
the original blast radius (a leaked `APP_SECRET` invalidates *every* JWT
type forever). Migrating the rest unifies the sign path on rotatable
per-server private keys without forcing any token reissue.
## Mechanical changes
### Sign side (8 services)
- `LoginTokenService.generateLoginToken`
- `TransientTokenService.generateTransientToken`
- `WorkspaceAgnosticTokenService.generateWorkspaceAgnosticToken`
- `ApplicationTokenService.signApplicationToken` (`APPLICATION_ACCESS` +
`APPLICATION_REFRESH`)
- `ApiKeyService.generateApiKeyToken`
- `FileUrlService.signFileByIdUrl` / `signWorkspaceLogoUrl`
- `ConnectionProviderOAuthFlowService.signState` (`APP_OAUTH_STATE`)
- `CodeInterpreterTool.generateSessionToken`
Each call site swaps `jwtWrapperService.sign(payload, { secret:
generateAppSecret(...), ... })` for `await
jwtWrapperService.signAsync(payload, { expiresIn, [jwtid] })`. The
`generateAppSecret` calls on the sign side are dropped (verifier-side
`generateAppSecret` stays in `resolveVerificationKey` for the HS256
fallback).
### Verifier side
- `WorkspaceAgnosticTokenService.validateToken` now goes through
`verifyJwtToken` instead of the bespoke `verify({ secret })` path, so
new ES256 tokens are accepted while the legacy HS256 fallback inside
`resolveVerificationKey` still serves the old shape.
- `JwtWrapperService.sign()` is kept (legacy compat / tests) but is now
strictly deprecated — there are no remaining production callers.
### Async ripple (`signFileByIdUrl` was synchronous)
- `FileUrlService.signFileByIdUrl` and `signWorkspaceLogoUrl` are now
`async`; the `signUrl` callback used by `getRecordImageIdentifier` is
widened to accept `Promise<string | null>`.
- Every direct/indirect caller is updated: admin panel (user lookup +
statistics + top workspaces), search service
(`computeSearchObjectResults`, `getImageIdentifierValue`), workspace
resolver (`logo` resolver, public workspace by domain/id),
`WorkspaceMemberTranspiler` (now `async toWorkspaceMemberDto[s]` /
`toDeletedWorkspaceMemberDto[s]` / `generateSignedAvatarUrl`),
`UserService.loadSignedAvatarUrlsByUserId`,
`UserWorkspaceService.castWorkspaceToAvailableWorkspace`,
workspace-invitation, approved-access-domain, agent-chat-streaming,
agent-message-part resolver, navigation-menu-item record identifier,
file-ai-chat / file-core-picture / file-email-attachment / file-workflow
/ files-field services, rich-text & files-field query result getters,
and the code-interpreter tool.
## Backward compatibility
- **Legacy HS256 tokens (no `kid`)** keep verifying via
`resolveVerificationKey` → `extractAppSecretBody` → `generateAppSecret`
for both `workspaceId`-bearing and `userId`-bearing payloads.
- The `API_KEY` HS256-via-ACCESS-secret fallback (#16504) still kicks in
inside `verifyJwtToken` for pre-2025-12-12 API keys.
- No payload shape changes, no DB writes, no env var changes — old
tokens issued by `main` continue to authenticate.
## Tests
### Unit (all green locally — 63/63)
Updated specs for every migrated service to mock `signAsync` instead of
`sign` and assert the new option shape:
- `login-token.service.spec.ts`, `transient-token.service.spec.ts`,
`workspace-agnostic-token.service.spec.ts`,
`application-token.service.spec.ts`, `api-key.service.spec.ts`,
`connection-provider-oauth-flow.service.spec.ts`.
### Integration (`jwt-key-rotation.integration-spec.ts`)
- Existing ACCESS coverage (current key, legacy HS256 fallback,
rotated-out key, revoked key, unknown kid) is preserved.
- New `it.each` assertion: `REFRESH`, `WORKSPACE_AGNOSTIC`, and `LOGIN`
tokens emitted by the real signUp → signUpInNewWorkspace →
getAuthTokensFromLoginToken pipeline are ES256 with a `kid` matching the
current signing key — proves end-to-end that the migration didn't
regress those flows.
## Open question (separate decision)
This PR keeps the legacy HS256 verification fallback **forever**. We may
eventually want to sunset it for `API_KEY` once telemetry shows
pre-migration tokens are gone, but that's a separate product/security
decision and not part of this change.
## Test plan
- [ ] CI green
- [ ] `npx nx lint:diff-with-main twenty-server` passes
- [ ] `npx nx typecheck twenty-server` passes
- [ ] `jwt-key-rotation` integration suite passes (new + existing
assertions)
- [ ] Manually verify: signing in issues an ES256 ACCESS / REFRESH
token, generating an API key issues an ES256 token with `kid`, signed
file URL JWT is ES256 with `kid`
- [ ] Pre-existing HS256 tokens still authenticate (covered by
integration test, but worth a manual check with a token from `main`)
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
**Overall strategy**
**1. Introduce “Billing V2” behind a workspace flag**
Gate the new model with FeatureFlagKey.IS_BILLING_V2_ENABLED so existing
workspaces stay on the old behavior until they’re migrated or explicitly
on V2.
**2. Replace workflow metered SKUs with a resource-credit product**
Conceptually, billable “workflow execution” usage is not the primary
subscription line item anymore. Add a RESOURCE_CREDIT product (and keep
WORKFLOW_NODE_EXECUTION as deprecated for the transition). Usage and
limits are expressed through credit buckets (e.g. price metadata like
credit_amount), so one product can represent pooled credits instead of a
narrow workflow-only meter.
**3. Migrate subscriptions in two layers**
Schema/catalog: persist extra price metadata (instance upgrade) so the
server knows credit amounts and can match Stripe prices to the new
model.
Per workspace: the registered workspace command
upgrade:2-2:migrate-to-billing-v2 finds subscriptions that still have
WORKFLOW_NODE_EXECUTION, swaps those items to the right RESOURCE_CREDIT
prices (using existing Stripe schedule +
BillingSubscriptionUpdateService stack), then treats the workspace as V2
(flag). Workspaces without that legacy item or without a subscription
are skipped.
**4. Unify subscription lifecycle + usage on the server**
**5. Refresh the product surface in Settings**
Test :
- [x] Subscribe v1 + Update subscribe + Migrate
- [x] Subscribe v2 + Update subscribe
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
## PR Description
### Summary
- Add AI chat thread actions: rename, archive (soft-delete via
`deletedAt`), and hard-delete with confirmation.
- Add chat thread filtering by status (active/archived/all), group-by
mode, and last activity.
- Rework drawer/side-panel thread lists to share thread sections, item
menus, archive icons, and empty-state behavior.
- Extend server chat thread model/API with `deletedAt`, mutations,
broadcasts, and archive-aware stream guards.
### Decisions
- Two-stage lifecycle: Archive sets `deletedAt` (soft); Delete is a
separate action on archived threads that hard-deletes the row. Aligns
with Twenty's soft-delete convention (Felix's suggestion).
- `lastMessageAt` is derived from `MAX(agentMessage.createdAt)` on read,
not stored. List query does inline aggregation for sort; `@ResolveField`
covers single-thread / mutation paths so the schema contract is honest
everywhere. Matches `timeline-messaging.service.ts` precedent and the
existing `totalInputCredits` / `totalOutputCredits` `@ResolveField`
pattern in the same resolver.
- Replaced auto-CRUD `chatThreads` (cursor-paginated Connection) with a
custom `[AgentChatThreadDTO!]` resolver. Frontend metadata-store treats
threads as a flat collection and filters/sorts client-side, so cursor
pagination was performative.
- Sending in an archived chat unarchives it optimistically on the client
and authoritatively on the server.
- Grouping and last-activity filtering use `lastMessageAt ?? updatedAt`
so archive/rename don't bump threads in the list.
- Kept metadata-store core API unchanged; AI chat uses the same local
cast pattern already used by other metadata-store partial updates.
https://github.com/user-attachments/assets/1b179b7b-1a2a-4a7a-aa0a-c88f6f051a87
## Background
The 2026-04-26 incident saw 716M Sonnet 4.6 tokens consumed in a single
trial workspace. Two causes: failed agent executions weren't billed
(addressed by #20065) and the credit-cap gate had been removed from
`WorkflowExecutorWorkspaceService.executeStep` in #19904, leaving no
enforcement point at all.
## Why not just revert #19904#19904 was right that gating at the workflow executor is too coarse.
When one user exhausted a workspace's credits via chat, *all* workflows
hard-failed mid-run — including cheap DB/CRUD/branch automations costing
essentially nothing. Reverting would re-introduce that cliff.
## New design: gate at the AI entry points
The chat resolver already gates this way
(`agent-chat.resolver.ts:137-148`). This PR replicates the same pattern
at every other point where the workspace can incur real AI cost:
- `executeAgent` in `agent-async-executor.service.ts`
- the REST handler in `ai-generate-text.controller.ts`
- `generateThreadTitle` in `agent-title-generation.service.ts`
In each, after auth/validation: skip if `IS_BILLING_ENABLED` is false;
otherwise call `BillingService.canBillMeteredProduct(workspaceId,
BillingProductKey.WORKFLOW_NODE_EXECUTION)`; on `false`, throw
`BillingException(BILLING_CREDITS_EXHAUSTED)`. No new method, no new
exception code, no new product key.
This matches industry convention (Lovable/Replit also gate at the
expensive-operation boundary, not at every cheap step).
## Deliberately not gated
- `WorkflowExecutorWorkspaceService.executeStep` — the design choice is
now intentional, so the #19904 TODO is replaced by a one-line
absolute-behavior comment explaining why the gate isn't here. Cheap
workflow steps (DB CRUD, branching, action steps) are not gated, so a
chat-driven cap exhaustion does not block non-AI automations.
- `repair-tool-call.util` — repair is a sub-call inside an already-gated
AI flow. If the parent is gated, repair will naturally not run. Adding a
gate here adds complexity without value.
## Net effect
A workspace that exhausts credits via chat or AI agent stops making AI
calls. Its non-AI workflows continue running normally. A workflow with
both AI and non-AI steps fails at the AI step with
`BILLING_CREDITS_EXHAUSTED`, but downstream non-AI steps that don't
depend on the AI output still run.
## Conflicts
This PR overlaps with three other in-flight PRs in the same files. None
of them touch the gate logic; rebasing on top of any of them is trivial:
- #20065 (agent-async-executor): adds `workspaceId` to `executeAgent`
args and bills in `finally`. The gate at the top of `executeAgent` from
this PR sits naturally above that.
- #20066 (REST controller): adds usage billing to the controller.
- #20067 (title gen): adds usage billing to title generation and
tool-call repair.
Recommend landing #20065/#20066/#20067 first; this PR rebases trivially
on top.
## Tests
Out of scope per the PR series convention. The existing chat-resolver
gate isn't unit-tested either; this PR follows the same precedent.
Follow-up: add integration coverage that exercises a workspace at
`hasReachedCurrentPeriodCap=true` against each of the three new gates
plus the pre-existing chat-resolver gate.
## Future follow-ups
- Per-user soft cap inside a workspace (the Lovable Business-tier
pattern), so one user can't exhaust the workspace's cap.
- Pre-flight cost estimate so the user sees an "approaching cap" warning
before the hard stop.
- Rename `BillingProductKey.WORKFLOW_NODE_EXECUTION` — the name predates
this design choice and is misleading now that it gates AI entry points
rather than workflow nodes.
## Test plan
- [ ] Trigger a workspace into `hasReachedCurrentPeriodCap=true`.
- [ ] Send a chat message — expect failure with
`BILLING_CREDITS_EXHAUSTED`.
- [ ] Run a workflow whose only AI step is an `ai-agent` action — expect
that step to fail with `BILLING_CREDITS_EXHAUSTED`, downstream non-AI
steps still run.
- [ ] POST to `/rest/ai/generate-text` — expect
`BILLING_CREDITS_EXHAUSTED`.
- [ ] Create a new chat thread (which kicks off `generateThreadTitle`) —
expect `BILLING_CREDITS_EXHAUSTED`.
- [ ] Run a workflow with no AI step (only DB CRUD/branching/actions) —
expect it to run unaffected.
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
EDIT :
- solving auto-caching from Anthropic by updating ai-sdk/anthropic +
adding providerOption at stream level
- concerning Bedrock, it needs breakpoint
**1. Breakpoints were only on the system prompt**
The code already placed a cache marker on the system prompt (~10K
tokens). But the conversation history — which can grow to hundreds of
thousands of tokens — had no marker, so Anthropic re-read it at full
price on every turn.
The fix adds a prepareStep hook inside streamText that stamps the last
message with a cache breakpoint before every LLM call. Anthropic then
caches the entire conversation prefix, and subsequent turns read it at
$0.30/M instead of $3/M.
prepareStep is used rather than a one-shot pre-processing step because
an agentic turn makes multiple internal LLM calls as tool results
accumulate — the hook refreshes the breakpoint before each one.
**2. Bedrock was using the wrong field**
The system prompt marker for Bedrock was set as cacheControl: { type:
'ephemeral' } — which is the Anthropic wire format. The Bedrock Converse
API expects cachePoint: { type: 'default' }. The system prompt was
silently not being cached on Bedrock at all.
Both the system prompt and the new prepareStep now go through a shared
getCacheProviderOptions helper that returns the correct field per
provider.
**3. Persisted cached token usage to monitor cache strat. efficiency**
## Summary
Adds `h1`–`h6` component overrides to `LazyMarkdownRenderer` so that
`[[record:...]]` references placed inside markdown headings in the AI
chat are parsed by `processChildrenForRecordLinks` and rendered as
clickable `RecordLink` chips, matching the behavior already in place for
`p`, `li`, `td`, `th`, and `a`.
Fixes#20072
## Test plan
- [ ] In AI chat, ask a question whose answer places a record reference
inside a markdown heading (e.g. `## Found [[person:uuid:John Doe]]`) and
confirm a clickable `RecordLink` chip renders instead of the raw
`[[...]]` text.
- [ ] Verify heading styling (sizes, weights, margins) is unchanged.
Generated with [Claude Code](https://claude.ai/code)
---------
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
Co-authored-by: nitin <ehconitin@users.noreply.github.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
- Update usageEvent clickhouse table, partitioning, indexing and
projection (auto materialized view) to optimize credit usage queries
- Add caching for available credits and billing subscription
To do in next PR: deprecate enforceCapUsage cron. Bonus : real-time on
billingSubscription
## Summary
`AgentAsyncExecutorService.executeAgent` consumes Anthropic tokens at
two points (the main `generateText` and the optional structured-output
sub-call). Billing was previously the **caller's** responsibility,
executed only after `executeAgent` returned. If `executeAgent` threw —
e.g. when `structuredResult.output == null` for a schema-mismatched
response, or anything caught by the catch-and-rethrow — we paid
Anthropic but never recorded a `usageEvent`. Likely the dominant source
of the 716M-token-vs-3.27-credits discrepancy seen on the affected
workspace in the 2026-04-26 incident.
## What changed
- Inject `AiBillingService` into `AgentAsyncExecutorService`. Add
`workspaceId` (required), `userWorkspaceId`, and `operationType`
(default `AI_WORKFLOW_TOKEN`) to `executeAgent`'s args.
- Capture `accumulatedUsage`, `cacheCreationTokens`, and
`nativeWebSearchCallCount` into mutable locals as each `generateText`
resolves. A throw between the main and structured-output calls still
bills the first call's tokens; the schema-validation throw still bills
the merged usage.
- Wrap the body in `try { ... } finally { ... }`. The finally calls
`calculateAndBillUsage` and `billNativeWebSearchUsage`, each guarded by
its own `try/catch + logger.error` so a billing exception can't mask the
original execution error or block the second emit.
- `ai-agent.workflow-action.ts`: pass the new args; drop the
now-redundant billing calls and `AiBillingService` injection.
`AiBillingModule` removed from this action's module imports.
- `run-evaluation-input.job.ts`: pass `workspaceId` (already in `data`)
and `userWorkspaceId: null`. **As a side effect, the eval pipeline now
bills correctly** — closing an additional billing leak from the audit
(`RunEvaluationInputJob` previously called `executeAgent` and discarded
`executionResult.usage`).
## Behavior change worth calling out
Previously, failed agent executions were silently free. They will now be
billed for the tokens Anthropic charged us. This is intentional and
correct.
## Test plan
- [ ] Trigger a workflow agent action that succeeds — `usageEvent` count
should match what was previously emitted.
- [ ] Trigger a workflow agent with a JSON response schema and ambiguous
input that produces a non-schema-conforming output
(`structuredResult.output == null`) — verify a `usageEvent` row is now
written for the consumed tokens (was 0 rows previously).
- [ ] Trigger a `runEvaluationInput` GraphQL mutation — verify a
`usageEvent` row is written (was 0 rows previously).
## Notes for review
- Conflicts trivially with the Sentry-context PR on
`run-evaluation-input.job.ts`. Recommend merging Sentry-context first;
this PR's 2-line argument addition rebases inside that PR's
`aiCallContextService.run(...)` callback wrapper.
- A small follow-up after this lands: thread `billingContext` through
the `experimental_repairToolCall` callback in
`agent-async-executor.service.ts:198` (currently marked with a TODO from
the title-gen+repair-tool PR).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
## Summary
`POST /rest/ai/generate-text` calls `generateText` and returns `usage`
to the client without emitting a `usageEvent`. Authenticated, gated only
by `PermissionFlagType.AI` — any workspace user with that permission
could call it in a loop without billing. Identified during the
2026-04-26 incident audit.
## What changed
- Inject `AiBillingService` into `AiGenerateTextController`.
- Add `@AuthUserWorkspaceId() userWorkspaceId: string` to source the
user-workspace identifier.
- Wrap the `generateText` call in `try { ... return ... } finally { ...
}` so billing fires even if the controller throws after Anthropic was
paid.
- Bill with `UsageOperationType.AI_WORKFLOW_TOKEN` and
`cacheCreationTokens: result.usage.inputTokenDetails?.cacheWriteTokens
?? 0`.
- Inner `try/catch` around the billing emit so a billing error can't
break the response.
- One-line module change: `AiGenerateTextModule` imports
`AiBillingModule` (NestJS DI requirement).
## Test plan
- [ ] Call `POST /rest/ai/generate-text` with a small prompt; verify a
`usageEvent` row appears in ClickHouse for the workspace with the
correct token count and `operationType = AI_WORKFLOW_TOKEN`.
- [ ] Call with a malformed model id that throws after the API key is
validated — verify no spurious billing call occurs (no Anthropic call
was made).
## Notes for review
- Response shape unchanged.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com>
## Summary
Two `generateText` call sites were unbilled — identified during the
2026-04-26 incident audit:
- **Thread title generation**
(`AgentTitleGenerationService.generateThreadTitle`) — fires once per new
chat thread; low-volume but completeness matters.
- **Tool-call repair** (`repairToolCall` util, used inside
`experimental_repairToolCall` callbacks) — can fire `MAX_STEPS` times
per agent turn if a model gets stuck producing malformed tool calls.
## What changed
- `AgentTitleGenerationService` — inject `AiBillingService`, expand
`generateThreadTitle` signature to accept `workspaceId` and
`userWorkspaceId`, bill in a `finally` block with
`UsageOperationType.AI_CHAT_TOKEN`. Restructured so the no-default-model
short-circuit happens before the `try`, avoiding a fake billing call.
- `repair-tool-call.util.ts` — added an optional `billingContext` arg
containing `aiBillingService`, `modelId`, `workspaceId`,
`userWorkspaceId`, `operationType`. Wraps the `generateText` call in
`try/finally`; bills in finally with the operation type provided by the
caller. Optional so existing callers keep compiling.
- `chat-execution.service.ts` — threads `billingContext`
(`AI_CHAT_TOKEN`) into the repair callback.
- `agent-chat.service.ts` — passes `workspaceId` and
`thread.userWorkspaceId` to `generateThreadTitle`.
- `agent-async-executor.service.ts` — TODO comment marking that the
repair callback should thread billing once `executeAgent` accepts
`workspaceId` (depends on the executeAgent-finally PR).
## Follow-up
After the executeAgent-finally PR lands, `workspaceId` is in scope at
the `experimental_repairToolCall` callback in
`agent-async-executor.service.ts:198`. Thread `billingContext` through
to fire repair-call billing for workflow agents too, and remove the
TODO. Tiny follow-up.
## Test plan
- [ ] Create a new chat thread; verify a `usageEvent` row is written for
the title generation call.
- [ ] Send a chat message that triggers tool-call repair (e.g. force a
malformed tool call); verify a `usageEvent` row for the repair sub-call.
## Notes for review
- `billingContext` arg made **optional** to allow incremental rollout —
the executeAgent-finally PR plus a small follow-up will close the
agent-async-executor side.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
## 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
Fixes Sentry errors of the form:
> \`messages.3: \`tool_use\` ids were found without \`tool_result\`
blocks immediately after: srvtoolu_…. Each \`tool_use\` block must have
a corresponding \`tool_result\` block in the next message.\`
### Root cause
When the model invokes a **provider-hosted tool** (e.g. Anthropic's
native \`web_search\` — note the \`srvtoolu_\` ID prefix), the AI SDK
marks the resulting \`UIMessagePart\` with \`providerExecuted: true\`.
\`convertToModelMessages\` uses that flag to emit the
tool_use/tool_result pair *inside the same assistant message* — the
format Anthropic requires for server-side tools.
Our \`AgentMessagePart\` persistence was dropping \`providerExecuted\`
on the way to the DB (and re-hydration didn't know to set it). On the
next turn, \`convertToModelMessages\` treated the rehydrated part as a
client-side tool call, splitting it into \`assistant(tool_use)\` +
\`user(tool_result)\` — which Anthropic then rejects with the error
above.
### Fix
- Add nullable \`providerExecuted BOOLEAN\` column on
\`core.agentMessagePart\` via a fast instance command.
- Surface the field on \`AgentMessagePartDTO\` (GraphQL).
- Preserve it through \`mapUIMessagePartsToDBParts\` (server) and both
\`mapDBPartToUIMessagePart\` mappers (server + frontend).
- Include it in \`GET_CHAT_MESSAGES\` and \`GET_AGENT_TURNS\`
selections.
- Regenerate \`generated-metadata/graphql.ts\`.
### Backwards compatibility
Existing rows have \`NULL providerExecuted\` and round-trip as the
omitted flag — which is exactly the pre-fix behaviour for tool parts
that were never provider-executed. Only *new* assistant messages using
\`web_search\` (or other provider-hosted tools) will write \`true\`, and
those are the only ones that were breaking.
## Test plan
- [x] \`npx tsgo\` typecheck — server + front clean
- [x] \`oxlint\` + \`prettier --check\` on all touched files — clean
- [x] \`npx nx run twenty-server:database:migrate:prod\` runs the new
instance command locally; \`providerExecuted\` column present on
\`core.agentMessagePart\`
- [x] Regenerated \`generated-metadata/graphql.ts\` —
\`providerExecuted\` wired into both queries and \`AgentMessagePart\`
type
- [ ] Manual: start a chat with Anthropic web_search enabled, invoke the
tool in turn 1, reply in turn 2 — should not throw the srvtoolu error
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
**PR 1 of 2.** Follow-up PR ships the Exa app, sets it as a default
pre-installed app, and removes the current `WebSearchTool` /
`WebSearchService` / `ExaDriver`. This PR adds the plumbing; no
user-visible change yet.
## Summary
- Server admins can declare a list of npm app packages to auto-install
on every new workspace and backfill onto existing workspaces via CLI.
- Server-level secrets (like Exa's API key) live on the
`ApplicationRegistration` (one row per server, encrypted) and are
injected into logic function execution env at runtime. No more
per-workspace storage of global secrets.
- A generic `POST /app/billing/charge` endpoint lets app logic functions
emit workspace usage events for metered features. Exa uses it in PR 2;
future apps (call recorder, etc.) reuse it.
- `LogicFunctionToolProvider` tool name prefix changes `logic_function_`
→ `app_`. Shorter, accurate (they come from installed apps).
## What's in this PR
**Logic function executor — server-level variables**
- `LogicFunctionExecutorService.getExecutionEnvVariables` now resolves
env vars in the order: hardcoded defaults →
`ApplicationRegistrationVariable[]` (server-level) →
`ApplicationVariable[]` (workspace-level override). The manifest
`serverVariables` schema has existed; this closes the loop.
**Config**
- `PRE_INSTALLED_APPS` — comma-separated list of npm packages. Default:
empty.
**\`PreInstalledAppsService\`** (new module)
- \`onApplicationBootstrap()\` — fetches each package's manifest from
the app registry CDN, upserts an \`ApplicationRegistration\`, and seeds
declared \`serverVariables\` from matching env vars (e.g.
\`EXA_API_KEY\` env → encrypted registration variable).
- \`installOnWorkspace(workspaceId)\` — installs all pre-installed apps
on a single workspace. Tolerates per-app failures.
**Auto-install on new workspace activation**
- \`WorkspaceService.prefillCreatedWorkspaceRecords\` invokes
\`installOnWorkspace\` after prefilling standard records. Non-blocking
on failure.
**Backfill CLI command**
- \`install-pre-installed-apps\` — iterates active and suspended
workspaces, installs pre-installed apps that aren't yet installed.
Idempotent. Run after changing \`PRE_INSTALLED_APPS\`.
**App billing endpoint**
- \`POST /app/billing/charge\`. Authenticated via \`APPLICATION_ACCESS\`
token (already injected into logic function execution env as
\`DEFAULT_APP_ACCESS_TOKEN\`). Body: \`{ creditsUsedMicro, quantity,
unit, operationType, resourceContext? }\`. Emits \`USAGE_RECORDED\` with
\`applicationId\` as \`resourceId\`. Generic — reusable by any app.
**Tool name prefix**
- \`LogicFunctionToolProvider.buildLogicFunctionToolName\` now produces
\`app_<name>\` instead of \`logic_function_<name>\`. Only affects tools
sourced from logic functions; other tool providers unchanged.
## Stats
- 16 files, +501 / −2
- 7 new files (1 command, 1 service × 2, 1 controller, 1 DTO, 2 modules)
- Typecheck: 7 pre-existing errors, zero new
- Prettier clean
## Behavior deltas
- **\`PRE_INSTALLED_APPS\` default = empty**: existing servers see no
change on merge.
- **\`ApplicationRegistrationVariable\` is now read by the executor**:
apps that were using manifest \`serverVariables\` but expecting them to
be ignored by the executor will now see them injected. No apps ship with
\`isTool: true\` logic functions today, so this is latent — first
consumer is Exa in PR 2.
- **Tool prefix**: currently no logic-function tools are named
\`logic_function_*\` in any production flow. The prefix change affects
only future tools emitted by \`LogicFunctionToolProvider\`.
## Risks
- **CDN unavailability at startup**: if the app registry CDN is down,
\`ensureRegistrationsExist\` logs warnings but doesn't block server
start. Installation on new workspaces during this window will find no
registrations and log a non-blocking error. Backfill command can retry
after CDN recovers.
- **Cold-start overhead**: \`ensureRegistrationsExist\` is called once
per process on bootstrap. Current configurable default is empty, so zero
overhead. When an admin sets \`PRE_INSTALLED_APPS\`, they accept one
HTTP call per package at boot.
- **Server-level variables flow**:
\`ApplicationRegistrationVariable.encryptedValue\` is shared by all
workspaces of a server. Appropriate for a single-tenant Exa key. Not
appropriate for per-tenant keys — those go in workspace-level
\`ApplicationVariable\` and override.
## Test plan
- [ ] \`npx nx typecheck twenty-server\` passes (verified: 7
pre-existing unrelated errors, zero new)
- [ ] Set \`PRE_INSTALLED_APPS=@twenty-apps/hello-world\` (or any real
npm-published app), \`HELLO_WORLD_API_KEY=xxx\`, restart server:
\`ApplicationRegistration\` row is upserted,
\`ApplicationRegistrationVariable\` for HELLO_WORLD_API_KEY is populated
(encrypted).
- [ ] Create a new workspace: the app is auto-installed,
\`ApplicationEntity\` row created, \`LogicFunctionEntity\` rows created.
- [ ] Existing workspace: run \`yarn nx run twenty-server:command
install-pre-installed-apps\`: apps install across all workspaces,
idempotent on re-run.
- [ ] Trigger a logic function that reads
\`process.env.HELLO_WORLD_API_KEY\`: value resolves from the
server-level \`ApplicationRegistrationVariable\`.
- [ ] Log a charge from the handler: \`POST /app/billing/charge\` with
\`Authorization: Bearer \$DEFAULT_APP_ACCESS_TOKEN\` body
\`{creditsUsedMicro: 1000, quantity: 1, unit: "INVOCATION",
operationType: "WEB_SEARCH"}\` → returns \`{success: true}\`,
\`USAGE_RECORDED\` event emitted with correct
\`resourceId=applicationId\`.
- [ ] Tool name generated by \`LogicFunctionToolProvider\` starts with
\`app_\`.
## What's NOT in this PR (PR 2 scope)
- The Exa app itself (\`packages/twenty-apps/...\` directory)
- Removing \`WebSearchTool\`, \`WebSearchService\`, \`ExaDriver\`,
\`web-search\` module
- Removing \`WEB_SEARCH_DRIVER\` config var
- Removing the current \`exa_web_search\` entry in
\`ActionToolProvider\`
- Chat preload list updated to \`app_exa_web_search\`
- Frontend \`getToolDisplayMessage\` branch for \`app_exa_web_search\`
- Setting \`PRE_INSTALLED_APPS\` default to include \`@twenty-apps/exa\`
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
- Today `WEB_SEARCH_PREFER_NATIVE` forces a **mutual exclusion**: either
the custom Exa tool preloads as `web_search` or the SDK-native
`web_search` binds. Same name, different backends.
- This PR lets them **coexist**. Custom Exa becomes `exa_web_search`;
native keeps `web_search`. The model picks based on tool descriptions.
- `WEB_SEARCH_PREFER_NATIVE` and `shouldUseNativeSearch()` are deleted.
Exa enablement follows `WEB_SEARCH_DRIVER` (existing). Native enablement
follows the agent's `modelConfiguration.webSearch.enabled` (existing).
## Key changes
**Config / service**
- Deleted `WEB_SEARCH_PREFER_NATIVE` (config-variables.ts)
- Deleted `WebSearchService.shouldUseNativeSearch()`
- `WebSearchService.isEnabled()` unchanged — still gates Exa
availability
**Custom tool rename**
- `ActionToolProvider.toolMap`: `'web_search'` → `'exa_web_search'`
- Descriptor name matches
- `WebSearchTool.description` rewritten to position Exa as
structured/entity-aware, complementary to native
**Native tool binder**
- `NativeToolBinder.bind()` drops the `shouldUseNativeSearch` gate.
Per-agent `modelConfiguration.webSearch.enabled` (inside
`getNativeModelTools`) stays authoritative.
**Chat**
- Preload list now always includes `exa_web_search` —
`ActionToolProvider` silently skips the descriptor when Exa is disabled,
so `getToolsByName` degrades gracefully
- Native tools always attempted; returns empty ToolSet when the model
doesn't support them
- `directTools = { ...preloadedTools, ...nativeSearchTools }` — both
present when both enabled
- `billNativeWebSearchUsage` called unconditionally (the function
already short-circuits on count ≤ 0)
**Workflow agent**
- Same unconditional billing pattern
- `WebSearchService` dependency removed
**System prompt**
- Dropped the special-cased `web_search` branch. Preloaded tools list
uniformly now.
**Frontend**
- `exa_web_search` reuses the same "Searching the web for X" display as
native
- Test coverage added
## Billing isolation (verified)
- `countNativeWebSearchCallsFromSteps` counts `toolName ===
'web_search'` only. After the rename, only native calls match. Exa calls
(`exa_web_search`) are billed separately via
`WebSearchService.emitUsageEvent` inside `search()`.
- No double-billing path.
## Behavior deltas (intended)
| Scenario | Before | After |
|---|---|---|
| Anthropic model + Exa enabled + PREFER_NATIVE=true | native only |
**both** |
| Anthropic + Exa enabled + PREFER_NATIVE=false | Exa only (as
`web_search`) | **both** |
| Non-native model + Exa enabled | Exa as `web_search` | Exa as
`exa_web_search` |
| Any model + Exa disabled + native supported | native only | native
only |
| Workflow agent with `webSearch.enabled=true` + Anthropic + Exa enabled
| native only | **both** |
## Known regression (accepted)
Customers who set `WEB_SEARCH_PREFER_NATIVE=false` to force Exa-only
will now **also** see native `web_search` if the model supports it.
There's no chat-level kill switch after this PR. Per discussion, this is
accepted — future model-level capability gating (in the model JSON) will
be the right place for that control.
## Stats
- 10 files, +63 / −73 (net deletion)
- Typecheck clean (server: 7 pre-existing unrelated, front: 13
pre-existing unrelated — zero new either side)
- Prettier clean
## Test plan
- [ ] `npx nx typecheck twenty-server` and `npx nx typecheck
twenty-front` pass
- [ ] With Anthropic + Exa enabled: chat shows both `web_search` and
`exa_web_search` in preloaded list; model can call either
- [ ] With Anthropic + Exa disabled: chat shows only native `web_search`
- [ ] With non-native model + Exa enabled: chat shows only
`exa_web_search`
- [ ] Workflow agent with `modelConfiguration.webSearch.enabled=true` +
Exa enabled: both available
- [ ] Billing: native calls billed via `billNativeWebSearchUsage`; Exa
calls billed via `WebSearchService.emitUsageEvent`; no double-billing
- [ ] Frontend: `exa_web_search` renders "Searching the web for X" the
same as `web_search`
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
**Stacked on top of #19960.**
## Summary
- `execute_tool` used to check `directTools[toolName]` first, falling
back to the registry. Same tool name, different wrapping: preloaded went
through `wrapToolsWithOutputSerialization`, fallback didn't. Silent
divergence — a model calling a CRUD tool via
\`learn_tools\`/\`execute_tool\` got raw output, while calling it as a
preloaded direct tool got compacted output.
- Now: `execute_tool` always routes through
`toolRegistry.resolveAndExecute`. One path, no fast-path.
- Output serialization (`compactToolOutput`) moves into the registry,
gated by a new `serializeOutput` flag on `hydrateToolSet` /
`resolveAndExecute` / `getToolsByName` / `getToolsByCategories` /
`ToolRetrievalOptions`. Chat passes `true`, MCP and workflow pass
`false`.
## Key changes
**Registry (`tool-registry.service.ts`)**
- `hydrateToolSet` options gain `serializeOutput?: boolean`; when true
the execute closure wraps dispatch result with `compactToolOutput`.
- `resolveAndExecute` signature: replaces unused \`_options:
ToolExecutionOptions\` with `{ serializeOutput?: boolean }`.
- `getToolsByName` and `getToolsByCategories` thread `serializeOutput`
through to `hydrateToolSet`.
**Meta-tool (`execute-tool.tool.ts`)**
- API changes from positional `(toolRegistry, context, directTools?,
excludeTools?)` to `(toolRegistry, context, options?: { excludeTools?,
serializeOutput? })`.
- `directTools` fallback removed. All invocations go to the registry.
**Chat (`chat-execution.service.ts`)**
- Passes `serializeOutput: true` to `getToolsByName` — preloaded tools
get compacted output from the hydrator, no external wrap needed.
- Drops the external `wrapToolsWithOutputSerialization(preloadedTools)`
call.
- `createExecuteToolTool` call now passes `{ serializeOutput: true }`.
Direct-tool and `execute_tool` paths produce identical output shape.
**MCP (`mcp-protocol.service.ts`)**
- `createExecuteToolTool` call updated to new options shape with `{
excludeTools: MCP_EXCLUDED_TOOLS }`. No `serializeOutput` flag → raw
output as today.
**Deletes**
- `output-serialization/wrap-tools-with-output-serialization.util.ts` —
sole caller removed.
## Behavior changes
- **Chat, `execute_tool` fallback path**: now produces compacted output
(matches direct path). Net effect: fewer tokens for CRUD results reached
via discovery. Intended improvement.
- **Chat, `execute_tool({toolName: 'web_search'})` edge**: today
silently hits the native tool via `directTools`; now returns \"tool not
found, use get_tool_catalog\". Self-correcting, rare — native tools are
always directly available to the model.
- **MCP**: no change. No `serializeOutput` flag → identical raw output.
- **Workflow agent**: no change. Doesn't use `execute_tool`.
## Test plan
- [ ] `npx nx typecheck twenty-server` passes (verified: 7 pre-existing
unrelated errors, zero new)
- [ ] \`npx jest
packages/twenty-server/src/engine/api/mcp/services/__tests__/mcp-protocol.service.spec.ts\`
passes in CI
- [ ] AI chat: call a preloaded tool (e.g. \`search_help_center\`)
directly → compacted output
- [ ] AI chat: call a non-preloaded CRUD tool via
\`learn_tools\`/\`execute_tool\` → compacted output (this is the
behavior change)
- [ ] AI chat: native \`web_search\` still works when model calls it
directly
- [ ] MCP: \`tools/call\` on a registry tool → raw output (nulls
preserved)
- [ ] Workflow AI agent: tool dispatch unchanged
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
## Summary
- AI is now GA, so the public/lab `IS_AI_ENABLED` flag is removed from
`FeatureFlagKey`, the public flag catalog, and the dev seeder.
- Drops every backend `@RequireFeatureFlag(IS_AI_ENABLED)` guard (agent,
agent chat, chat subscription, role-to-agent assignment, workflow AI
step creation) and the now-unused `FeatureFlagModule`/`FeatureFlagGuard`
wiring in the AI and workflow modules.
- Removes frontend gating from settings nav, role
permissions/assignment/applicability, command menu hotkeys, side panel,
mobile/drawer nav, and the agent chat provider so AI UI is always on.
Tests and generated GraphQL/SDK schemas updated accordingly.
## Test plan
- [x] `npx nx typecheck twenty-shared`
- [x] `npx nx typecheck twenty-server`
- [x] `npx nx typecheck twenty-front`
- [x] `npx nx lint:diff-with-main twenty-server`
- [x] `npx nx lint:diff-with-main twenty-front`
- [x] `npx jest --config=packages/twenty-server/jest.config.mjs
feature-flag`
- [x] `npx jest --config=packages/twenty-server/jest.config.mjs
workspace-entity-manager`
- [ ] Manual smoke test: AI features still accessible without any flag
row in `featureFlag`
🤖 Generated with [Claude Code](https://claude.com/claude-code)
---------
Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.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
- The exception class under `ai-agent/` was serving every AI surface
(agent, chat, role, models, generate-text), so `Agent` was a misnomer.
Promoted to the `ai/` namespace; renamed `AgentException` →
`AiException`, `AgentExceptionCode` → `AiExceptionCode`, and related
interceptor / filter / handler / file names accordingly.
- Split the single `AGENT_NOT_FOUND` code into entity-specific codes.
Chat-thread lookups no longer reuse the agent identifier.
- **Fixes Sentry 500s on `GetChatMessages` / `chatThread`.** Every
"Thread not found" and "Queued message not found" throw site in ai-chat
was previously wired to `AGENT_EXECUTION_FAILED`, which maps to
`InternalServerError` (HTTP 500). They now use `THREAD_NOT_FOUND` /
`MESSAGE_NOT_FOUND`, both of which map to `NotFoundError` (HTTP 404) in
the GraphQL and REST handlers.
The underlying cause of *why* clients are asking for threads that no
longer resolve for them — per-user chat-thread create events being
broadcast workspace-wide — is addressed separately in a follow-up PR.
### Code map
- Added: `ai/ai.exception.ts`,
`ai/utils/ai-graphql-api-exception-handler.util.ts` (+ spec with new
THREAD/MESSAGE cases),
`ai/interceptors/ai-graphql-api-exception.interceptor.ts`,
`ai/filters/ai-api-exception.filter.ts`
- Deleted: `ai/ai-agent/agent.exception.ts`,
`ai/ai-agent/utils/agent-graphql-api-exception-handler.util.ts` (+
spec),
`ai/ai-agent/interceptors/agent-graphql-api-exception.interceptor.ts`,
`ai/ai-agent/filters/agent-api-exception.filter.ts`
- Updated: 21 call sites across ai-agent, ai-agent-execution,
ai-agent-role, ai-chat, ai-generate-text, ai-models, role, and
workspace-migration validators.
## Test plan
- [x] `npx nx typecheck twenty-server`
- [x] `npx jest ai-graphql-api-exception-handler` (3/3 including new
THREAD_NOT_FOUND and MESSAGE_NOT_FOUND cases)
- [x] `npx jest agent-role.service` (9/9)
- [x] `npx oxlint --type-aware` on all changed files (0 warnings/errors)
- [x] `npx prettier --check` on all changed files
- [ ] CI
## Summary
Fixes cross-user cache contamination for AI chat threads in multi-user
workspaces.
`agentChatThread` is the only user-scoped (`userWorkspaceId`-filtered)
entry in `METADATA_NAME_TO_ENTITY_KEY`, but its create/update events
were going through `WorkspaceEventBroadcaster`, which fans out to every
active SSE stream in the workspace. Every client therefore received
other users' threads into their local `agentChatThreads` metadata store
(which is persisted to localStorage). On a subsequent session,
`AgentChatThreadInitializationEffect` would pick the
most-recently-updated thread — potentially another user's — and fire
`GetChatMessages` against it; the server's `userWorkspaceId` filter
didn't match, producing "Thread not found" errors (now 404 thanks to
twentyhq/twenty#19831).
### The fix mirrors the RLS pattern already used for object records
`ObjectRecordEventPublisher` already reads
`streamData.authContext.userWorkspaceId` to filter per subscriber.
Metadata events had no equivalent. This PR closes that gap with a
minimal, opt-in change:
- Add optional `recipientUserWorkspaceIds?: string[]` to
`WorkspaceBroadcastEvent`. Omit → workspace-wide (unchanged for views,
objects, fields, etc.). Set → delivered only to streams whose
`authContext.userWorkspaceId` is in the list.
- `WorkspaceEventBroadcaster.broadcast` builds the payload per stream
and skips events whose recipient doesn't match.
- Both `agentChatThread` broadcast call sites in `AgentChatService` now
pass `recipientUserWorkspaceIds: [userWorkspaceId]`.
### Scope
3 files, +40/-11 LOC:
-
`subscriptions/workspace-event-broadcaster/types/workspace-broadcast-event.type.ts`
-
`subscriptions/workspace-event-broadcaster/workspace-event-broadcaster.service.ts`
- `metadata-modules/ai/ai-chat/services/agent-chat.service.ts`
### Stale cache note
Existing clients still carry poisoned localStorage from before this
lands. They self-heal on sign-out/in because
`clearAllSessionLocalStorageKeys` already drops `agentChatThreads`. If
we want to actively flush on deploy, a frontend cache-version bump can
follow in a separate PR.
### Related
- twentyhq/twenty#19831 turns the resulting "Thread not found" from 500
to 404. This PR addresses the root cause; #19831 stops the Sentry noise.
## Test plan
- [x] `npx nx typecheck twenty-server`
- [x] `npx oxlint --type-aware` on all 3 files (0 warnings/errors)
- [x] `npx prettier --check` on all 3 files
- [ ] Manual: two users in the same workspace, user A creates/sends a
chat thread — confirm user B's session does not receive the event and
their `chatThreads` sidebar is unaffected
- [ ] CI
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
Automated daily sync of `ai-providers.json` from
[models.dev](https://models.dev).
This PR updates pricing, context windows, and model availability based
on the latest data.
New models meeting inclusion criteria (tool calling, pricing data,
context limits) are added automatically.
Deprecated models are detected based on cost-efficiency within the same
model family.
**Please review before merging** — verify no critical models were
incorrectly deprecated.
Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com>
github-actions[bot]