027a09809928d92b28ea234007bd59e40b8e3036
218 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
 Félix Malfait
|
027a098099 |
fix(ai): fully roll back a question answer when the resume enqueue fails (#22492)
## Rationale `answerAgentChatQuestion` is a three-step transaction without the transaction: resolve the question (flip tool part to `answered`, clear `pendingQuestionMessageId`, claim the stream), then enqueue the resume job. If the **enqueue fails**, the catch restores only `activeStreamId`. What's left behind: tool part says `answered`, `pendingQuestionMessageId` is `null`, no job will ever run. The client's own error handler rolls its card back to *pending* — so the user sees an answerable question whose re-submission deterministically throws `QUESTION_NOT_PENDING`. The turn is stuck and state is divergent on three surfaces (DB part, DB thread, client). ## Why this is the root cause, not a symptom patch The failure path was rolling back one of three writes. This makes the rollback total and **exact**: `resolvePendingQuestion` now returns the part's precise previous `toolOutput` (no reconstruction guesswork — question tools can carry arbitrary output fields), and the failure path restores the part verbatim plus the thread's pending-question state, guarded on the observed streamId so a competing claim is never clobbered. After rollback, server and client agree again: the question is pending, answering retries cleanly. The audit's alternative — forward recovery (keep the answers, mark the turn interrupted, resume via Retry) — has nicer UX in isolation but contradicts the client's existing rollback-to-pending behavior; matching the established contract wins until the client changes. ## User impact A transient Redis/queue hiccup at answer time currently bricks the question turn permanently. With this, the user sees the question again and can just re-answer. ## Test plan - [ ] CI green - [ ] Manual: fail the enqueue (kill Redis briefly) at answer time → question card returns to pending, re-answer succeeds https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22492?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. --> |
||
|
Félix Malfait
|
087bee0036 |
fix(ai): notify all tabs when a pending question is answered (#22491)
## Rationale `resolvePendingQuestion` updates the question tool-part to `answered` and re-claims the thread — but publishes **nothing**. The answering tab converges via a local browser event; every other tab keeps rendering the question card as interactive until the resumed stream's first chunk happens to arrive. A second tab (or teammate view on shared context) can attempt to answer an already-answered question and hit a confusing `QUESTION_NOT_PENDING` error. ## Why this is the root cause, not a symptom patch Answering a question is a state transition every subscriber cares about — exactly like queue promotion, message persistence, and stream errors, all of which publish. This transition just never did. The fix publishes the existing refetch-trigger event (`queue-updated`, which every tab already handles by refetching messages + thread state) right after resolution — no new event type, no new client code path, consistent by construction with how every other transition converges tabs. A dedicated `question-answered` event carrying the answers would save one refetch round-trip; the audit's verdict was that's over-engineering for a rare interaction. Publishing *before* the resume-enqueue is deliberate: even if the enqueue fails, the question **is** answered server-side, and tabs should reflect server truth. ## User impact Second tabs stop offering an interactive question that will error when submitted; everyone sees the answered state within a refetch instead of whenever the stream resumes. ## Test plan - [ ] CI green - [ ] Manual: two tabs on one thread, answer the question in tab A → tab B's card flips to answered without interaction https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22491?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. --> |
||
|
Félix Malfait
|
d1a4e250cf |
fix(ai): converge every tab's transcript when a stream fails with partial output (#22494)
## Rationale When a turn fails after emitting output, the partial assistant message **is persisted** — but only the success path publishes the event (`message-persisted`) that makes other tabs refetch. On failure, tabs that weren't watching the live stream keep a stale transcript until manual reload. (Error *visibility* itself already works — `stream-error` reaches every subscribed tab — the gap is purely the persisted-transcript sync. The original gap-analysis framing of this as an error-visibility problem was wrong; this is the corrected scope.) ## Why this is the root cause, not a symptom patch Turn settlement should converge subscribers regardless of *how* the turn settled — success and failure both persist state that tabs need. The failure path now publishes the same refetch-trigger event the rest of the lifecycle uses, right after the terminal `stream-error`. No new event type, no client changes: the existing guarded replay (`firstLiveSeq === null`) already ensures tabs with a live view keep their in-place error rendering while background tabs pick up the persisted partial message and error state. ## User impact Open the same thread in two tabs, have the turn die mid-answer in one: the other tab currently shows the conversation frozen pre-turn until reload. Now both converge to the persisted partial output plus the failed-turn state within one refetch. ## Test plan - [ ] CI green - [ ] Manual: two tabs, kill the provider mid-stream in tab A → tab B shows the partial message + error without reload https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22494?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. --> |
||
|
Félix Malfait
|
a505ed3245 |
feat(ai): typed CONTEXT_WINDOW_EXCEEDED error that hides the pointless Retry (#22488)
## Rationale When message pruning can't fit the conversation into the model's context window, `chat-execution.service.ts` throws a **raw `Error`**. `mapErrorToStreamError` classifies it as generic `STREAM_EXECUTION_FAILED`, so the client renders a standard failure with a **Retry button that deterministically fails again** — the conversation doesn't get shorter by retrying. Users loop on Retry against a permanently-failing thread. ## Why this is the root cause, not a symptom patch The failure is *terminal for the thread by construction*, and the error channel already distinguishes terminal-vs-retryable via typed `AiExceptionCode`s — this failure just never got one. Adding `CONTEXT_WINDOW_EXCEEDED` (typed exception → `UserInputError` mapping instead of a 500 → both error surfaces render the start-a-new-thread message without `onRetry`) puts it on the same rails as `API_KEY_NOT_CONFIGURED` and the other special-cased codes. Both frontend error surfaces route through `AiChatErrorRenderer`, so one case covers the in-message and under-list renderings. The deeper endgame (auto-summarize/compact older turns so threads never brick) is a multi-week feature — and this typed error remains necessary even then, as its terminal fallback. ## User impact Instead of an opaque error and a Retry that never works, users hitting the context limit get told exactly what happened and what to do (start a new thread), and monitoring stops counting a user-condition as a server error. ## Test plan - [ ] CI green - [ ] Manual: fill a thread past the model limit → typed message, no Retry on either error surface https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22488?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. --> |
||
|
Félix Malfait
|
56a6f9419b |
chore(server): fix stream-agent-chat job spec formatting breaking main's lint (#22493)
The stacked merges of #22479 and #22480 left `stream-agent-chat.job.spec.ts` on main failing the `oxfmt --check` gate (2 stray blank lines), which currently fails `server-lint-typecheck` on **every** open PR's merge ref. Two-line whitespace fix, no behavior change. Merging this first unblocks the rest of the queue. https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22493?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. --> |
||
|
Félix Malfait
|
3b76ec528f |
fix(ai): route missing-workspace stream failures through the standard error path (#22480)
## Rationale When `StreamAgentChatJob` can't find the workspace, it publishes a transient `stream-error` event and **returns before the try/finally exists** (`stream-agent-chat.job.ts`). Consequences on main: - `activeStreamId` is never cleared → every subsequent send in that thread queues behind a dead claim, forever; - no `lastStreamError` is persisted → nothing renders after a reload, and Retry has nothing to retry; - nothing throws → **zero telemetry**. Sentry confirms: the "Workspace not found" issues that exist are all auth/Stripe paths — this path fails in complete silence. ## Why this is the root cause, not a symptom patch The job's catch/finally already implement the correct failure contract for *every other* error: persist a typed `lastStreamError`, publish the typed event, release the claim guarded on the observed streamId. The bug is that one code path bypasses that contract via an early return. The fix removes the bypass — the lookup moves inside the `try` and throws a typed `AiException(WORKSPACE_NOT_FOUND)` — rather than duplicating cleanup in the early-return branch (which would be the symptom patch, and would drift the next time the contract changes). The alternative "prevent the job from existing when the workspace is gone" isn't achievable: workspace deletion between enqueue and pickup is an inherent race, so the job must handle it regardless. ## User impact A workspace deleted/deactivated mid-flight currently bricks the thread silently (the user just sees sends vanish into a queue). With this, the failure is visible (typed error message), recoverable (standard failed-turn state), and observable (real exception in monitoring). ## Stack Based on #22479 (spec harness) — it extends the same spec file with the regression test. `WORKSPACE_NOT_FOUND` is a TypeScript enum member, not a GraphQL schema change: no client-sdk regeneration needed. ## Test plan - [x] Regression test: missing workspace → typed rejection, `lastStreamError` persisted, terminal event published, claim released - [ ] CI green https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22480?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. --> |
||
|
Félix Malfait
|
ca8ab32253 |
fix(ai): delete the Redis chunk list on credits-exhausted terminal events (#22477)
## Rationale The AI chat stream keeps every published chunk in a Redis list (`agent-chat-stream-chunks:<threadId>`, 1h TTL) so late subscribers can catch up. On `message-persisted` the list is deleted. On `credits-exhausted` — the *other* successful terminal event — it wasn't. Any reload/refetch within the TTL replayed the orphaned chunks, flipping the thread into a "streaming" state that no terminal event ever closes: an endless spinner until the user sends another message. **Production evidence (Sentry):** `Billing Credits Exhausted` fired for **290 users / 937 events in 90 days**, ongoing ([TWENTY-SERVER-G42](https://twenty-v7.sentry.io/issues/TWENTY-SERVER-G42)) — every one of those users who reloads the chat within an hour hits this. ## Why this is the root cause, not a symptom patch The chunk list's lifecycle contract is "cleared when the turn settles". `credits-exhausted` resolves the job successfully **without** persisting a `lastStreamError`, so unlike `stream-error` there is no persisted terminator for catchup to replay after the chunks — the replay is unconditionally un-closeable. Deleting on both settle events restores the contract exactly where it's already enforced for `message-persisted`. `stream-error` deliberately keeps the list: the persisted error acts as the replay terminator, letting a reloading client still see the failed turn's partial output. ## User impact Users who hit their billing cap mid-answer (~100/month) no longer come back to a permanently spinning thread after a reload — they see the settled conversation and the billing state. ## Test plan - [x] Unit spec: chunk accumulation with 1-based seq, deletion on both terminal events (`it.each`), retention on `stream-error` - [ ] CI green https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22477?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. --> |
||
|
Félix Malfait
|
6ac8ebcd18 |
test(ai): pin StreamAgentChatJob stream lifecycle with a reusable spec harness (#22479)
## Rationale `StreamAgentChatJob` is the most failure-sensitive path in the AI chat stack — it coordinates the model stream, Redis event publishing, message persistence, the thread's stream claim, and the queued-message flush — and it had **zero unit coverage**. Both historical hangs lived here: - a throw before the model stream merges bypassed `onFinish` entirely and hung the job until the **10-minute BullMQ lock** expired (thread stuck the whole time); - a throw inside `onFinish` (persistence failure) left trailing chunks published with **no terminal event** — the client spinner ran forever. Production still shows this class is live: `Query read timeout` thrown from inside `handleStreamFinish` ([TWENTY-SERVER-GV7](https://twenty-v7.sentry.io/issues/TWENTY-SERVER-GV7)). ## Why this shape, not something else Tests-only PR, zero production risk. The fake chat stream mirrors the one AI SDK contract the job's coordination depends on — verified against the installed `ai@6.0.97` dist: `toUIMessageStream` converts mid-stream errors into error parts and **always** fires `onFinish` when the stream ends (`handleUIMessageStreamFinish` invokes it from both `flush()` and `cancel()`). Pinning that contract in the fake means a future SDK upgrade that breaks it fails these tests instead of production. Six tests pin current behavior: chunk ordering with `message-persisted` last, opaque error-chunk suppression, mid-stream failure persisting `lastStreamError` + releasing the claim, the two hang regressions above, and cancel skipping the queued flush. Three sibling PRs extend this exact spec file (missing-workspace routing, halted queue, and — later — auto-retry), which is why the harness lands first. ## User impact None directly; it makes the two worst historical user-facing hangs (10-minute dead thread, infinite spinner) regression-proof before the stuck-state fix series touches this code. ## Test plan - [x] 6 unit tests, no production code changed - [ ] CI green https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38 --- _Generated by [Claude Code](https://claude.ai/code/session_01Lyi6zTema2FMVVh8MD6c38)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22479?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. --> |
||
|
Félix Malfait
|
4aaf171d63 |
feat(ai): add ask_questions interactive clarifying-question tool (#22346)
## What & why Adds an `ask_questions` tool that lets the in-app **Ask AI** assistant **pause a turn to ask the user one or more multiple-choice questions** (per the [Figma design](https://www.figma.com/design/xt8O9mFeLl46C5InWwoMrN/Twenty?node-id=105959-117153)) and resume once answered — instead of guessing on ambiguous/consequential decisions. The tool is **harness-only**: an interactive question UI is meaningless without a user to answer it, so it must be absent from MCP and from head-less workflow agents. ## Design — true tool-result resume (not a synthetic user message) The user's answer is a **structured tool result bound to the `toolCallId`**, and the **same agent turn resumes** — exactly how Anthropic (`tool_result` by `tool_use_id`) and OpenAI (`function_call_output`) model human-in-the-loop. The naive form of this (leave the tool call in `input-available` to mean "pending") is **impossible** here: `finalizeDanglingToolParts` rewrites `input-available` → `output-error` ("Tool execution was interrupted") on both the persist path (`addMessage`) and the model-reload path (`chat-execution.service.ts`). That util is a load-bearing safety net, so weakening it is the wrong move. Instead: - `ask_questions` is an **inline, chat-only tool with an `execute` that returns a `status: 'pending'` result immediately**, so the tool part is always `output-available` and **immune to `finalizeDanglingToolParts`**. `stopWhen(hasToolCall('ask_questions'))` halts the turn right after the call (the model never sees the placeholder). - A nullable **`thread.pendingQuestionMessageId`** marker records that a turn is awaiting an answer. - The new **`answerAgentChatQuestion`** mutation atomically *claims* the question (clears the marker, marks the thread streaming), **writes the answer onto the same tool part** (`status: 'answered'`), and **re-enqueues the turn via the existing `existingTurnId` plumbing** (`isResume` bypasses the per-turn dedup guard). On resume `finalizeDanglingToolParts` leaves the `output-available` part untouched and `convertToModelMessages` emits `assistant(tool_use)` + `tool_result(answers)`, so the model continues. This achieves the platform-aligned semantics **without** weakening the finalize safety net or inventing a fragile new part state. ### Meets the two requirements - **Survives refresh, scoped per-thread** — the pending state is a normal persisted `output-available` part + the thread marker; the frontend card is derived per-thread from the loaded messages, so it re-appears on reload and only on its own thread. - **Takes priority over the queue** — a unified `isBlocked = activeStreamId || pendingQuestionMessageId` gate is applied in both `sendChatMessage` (new messages queue) and `flushNextQueuedMessage` (the drain). The queue cannot unpile until the question is answered and the resumed turn completes. ### Harness-only by construction `ask_questions` is added **only** to the chat's inline `activeTools` (like `learn_tools`/`execute_tool`/`load_skills`). It never enters the tool registry/catalog, so it is invisible to MCP and to workflow agents — no `MCP_EXCLUDED_TOOL_NAMES` entry needed. ## UX While a question is pending, the **composer is replaced by the question card** (matching the Figma): question title + pager (`1/2`), numbered option rows (`IconSquareNumber*`) with per-option info-icon descriptions and a "Recommended" badge, and the normal composer as the free-text fallback ("Type anything to do differently."). The transcript shows a compact "Asking questions…" status line that becomes an answered summary. ## Changes **twenty-shared** - `ai/types/AskQuestionsToolTypes.ts` — `AskQuestionItem/Option/Answer/Result`, `ASK_QUESTIONS_TOOL_NAME`. **twenty-server** - `ai-chat/tools/ask-questions.tool.ts` — inline tool factory (pending-result `execute`, zod schema, 1–4 questions × 2–4 options). - `chat-execution.service.ts` — add to `activeTools` + `preloadedToolNames`; `hasToolCall` in `stopWhen`. - `chat-system-prompts.const.ts` — when-to-use guidance. - `entities/agent-chat-thread.entity.ts` — `pendingQuestionMessageId` column. - `stream-agent-chat.job.ts` — set the marker on a question pause; bypass the dedup guard on resume; suppress the no-text warning for question pauses. - `agent-chat-streaming.service.ts` — gate `flushNextQueuedMessage`; `enqueueResumeStream`. - `agent-chat.resolver.ts` — gate `sendChatMessage`; `answerAgentChatQuestion` mutation. - `agent-chat.service.ts` — `resolvePendingQuestion` (atomic claim + write answer). - `dtos/agent-chat-question-answer.input.ts`, `ai.exception.ts` (`QUESTION_NOT_PENDING`), `utils/find-pending-question-part.util.ts`. **twenty-front** - `components/AiChatQuestionCard.tsx` — the interactive card (matches Figma tokens) + `__stories__/AiChatQuestionCard.stories.tsx`. - `components/AiChatEditorSection.tsx` — swap the composer for the card while pending. - `components/AiChatQuestionStatusRenderer.tsx` + branch in `AiChatAssistantMessageRenderer.tsx`. - `states/selectors/agentChatPendingQuestionComponentSelector.ts`, `types/AgentChatPendingQuestion.ts`. - `hooks/useSubmitQuestionAnswer.ts` + `utils/markQuestionAnswered.ts` (optimistic) + `graphql/mutations/answerAgentChatQuestion.ts`. A design doc lives at `packages/twenty-server/docs/ASK_USER_QUESTION_TOOL_PLAN.md`. ## Migration Adds a nullable `pendingQuestionMessageId` (uuid) column to `core.agentChatThread`. Needs a generated **fast instance command** (`database:migrate:generate --name addThreadPendingQuestion --type fast`) — see "Verification status". ## Tests - Server: `ask-questions.tool.spec.ts` (pending echo + schema bounds), `find-pending-question-part.util.spec.ts`. - Front: `markQuestionAnswered.test.ts`, plus the Storybook story. ## Verification status (please read) This branch was authored in an environment where the monorepo `yarn install` repeatedly failed on transient TLS resets from the package registry, so I could **not** locally run the mechanical gates. The logic was reviewed by hand and the `ai@6.0.97` exports used (`hasToolCall`, `stepCountIs`, `generateId`) were confirmed against the package's type defs. Still **TODO** (will rely on CI / a follow-up once deps install): - [ ] `nx run twenty-shared:generateBarrels` (the `ai/index.ts` export was added by hand; regen to reconcile) - [ ] `nx run twenty-front:graphql:generate` (new mutation + input type) - [ ] generate the fast instance command (migration) for the new column - [ ] `typecheck` + `lint:diff-with-main` (front + server) — expect minor import-ordering autofixes - [ ] run the unit tests **Screenshots:** reproducing the live flow needs an AI provider API key (to get the model to actually call `ask_questions`), which isn't available here. The card can be screenshotted from its **Storybook story** (`AiChatQuestionCard.stories.tsx`) with no API key — I'll add that image once deps install, or a reviewer can run `nx storybook twenty-front`. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01AArS8H3y3Z1Qwm763xhPLB --- _Generated by [Claude Code](https://claude.ai/code/session_01AArS8H3y3Z1Qwm763xhPLB)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22346?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. --> |
||
|
Félix Malfait
|
d709467902 |
feat(ai): surface AI chat stream failures through one typed error channel (#22434)
## Context Investigating a report where the AI chat showed only a `...` spinner while the network response clearly contained `No AI models are available`. Root cause: terminal stream failures reach the client on **two mismatched channels**. | Representation | Persisted (survives reload) | Rendered by client | |---|---|---| | AI-SDK `error` chunk (inside `stream-chunk`) | ✅ RPUSH'd to Redis | ❌ dropped by `readUIMessageStream` (no message part, no error state) | | typed `stream-error` event | ❌ never persisted | ✅ sets the error atom | Live, the `stream-error` event renders. But on reload, `chatStreamCatchupChunks` replays only the persisted **error chunk** — which the reducer discards — and the streaming indicator never clears. ## Change Collapse to a single typed error contract: - **Suppress the opaque `error` chunk** in the stream job; every failure is surfaced through the typed `stream-error` event. Errors are mapped via `mapErrorToStreamError` so an `AiException` keeps its `AiExceptionCode` (e.g. `API_KEY_NOT_CONFIGURED` → the existing "AI not configured" banner) instead of leaking a raw string. - **Persist the terminal error** next to the accumulated chunks and expose it as an explicit `error { code message }` field on `ChatStreamCatchupChunks`, so a client catching up after a reload recovers it — no dependency on the AI SDK's internal chunk shape. - **Reset per-thread stream state at job start**, so a failed turn's leftover chunks/error never replay on the next stream. - **Client replays the catchup error** as a terminal `stream-error` event, which clears the streaming indicator and renders the error (fixes the infinite spinner on a stream that ended in error). ## Notes - `ChatStreamError` is a new metadata GraphQL type; generated types (twenty-front metadata + client-sdk) were hand-updated to keep the tree consistent and will be reconciled by CI's `graphql:generate` check if anything differs. - Server unit test added for the error mapping. No schema/DB migration. ## Test plan - [ ] With no AI provider configured, send a chat message → error renders immediately (not a spinner). - [ ] Reload the thread → the error still renders (recovered from catchup), indicator not spinning. - [ ] Configure a provider and send again → normal streaming; no stale error from the previous failed turn. <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22434?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. --> |
||
|
github-actions[bot]
|
18c0d117a3 |
chore: sync AI model catalog from models.dev (#22387)
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. <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22387?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. --> Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com> |
||
|
Etienne
|
56deba351b |
feat(ai): reliable bulk data import via code-interpreter (#22209)
## Summary Makes AI-assisted bulk data import (CSV/Excel/spreadsheets) reliable and token-efficient by letting an entire import run inside a single code-interpreter call, with a persistent sandbox session and server-side bulk helpers. Also includes supporting improvements to attachment handling, upsert reporting, and field-permission error messages. ## Changes ### Code interpreter - **Persistent per-session kernel** in `LocalDriver`: a long-lived Python process per `sessionId` keeps variables, imports, and files alive across calls (matching E2B behavior). Falls back to the existing ephemeral per-call path when no session is provided. CAN BE REMOVED, INTERESTING FOR DEV X - Idle watchdog that self-terminates the kernel, configurable via the new `CODE_INTERPRETER_IDLE_TIMEOUT_MS` config variable; the process also exits on parent shutdown (EOF on control fd). CAN BE REMOVED, INTERESTING FOR DEV X - New `bulk_upsert` and `lookup_by` helpers on the sandbox `twenty` object for idempotent batched writes (≤200/batch) and bounded relation-ID resolution. ### Records - `upsert_many_*` now reports a `created` / `updated` / `total` split in its result and log line (new `isFreshlyCreatedRecord` util). ### AI chat - `replaceUnsupportedFileParts`: user-attached files whose MIME type the model can't handle natively (and that aren't code-interpreter-supported) are downgraded to a descriptive text note instead of being sent as unsupported file parts. Modality→MIME mapping drives native support detection. - Finalize dangling tool parts before `convertToModelMessages` to avoid malformed model messages. - Extracted shared types/constants for code-interpreter file extraction. ### Permissions - Field permission-denied exceptions now include the field name and entity name for easier debugging. ### Skill docs - Added the bulk-import recipe ## To do in following PR - [ ] Skill command migration ## Test plan - [x] Unit tests for `getNativeMimeTypesForModalities` and `replaceUnsupportedFileParts` pass - [x] Run a bulk import (>50 rows) end-to-end through the code interpreter and verify a single sandbox call handles read → resolve relations → upsert → summary - [x] Verify session persistence: define a variable in one call, use it in the next within the same session - [x] Verify the kernel self-terminates after `CODE_INTERPRETER_IDLE_TIMEOUT_MS` - [x] Verify unsupported attachments are replaced with a text note for models lacking the modality - [x] Verify `upsert_many_*` returns correct created/updated counts - [x] Verify field-restricted role triggers a permission error naming the field and entity - [ ] Test with [hotel_business.xlsx](https://github.com/user-attachments/files/29376307/hotel_business.xlsx) and simple "import record" prompt <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22209?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. --> |
||
|
github-actions[bot]
|
f041f6dfb6 |
chore: sync AI model catalog from models.dev (#22242)
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. <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22242?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. --> Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com> |
||
|
Etienne
|
3525187321 |
fix(ai) - fixes (#22227)
- ai chat author fix (before : "workflow", after : "user") - https://discord.com/channels/1130383047699738754/1496872385687584768 <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22227?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. --> |
||
|
Amresh Chaurasiya
|
e1120d38b6 |
fix: stamp MCP and AI Agent writes with FieldActorSource.AGENT (#22215)
## Description Fixes #21437 MCP and AI Agent writes now correctly stamped with ### Problem Records created through MCP server were stamped as `WORKFLOW`, making them indistinguishable from workflow-created records. This breaks loop-protection filters that skip workflow-originated records. ### Solution - MCP writes now correctly stamped with `createdBy.source = AGENT` - AI Agent execution now uses `AGENT` instead of `MANUAL` - Added `WorkspaceCacheModule` to MCP module - Updated tests to verify AGENT source ### Files Changed - `mcp.module.ts`: Added WorkspaceCacheModule import - `mcp-protocol.service.ts`: Set AGENT source in buildMcpToolSet - `mcp-protocol.service.spec.ts`: Updated tests - `agent-actor-context.service.ts`: Changed MANUAL → AGENT - ## Type of Change - [x] Bug fix (non-breaking change) ## Checklist - [x] Code follows project style - [x] Tests added/updated - [x] Issue linked Fixes #21437 <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22215?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. --> |
||
|
Félix Malfait
|
da6a2ee300 |
fix(ai-chat): keep streams alive on silent SSE death + make the stream job idempotent (#22201)
## Problem In production, an AI-chat assistant response sometimes freezes mid-stream (partial text, looks hung), then "picks up again on its own" later without the user resending and without a known worker restart. Root cause: the **agent-chat SSE subscription has no keepalive and no silent-death detection**. - Delivery is fire-and-forget Redis pub/sub (`SubscriptionService.publishToAgentChat`) and the resolver returns the **raw** iterator — unlike `EventStreamResolver`, which heartbeats every 30s via `wrapAsyncIteratorWithLifecycle`. - During a quiet model/tool gap the connection sends no bytes, so a proxy/LB/NAT can silently drop it mid-stream. `graphql-sse` neither surfaces an error nor resumes with `Last-Event-ID`, and **nothing re-pulls the existing Redis chunk catch-up on reconnect** (it only runs on thread (re)mount / `message-persisted` refetch). - So the live view freezes; recovery only happens when the terminal `message-persisted` fires a full refetch from the DB — the observed "self-recovery". This is the **same silent-SSE-death class fixed for the DB event stream in #21061**, which was never applied to the agent-chat path. The symptom also matches #21096 (worker logs the job finishing, client never updates, reload shows the message). It is **not** queue prioritization, and it is **not** addressed by #22193 (which only stabilizes the assistant message id and removes end-of-stream flicker). A secondary, independent self-recovery path also existed: BullMQ stalled-job re-run (default 30s `lockDuration`, no idempotency guard) re-streaming the whole turn → duplicate assistant messages / double billing. ## Changes ### Commit 1 — keepalive + silent-death recovery (ports the #21061 pattern to agent chat) - **Shared:** new `keepalive` variant on `AgentChatSubscriptionEvent`. - **Server:** wrap the agent-chat subscription iterator with `wrapAsyncIteratorWithLifecycle` — emit a `keepalive` on connect and every `APPLICATION_KEEPALIVE_INTERVAL_MS` (30s) so the connection keeps flushing bytes and a dead connection becomes detectable. - **Client:** track the last received event timestamp (refreshed on every chunk/keepalive in the SSE `next` sink); new `AgentChatStreamKeepAliveEffect` forces a resubscribe + messages refetch after 90s of silence, so the durable Redis chunk list backfills the gap (`firstLiveSeq` is reset on resubscribe). ### Commit 2 — stream-job idempotency + lockDuration - Thread a `lockDuration` option through `MessageQueueWorkerOptions` + the BullMQ driver; set `aiStreamQueue` to 10 min so long streams aren't falsely stalled. - Guard `StreamAgentChatJob.handle` with a `streamId`-scoped Redis lock (`SET NX PX` + compare-and-delete release) so a stalled re-run is skipped instead of double-processing. ## Verification ⚠️ I could **not run typecheck/lint locally** — `yarn install` could not complete in this environment (transient registry network aborts before the link step, so `node_modules` never populated). **Please rely on CI for type/lint verification.** The changes are written to match existing conventions; the points most worth a reviewer's eye are the resolver's iterator typing and the ioredis `set(..., 'PX', ttl, 'NX')` overload. How to confirm the root cause in prod: a frozen client with the worker logging `StreamAgentChatJob processed in …ms` and no `[AI_CHAT_NO_TEXT]` is the silent-death signature (check reverse-proxy idle/buffering). For the secondary path, watch `aiStreamQueue` `stalled`/re-processed metrics and duplicate turns around worker restarts. ## Notes / trade-offs - The 10-min `lockDuration` means a genuinely crashed worker's job isn't reclaimed for up to 10 min; the client-side keepalive/catch-up recovers the view independently, and the idempotency lock prevents duplicates. Faster dead-worker recovery could be a follow-up. - Touches `useAgentChatSubscription.ts` / `AgentChatRuntimeEffects.tsx` / `stream-agent-chat.job.ts`, which #22193 also touches — trivial rebase expected. Opened as **draft** pending CI. https://claude.ai/code/session_018dF82A1VcsuWMxPLmdY3dm --- _Generated by [Claude Code](https://claude.ai/code/session_018dF82A1VcsuWMxPLmdY3dm)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22201?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. --> |
||
|
Etienne
|
b625bd1995 |
fix(ai-chat) - improvements (#22193)
- remove flickering at assistant message streamed end - add copy code - leave chat history when navigating to settings <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/22193?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. --> |
||
|
Etienne
|
5ca41d55fb |
feat(ai): humanize tool-call (#21976)
# Humanize tool-call labels cc: https://github.com/twentyhq/twenty/pull/21462 ## Preview <img width="459" height="156" alt="Screenshot 2026-06-22 at 19 13 11" src="https://github.com/user-attachments/assets/e7a2f5f5-cd09-4ec6-920b-5eb16b98285c" /> <img width="461" height="156" alt="Screenshot 2026-06-22 at 19 14 54" src="https://github.com/user-attachments/assets/c2114d2e-2aa8-499a-9801-68e3bb7c45f8" /> <img width="461" height="505" alt="Screenshot 2026-06-22 at 19 15 01" src="https://github.com/user-attachments/assets/ee9ca5d0-8e79-4c63-a2ff-ed5e359a9a9c" /> ## Why In the AI chat, tool steps were displayed using raw tool identifiers (`find_many_companies`, `create_one_task`, `send_email`...) and labels were partially reconstructed/humanized on the frontend. This was hard to localize and inconsistent across tool categories. This PR makes the **backend the single source of truth for human-readable, localized tool labels**, exposes them through `getToolIndex`, and reduces the frontend to a thin resolver that picks the right label for the current status (in-progress / completed). ## What changed ### Backend - `ToolIndexEntry` (and the `getToolIndex` GraphQL DTO) now carry `label`, `inProgressLabel?`, `completedLabel?`. - New `getCrudToolLabels(operation, objectLabel, i18nService, locale)` builds CRUD labels from a verb table (Search / Find / Group / Create / Update / Upsert / Delete × imperative / in-progress / completed) + the (translated, lowercased) object label. - New `translate-tool-label.util.ts` translates a source label via `I18nService` (`generateMessageId` → fallback to source when no translation exists). - Action tools: labels extracted to the `ACTION_TOOL_LABELS` constant (`msg` + `i18nLabel`) and translated in `ActionToolProvider.buildDescriptor`. - Logic-function tools use the function name as label; `toolSetToDescriptors` (workflow / view / metadata / dashboard) accepts an optional `labels` map and falls back to a humanized tool name. - Labels are localized server-side using the request locale (`@RequestLocale` → `buildToolIndex` → `context.locale`, threaded through `ToolContext` / `ToolProviderContext`). - `code_interpreter` schema now asks the model for `loadingMessage` (present tense) and `completedMessage` (past tense), so its status text is model-generated. - Removed the old generic `loadingMessage` injection mechanism (`wrap-tool-for-execution.util.ts` deleted; `wrapJsonSchemaForExecution` / `stripLoadingMessage` no longer wrap every tool). ### Frontend - New `useToolLabelMap()` hook builds a `Map<name, { label, inProgressLabel, completedLabel }>` from `getToolIndex`. - `getToolDisplayMessage` → `resolveToolDisplayMessage({ input, toolName, isFinished, labelMap, output })`: a small resolver registry keyed by tool name (`execute_tool`, `web_search`, `learn_tools`, `load_skills`, `code_interpreter`, default). - Default resolver prefers backend `completedLabel` / `inProgressLabel`, falling back to `Ran X` / `Running X`. - `learn_tools` / `load_skills` resolve their inner tool/skill names to labels (label map → tool output labels via `getToolOutputLabelEntries` → raw name). - `code_interpreter` step is now expandable to show the code even while running. ## How tool labelling flows (BE → FE) ```text BACKEND ┌───────────────────────────────────────────────────────────────────────────┐ │ Tool providers (per category) → ToolIndexEntry │ │ │ │ DatabaseToolProvider │ │ getCrudToolLabels(operation, object.labelPlural/Singular, i18n, locale) │ │ verb table (Search/Create/Update/Delete…) + translateToolLabel(object) │ │ → { label, inProgressLabel, completedLabel } │ │ │ │ ActionToolProvider │ │ ACTION_TOOL_LABELS[toolId] (msg) → translateToolLabel(…, locale) │ │ → { label, inProgressLabel?, completedLabel? } │ │ │ │ LogicFunctionToolProvider → label = logicFunction.name │ │ toolSetToDescriptors → label = labels[name] ?? humanize(name) │ │ (workflow / view / metadata / dashboard) │ └───────────────────────────────────────────────────────────────────────────┘ │ ▼ ┌───────────────────────────────────────────────────────────────────────────┐ │ GraphQL Query getToolIndex : [ToolIndexEntry] │ │ { name, label, inProgressLabel, completedLabel, description, │ │ category, objectName, icon } │ └───────────────────────────────────────────────────────────────────────────┘ │ ▼ FRONTEND ─ resolve the right label for the current status ┌───────────────────────────────────────────────────────────────────────────┐ │ useGetToolIndex() → useToolLabelMap() │ │ Map<name, { label, inProgressLabel?, completedLabel? }> │ └───────────────────────────────────────────────────────────────────────────┘ │ ▼ ┌───────────────────────────────────────────────────────────────────────────┐ │ resolveToolDisplayMessage({ input, toolName, isFinished, labelMap, output })│ │ │ │ TOOL_LABEL_RESOLVERS[toolName] ?? defaultResolver │ │ ├─ execute_tool → unwrap { toolName, arguments } then re-resolve │ │ ├─ web_search → "Searching/Searched the web for <query>" │ │ ├─ learn_tools → "Learning/Learned <labels>" │ │ ├─ load_skills → "Loading/Loaded <labels>" │ │ │ inner names resolved via: labelMap → output labels → raw name │ │ ├─ code_interpreter → model's loadingMessage / completedMessage │ │ └─ default → isFinished │ │ ? completedLabel ?? "Ran <label>" │ │ : inProgressLabel ?? "Running <label>" │ └───────────────────────────────────────────────────────────────────────────┘ │ ▼ Rendered by ThinkingStepsDisplay / ToolStepRenderer ``` ## Localization notes - Standard object labels and action/CRUD verbs are translated server-side via `I18nService` using the requester's locale. - Custom object labels are not translated unless a workspace custom translation exists (matched by `generateMessageId`); otherwise the source label is used as-is. ## Tests - **FE:** `resolveToolDisplayMessage` / `getToolOutputLabelEntries` (status selection, inner-name resolution, `code_interpreter` model labels, fallbacks). - **BE:** `toolSetToDescriptors` (label map + humanized fallback) and `database-tool.provider` label generation. <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/21976?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. --> |
||
|
Etienne
|
7b45380777 |
feat(ai): large tool output handling + navigation tools (#21982)
## Summary
Large tool outputs (e.g. a workflow run that serializes to ~70k tokens)
blow the chat context budget and force per-tool "raw" variants. This PR
handles oversized outputs generically in one place:
1. **Producer:** when a tool result exceeds a byte budget, it is spilled
to a `FileFolder.AgentChat` file and replaced with a compact `{ spilled,
outputRef, shape, hint }` envelope.
2. **Consumer:** two bounded, in-server navigation tools —
`extract_json_path` and `search_output` — let the model dig into the
spilled file by `fileId` without spinning up `code_interpreter`.
Together they add a fast, auditable middle tier between "truncated
inline preview" and "full code_interpreter relay," and enable an
enterprise "restricted" mode (spill + navigation, no sandbox).
## Data flow
```mermaid
flowchart TD
exec["resolveAndExecute / hydrateToolSet closure"] --> compact[compactToolOutput]
compact --> enabled{"spillLargeOutput enabled? (chat only)"}
enabled -->|no| inlineRaw["inline raw (MCP, workflow, sandbox bridge)"]
enabled -->|yes| size{"bytes > MAX_INLINE_TOOL_OUTPUT_BYTES?"}
size -->|no| inline["inline result"]
size -->|yes| skeleton["jsonShapeSkeleton + largeOutputHint"]
skeleton --> write["writeFile(AgentChat)"]
write --> envelope["return { spilled, outputRef, shape, hint }"]
envelope --> model[Model]
model --> nav["extract_json_path / search_output / code_interpreter (by fileId)"]
```
## Part 1 — Navigation tools (consumer)
- `extract_json_path`: extracts a sub-tree from a spilled JSON file by a
JSONPath-lite expression (dot/bracket access, array slicing,
single-level wildcard), with `maxItems`/`maxDepth` bounding. No filters
or recursive descent — those belong to `code_interpreter`.
- `search_output`: grep-like line search with context lines and
stateless `offset` pagination (`{ matches, totalMatches, hasMore }`).
- Both read from `FileFolder.AgentChat` by `fileId`, enforce their own
output byte cap, and are registered in `ActionToolProvider` (always
available; read-only).
## Part 2 — Spill producer
- Spilling slots in right after the existing `compactToolOutput` step at
the two seams in `ToolRegistryService` (`resolveAndExecute` and the
`hydrateToolSet` execute closure).
- `ToolOutputSpillService.spillIfTooLarge()` measures
`Buffer.byteLength`; over `MAX_INLINE_TOOL_OUTPUT_BYTES` (16 KB ≈ 4k
tokens) it writes the full payload and returns the envelope. Spill
failures never block the call (inline + warning).
- `jsonShapeSkeleton` computes a bounded structural map (depth 4, arrays
as `"array[N] of <type>"`, id-keyed maps collapsed, long leaves as size
markers, hard-capped at 1024 bytes) so the model knows the key paths in
one pass.
- Optional per-tool `largeOutputHint` (on the `Tool` type, threaded via
the descriptor) is used as the hint when present, else a generic hint.
The `shape` is always computed generically.
## Surfaces
Spilling is an opt-in flag (`spillLargeOutput`) mirroring
`compactOutput`:
| Surface | `spillLargeOutput` | Behavior |
| --- | --- | --- |
| AI chat / agent | `true` (in `chat-execution.service.ts`) | Spill on;
nav tools + `code_interpreter` in catalog |
| External MCP clients | unset | Raw output |
| Workflow agents | unset | Raw output |
| `code_interpreter` sandbox bridge | unset (it's an MCP call) | Raw
output |
The sandbox bridge inherits "no spill" for free via the MCP path — no
header sniffing, no `ToolContext.source` field.
## Design constraints (anti-micro-OS)
Exactly two navigation tools, no composition/piping, read-only, bounded
output. The boundary is: expressible as a single path lookup or text
search → nav tool; aggregation/correlation/transform →
`code_interpreter`.
## Notes / deviations from the plan
- `jsonShapeSkeleton` and `ToolOutputSpillService` live under the `tool`
module (not `tool-provider/output-transforms`) to avoid a `tool →
tool-provider` import cycle.
- Spill files use `{ isTemporaryFile: false, toDelete: false }` (same as
`code_interpreter`); `isTemporaryFile` here means files-field promotion,
not a TTL.
## Test plan
- [x] `extract-json-path` + `search-output` util unit tests (23 cases)
- [x] `jsonShapeSkeleton` unit tests (6) and `ToolOutputSpillService`
unit tests (4)
- [x] oxlint + oxfmt clean on changed files; `twenty-server` typecheck
clean (pre-existing unrelated errors aside)
- [ ] Manual: trigger an oversized tool result in chat, confirm the
envelope is returned and `extract_json_path` / `search_output` read the
spilled file by `fileId`
## Why no automated e2e
Spilling is chat-only and the chat path runs a live model, so the
black-box MCP integration harness can't deterministically trigger a
spill (MCP intentionally doesn't spill). The seam is small, explicit
flag-threading mirrored on `compactOutput`, covered by the unit suites.
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/21982?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. -->
|
||
|
mfamularopsyc
|
d2083e7a1b |
Set OpenAI Responses store false for AI chat and agents (#20888)
## Summary This PR sets `openai.store = false` for Twenty's `@ai-sdk/openai` AI calls. This follows the approach discussed in #20877: instead of adding a new Twenty-specific Zero Data Retention config variable, OpenAI Responses calls no longer rely on OpenAI-stored response/item references. This should help Zero Data Retention organizations and may also avoid stale persisted-item replay errors for non-ZDR OpenAI users. Changes included: - Adds a shared OpenAI provider-options helper that merges `openai.store = false` for `@ai-sdk/openai` models. - Applies the helper to AI chat `streamText` calls. - Applies the helper to workflow/agent `generateText` calls. - Preserves OpenAI encrypted reasoning metadata through DB/UI message mappers so reasoning context can be replayed without stored OpenAI item references. - Does not add a new env/config variable. Related to issue #20877. ## Behavior / Tradeoffs This changes OpenAI Responses behavior for all Twenty OpenAI users, not only ZDR users. The intended benefit is that Twenty no longer depends on OpenAI-stored response/item references. The main tradeoff is reduced provider-side item-reference reuse for non-ZDR OpenAI users. To reduce the impact for reasoning models, this PR preserves `providerMetadata.openai.reasoningEncryptedContent` through message persistence/replay so reasoning context can still be provided without stored OpenAI item references. ## Tests - Focused server Jest tests for OpenAI provider-options merging and reasoning metadata mapping. - Focused frontend Jest test for reasoning metadata mapping. - `oxlint` and `oxfmt --check` on changed files. - `git diff --check`. --------- Co-authored-by: Charles Bochet <charles@twenty.com> Co-authored-by: Etienne <45695613+etiennejouan@users.noreply.github.com> |
||
|
Abdul Rahman
|
1b7dc0367e |
fix(ai): gemini not working in ask ai (#21898)
Upstream issue: https://github.com/vercel/ai/issues/14369 Gemini 400s whenever a tool result contains JSON Schema `$ref`/`$defs` (it reads `$ref` as a function declaration name and finds no match). We hit this because `learn_tools` returns tool input schemas, and our recursive filter schema emits `$ref`/`$defs`. Other providers accept it fine, so this only blocks Gemini. Adds a Google-only `wrapLanguageModel` middleware that serializes ref-bearing tool results to text before they reach Gemini, so the pointers travel as a string instead of structured keys. The model still reads the full schema (same as the MCP path). Guarded so normal tool results pass through untouched. <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/21898?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. --> --------- Co-authored-by: Charles Bochet <charles@twenty.com> |
||
|
Abdul Rahman
|
ccc77932a0 |
Tool execution metrics (#21587)
<!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/21587?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. --> |
||
|
Félix Malfait
|
f96e36d3e6 |
fix(ai): prevent chat thread bricking from tool parts with null input (#21752)
## Problem Fixes #21695. An AI chat thread became **permanently unusable** — every subsequent message failed with `AI_APICallError: Internal server error` from Anthropic — when the thread history contained a tool part in `output-error` state with a **null input** (e.g. a tool call that failed input validation before execution, so neither `toolInput` nor `toolOutput` was ever captured). ## Validation of the reported findings I reproduced and confirmed the root cause empirically against the pinned `ai@6.0.97` SDK before writing the fix. **Root cause (confirmed from SDK source).** `convertToModelMessages` serializes every non-`input-streaming` tool part into a provider `tool_use` block, and for errored parts it uses: ```ts input: part.state === 'output-error' ? (part.input ?? ('rawInput' in part ? part.rawInput : undefined)) : part.input, ``` When both `input` and `rawInput` are nullish, the block is built with `input: undefined`, which `JSON.stringify` drops — so the HTTP payload carries a `tool_use` with **no `input` field**. This matches the reporter's minimal repro exactly (no `input` → `400 Field required`; `input: {}` → `200`). Inside a large streamed conversation the same malformed block surfaces as the generic `500`, and because the bad part is replayed on every turn the thread stays bricked. **Why #21276 didn't catch it.** `finalizeDanglingToolParts` only rewrote `input-available` parts; a part that arrives already in `output-error` with a null input was passed through untouched. **Note on current `main`.** A read-path default added recently (`mapDBPartToUIMessagePart`: `input: part.toolInput ?? {}`) already masks the live 500 on the standard reload path. However the gap is real and worth closing: the persist path still writes `toolInput = NULL` (the exact malformed rows the reporter found in `core."agentMessagePart"`), `finalizeDanglingToolParts` still doesn't normalize this case, and the protection rested on a single implicit default with no regression coverage. A small repro harness confirmed all of this: persisted `toolInput` was `undefined`, and a raw (non-defaulted) `output-error` part produced a `tool-call` whose `input` value was `undefined`. ## Fix Defense-in-depth so the invariant *"a tool part always carries a defined input"* holds at both the finalize and storage boundaries: - **`finalizeDanglingToolParts`** now backfills `input: {}` for `output-error` parts whose input is null, while preserving the original error message. This is the natural chokepoint (it already runs immediately before every persist). - **`mapUIMessagePartsToDBParts`** defaults a nullish tool input to `{}` so malformed rows are never persisted, independent of the caller. The existing read-path `?? {}` default is kept as a third safety net. ## Tests - Unit tests for `finalizeDanglingToolParts`: backfills `{}` for an `output-error` part missing its input, and preserves the existing validation error message. - Persistence test: `mapUIMessagePartsToDBParts` stores `{}` (never `null`) for a missing input. - End-to-end round-trip test: after finalize → persist → reload, `convertToModelMessages` produces a `tool-call` with a defined input and the errored call stays resolved. All three new core assertions were verified to **fail without the fix** and pass with it. Full AI module suite (97 tests) passes; `oxlint --type-aware`, `oxfmt`, and `tsgo` typecheck are clean. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_01SpuX6Pp2yTevk1zKTRiB9G --- _Generated by [Claude Code](https://claude.ai/code/session_01SpuX6Pp2yTevk1zKTRiB9G)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/21752?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. --> --------- Co-authored-by: Claude <noreply@anthropic.com> |
||
|
Etienne
|
d99e479be8 |
feat(billing) - facilitate top up in ai chat (#21645)
Today, when a trialing user hits their AI usage cap inside the Ask AI chat, ending the trial bounces them to the Stripe billing portal (and, for card-less users, loses their place in the conversation). This PR makes activating a paid plan / topping up credits feel seamless from within the chat: Trial users with a card on file activate their subscription in place, without leaving the app. Trial users without a card are sent to the Stripe payment-method portal and, on return, the trial is ended automatically and they're dropped back into the exact Ask AI thread they came from. Credit-exhaustion and trial banners now reflect whether a payment method exists (Add Credit Card vs Subscribe Now / End Trial Period) and upgrade inline via a confirmation modal instead of redirecting to Settings. Uploading Screen Recording 2026-06-16 at 07.51.12.mov… https://github.com/user-attachments/assets/4ea77273-da63-4b32-b6f1-5ac9e9560651 <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/21645?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. --> |
||
|
Félix Malfait
|
02a3a3c47c |
fix(ai): handle dynamic-tool message parts in chat persistence (#21740)
## Summary Fixes #20558. AI chat streams crashed with `Unsupported part type: dynamic-tool` whenever the model emitted a *dynamic* tool call (a tool that isn't part of the bound schema). The assistant message never persisted, so the user saw a hard failure mid-stream. ## Root cause The AI SDK v6 emits two flavors of tool parts: - **Static** — `type: "tool-<toolName>"` (e.g. `tool-execute_tool`) - **Dynamic** — `type: "dynamic-tool"`, with the name on `part.toolName` `mapUIMessagePartsToDBParts` recognised tool parts with a homegrown check: ```ts part.type.includes('tool-') && 'toolCallId' in part ``` That returns `false` for `'dynamic-tool'` (it contains `-tool`, not `tool-`), so dynamic parts fell through to `throw new Error(\`Unsupported part type: ${part.type}\`)` during the `handleStreamFinish` persistence step. Stack trace from the issue matches exactly. The same broken heuristic was duplicated in: - `packages/twenty-server/.../mapDBPartToUIMessagePart.ts` (reverse mapper) - `packages/twenty-front/.../utils/mapDBPartToUIMessagePart.ts` (frontend mirror — would also throw on a `dynamic-tool` row reloaded from history) Meanwhile, two other call sites in the codebase (`finalize-dangling-tool-parts.util.ts`, `isThinkingStepPart.ts`) already correctly use the SDK's `isToolUIPart`, which natively recognises both flavors. ## What this PR does 1. **Switches all three mappers to the SDK's canonical check** (`isToolUIPart` on the forward path; explicit `dynamic-tool` + `tool-` startsWith on the reverse paths, where the input is an entity/DTO, not a UI part). 2. **Persists `toolName`** — the column already existed on the entity, DTO and GraphQL fragment but nothing wrote it. For static parts the name is recoverable from `type`; for dynamic parts it's the only place the name lives, so without it the round-trip is impossible. The shared denormalisation also helps existing per-tool analytics (`count-native-web-search-calls-from-steps.util.ts`). 3. **Reconstructs `dynamic-tool` parts on read** (with `toolName`) so they survive a DB round-trip both on the server and on the frontend history view. 4. **Adds a round-trip unit test** covering both `dynamic-tool` and a static tool part to lock the behavior in. ## Architecture notes (called out for review) - `mapDBPartToUIMessagePart` is duplicated frontend + backend because the input shape differs (TypeORM entity vs. GraphQL DTO). Out of scope to consolidate here, but they're drifting — this PR is what that drift looked like in production. Worth a follow-up to express the shared logic once over a unified row type. - I left the existing renderer guard `part.type !== 'dynamic-tool'` in `AiChatAssistantMessageRenderer.tsx` alone — it's a reasonable UI-side decision to not attempt to render an unknown dynamic tool generically. Persistence and history reload now work; rendering of dynamic tool calls is a separate UX decision. - No DB migration needed — the `toolName` column already exists. Old static rows have `toolName: null`; the reverse mapper recovers their name from the `type` column as before. Old dynamic-tool rows don't exist (they all threw on write). ## Test plan - [x] `yarn workspace twenty-server jest map-message-parts.dynamic-tool` — 5 passed - [x] `yarn workspace twenty-server jest finalize-dangling-tool-parts.roundtrip` — still 4 passed (no regression) - [x] `yarn nx typecheck twenty-server` — clean - [x] `yarn nx typecheck twenty-front` — clean - [x] `yarn nx lint:diff-with-main twenty-server` — clean - [x] `yarn nx lint:diff-with-main twenty-front` — clean - [ ] Manual: trigger an AI chat that exercises a dynamic tool (e.g. via an MCP server returning a tool not in the bound schema) and confirm the stream finishes and the message persists. 🤖 Generated with [Claude Code](https://claude.com/claude-code) https://claude.ai/code/session_013EE11eVWtyxmdcbEHVJKoc --- _Generated by [Claude Code](https://claude.ai/code/session_013EE11eVWtyxmdcbEHVJKoc)_ <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/21740?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. --> --------- Co-authored-by: Claude <noreply@anthropic.com> |
||
|
Félix Malfait
|
bb6da7b7d1 |
feat(code-interpreter): reuse a warm sandbox per conversation (E2B) (#21664)
## What
The E2B code-interpreter driver created a **fresh sandbox on every
execution** and killed it in `finally`, so every call in a conversation
paid full cold-start and started blank. This PR keeps **one warm sandbox
per conversation** and, on idle, **pauses** it rather than killing it.
## How
- **Discovery without a registry:** the sandbox is tagged with the chat
`threadId` (scoped `workspaceId:threadId`) via E2B **metadata**, found
with `Sandbox.list({ query: { state: ['running','paused'], metadata }
})` and resumed with `Sandbox.connect()` (which auto-resumes a paused
sandbox). E2B is the source of truth — no Redis/DB mapping.
- **Pause/resume (E2B 2.x):** session sandboxes are created with
`lifecycle: { onTimeout: 'pause', autoResume: true }`. When idle they
**pause** — compute billing stops, filesystem **and** kernel/memory
state are preserved — and resume in ~1s on the next call. This replaces
the earlier keepalive approach.
- **No premature pause mid-run:** the sandbox is kept alive for
`max(execution timeout, idle window)`, so a long execution is never
paused underneath itself.
- **Tenant isolation:** discovery filters by the `twentySessionId` tag
and **re-checks it client-side**, so a loose server-side match can never
hand one conversation's warm sandbox (with its files, kernel state,
token) to another.
- **Concurrency:** executions sharing a session are serialized
in-process (one active stream per thread, run as a single job — the chat
resolver queues concurrent messages), so parallel tool calls can't race
the shared kernel.
- **Output isolation:** `/home/user/output` is reset at the start of
each reused run, so a call only returns the artifacts it actually
produced; durable state lives elsewhere and persists.
## SDK upgrade
`@e2b/code-interpreter` **`^1.0.4` → `^2.6.0`** (pulls `e2b@2.x`). The
typed pause/resume API, `lifecycle`, and the `state`/`metadata` list
filter only exist in the 2.x line; 1.x exposed them only as untyped
OpenAPI internals. `Sandbox.list()` is now a paginator (handled).
## Config
| Var | Default | Purpose |
|---|---|---|
| `CODE_INTERPRETER_TIMEOUT_MS` | `300000` | Max single-execution
duration. |
| `CODE_INTERPRETER_IDLE_TIMEOUT_MS` | `300000` | Idle window before the
warm sandbox auto-pauses. |
Reuse is always-on when a session id is present (chat path). The
workflow-agent path and the dev-only `LocalDriver` are unaffected.
## ⚠️ Open item before merge: paused-sandbox GC
E2B retains paused sandboxes **indefinitely** (no TTL). Unlike the old
keepalive path (which auto-killed on idle), pause means a conversation's
sandbox persists after the chat ends — so without garbage collection,
paused sandboxes accumulate (≈ one per historical conversation) and
consume storage. A GC policy is required; the approach + retention
window are being decided (see PR discussion). Also: the E2B runtime path
can't run in CI, so this still needs a **live smoke test** (reuse hit,
idle→pause, resume) and confirmation of paused-storage pricing before
rollout.
## Tests / checks
- Resolver unit tests (`getOrCreateSessionSandbox`): reuse+extend,
create-when-absent, duplicate reaping, connect-failure fallback,
keep-first-connectable-when-earlier-dead, **ignore cross-tenant
metadata**, and **kill-on-timeout-refresh-failure**.
- `nx typecheck twenty-server` (against e2b 2.x), `oxlint --type-aware`,
`oxfmt --check` all clean.
---------
Co-authored-by: Claude <noreply@anthropic.com>
|
||
|
Charles Bochet
|
9bb98fa5b5 |
fix(billing): don't crash when workspace has no active subscription (#21510)
## Problem
Sentry (high severity, SLA-breaching): `Billing Subscription Not Found:
No active subscription found for workspace …`
The `billingSubscription` workspace-cache provider
(`WorkspaceBillingSubscriptionCacheService.computeForCache`) called
`getCurrentBillingSubscriptionOrThrow`. For a workspace whose
subscription is fully canceled, `getCurrentBillingSubscription` filters
out `Canceled` and returns `undefined`, so the provider **threw**
`BILLING_SUBSCRIPTION_NOT_FOUND`.
That cache key is read on every usage-recording path:
- workflow execution
(`WorkflowExecutorWorkspaceService.sendWorkflowNodeRunEvent`)
- AI usage (`AiBillingService`)
- logic-function execution (`LogicFunctionExecutorService`)
- app charges (`AppBillingService`)
- the gate `BillingUsageService.canFeatureBeUsed` /
`hasAvailableCredits` / `decrementAvailableCreditsInCache`
- the cancellation webhook
(`invalidateAndRecompute('billingSubscription')`)
So any of these throws an unhandled exception for a
no-active-subscription workspace. The intent was clearly to tolerate
this state — `canFeatureBeUsed` already guards with
`isDefined(billingSubscription)` and the workflow runner logs *"there is
no subscription for this workspace"* — but the throwing provider made
those guards unreachable.
## Fix
- `computeForCache` now returns `FlatBillingSubscription | null` via the
non-throwing `getCurrentBillingSubscription`, and the cache type allows
`null`.
- Every consumer guards the absent case (`isDefined` / optional
chaining) and no-ops: usage events still emit with an undefined
`periodStart`, credits aren't decremented, `hasAvailableCredits` returns
`false`.
- `getCurrentBillingSubscriptionOrThrow` is **left untouched** for the
many callers (resolver, subscription-update, etc.) that genuinely
require a subscription.
## Test
Adds `workspace-billing-subscription-cache.service.spec.ts`: the
provider returns `null` when there's no active subscription (regression)
and the flattened subscription when one exists.
All 142 tests across the billing / ai-billing / workflow-executor suites
pass; `oxlint --type-aware` and `oxfmt` are clean.
<!-- This is an auto-generated description by cubic. -->
<a
href="https://cubic.dev/pr/twentyhq/twenty/pull/21510?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. -->
|
||
|
github-actions[bot]
|
e6d730cd75 |
chore: sync AI model catalog from models.dev (#21476)
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. <!-- This is an auto-generated description by cubic. --> <a href="https://cubic.dev/pr/twentyhq/twenty/pull/21476?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. --> Co-authored-by: FelixMalfait <6399865+FelixMalfait@users.noreply.github.com> |
||
|
Etienne
|
303c415dd1 |
fix(ai) - add logs + remove dashboard building (#21440)
- add logs for thread finishing without agent message - add logs to monitor toolCall token usage - remove dashboard building via AI (before fixing it) - fix Anthropic compute |
||
|
github-actions[bot]
|
3c81566d65 |
chore: sync AI model catalog from models.dev (#21392)
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]
|
441fe73be5 |
chore: sync AI model catalog from models.dev (#21353)
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> |
||
|
Etienne
|
b56fea69aa |
fix(ai) - optimize metadata CRUD tools (#21235)
Reduces output tokens for all 13 metadata tools by (~49%) based on
production sampling data.
GET tools (field + object metadata)
System fields are now returned as compact {id, name, type} instead of
the full ~20-key payload (opt-in includeFullSystemFields to get full
payload). System objects are similarly compacted to {id, nameSingular,
namePlural}.
Internal fields the agent never uses (searchVector, deletedAt, position,
updatedBy) are excluded entirely.
workspaceId and applicationId are hoisted into a response envelope
instead of being repeated on every record.
Null/default-false properties are stripped from custom field and object
payloads (e.g. options: null, settings: null, isUIReadOnly: false).
CUD tools (create/update/delete)
Create and update field tools now return {id, name, type, label} instead
of the full DTO.
Create and update object tools now return {id, nameSingular,
labelSingular} instead of the full DTO.
Delete tools return {id, success: true} instead of the full DTO of the
deleted entity.
Validation errors are grouped by message — e.g. 10 fields failing the
same check produce one line with all names instead of 10 identical
lines.
Learn schemas (all tools)
UUID pattern regex stripped from JSON schemas (keeps format: "uuid").
$schema and additionalProperties: false stripped from all generated
schemas.
All Zod .describe() annotations and tool descriptions shortened.
Skill & tool description updates:
All references to the removed list_object_metadata_items tool replaced
with get_object_metadata / get_field_metadata across skill instructions,
dashboard tools, view filter/sort tools, and MCP server instructions.
|
||
|
Félix Malfait
|
898713bd49 |
fix(server): finalize dangling tool calls when persisting agent chat messages (#21276)
## Problem
Interrupting an AI chat turn mid tool-call batch permanently bricks the
thread. Every subsequent message fails with:
> Tool results are missing for tool calls toolu_…, toolu_…
## Root cause
When the model fires a parallel batch of tool calls, it streams all the
calls first, then results come back one by one. If the stream is aborted
(user hits stop, credit cutoff, etc.) after only some have resolved, the
AI SDK's `onFinish` still fires with the partial assistant message —
including tool parts left in `input-available` state (a tool call with
no result).
`addMessage` persists that message verbatim. On the next turn the
history is rebuilt and `streamText` validates it: every `tool-call` must
be cleared by a `tool-result` before the next user message, or it throws
`MissingToolResultsError` (`ai/dist`, the `MissingToolResultsError`
check). The orphaned calls are now in the DB, so the thread fails on
every turn from then on.
## Fix
Enforce the invariant at the single write chokepoint. Every chat message
is persisted through `AgentChatService.addMessage`, so
`finalizeDanglingToolParts` runs there once: any tool part still in
`input-available` is rewritten to `output-error` ("Tool execution was
interrupted.") before mapping to DB rows.
`output-error` converts to a real `tool-result`, so the persisted turn
is always self-consistent and the next request is valid. Interrupted
calls are kept (not dropped) and surfaced as errored rather than
perpetually "running" — honest, since a partially-executed call may have
committed side effects the model should be able to reconcile.
One guard at one point covers every abort source — no read-side
patching, no migration, no schema change.
## Caching impact
None on the happy path. A completed turn has no `input-available` parts,
so the helper is a no-op and the persisted bytes (and therefore the
cached prefix) are identical to before. For an interrupted turn, the
finalized content is deterministic and written once, so it caches
cleanly on the following request and stays stable across later turns —
there is no scenario where this invalidates an existing cache entry. Net
effect: turns a hard failure into a normally-cached continuation.
## Testing
- New unit test covering finalize / no-op cases (7 cases, passing)
- `oxlint --type-aware` + `oxfmt` clean on changed files
|
||
|
Félix Malfait
|
91f2f08995 |
feat(server): unify workspace-event ingestion behind one EventSink pipeline (#21197)
## Why The five event-log streams (`workspaceEvent`, `pageview`, `objectEvent`, `usageEvent`, `applicationLog`) each wrote to ClickHouse through their own fire-and-forget writer (`AuditService`, `UsageEventWriterService`, and the `application-logs` driver), with the per-type knowledge (table names, normalization, access rules) spread across several modules. Three of them reimplemented the same ClickHouse insert, and the read side, the live stream, and the producers lived in different modules under two different names. This consolidates them into one `core-modules/event-logs/` subsystem (emit, write, live, read), with the per-type config in a single registry so adding an event type is roughly one file. The base Logs settings tab and free application logs shipped separately in #21180 (merged). This PR adds the unified backend, the registry, and the viewer's live mode and entitlement gating. ## Pipeline ```mermaid flowchart TB subgraph PROD["Producers"] A["auth, billing, impersonation,<br/>webhook, custom-domain"] U["usage listener"] F["logic-function executor (app logs)"] R["record CRUD (entity events)"] end EM["EventLogEmitterService<br/>createContext().insert* / dispatch()"] EQ(["entityEventsToDbQueue<br/>(existing, shared with timeline)"]) CIE["CreateEventLogFromInternalEvent"] SINK["WorkspaceEventSinkService.ingest()"] C1["ClickHouseEventSink"] C2["ConsoleEventSink"] LIVE["EventLogLiveService.publishWatched()<br/>(presence-gated)"] CH[("ClickHouse, 5 tables, async_insert")] CHAN(["WORKSPACE_EVENTS_CHANNEL"]) RS["EventLogsService (registry-driven read)"] LR["EventLogsLiveResolver"] UI["Settings > Logs"] A --> EM U --> EM F --> EM EM -->|direct| SINK R --> EQ --> CIE -->|ingest| SINK SINK --> C1 --> CH SINK --> C2 SINK --> LIVE -.->|if a viewer is watching| CHAN --> LR --> UI CH --> RS --> UI ``` ## What it does - Producers call `EventLogEmitterService.createContext().insert*()`, which builds a typed `WorkspaceEventEnvelope` and writes it through `WorkspaceEventSinkService` to the configured sinks (ClickHouse, Console) plus a presence-gated live fan-out. Record/CRUD events reach the same sink through the existing `entityEventsToDbQueue`. There is no dedicated queue; ClickHouse `async_insert` batches server-side. Writes are best-effort, as on main today. - `EVENT_LOG_TYPES[table]` is the per-type source of truth: the ClickHouse table, the required entitlement, the free-text filter column, and the row-to-GraphQL mapping. Read row shapes derive from the write rows. - Four modules along their dependency boundaries: `EventLogEmitterModule` (producer API), `EventLogIngestionModule` (sink layer), `EventLogLiveModule` (fan-out), and `EventLogsViewerModule` (the entitlement-gated GraphQL read, which is where billing/enterprise/permissions stay so producers stay light). - Logs viewer: per-table columns, filters (text, date, record), live mode, and an upgrade card that points to Billing on Cloud or the Admin Panel on self-hosted. Application logs are free on every plan; the other four require the `AUDIT_LOGS` entitlement (with a `NO_ENTITLEMENT` fallback to the upgrade card). - Renames `AuditService` to `EventLogEmitterService`, and the generic `Monitoring` event to a typed `Impersonation` event (`level` + `action`). - Removes `UsageEventWriterService`, the `application-logs` driver/module, and `AuditService`'s direct inserts. ## Durability Writes are best-effort, the same as main today (the old writers were fire-and-forget). A dedicated queue was tried mid-PR and removed: `async_insert` already batches server-side, so the queue only added durability, which isn't a requirement right now. The `EventSink` seam keeps a durable transport (e.g. a Redis-Streams buffer) easy to add later without touching producers. ## Out of scope S3 peer sink (seam only), Postgres or any second read path, `ReplicatedMergeTree`, ClickHouse table-schema changes, and the record-data `EVENT_STREAM_CHANNEL` (unchanged, separate concern). ## Testing Unit tests cover the registry definitions and row normalization, the entitlement gating, the envelope builders, and the producers. Integration tests cover the write paths (record create produces an `objectEvent`; the track mutation produces a `workspaceEvent`) and the read/query path across all five tables. Verified with typecheck, lint, a server boot, and GraphQL/SDK codegen. |
||
|
Félix Malfait
|
1b30983307 |
fix(settings): gate the AI settings page on AI_SETTINGS, not the chat flag (#21239)
## Summary Closes #21229. The two AI role permissions behaved **opposite to their labels**. The trap is that the flag's code name is the inverse of its UI label: | `PermissionFlagType` | UI label | Section | Means | |---|---|---|---| | `AI` | **"Ask AI"** | Actions | End-user: chat with AI | | `AI_SETTINGS` | **"AI"** | Member / settings | Admin: configure AI agents | Before this PR (on `main`): - `AI` ("Ask AI", chat) gated **both** the AI chat **and** the AI settings page. - `AI_SETTINGS` ("AI", configure agents) gated **nothing** the user could see. So a chat-only user could reach the whole AI **configuration** page, and toggling the "AI" settings permission did nothing — exactly the misalignment reported in #21229. ## Root cause `PermissionFlagType.AI` *reads* like "the AI permission", so it looks like the natural gate for the AI settings page — but it's actually the **chat** flag. The settings page (nav item + route) had been pointed at `AI` in #21072 to match the Overview stats query (`findWorkspaceAiStats`), which was itself mis-gated on `AI`. Both the stats query and the rest of the settings surface are admin/config features, so they belong on `AI_SETTINGS`. ## Changes All three move the **AI settings surface** from the chat flag (`AI`) to the settings flag (`AI_SETTINGS`); chat keeps following `AI`: - `useSettingsNavigationItems.tsx` — AI nav item → `AI_SETTINGS` - `SettingsRoutes.tsx` — AI settings route group → `AI_SETTINGS` - `ai-workspace-stats.resolver.ts` — `findWorkspaceAiStats` (settings-only, drives the Overview tab) → `AI_SETTINGS` After this: the "AI" permission controls the AI settings page + its Overview; the "Ask AI" permission controls the chat. Both toggles now match their labels. ## Test plan - [ ] Role with **only "Ask AI"** (`AI`): AI chat tabs/pane visible; **Settings → AI is hidden** and the route is not reachable. - [ ] Role with **only "AI"** (`AI_SETTINGS`): Settings → AI is visible, Overview stats load; chat nav is hidden. - [ ] Admin (both flags): everything works as before. ## Known follow-ups (out of scope — pre-existing, shared endpoints) These remain on `AI` because they're shared with non-settings surfaces and need either OR-gating or a resolver split, so a role with `AI_SETTINGS` but **not** `AI` still can't use them yet: - `getAiSystemPromptPreview` (Models/Prompts tabs) lives in the chat resolver, class-gated `AI`; NestJS guards are additive so it can't be cleanly method-overridden — it should be pulled into a settings resolver. - Agent reads `findManyAgents` / `findOneAgent` (agent create/edit forms) are class-gated `AI` and shared with the **Workflow** editor and **Roles** pages; these want a guard that accepts `AI ∨ AI_SETTINGS ∨ WORKFLOWS`. |
||
|
github-actions[bot]
|
ff9b5a5cad |
chore: sync AI model catalog from models.dev (#21242)
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> |
||
|
Etienne
|
8bd4cbc3fd |
fix(ai) - optim (#21233)
1. tool-registry.service.ts, Pass precomputed catalog to resolveSchemas() resolveSchemas() now accepts an optional precomputedCatalog parameter. Both getToolsByName() and getToolInfo() pass the catalog they already fetched, eliminating a redundant getCatalog() rebuild inside resolveSchemas(). 2. database-tool.provider.ts, Skip field lookup when schemas=false When building the catalog index (includeSchemas=false), getFlatFieldsFromFlatObjectMetadata() is no longer called for each of the 25 objects. The hasGroupByToolInputSchema() check is also skipped, group_by tools are always included in the index, with the real eligibility check deferred to learn_tools time. --> 100/150ms gain on learn/execute_tool execution |
||
|
martmull
|
c2ca90c255 |
feat(sdk): add runAgent() to run app agents from logic functions (#21157)
<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> |
||
|
Paul Rastoin
|
3d49642d12 |
[AUDIT] Run knip over twenty-server (#21159)
# Introduction Run [knip](https://knip.dev/) over twenty-server Used config: ```json { "$schema": "https://unpkg.com/knip@5/schema.json", "workspaces": { "packages/twenty-server": { "entry": [ "src/main.ts", "src/command/command.ts", "src/queue-worker/queue-worker.ts", "src/database/scripts/setup-db.ts", "src/database/scripts/truncate-db.ts", "src/database/clickHouse/migrations/run-migrations.ts", "src/database/clickHouse/seeds/run-seeds.ts", "src/instrument.ts", "lingui.config.ts", "test/integration/graphql/codegen/index.ts", "test/integration/utils/setup-test.ts", "test/integration/utils/teardown-test.ts", "scripts/**/*.ts", "**/*.spec.ts", "**/*.integration-spec.ts" ], "project": ["src/**/*.ts", "test/**/*.ts", "scripts/**/*.ts"], "ignore": [ "src/database/typeorm/**/migrations/**", "src/database/typeorm/**/*.entity.ts", "**/*.workspace-entity.ts", "**/logic-function-resource/constants/seed-project/**" ], "ignoreDependencies": ["@types/psl", "@types/aws-lambda"], "ignoreBinaries": ["nest", "lingui", "typeorm"] } } } ``` |
||
|
github-actions[bot]
|
e99116b0a5 |
chore: sync AI model catalog from models.dev (#21212)
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> |
||
|
Etienne
|
15eaabdbc1 |
fix(ai) - optimize crud tools (#21133)
- **Add delete many**, `delete_many_{object}` added alongside the
existing `delete_one_{object}`.
- **Uniformize naming**, crud module, type names, and MCP helper
constants renamed for consistency.
- **Optimize tool schema (learn phase)**
- `find_many(_companies)`: **7 158 → 2 700 tokens**
- `find_one(_company)`: **280 → 126 tokens**
- ....
- Main mechanism: `reused: 'ref'` (line 7 of
`to-tool-json-schema.util.ts`). Zod walks the schema tree, tracks which
Zod schema instances appear more than once, and emits each reused
instance exactly once in `$defs`, replacing all subsequent occurrences
with a `$ref`. Works because filter and value schemas are now extracted
as shared objects.
- **Optimize system prompt (tool catalog)**, DATABASE_CRUD section
restructured to list operation patterns (`find_many_{object}`, …) once +
objects once, instead of the full N×M cross-product of tool names.
- **Optimize execute_tool**, shared record-properties schema (same
`$defs` deduplication applies at call time); introduced `upsert_many`;
added `selectedFields` to `find_*` so the agent only fetches the fields
it needs.
|
||
|
Etienne
|
dd0039ca1c |
feat(mcp) - optimize instruction prompt and hide get_tool_catalog (#21183)
Workspace-aware initialize.instructions - Deleted the static mcp-server-instructions.const.ts - Created build-mcp-server-instructions.util.ts — a comprehensive system prompt with identity, object list, tool grammar, routing decision tree, intent mapping, skills vs tools, safety constraints, and data efficiency guidelines - Created McpInstructionBuilderService — fetches workspace-specific object names + skill names and injects them into the instructions Hide/deprecate get_tool_catalog Benefit : skip first MCP call (tools are included in instruction) |
||
|
Marie
|
4b15b949f3 |
Provide additional logsobservability to workflow runs (per node) (#21142)
Surfaces per-step "Logs" tabs in the workflow run side panel so users can see what each step actually did (model + tokens + tool calls for AI, console output for serverless functions, request/response for HTTP, recipients/body for Email). <img width="546" height="501" alt="ai_agent_without_websearch" src="https://github.com/user-attachments/assets/c6ca3518-9489-4484-a570-3d0569ff3b03" /> ## Storage - New `stepLogs` JSONB column on the `workflowRun` workspace entity, typed as `Record<string, WorkflowRunStepLog>` (keyed by step id). - Schema lives in `twenty-shared`: `workflowRunStepLogSchema` with a discriminated `details.type` union for `AI_AGENT | CODE | HTTP_REQUEST | EMAIL` — frontends and backends consume the same Zod-inferred type. - Field is added to existing workspaces via a workspace upgrade command (`2-9 add-workflow-run-step-logs-field`); the standard-object metadata declares it for new workspaces. - Writes happen atomically per step in `WorkflowRunStepLogWorkspaceService.setStepLog` using `jsonb_set`. That lets concurrent steps in the same run write their own keys without contending with the existing lock around `workflowRun.state`. - Per-step payload is hard-capped at 256 KB; anything larger is dropped with a `logger.warn`, so a pathological tool call can never bloat a row. See below for more information. ## How logs are produced **Aalmost everything was already being collected; this PR mostly persists and renders it.** - **AI agent** — `AgentAsyncExecutorService` already tracked token usage, model id, native web-search count, and the AI SDK's `steps[]`. We map those into the log via `mapAiStepsToToolCallLogs` (`searchVector` stripped from record outputs, per-call input/output capped at 32/64 KB, max 200 tool calls per step). The only new measurement is a wall-clock `durationMs` taken around `executeAgent`, and we now fold native web-search cost into the displayed `totalCostInDollars` (it was already billed, just not shown). - **Code / serverless function** — reuses the `console.log` output the function runner already returns (`logsByLevel`); `build-code-step-log.util` only repackages it. - **HTTP request** — built from the action's existing input/output via `build-http-request-step-log.util`. No new signals collected. - **Email (send / draft)** — added `sanitizedHtmlBody` + `plainTextBody` to the existing tool outputs (a small additive change), then `build-email-step-log.util` consumes them. No additional AI inference or external calls are made for logging — the cost is a small CPU overhead per step plus the JSONB write. ## Security The log surface intentionally shows whatever the workflow touched, which made redaction and sanitization the main design concern. - **HTTP — secrets in headers**: existing `SENSITIVE_HEADER_NAMES` set (Authorization, Cookie, …) replaced with `[redacted]` in both request and response. - **HTTP — secrets in URLs**: `SENSITIVE_URL_PARAM_NAMES` (e.g. `api_key`, `token`, `access_token`) replaced in the query string via `URL`-based parsing. - **HTTP — secrets in bodies**: `SENSITIVE_BODY_KEY_REGEX` deep-walks JSON request/response bodies (object input or stringified JSON) and redacts matching keys. Applied to the `error` field too, since transport-layer errors sometimes embed structured payloads. - **Email — XSS risk in body preview**: tool outputs now expose a server-side `sanitizedHtmlBody`; the log builder prefers it over the raw user-authored `input.body`, with `plainTextBody` as a second fallback. The original raw body is only used if sanitization didn't happen (e.g. tool failed before composing). - **AI — internal/noisy data**: `searchVector` (Postgres tsvector strings) is stripped from record outputs returned by Twenty tools to avoid leaking internal full-text-search payloads. - **DB bloat / runaway agents**: 256 KB per-step cap + 32 KB / 64 KB per-tool-call input/output cap + 200 tool calls per step. <img width="547" height="307" alt="logic_function" src="https://github.com/user-attachments/assets/dd4a3d16-67f2-434b-95b3-bdcaf9ed053d" /> ## More details on Log size & truncation Logs are stored in `workflowRun.stepLogs` (JSONB), keyed by `stepId`. ### Per-step cap Each step's log is hard-capped at **256 KB** (`MAX_STEP_LOG_BYTES` in `WorkflowRunStepLogWorkspaceService.setStepLog`). For ~99% of workflows this is roomy — typical real-world sizes: - Code / serverless function: 1–20 KB - HTTP request: 5–70 KB - Email: 5–30 KB - AI agent (a handful of tool calls): 5–50 KB ### Two layers of bounding 1. **Per-field truncation** in each builder (before writing): - **Code**: ≤ 500 entries, ≤ 4 KB per message, ≤ 8 KB stack trace - **HTTP**: ≤ 32 KB per body (request + response), UTF-8 byte-aware - **Email**: ≤ 8 KB body preview, UTF-8 byte-aware - **AI agent**: ≤ 32 KB tool input, ≤ 64 KB tool output, ≤ 200 tool calls/step 2. **Global per-step safety net** at write time: if the assembled `stepLog` still exceeds 256 KB, the write is **dropped entirely** with a `logger.warn`. The workflow itself keeps running unaffected. ### What this means in practice - **Safe**: workflow execution, step results, downstream steps — never blocked by log size. - **Safe**: iterators (each iteration overwrites the previous log for that `stepId`, so they can't accumulate). - **Safe**: step retries (same `stepId` is overwritten, not appended). - **Possible**: an AI agent step with many large tool outputs (e.g., 50+ heavy `web_search` calls) can exceed 256 KB → the **entire** step's log is dropped, side panel shows "No logs were recorded for this step". The user has no explicit signal that the log was dropped due to size (only server-side warn). - **Possible** (theoretical): a workflow with hundreds of distinct steps could push the row toward Postgres's internal ~256 MB jsonb limit. Beyond that, individual `jsonb_set` writes would error and be swallowed by the action's try/catch — workflow still completes. ### Possible future hardening (not in this PR) - Replace "drop entire log" with a stub that preserves the summary card (cost, duration, status) and marks `truncated.reason = 'size_cap'`. - Surface size-drops in the UI (similar to the existing `<StyledTruncatedNotice>`). - Emit a metric so dropped logs are observable in dashboards. |
||
|
github-actions[bot]
|
0d9f7673c9 |
chore: sync AI model catalog from models.dev (#21170)
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> |
||
|
Félix Malfait
|
b338a7a1d2 |
feat(settings): discovery hero rollout + ephemeral playground token (#21072)
## Summary
Two intertwined streams of work:
### UI — discovery hero pattern, settings shell, AI/API redesign
- **Generalize `SettingsDiscoveryHeroCard`** and use it on Layout, Data
Model, Apps, AI, API/Webhooks, Members. Drops 4 per-page wrapper files
(`SettingsObjectCoverImage`, `SettingsLayoutCoverImage`,
`SettingsLayoutCustomizeVideoModal`,
`SettingsDataModelVisualizeVideoModal`). Each page now supplies cover
src, modal id, and tab list.
- **Modal**: swap `<video>` placeholder for the Vimeo iframe pattern
from `twenty-docs`, per-tab `vimeoId`. Drop the parallel border-bottom
on the header (TabList draws its own baseline) and the grey background
behind the video. Note: Vimeo's embed allowlist applies — the iframes
load with the correct URL on `localhost` but the player itself requires
the video owner to allow the dev/staging domains in Vimeo settings.
- **AI page** rebuilt into a Cockpit pattern (Overview / Models / Skills
/ Tools / Usage). New `SettingsAiOverviewTab` with default Smart/Fast
pickers, at-a-glance stats, and an MCP signpost that deep-links to
`/settings/api-webhooks#mcp`. System Prompt link moved under Models.
Advanced tab removed.
- **API & Webhooks** now has 4 tabs (Playground / MCP / API Keys /
Webhooks). Hero card above tabs. Playground tab inverted to "Core API" /
"Metadata API" sections, each containing REST + GraphQL cards — schema
is the meaningful axis, protocol is secondary. Hash deep-link sync
delegated to the shared `TabListFromUrlOptionalEffect`.
- **Settings shell**: unified drawer outer padding (kill `isSettings`
branch), extract `CollapsibleNavigationDrawerSection`, add `iconColor`
on settings nav items, fix Exit Settings button alignment, 880px content
cap.
### Backend — strategy C: ephemeral playground token
The legacy paste-your-API-key flow is replaced by an on-demand
short-lived token scoped to the calling user's permissions. No shared
"Playground" API key to manage or revoke.
- New `JwtTokenTypeEnum.PLAYGROUND`. `PlaygroundTokenJwtPayload =
Omit<AccessTokenJwtPayload, 'type' | impersonation fields>` so any
future ACCESS claim flows through automatically.
- `AccessTokenService.generatePlaygroundToken` signs an access-shaped
JWT with `type: PLAYGROUND` and a configurable short TTL. A shared
private `resolveTokenSubject` helper parallelizes the user / workspace /
userWorkspace lookups for both generators.
- `JwtAuthStrategy.validateAccessToken` widened to accept
`AccessTokenJwtPayload | PlaygroundTokenJwtPayload`; impersonation gated
on `payload.type === ACCESS` so the union narrows without `as unknown
as` casts. The two branches in `validate()` collapse into one.
- New `PLAYGROUND_TOKEN_EXPIRES_IN` config var (default `2h`).
- New `generatePlaygroundToken` mutation (`WorkspaceAuthGuard`, no args,
returns `AuthToken`).
- Frontend `useOpenPlayground` hook centralizes mint → atom write →
navigate, with Apollo `onError` snackbar and a "use cached PLAYGROUND
token if still fresh" short-circuit (decodes via `jwt-decode`, checks
both `type` AND `exp`). Old API_KEY tokens left in localStorage from the
prior paste-form flow are rejected on `type` alone and force a re-mint —
this is what was causing the "This API Key is revoked" symptom on stale
browsers.
### Drive-by cleanups
- `PlaygroundToken` DTO removed (identical shape to `AuthToken` already
in use).
- 5 `customize-sidebar.webm` imports and the dead placeholder pipeline
removed.
## Test plan
### Discovery hero
- [ ] `/settings/layout`, `/settings/data-model`,
`/settings/applications`, `/settings/ai`, `/settings/api-webhooks`,
`/settings/members` each render the discovery hero card with its
illustration + play button + tabbed modal
- [ ] Modal tabs show the correct Vimeo embed URL per tab; aspect ratio
stays at 1440/900; no parallel border-bottom jog at the tab baseline
- [ ] AI Overview tab shows Smart/Fast model pickers + stats grid + MCP
signpost card; the MCP card lands on `/settings/api-webhooks#mcp` with
the MCP tab active
### API playground (ephemeral token)
- [ ] With an empty `playgroundApiKeyState` in localStorage, clicking
REST or GraphQL playground card opens the playground and the cached
token has `type: "PLAYGROUND"` with ~2h exp
- [ ] Clicking the card again within the freshness window does **not**
re-mint (`iat` / fingerprint stable across visits)
- [ ] Planting a fake API_KEY-shaped JWT in localStorage and clicking
the card forces a fresh mint (old token rejected on `type`)
- [ ] `GET /rest/companies?limit=1` with the cached token returns 200 +
real data
- [ ] `POST /graphql { __typename }` returns 200
### Settings shell
- [ ] Settings nav matches main app drawer padding; sections collapse;
Exit Settings button aligns with the workspace links above
- [ ] Active nav items have a right-gap (cleaner active state)
- [ ] Content area capped at 880px
### Verify
- [ ] `npx nx typecheck twenty-front` passes
- [ ] `npx nx typecheck twenty-server` passes
- [ ] `npx nx lint:diff-with-main twenty-front` passes
- [ ] `npx nx lint:diff-with-main twenty-server` passes
|
||
|
nitin
|
e430e4ea0a |
fix(ai): route xAI search through Responses API as native tools (#21037)
xAI deprecated Live Search, so the `searchParameters` provider option now returns 410. This routes all xAI models through the Responses API and binds web/X search as native agent tools, matching how Anthropic/OpenAI expose search. - xAI provider now uses `provider.responses()` — its `webSearch()`/`xSearch()` tools only run against the Responses endpoint, not chat completions - web/X search migrated from the `provider-option` variant to `sdk-tool` (`web_search`/`x_search`); deleted the dead `searchParameters` path, the `provider-option` variant, and `providerOptions` on `NativeModelBinding` - dropped a dead `rolePermissionConfig` param on `getAgentRoleId`, left over from #20331 --------- Co-authored-by: claude[bot] <41898282+claude[bot]@users.noreply.github.com> Co-authored-by: Félix Malfait <FelixMalfait@users.noreply.github.com> |
||
|
github-actions[bot]
|
3041ed3b6e |
chore: sync AI model catalog from models.dev (#21041)
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> |
||
|
nitin
|
996cdaf3ff |
refactor(agents): split tool resolution into native and action rails (#20331)
## Summary
Splits AI agent tool resolution into two independent rails:
- **Native tools** — capabilities baked into the model SDK
(Anthropic/OpenAI `web_search`, xAI `web`/`x` provider options). Bound
by `NativeToolBinderService`, controlled by per-agent
`modelConfiguration` toggles. Opaque to Twenty — executed on the model
provider's servers.
- **Action tools** — registry-scoped tools from `ToolRegistryService`
(code interpreter, send email, record CRUD, etc.). Permission-gated via
the agent's role. Executed on Twenty's server.
Both rails merge into a single `ToolSet` at call time. When both
surfaces expose a search tool the model picks at runtime — coexistence
is intentional (relevant once Exa returns as an action, see below).
## Notable changes worth calling out
**Contract change: `AgentAsyncExecutorService.executeAgent` no longer
accepts `rolePermissionConfig`.** Workflow agents now scope exclusively
by the agent's own permission-tab role (`unionOf: [agentRoleId]`). The
previous role-merging path (caller role intersected with agent role) is
removed. No agent role → no registry tools (fail-closed by design).
**`NativeToolBinderService` relocated** from
`core-modules/tool-provider/native/` →
`metadata-modules/ai/ai-models/services/`. The binder needs SDK-package
knowledge, which lives in `ai-models`. Old location created a backwards
module dependency.
**`NATIVE_MODEL_TOOLS_BY_SDK_PACKAGE` is exhaustive over
`AiSdkPackage`** (`Record<>`, not `Partial<Record<>>`). Adding a new SDK
without thinking about native tools now fails the build. SDKs without
native tools (Bedrock, Google, Mistral, Azure, OpenAI-compatible) get
explicit `{}` entries.
**Discriminated union `kind: 'sdk-tool' | 'provider-option'`** lets one
registry describe both function tools (Anthropic/OpenAI) and runtime
sources (xAI). Follows the local `tool-provider` convention from #19321.
## Deferred to follow-ups
- **Exa web search is dropped from this PR** (along with its
`WEB_SEARCH_TOOL` permission flag and the Exa-specific gating). Exa
comes back as an **action/app tool** once apps can define permission
flags through the SDK — ongoing work in #20481.
- **xAI native search currently errors.** xAI deprecated its Live Search
API (the `web`/`x` provider-option sources this rail maps to), so xAI
returns `410` when native search is actually exercised. The code path
itself is clear — it's only hit if you test xAI native tools. Fixed
separately alongside the broader xAI model fixes.
## Conscious non-decisions
- **No "twenty-native" category.** `native` is reserved for
model/provider SDK features; everything Twenty-owned is just a
tool/action.
- **Coexistence over precedence.** No rule forcing an action search tool
to override native search (or vice-versa) — when both exist, it's the
user's choice in workflow agents and the model's choice in chat.
---------
Co-authored-by: Félix Malfait <felix.malfait@gmail.com>
Co-authored-by: Félix Malfait <felix@twenty.com>
|
||
|
Félix Malfait
|
f4ead89956 |
refactor(twenty-orm): migrate 23 grandfathered entities to WorkspaceScopedRepository (#20987)
## Summary Follow-up to #20953. Migrates 23 of the 30 entities that were left in `WORKSPACE_SCOPED_EXEMPTIONS` last time, so the lint rule's workspaceId-enforcement default now covers most of the core/metadata schema. ### Migrated (23 entities, 88 files, 22 commits) | Family | Entities | |---|---| | Trivial caches | `NavigationMenuItem`, `Skill`, `DataSource`, `Webhook`, `CommandMenuItem`, `IndexMetadata` | | Views | `View`, `ViewField`, `ViewFieldGroup`, `ViewFilter`, `ViewFilterGroup`, `ViewGroup`, `ViewSort` | | Layouts | `PageLayout`, `PageLayoutTab`, `PageLayoutWidget` | | Roles & permissions | `Role`, `RoleTarget`, `PermissionFlag`, `ObjectPermission`, `FieldPermission`, `RowLevelPermissionPredicate`, `RowLevelPermissionPredicateGroup` | For each entity: swap `@InjectRepository(X)` → `@InjectWorkspaceScopedRepository(X)` (and the field type → `WorkspaceScopedRepository<X>`); rewrite every call site to pass `workspaceId` as the first arg (stripped from `where`/criteria — the wrapper throws if you include it now); register `provideWorkspaceScopedRepository(X)` in every owning NestJS module; update affected spec providers to `getWorkspaceScopedRepositoryToken(X)`. ### Rule update - `ApplicationRegistrationVariableEntity` was misclassified — moved to `STRUCTURAL_EXEMPTIONS` (no `workspaceId` column; it's keyed on `applicationRegistrationId` at the instance level). - 22 of the 23 migrated entities removed from `WORKSPACE_SCOPED_EXEMPTIONS` entirely (zero remaining raw `@InjectRepository` sites). - `RoleTargetEntity` also removed; one call site in `user-workspace.service.ts` keeps a raw injection with an `eslint-disable` + reason because `softRemove(...)` is not on the wrapper API yet (the migration would require threading `workspaceId` through `deleteUserWorkspace`'s three callers). ### Still exempted (7 entities, follow-up PRs) | Entity | Why deferred | |---|---| | `ApplicationEntity` | ~50 sites with several cross-workspace lookups by id (auth, OAuth, file-storage, cleanup) | | `CalendarChannelEntity` / `MessageChannelEntity` | Use `.increment(...)` (not on wrapper) and `repository.manager.transaction(...)` — wrapper needs to grow `.increment` + the transaction sites need `withManager` or dual-inject | | `FieldMetadataEntity` / `ObjectMetadataEntity` | The metadata services `extends TypeOrmQueryService<X>` and `super(rawRepo)` — requires dual-inject or reworking the inheritance | | `KeyValuePairEntity` | Allows `workspaceId: IsNull()` for instance-level config; wrapper rejects null | | `UpgradeMigrationEntity` | Same — instance-level + cross-workspace ledger | ## Test plan - [x] `npx nx typecheck twenty-server` — clean - [x] `npx nx lint twenty-server` — clean (0/0) - [x] All 10 affected unit specs pass (115 tests) — api-key, agent-role, permissions, workspace-roles-permissions-cache, view-filter-group, workflow-version-step-operations, two-factor-authentication (service + resolver), user-workspace, file - [ ] Server integration tests in CI |
||
|
Félix Malfait
|
865ca697ca |
Fix AI permission gating: use Ask AI for chat UI, AI Settings for admin endpoints (#21030)
## Summary Closes #20662. Two AI permission flags exist: - **`AI`** (label "Ask AI") — user-facing: chat with AI agents, use AI features - **`AI_SETTINGS`** (label "AI") — admin: create and configure AI agents After auditing every use of these flags I found: ### Frontend — chat UI gated by the admin permission (user-facing bug from the issue) A user granted only `Ask AI` could not see chat tabs, the "new chat" button (desktop & mobile), or the chat content pane; thread initialization was also skipped, leaving the chat in a half-initialized state and producing intermittent `THREAD_NOT_FOUND` errors. Switched these to `AI`: - `MainNavigationDrawerTabsRow.tsx` - `MainNavigationDrawer.tsx` - `MobileNavigationBar.tsx` - `AgentChatThreadInitializationEffect.tsx` ### Backend — admin-only resolvers gated by the user permission (privilege escalation) Two resolvers had a class-level guard of `AI`, letting any user with the user-facing flag reach admin endpoints (skill CRUD, eval runs). Switched the class-level guards to `AI_SETTINGS`: - `SkillResolver` — create/update/delete/activate/deactivate skills - `AgentTurnResolver` — read turns, run/grade evaluations ### Left as-is (already correct) - `AgentResolver` — class-level `AI` for reads (workflow editors and admin pages both need them), mutation-level `AI_SETTINGS` overrides for writes - `AgentChatResolver` & `AgentChatSubscriptionResolver` — already `AI` - `AiGenerateTextController` — already `AI` - Workspace AI config fields in `workspace.service.ts` — already `AI_SETTINGS` ## Test plan - [ ] As a user with `Ask AI` only (no `AI_SETTINGS`): chat tabs, "new chat" button, and chat history pane are visible on desktop + mobile; sending a message works; no `THREAD_NOT_FOUND` errors - [ ] As a user with `AI_SETTINGS` but no `Ask AI`: chat UI is hidden - [ ] As a user with `Ask AI` only: calling `skills` / `createSkill` / `agentTurns` / `runEvaluationInput` via GraphQL returns permission denied - [ ] As an admin (`AI_SETTINGS`): skill settings and agent eval pages still work |
