feat(aiobservability): read the messages the Langfuse SDK writes (#13331)

#### Description

- Langfuse SDK spans reached the thread with empty formatted messages:
the SDK writes them to `langfuse.observation.input/output`, which the
mapper never moved onto `gen_ai.*.messages`.
- `gen_ai.llm` v5 moves `langfuse.observation.input/output` onto
`gen_ai.input/output.messages` and copies `langfuse.observation.type`
onto `gen_ai.operation.name`, above every existing source. New spans
only.
- Formatter handles the shapes these spans carry: a superjson `{json}`
envelope, a single message object (incl. role-less `{tool_calls}` from
the LangChain handler), a bare list of Responses items (LiteLLM
`langfuse_otel` tool turns), and falls back to the raw output when the
OpenInference bucket is empty. Mixed-format warning dropped when one
side is generic or OpenInference.

  #### Additional Information

- Verified with unit tests, the shipped definition through the
in-process collector pipeline.
- Why Langfuse keeps its own keys instead of emitting semconv, checked
in the Langfuse repos:
- Docs ([OpenTelemetry
integration](https://langfuse.com/integrations/native/opentelemetry)):
`langfuse.*` attributes "always take precedence over the generic
OpenTelemetry conventions" and are "recommended for all users that are
manually instrumenting"; semconv is "still evolving", so OTel traces are
mapped into the Langfuse model.
- [Discussion #2509](https://github.com/orgs/langfuse/discussions/2509),
Sep 2024, maintainer: Langfuse features "not yet covered by the
conventions" and the risk of "breaking backwards compatibility once
conventions are in place" are the reasons for an own namespace.
- Ingestion (`OtelIngestionProcessor.ts`, `extractInputAndOutput`): own
keys first, then vendor keys, then semconv `gen_ai.*.messages`, then
events. Semconv is one vendor format among many.
- SDK sources: input and output are any JSON on any observation type
(span, agent, tool, generation, …), looser than `gen_ai.input.messages`.
- Still unmapped: `usage_details`, `cost_details`, `model.parameters`
are JSON objects; the mapper cannot extract fields, so pure Langfuse
spans have no tokens or cost.


Closes https://github.com/SigNoz/nerve-pod/issues/579

.claude/rules/comments.md

@@ -4,9 +4,9 @@ Applies to everything in the repo — code, config, workflows.
 
 - **No unnecessary comments.** Do not comment where the code is self-explanatory; never restate what the code already says.
 - **Document only** non-obvious behavior, constraints, formats, and edge cases.
-- **Rationale goes in prose, not source.** Why a job exists, how a subsystem fits together: that belongs in the README or the PR. Exception: a config entry that would look like a mistake without its reason (a pinned version, a trust or lint exclusion) keeps a short reason next to it.
-- **Never remove pre-existing comments** when editing code. The bar above applies to comments you write, not comments already there. When the code a comment describes changes, update the comment; when that code is deleted, delete the comment with it.
-- **Never talk to the reviewer.** No comments about where a change came from, what was changed, or why the change is correct — that belongs in the PR description and is noise the moment it merges. A banner a skill requires (the AI-owned banner on story mocks and `tests/e2e/drivers/`) tells maintainers how the file is made; keep it.
+- **Rationale goes in prose, not source.** Why a version is pinned, why a job exists, how a subsystem fits together — that belongs in the README or the PR.
+- **Never remove pre-existing comments** when editing code. The bar above applies to comments you write, not comments already there.
+- **Never talk to the reviewer.** No comments about where a change came from, what was changed, or why the change is correct — that belongs in the PR description and is noise the moment it merges.
 - **Less is more.** When writing something intended for human consumption, (comment, commit message, reply to prompt) use as few words as possible. Pick every word meticulously to reduce the volume to a strict minimum. Be down to the point. Less is more.
 
 Language rules build on this one: [`go-comments`](go-comments.md), [`py-comments`](py-comments.md).

.claude/rules/frontend-state.md

@@ -1,40 +0,0 @@
----
-paths:
-  - "frontend/src/**/*.ts"
-  - "frontend/src/**/*.tsx"
----
-
-# Frontend state
-
-Pick the store by what the state is. Never add Redux or React Context for shared state; to move existing Redux or Context state, use the `migrate-state-management` skill ([`frontend/.claude/skills/migrate-state-management/SKILL.md`](../../frontend/.claude/skills/migrate-state-management/SKILL.md)).
-
-| State | Use | Never |
-|---|---|---|
-| Server / API data | React Query, generated hooks from `src/api/generated` | a copy in Redux, Context or Zustand |
-| Shareable: filters, time range, selection, page | nuqs (`useQueryState` with a parser) | Redux, Context, hand-parsed `useSearchParams` |
-| Global client state | a small Zustand store per domain | Redux, Context |
-| One component's UI state | `useState` / `useReducer` | any store |
-
-## Server state
-
-- Expose a query through a hook that returns its result plus memoized derived values. Never dispatch a response into another store.
-- **The query key holds every value the request reads.** A field change that alters the request (name, context, data type) must change the key; a lossy signature serves stale rows.
-
-## URL state
-
-- Keep the URL small (Chrome trims near 2k characters). No large datasets, no sensitive data.
-- **A rewrite touches only the params it owns.** Repairing or normalising one param keeps the rest (`activeLogId`, pagination) as they were.
-- **Links carry the current time range.** A row link, "open in new tab" and "copy link" land on the same period as an in-app click.
-- **A preset read from the URL applies once.** A later user change or reset must not re-apply it.
-
-## Zustand
-
-- One `create()` per module; split a large store into slice factories combined in that one `create()`.
-- Always read through a selector: `useDashboardStore((s) => s.isDashboardLocked)`, never `useDashboardStore()`.
-- Update only through `set`; never mutate state. State fields before actions in the store object.
-- Hooks never call `useStore.getState()` (`signoz/no-zustand-getstate-in-hooks`); export standalone actions instead.
-
-## Local and form state
-
-- **Editable state starts from saved data once.** Initialise the form from the saved value, then let the user own it; never re-derive it from saved data on each render.
-- **A mode switch owns its in-flight work.** While a request from one mode runs, disable the switch or drop the result when the mode changes, so it can't close or overwrite the new mode.

.claude/rules/frontend-tests.md

@@ -1,20 +0,0 @@
----
-paths:
-  - "frontend/src/**/*.test.ts"
-  - "frontend/src/**/*.test.tsx"
----
-
-# Frontend tests
-
-Examples and the harness API live in [`frontend/src/tests/README.md`](../../frontend/src/tests/README.md).
-
-- **Import from the harness.** `render`, `screen`, `userEvent`, `waitFor` and the rest come from `tests/test-utils`, never from `@testing-library/react`. The harness mounts the router, React Query, Redux, app context and preferences. When it can't do what a test needs, extend `tests/test-utils` instead of importing around it.
-- **Find elements by test ID.** Query with `getByTestId`/`findByTestId` on the `data-testid` or `testId` the component exposes, and add one to the component when it has none. Fall back to role or text only for markup the app doesn't own.
-- **Interact like a user.** `const user = userEvent.setup({ pointerEventsCheck: 0 })`, then an awaited `user.click`, `hover`, `keyboard` or `type`. `fireEvent` only for events `userEvent` has no equivalent for: scroll, resize, setting `scrollTop`.
-- **Fake timers per test only**, for debounce or throttle, with `userEvent.setup({ advanceTimers: (ms) => jest.advanceTimersByTime(ms) })`. Never global.
-- **Network through MSW.** The global server runs for every test; override per test with `server.use(...)` from `mocks-server/server`. Large fixtures go in `mocks-server/__mockdata__`.
-- **Router through the harness.** Pass `initialRoute`/`routePath` to `render` and assert with `getCurrentLocation()` from `lib/router/navigation`. Never mock `react-router`, `react-router-dom` or `lib/history`; when a test can't render at a route, use the `lib/router` spies in `tests/router-mocks`.
-- **Mock only business boundaries.** Local `jest.mock` for API modules, custom hooks and domain components. Never re-mock what the harness provides or what `jest.config.ts` maps (`uplot`, `@signozhq/icons`, `useSafeNavigate`, `logEvent`).
-- **Type mocks and fixtures.** `jest.mocked(fn)` for module mocks, `jest.fn() as jest.MockedFunction<T>` for a standalone mock you configure or assert on. Fixtures carry their real type, from `src/api/generated` where it exists. No `any`, no `as unknown as`. An inline `onX={jest.fn()}` on a typed prop that is never asserted needs no annotation.
-- **A test fails without its change.** Revert the change, or break the line under test, and run the test before you finish: it must fail. Assert the specific outcome (exact text, request payload, call arguments), never only that an element exists.
-- **No limit on tests per file.** Split a file when it mixes unrelated behaviour, not to hit a count.

.claude/rules/pull-requests.md

@@ -9,5 +9,3 @@
 - **Keep the commit body short and human readable** focused on decision made if any. Commit body must not re-iterate the changes done, skip if title is sufficient in conveying the change.
 - **Use convensional commit format** for commits and PR title.
 - **Do not amend the commits once pushed.** Always create a new commit once changes are pushed to remote.
-- **Review fixes are new commits too.** A force push hides what changed from human reviewers and makes Greptile post its findings again. Bring in `main` with a merge; only a stacked branch (`gh stack`) rebases, since the stack moves as a unit.
-- **Answer every review thread.** Reply with the commit that fixes it (`Fixed in 4df6f3198d: <what changed>`) or the reason it stays, then resolve. Never resolve a thread with neither: Greptile withdraws a finding, and remembers it, only when told why.

.claude/rules/storybook.md

@@ -10,9 +10,6 @@ Page stories are written with the `signoz-page-story` skill ([`.claude/skills/si
 - **A UI change lands with its story.** Any change that alters what a page renders — a new page, a new state, a new branch, a changed layout — updates the story of every page that renders it, or creates one where the page has none. Go through the `signoz-page-story` skill; never hand-write a page story outside it.
 - **New page, new story. New state, new control or variant.** The skill decides which: a control on a response, a variant worth linking to, or a `play` story for a state only a click reaches.
 - **A shared component is covered through its pages.** A change under `components/` or `container/` with no story of its own updates the stories of the pages that render it.
-- **Plays live with the mocks.** A `play` is a named function exported from `stories/__story_mockdata__/<page>.play.ts`, which is collapsed for review like the rest of the AI-owned mocks. The story file only names it: `play: openEditDrawer`. A story file that still holds play bodies moves them there the first time you edit it.
-- **A play passes for every control value.** Branch each check on the args it depends on (panel kind, data state, row count, hidden header). A value the page supports must never fail the play.
-- **Mocks answer every request a reachable state makes.** Selecting a row, opening a view or switching a tab that fetches more hits a handler, never the 501 catch-all.
 - **Already covered, say so.** When the existing story and its controls already reach the changed state, no story edit is needed; state that in the PR description.
 - **Opt out only when there is no visual diff.** Skip stories only when the change renders nothing differently: types, api clients, hooks with no UI branch, tests. Deleting a page deletes its stories.
 - **Prove it with a visual diff when asked.** Before/after screenshots and "did this touch other screens" go through `sbshot-visual-diff`, which can only show what a story renders.

.claude/skills/signoz-page-story/SKILL.md

@@ -115,8 +115,8 @@ await expect(toggle.getAttribute('aria-checked')).toBe(before);
 - **A variant earns a story only when it is worth linking to**: an empty workspace, a viewer, a page mid-load. Everything else stays a control.
 - **Endpoints the page owns go through `response.json`**, so the Data control covers loaded, loading and failed in one declaration. Endpoints the page cannot render without (ingestion detection, preferences, feature payloads) take a plain resolver so the shell survives the loading and error states.
 - **Query-param state starts from `route`** (`/logs?tab=explorer`). In-page param navigation works inside a story; a different pathname is blocked and reported by the overlay. A control for a param is worth it only when the param is a page mode someone would want to flip.
-- **File layout**: every story file for a page lives under `src/pages/<Page>/stories/`: `<Page>.stories.tsx`, `<Page>.stories.mocks.tsx`, payload builders in `stories/__story_mockdata__/<page>.ts`, play functions in `stories/__story_mockdata__/<page>.play.ts`. Nothing page-specific in `src/storybook/controls/`. A page that is a tab strip over several routes gets one story file per tab, in its own folder under the module page (`LogsModulePage/Pipelines/stories/Pipelines.stories.tsx`), each with its own mocks and `__story_mockdata__/`; the builders more than one tab needs stay in the module page's own `stories/__story_mockdata__/` (`AlertList/stories/__story_mockdata__/alerts.ts`), which a tab reaches as `../../stories/__story_mockdata__/alerts`. Every one of them renders the module page, so the tab strip is there, and the `route` its mocks return decides which tab is open. A page's permission stories go one level further down, in `stories/authz/<Page>.authz.stories.tsx`, on the page's own mocks: see **Permission stories**.
-- **A state only a click reaches is a story with a `play` function**, not a control: a drawer, a modal, an edit mode the page holds in component state. Drive it with `userEvent` and the queries from `storybook/test`, take the first of a repeated row action, and wait on the state's own text. The page fetches before it renders a row, so the finder needs a timeout past the 1s default. The play is a named function exported from `__story_mockdata__/<page>.play.ts` and the story file only names it (`play: openEditDrawer`), so the reviewer reads states, not clicks; a story file that still holds play bodies moves them there the first time you edit it. It branches every check on the args it depends on: a control value the page supports never fails it. A state the app drops again on its own, such as one keyed on an array identity that a refetch replaces, does not get a story: it would not survive being looked at. A _sequence_ of such states, a wizard's steps or a questionnaire's pages, is still a control: declare the steps in the mocks module and walk them from a `play` on the meta that destructures `mount`, which is what makes Storybook replay it on an arg change. See [references/controls.md](references/controls.md).
+- **File layout**: every story file for a page lives under `src/pages/<Page>/stories/`: `<Page>.stories.tsx`, `<Page>.stories.mocks.tsx`, payload builders in `stories/__story_mockdata__/<page>.ts`. Nothing page-specific in `src/storybook/controls/`. A page that is a tab strip over several routes gets one story file per tab, in its own folder under the module page (`LogsModulePage/Pipelines/stories/Pipelines.stories.tsx`), each with its own mocks and `__story_mockdata__/`; the builders more than one tab needs stay in the module page's own `stories/__story_mockdata__/` (`AlertList/stories/__story_mockdata__/alerts.ts`), which a tab reaches as `../../stories/__story_mockdata__/alerts`. Every one of them renders the module page, so the tab strip is there, and the `route` its mocks return decides which tab is open. A page's permission stories go one level further down, in `stories/authz/<Page>.authz.stories.tsx`, on the page's own mocks: see **Permission stories**.
+- **A state only a click reaches is a story with a `play` function**, not a control: a drawer, a modal, an edit mode the page holds in component state. Drive it with `userEvent` and the queries from `storybook/test`, take the first of a repeated row action, and wait on the state's own text. The page fetches before it renders a row, so the finder needs a timeout past the 1s default. A state the app drops again on its own, such as one keyed on an array identity that a refetch replaces, does not get a story: it would not survive being looked at. A _sequence_ of such states, a wizard's steps or a questionnaire's pages, is still a control: declare the steps in the mocks module and walk them from a `play` on the meta that destructures `mount`, which is what makes Storybook replay it on an arg change. See [references/controls.md](references/controls.md).
 - **The mocks are AI-owned and say so.** `<Page>.stories.mocks.tsx` and every file under a `__story_mockdata__/` open with this banner, above the imports:
 
   ```ts
@@ -143,8 +143,6 @@ await expect(toggle.getAttribute('aria-checked')).toBe(before);
 - [ ] a page tagged `authz` has its `Authz` folder: one story per permission it reads, each reached by `revoked`, none of them a role preset, and each one checked by the `disabled` state of what the permission gates
 - [ ] the mocks module and every `__story_mockdata__` file carry the AI-owned banner
 - [ ] every control flipped once, its effect seen on screen
-- [ ] every play passes with each control value its story exposes, and the story file holds no play body
-- [ ] every request a reachable state makes (a selected row, an opened view, a switched tab) has a handler
 - [ ] console clean: no `[storybook] no msw handler`, no 501, no msw unhandled request, no React warning
 - [ ] no navigation overlay on mount
 - [ ] `pnpm tsgo --noEmit`, `pnpm exec oxlint <files>`, `pnpm exec oxfmt --check <files>` all clean. The repo has no prettier: `pnpm exec prettier` prints a pass while exiting 254

docs/contributing/go/spanmapper.md

@@ -17,7 +17,7 @@ Condition: `model`, `llm.input_messages` (OpenInference sets `llm.model_name` fr
 | `gen_ai.response.model`, last source for the request model | vLLM, which writes only the response model |
 | `ai.model.id`, `ai.model.provider`, `ai.response.model`, `ai.response.finishReason`, `ai.prompt.messages`, `ai.response.text`, `ai.response.toolCalls`, `ai.usage.promptTokens`, `ai.usage.completionTokens` | Vercel AI SDK v4, Mastra |
 | `ai.usage.inputTokens`, `ai.usage.outputTokens`, `ai.usage.cachedInputTokens`, `ai.usage.reasoningTokens` | Vercel AI SDK v5 |
-| `langfuse.observation.model.name`, `langfuse.session.id` | Langfuse SDK v3 |
+| `langfuse.observation.model.name`, `langfuse.observation.input`, `langfuse.observation.output`, `langfuse.observation.type`, `langfuse.session.id` | Langfuse SDK v3 (Python) and v4 (JS), LiteLLM `langfuse_otel` callback |
 | `session.id` | OpenInference, Langfuse, Claude Code |
 | `gen_ai.client.session_id` | opentelemetry-hooks runner (Cursor, Copilot, Claude Code hooks) |
 | `model`, `input_tokens`, `output_tokens`, `cache_read_tokens`, `cache_creation_tokens` | Claude Code `claude_code.llm_request` spans ([docs](https://code.claude.com/docs/en/monitoring-usage#span-attributes)) |

frontend/.cursor/rules/state-management.mdc

@@ -0,0 +1,97 @@
+---
+globs: **/*.store.ts
+alwaysApply: false
+---
+# State Management: React Query, nuqs, Zustand
+
+Use the following stack. Do **not** introduce or recommend Redux or React Context for shared/global state.
+
+## Server state → React Query
+
+- **Use for:** API responses, time-series data, caching, background refetch, retries, stale/refresh.
+- **Do not use Redux/Context** to store or mirror data that comes from React Query (e.g. do not dispatch API results into Redux).
+- Prefer generated React Query hooks from `frontend/src/api/generated` when available.
+- Keep server state in React Query; expose it via hooks that return the query result (and optionally memoized derived values). Do not duplicate it in Redux or Context.
+
+```tsx
+// ✅ GOOD: single source of truth from React Query
+export function useAppStateHook() {
+  const { data, isError } = useQuery(...)
+  const memoizedConfigs = useMemo(() => ({ ... }), [data?.configs])
+  return { configs: memoizedConfigs, isError, ... }
+}
+
+// ❌ BAD: copying React Query result into Redux
+dispatch({ type: UPDATE_LATEST_VERSION, payload: queryResponse.data })
+```
+
+## URL state → nuqs
+
+- **Use for:** shareable state, filters, time range, selected values, pagination, view state that belongs in the URL.
+- **Do not use Redux/Context** for state that should be shareable or reflected in the URL.
+- Use [nuqs](https://nuqs.dev/docs/basic-usage) for typed, type-safe URL search params. Avoid ad-hoc `useSearchParams` encoding/decoding.
+- Keep URL payload small; respect browser URL length limits (e.g. Chrome ~2k chars). Do not put large datasets or sensitive data in query params.
+
+```tsx
+// ✅ GOOD: nuqs for filters / time range / selection
+const [timeRange, setTimeRange] = useQueryState('timeRange', parseAsString.withDefault('1h'))
+const [page, setPage] = useQueryState('page', parseAsInteger.withDefault(1))
+
+// ❌ BAD: Redux/Context for shareable or URL-synced state
+const { timeRange } = useContext(SomeContext)
+```
+
+## Client state → Zustand
+
+- **Use for:** global/client state, cross-component state, feature flags, complex or large client objects (e.g. dashboard state, query builder state).
+- **Do not use Redux or React Context** for global or feature-level client state.
+- Prefer small, domain-scoped stores (e.g. DashboardStore, QueryBuilderStore).
+
+### Zustand best practices (align with eslint-plugin-zustand-rules)
+
+- **One store per module.** Do not define multiple `create()` calls in the same file; use one store per module (or compose slices into one store).
+- **Always use selectors.** Call the store hook with a selector so only the used slice triggers re-renders. Never use `useStore()` with no selector.
+
+```tsx
+// ✅ GOOD: selector — re-renders only when isDashboardLocked changes
+const isLocked = useDashboardStore(state => state.isDashboardLocked)
+
+// ❌ BAD: no selector — re-renders on any store change
+const state = useDashboardStore()
+```
+
+- **Never mutate state directly.** Update only via `set` or `setState` (or `getState()` + `set` for reads). No `state.foo = x` or `state.bears += 1` inside actions.
+
+```tsx
+// ✅ GOOD: use set
+increment: () => set(state => ({ bears: state.bears + 1 }))
+
+// ❌ BAD: direct mutation
+increment: () => { state.bears += 1 }
+```
+
+- **State properties before actions.** In the store object, list all state fields first, then action functions.
+- **Split into slices when state is large.** If a store has many top-level properties (e.g. more than 5–10), split into slice factories and combine with one `create()`.
+
+```tsx
+// ✅ GOOD: slices for large state
+const createBearSlice = set => ({ bears: 0, addBear: () => set(s => ({ bears: s.bears + 1 })) })
+const createFishSlice = set => ({ fish: 0, addFish: () => set(s => ({ fish: s.fish + 1 })) })
+const useStore = create(set => ({ ...createBearSlice(set), ...createFishSlice(set) }))
+```
+
+- **In projects using Zustand:** add `eslint-plugin-zustand-rules` and extend `plugin:zustand-rules/recommended` to enforce these rules automatically.
+
+## Local state → React state only
+
+- **Use useState/useReducer** for: component-local UI state, form inputs, toggles, hover state, data that never leaves the component.
+- Do not use Zustand, Redux, or Context for state that is purely local to one component or a small subtree.
+
+## Summary
+
+| State type        | Use              | Avoid              |
+|-------------------|------------------|--------------------|
+| Server / API      | React Query      | Redux, Context     |
+| URL / shareable   | nuqs             | Redux, Context     |
+| Global client     | Zustand          | Redux, Context     |
+| Local UI          | useState/useReducer | Zustand, Redux, Context |

frontend/.cursor/rules/testing-conventions.mdc

@@ -0,0 +1,74 @@
+---
+description: Core testing conventions - imports, rendering, MSW, interactions, queries
+globs: **/*.test.{ts,tsx}
+alwaysApply: false
+---
+
+# Testing Conventions
+
+## Imports
+
+Always import from the test harness, never directly from `@testing-library/react`:
+
+```ts
+import { render, screen, userEvent, waitFor } from 'tests/test-utils';
+import { server, rest } from 'mocks-server/server';
+```
+
+## Router
+
+Use the built-in router in `render`:
+
+```ts
+render(<Page />, undefined, { initialRoute: '/traces-explorer' });
+```
+
+Only mock `useLocation` / `useParams` if the test depends on their values.
+
+## MSW
+
+Global MSW server runs automatically. Override per-test:
+
+```ts
+server.use(
+  rest.get('*/api/v1/foo', (_req, res, ctx) =>
+    res(ctx.status(200), ctx.json({ ok: true })))
+);
+```
+
+Keep large response fixtures in `mocks-server/__mockdata_`.
+
+## Interactions
+
+- Prefer `userEvent` for real user interactions (click, type, select, tab).
+- Use `fireEvent` only for low-level events not covered by `userEvent` (e.g., scroll, resize). Wrap in `act(...)` if needed.
+- Always `await` interactions:
+
+```ts
+const user = userEvent.setup({ pointerEventsCheck: 0 });
+await user.click(screen.getByRole('button', { name: /save/i }));
+```
+
+## Timers
+
+No global fake timers. Per-test only, for debounce/throttle:
+
+```ts
+jest.useFakeTimers();
+const user = userEvent.setup({ advanceTimers: (ms) => jest.advanceTimersByTime(ms) });
+await user.type(screen.getByRole('textbox'), 'query');
+jest.advanceTimersByTime(400);
+jest.useRealTimers();
+```
+
+## Queries
+
+Prefer accessible queries: `getByRole` > `findByRole` > `getByLabelText` > visible text > `data-testid` (last resort).
+
+## Anti-patterns
+
+- Never import from `@testing-library/react` directly
+- Never use global fake timers
+- Never wrap `render` in `act(...)`
+- Never mock infra dependencies locally (router, react-query)
+- Limit to 3-5 focused tests per file

frontend/.cursor/rules/testing-mock-strategy.mdc

@@ -0,0 +1,54 @@
+---
+description: When to use global vs local mocks in tests
+globs: **/*.test.{ts,tsx}
+alwaysApply: false
+---
+
+# Mock Strategy
+
+## Use Global Mocks For
+
+High-frequency dependencies (20+ test files):
+- Core infrastructure: react-router-dom, react-query, antd
+- Browser APIs: ResizeObserver, matchMedia, localStorage
+- Utility libraries: date-fns, lodash
+
+Available global mock files (from jest.config.ts):
+- `uplot` -> `__mocks__/uplotMock.ts`
+
+## Use Local Mocks For
+
+- Business logic dependencies (API endpoints, custom hooks, domain components)
+- Test-specific behavior (different data per test, error scenarios, loading states)
+
+## Decision Tree
+
+```
+Used in 20+ test files?
+  YES -> Global mock
+  NO  -> Business logic or test-specific?
+    YES -> Local mock
+    NO  -> Consider global if usage grows
+```
+
+## Correct Usage
+
+```ts
+// Global mocks are already available - just import
+import { useLocation } from 'react-router-dom';
+
+// Local mocks for business logic
+jest.mock('../api/tracesService', () => ({
+  getTraces: jest.fn(() => mockTracesData),
+}));
+```
+
+## Anti-patterns
+
+```ts
+// Never re-mock globally mocked dependencies locally
+jest.mock('react-router-dom', () => ({ ... }));
+
+// Never put test-specific data in global mocks
+jest.mock('../api/tracesService', () => ({ getTraces: jest.fn(() => specificTestData) }));
+```

frontend/.cursor/rules/testing-type-safety.mdc

@@ -0,0 +1,54 @@
+---
+description: TypeScript type safety requirements for Jest tests
+globs: **/*.test.{ts,tsx}
+alwaysApply: false
+---
+
+# TypeScript Type Safety in Tests
+
+All Jest tests must be fully type-safe. Never use `any`.
+
+## Mock Function Typing
+
+```ts
+// Use jest.mocked for module mocks
+import useFoo from 'hooks/useFoo';
+jest.mock('hooks/useFoo');
+const mockUseFoo = jest.mocked(useFoo);
+
+// Use jest.MockedFunction for standalone mocks
+const mockFetch = jest.fn() as jest.MockedFunction<(id: number) => Promise<User>>;
+```
+
+## Mock Data
+
+Define interfaces for all mock data:
+
+```ts
+const mockUser: User = { id: 1, name: 'John', email: 'john@example.com' };
+
+const mockProps: ComponentProps = {
+  title: 'Test',
+  data: [mockUser],
+  onSelect: jest.fn() as jest.MockedFunction<(user: User) => void>,
+};
+```
+
+## Hook Mocking Pattern
+
+```ts
+import useFoo from 'hooks/useFoo';
+jest.mock('hooks/useFoo');
+const mockUseFoo = jest.mocked(useFoo);
+mockUseFoo.mockReturnValue(/* minimal shape */);
+```
+
+Prefer helpers (`rqSuccess`, `rqLoading`, `rqError`) for React Query results.
+
+## Checklist
+
+- All mock functions use `jest.MockedFunction<T>` or `jest.mocked()`
+- All mock data has proper interfaces
+- No `any` types in test files
+- Component props are typed
+- API response types are defined

frontend/.cursor/rules/ui-components-and-icons.mdc

@@ -0,0 +1,19 @@
+---
+description: Prefer SigNoz UI and icons across frontend code
+globs: **/*.{ts,tsx,js,jsx}
+alwaysApply: true
+---
+
+# UI Components and Icons Source of Truth
+
+For all frontend implementation work in this repository:
+
+- Always use UI primitives/components from `@signozhq/ui`.
+- Always use icons from `@signozhq/icons`.
+- Do not introduce new usage of icon libraries directly (for example `lucide-react`) in app code.
+- Do not mix multiple component systems for the same UI surface when an equivalent exists in `@signozhq/ui`.
+
+## Migration guidance
+
+- If touching a file that already uses non-`@signozhq/icons` icons, prefer migrating that file to `@signozhq/icons` as part of the same change when practical.
+- If a required component or icon is missing from SigNoz packages, call this out explicitly in the PR/summary before introducing alternatives.

frontend/.cursor/skills/migrate-state-management/SKILL.md

@@ -146,5 +146,5 @@ Do not use Zustand, Redux, or Context for purely local UI state.
 
 ## Additional resources
 
-- Project rule: [.claude/rules/frontend-state.md](../../../../.claude/rules/frontend-state.md)
+- Project rule: [.cursor/rules/state-management.mdc](../../rules/state-management.mdc)
 - Detailed patterns and rationale: [reference.md](reference.md)

frontend/AGENTS.md

@@ -25,14 +25,8 @@ You are operating within a constrained context window and strict system prompts.
 - Never create barrel files.
 - When writing new css, prefer CSS Modules
   - Use ./docs/css-modules-guide.md as reference on how to write good CSS Modules.
-- UI primitives come from `@signozhq/ui` (subpath imports) and icons from `@signozhq/icons`. No new `antd` component or direct icon library (`lucide-react`) where `@signozhq/ui` has an equivalent; `signoz/no-antd-components` enforces part of this list.
-  - Exception: a control added beside `antd` siblings of the same kind (a tab in an `antd` tab strip) matches them until that surface migrates. Say so in the PR.
-  - When `@signozhq/ui` or `@signozhq/icons` lacks what you need, say so in the PR before reaching for another library.
 - Before styling a `@signozhq/ui` component, check its props in `node_modules/@signozhq/ui/dist/<component>/*.d.ts` for `className`/`style` support; some omit them. Avoid overriding the component's CSS vars unless explicitly needed.
 - When writing code that could need authorization checks, read ./src/lib/authz/README.md
-  - A flow that gains a request also gates on that request's permission: a rename that now reads the view first needs `read` as well as `update`.
-- Identity keys (legend groups, color groups, CSV headers, React keys) come from structured values, never from joined display text: `['us, east', 'prod']` and `['us', 'east, prod']` both join to `us, east, prod`.
-- A payload saved under another kind or schema copies only the fields that schema accepts, and drops bindings to columns the user removed. The server rejects unknown or stale fields, so the preview works and the save fails.
 
 3. FORCED VERIFICATION: Your internal tools mark file writes as successful even if the code does not compile. You are FORBIDDEN from reporting a task as complete until you have:
 - Run `pnpm tsgo --noEmit`
@@ -40,8 +34,6 @@ You are operating within a constrained context window and strict system prompts.
 - Run `pnpm oxlint <file1> <file2>` and fix all warnings
 - Run `pnpm build`
 - Find if the file has tests for it, or if there's `__test__` folder or the parent folder has tests, and run. 
-- For every test you added or changed, revert the change it covers and confirm the test fails
-- On a stacked branch (`gh stack`), run `pnpm tsgo --noEmit` on every branch above it before pushing: a renamed prop or moved export breaks them
 - Fixed ALL resulting errors
 
 4. BEHAVIOR CHANGE DETECTION: When modifying existing behavior:
@@ -49,14 +41,6 @@ You are operating within a constrained context window and strict system prompts.
 - Update test assertions to match new behavior
 - If no tests exist, add them BEFORE changing behavior
 
-## Repo Rules
-
-Claude Code loads `../.claude/rules/` on its own. Other agents read the matching file before editing:
-- Every file: [comments](../.claude/rules/comments.md), [pull requests](../.claude/rules/pull-requests.md)
-- `src/**/*.ts(x)` that holds state, a query, URL params or a form: [frontend state](../.claude/rules/frontend-state.md)
-- `*.test.ts(x)`: [frontend tests](../.claude/rules/frontend-tests.md)
-- `*.tsx` that changes what a page renders: [storybook](../.claude/rules/storybook.md)
-
 ## Context Management
 
 1. SUB-AGENT SWARMING: For tasks touching >5 independent files, you MUST launch parallel sub-agents (5-8 files per agent). Each agent gets its own context window. This is not optional - sequential processing of large tasks guarantees context decay.

frontend/package.json

@@ -70,7 +70,7 @@
     "antd": "5.11.0",
     "antd-table-saveas-excel": "2.2.1",
     "antlr4": "4.13.2",
-    "axios": "1.18.0",
+    "axios": "1.21.0",
     "babel-jest": "^29.6.4",
     "chart.js": "3.9.1",
     "chartjs-adapter-date-fns": "^2.0.0",

frontend/pnpm-lock.yaml

@@ -127,8 +127,8 @@ importers:
         specifier: 4.13.2
         version: 4.13.2
       axios:
-        specifier: 1.18.0
-        version: 1.18.0
+        specifier: 1.21.0
+        version: 1.21.0
       babel-jest:
         specifier: ^29.6.4
         version: 29.6.4(@babel/core@7.29.7)
@@ -4593,8 +4593,8 @@ packages:
     resolution: {integrity: sha512-UzGt8zg7Ny8djbYMhxl2zuEevVa7r2gJjYY5Lwr1xM7+XU2nd6CkIWFTVcCIbAP63vSz71NaVyyuSk9lHKcy0A==}
     engines: {node: '>=4'}
 
-  axios@1.18.0:
-    resolution: {integrity: sha512-E32NzpYKp++W7XRe52rHiXV2ehxmh3wbdgO7MHeFM+vqxLBYHzt0ElkiImtOBxtOmyp0yoC8C6uESVV84Y2/hw==}
+  axios@1.21.0:
+    resolution: {integrity: sha512-YCzXMB53Q9PZBsh7yN7wtapit4WepwcHS2fP8QhGsBxSz5YtgYobV7GGlWlAJi+RxfU77GQTQi2DSLUma42EYA==}
 
   babel-jest@29.6.4:
     resolution: {integrity: sha512-meLj23UlSLddj6PC+YTOFRgDAtjnZom8w/ACsrx0gtPtv5cJZk0A5Unk5bV4wixD7XaPCN1fQvpww8czkZURmw==}
@@ -14283,7 +14283,7 @@ snapshots:
 
   axe-core@4.13.0: {}
 
-  axios@1.18.0:
+  axios@1.21.0:
     dependencies:
       follow-redirects: 1.16.0
       form-data: 4.0.6
@@ -20310,7 +20310,7 @@ snapshots:
 
   wait-on@7.2.0:
     dependencies:
-      axios: 1.18.0
+      axios: 1.21.0
       joi: 17.13.7
       lodash: 4.18.1
       minimist: 1.2.8

frontend/src/lib/authz/components/withAuthZ/withAuthZ.authz.test.tsx

@@ -262,21 +262,17 @@ describe('withAuthZ router context stability', () => {
 		server.use(setupAuthzGrantByPrefix('read||__||role'));
 	});
 
-	it('keeps the router context when useParams returns new object with same values', async () => {
+	it('does not re-render when useParams returns new object with same values', async () => {
 		mockAppParams({ roleId: 'r-1' });
 		mockAppLocation({ pathname: '/', search: '' });
-		const seen: RouterContext[] = [];
 
 		const Guarded = withAuthZ(RenderCounter, {
-			checks: (_props, router: RouterContext) => {
-				seen.push(router);
-				return [
-					buildPermission(
-						'read',
-						buildObjectString<'read'>('role', router.params.roleId ?? '*'),
-					),
-				];
-			},
+			checks: (_props, router: RouterContext) => [
+				buildPermission(
+					'read',
+					buildObjectString<'read'>('role', router.params.roleId ?? '*'),
+				),
+			],
 		});
 
 		const { rerender } = render(<Guarded />);
@@ -285,24 +281,24 @@ describe('withAuthZ router context stability', () => {
 			expect(screen.getByTestId('render-count')).toBeInTheDocument();
 		});
 
-		const checksBefore = seen.length;
+		const initialCount = renderCount;
 
-		// Return NEW object with SAME values: the router context stays the same
+		// Return NEW object with SAME values — should not cause re-render
 		mockAppParams({ roleId: 'r-1' });
 		rerender(<Guarded />);
 
-		expect(seen.length).toBeGreaterThan(checksBefore);
-		expect(seen[seen.length - 1]).toBe(seen[0]);
+		// Allow any pending effects to flush
+		await waitFor(() => {
+			expect(renderCount).toBe(initialCount);
+		});
 	});
 
-	it('keeps the router context when useLocation returns new object with same pathname', async () => {
+	it('does not re-render when useLocation returns new object with same pathname', async () => {
 		mockAppParams({});
 		mockAppLocation({ pathname: '/roles', search: '' });
-		const seen: RouterContext[] = [];
 
 		const Guarded = withAuthZ(RenderCounter, {
 			checks: (_props, router: RouterContext) => {
-				seen.push(router);
 				const match = router.matchPath<{ id: string }>('/roles/:id');
 				return match
 					? [buildPermission('read', buildObjectString<'read'>('role', match.id))]
@@ -316,24 +312,23 @@ describe('withAuthZ router context stability', () => {
 			expect(screen.getByTestId('render-count')).toBeInTheDocument();
 		});
 
-		const checksBefore = seen.length;
+		const initialCount = renderCount;
 
-		// Return NEW object with SAME pathname: the router context stays the same
+		// Return NEW object with SAME pathname — should not cause re-render
 		mockAppLocation({ pathname: '/roles', search: '' });
 		rerender(<Guarded />);
 
-		expect(seen.length).toBeGreaterThan(checksBefore);
-		expect(seen[seen.length - 1]).toBe(seen[0]);
+		await waitFor(() => {
+			expect(renderCount).toBe(initialCount);
+		});
 	});
 
-	it('keeps the router context when useLocation returns new object with same search params', async () => {
+	it('does not re-render when useLocation returns new object with same search params', async () => {
 		mockAppParams({});
 		mockAppLocation({ pathname: '/', search: '?tab=keys' });
-		const seen: RouterContext[] = [];
 
 		const Guarded = withAuthZ(RenderCounter, {
 			checks: (_props, router: RouterContext) => {
-				seen.push(router);
 				// Access searchParams to ensure it's part of the dependency chain
 				void router.searchParams.get('tab');
 				return [readPerm];
@@ -346,14 +341,15 @@ describe('withAuthZ router context stability', () => {
 			expect(screen.getByTestId('render-count')).toBeInTheDocument();
 		});
 
-		const checksBefore = seen.length;
+		const initialCount = renderCount;
 
-		// Return NEW object with SAME search: the router context stays the same
+		// Return NEW object with SAME search — should not cause re-render
 		mockAppLocation({ pathname: '/', search: '?tab=keys' });
 		rerender(<Guarded />);
 
-		expect(seen.length).toBeGreaterThan(checksBefore);
-		expect(seen[seen.length - 1]).toBe(seen[0]);
+		await waitFor(() => {
+			expect(renderCount).toBe(initialCount);
+		});
 	});
 
 	it('re-renders when params values actually change', async () => {

frontend/src/storybook/README.md

@@ -509,34 +509,16 @@ component state, is a story with a `play` function rather than a control. Use
 row action, and wait on the state's own text. The page fetches before it renders
 a row, so the finder needs a timeout past the 1s default.
 
-The play is a named function in `stories/__story_mockdata__/<page>.play.ts`,
-collapsed for review with the other AI-owned files, and the story file only names
-it. Branch each check on the args it reads, so every control value the page
-supports passes. A story file that still holds play bodies moves them there the
-first time it is edited.
-
 A *sequence* of those states, such as the steps of a wizard or the pages of a
 questionnaire, is still a control. Declare the steps in the page's mocks and walk
 them from a `play` on the meta, so every story of the page inherits the walk and
 only sets `args`:
 
 ```tsx
-// stories/__story_mockdata__/<page>.play.ts
-export async function walkToStep({
-	mount,
-	args,
-	canvasElement,
-}: {
-	mount: () => Promise<unknown>;
-	args: { step: WizardStep };
-	canvasElement: HTMLElement;
-}): Promise<void> {
+play: async ({ mount, args, canvasElement }): Promise<void> => {
 	await mount();
 	await advanceToStep(canvasElement, args.step);
-}
-
-// <Page>.stories.tsx, on the meta
-play: walkToStep,
+},
 ```
 
 Destructuring `mount` is what makes the panel row work. Storybook replays a play

frontend/src/tests/README.md

@@ -10,7 +10,7 @@
 #### Unit Testing: What, Why, How
 - What: Small, isolated tests for components, hooks, and utilities to verify behavior and edge cases.
 - Why: Confidence to refactor, faster feedback than E2E, catches regressions early, documents intended behavior.
-- How: Use our test harness with providers, mock external boundaries (APIs), find elements by test ID, and assert on visible behavior, not implementation details.
+- How: Use our test harness with providers, mock external boundaries (APIs, router), assert on visible behavior and accessible roles, not implementation details.
 
 #### Basic Template
 ```ts
@@ -29,20 +29,18 @@ describe('MyComponent', () => {
     render(<MyComponent />, undefined, { initialRoute: '/foo' });
 
     expect(await screen.findByText(/value: 42/i)).toBeInTheDocument();
-    await user.click(screen.getByTestId('refresh-button'));
+    await user.click(screen.getByRole('button', { name: /refresh/i }));
     await waitFor(() => expect(screen.getByText(/loading/i)).toBeInTheDocument());
   });
 });
 ```
 
-#### Rules (highlights)
-The full list is [`.claude/rules/frontend-tests.md`](../../../.claude/rules/frontend-tests.md).
-
+#### .cursorrules (Highlights)
 - Import from `tests/test-utils` only.
 - Prefer `userEvent` for real interactions; use `fireEvent` only for low-level events (scroll/resize/setting `scrollTop`).
 - Use MSW to mock network calls; large JSON goes in `mocks-server/__mockdata__`.
 - Keep tests type-safe (`jest.MockedFunction<T>`, avoid `any`).
-- Find elements by test ID (`getByTestId`, `findByTestId`); role or text only for markup the app doesn't own.
+- Prefer accessible queries (`getByRole`, `findByRole`) before text and `data-testid`.
 - Pin time only when asserting relative dates; avoid global fake timers otherwise.
 
 Repo-specific reasons:
@@ -71,10 +69,10 @@ jest.mock('hooks/queryBuilder/useQueryBuilder', () => ({
 }));
 ```
 
-Interact via test IDs:
+Interact via accessible roles:
 ```ts
 const user = userEvent.setup({ pointerEventsCheck: 0 });
-await user.click(screen.getByTestId('save-changes'));
+await user.click(screen.getByRole('button', { name: /save changes/i }));
 expect(screen.getByText(/ADDED FILTERS/i)).toBeInTheDocument();
 ```
 

frontend/src/tests/test-utils.tsx

@@ -190,7 +190,7 @@ const customRender = (
 	} = providerProps;
 
 	return render(ui, {
-		wrapper: ({ children }) => (
+		wrapper: () => (
 			<AllTheProviders
 				role={role}
 				appContextOverrides={appContextOverrides}
@@ -198,7 +198,7 @@ const customRender = (
 				initialRoute={initialRoute}
 				routePath={routePath}
 			>
-				{children}
+				{ui}
 			</AllTheProviders>
 		),
 		...options,

pkg/modules/spanmapper/implspanmapper/fs/definitions/gen_ai.llm.json

@@ -1,5 +1,5 @@
 {
-  "version": 4,
+  "version": 5,
   "definition": {
     "name": "gen_ai.llm",
     "condition": {
@@ -134,6 +134,12 @@
         "fieldContext": "attribute",
         "config": {
           "sources": [
+            {
+              "key": "langfuse.observation.type",
+              "context": "attribute",
+              "operation": "copy",
+              "priority": 40
+            },
             {
               "key": "openinference.span.kind",
               "context": "attribute",
@@ -356,6 +362,12 @@
         "fieldContext": "attribute",
         "config": {
           "sources": [
+            {
+              "key": "langfuse.observation.input",
+              "context": "attribute",
+              "operation": "move",
+              "priority": 40
+            },
             {
               "key": "gen_ai.prompt",
               "context": "attribute",
@@ -382,6 +394,12 @@
         "fieldContext": "attribute",
         "config": {
           "sources": [
+            {
+              "key": "langfuse.observation.output",
+              "context": "attribute",
+              "operation": "move",
+              "priority": 40
+            },
             {
               "key": "gen_ai.completion",
               "context": "attribute",

pkg/types/aiobservabilitytypes/message_format.go

@@ -121,10 +121,13 @@ func (f *formatting) messages() {
 
 // OpenRouter's answer object or an Anthropic response beside the OpenAI chat request it answers is one
 // format, not a mix: an Anthropic request without system and without Anthropic-only blocks reads as an
-// OpenAI chat request.
+// OpenAI chat request. A side kept as generic already warned, and a side read from OpenInference's
+// buckets beside the raw payload of the other is the documented fallback.
 func (f *formatting) warnMixedFormats(inputLabel, outputLabel string) {
 	switch {
 	case inputLabel == "" || outputLabel == "" || outputLabel == inputLabel:
+	case inputLabel == FormatterGeneric || outputLabel == FormatterGeneric:
+	case inputLabel == FormatterOpenInference || outputLabel == FormatterOpenInference:
 	case inputLabel == FormatterOpenAIChat && (outputLabel == FormatterOpenRouter || outputLabel == FormatterAnthropic):
 	default:
 		f.warn("output formatted as %s, input as %s", outputLabel, inputLabel)
@@ -183,11 +186,16 @@ func (f *formatting) side(value any, role Role) (OutputMessages, string) {
 				return f.semconv(v), FormatterSemconv
 			case isAnthropicMessages(v):
 				return f.anthropicMessages(v), FormatterAnthropic
+			case isResponsesItems(v):
+				return foldAssistant(f.responsesItems(v)), FormatterOpenAIResponses
 			case first["role"] != nil:
 				return f.openAIChatMessages(v), FormatterOpenAIChat
 			}
 		}
 	case map[string]any:
+		if inner, ok := superJSON(v); ok {
+			return f.side(decode(inner), role)
+		}
 		switch {
 		case isLangChainRun(v):
 			return f.langChainRun(v), FormatterLangChain
@@ -214,6 +222,9 @@ func (f *formatting) side(value any, role Role) (OutputMessages, string) {
 		if messages, ok := v["messages"].([]any); ok {
 			return f.openAIChatMessages(messages), FormatterOpenAIChat
 		}
+		if (v["role"] != nil && hasKeys(v, "content")) || v["tool_calls"] != nil {
+			return OutputMessages{f.openAIChatSingle(v)}, FormatterOpenAIChat
+		}
 	case string:
 		if v == "" {
 			return OutputMessages{}, FormatterText
@@ -309,6 +320,23 @@ func foldAssistant(msgs OutputMessages) OutputMessages {
 	return out
 }
 
+// superJSON unwraps {json, meta?}, the envelope superjson, the transformer tRPC apps serialise with,
+// puts around a value; a Langfuse generation records it around the request when the app passes the
+// serialised form:
+// https://github.com/flightcontrolhq/superjson/blob/aaa65e36e8e31da740265a56b27c965ca1c7b754/README.md
+func superJSON(m map[string]any) (any, bool) {
+	inner, ok := m["json"]
+	if !ok {
+		return nil, false
+	}
+	for key := range m {
+		if key != "json" && key != "meta" {
+			return nil, false
+		}
+	}
+	return inner, true
+}
+
 // decode parses a JSON-encoded attribute value, twice when an SDK encoded the string again.
 // Anything that is not JSON is returned as is.
 func decode(value any) any {

pkg/types/aiobservabilitytypes/message_format_openai_chat.go

@@ -4,8 +4,9 @@ import (
 	"strings"
 )
 
-// Format: OpenAI Chat Completions, a request {messages}, a response {choices}, or a bare message list.
-// Written by: bifrost, openrouter requests, the raw OpenAI calls OpenInference records.
+// Format: OpenAI Chat Completions, a request {messages}, a response {choices}, a bare message list, or a
+// single message.
+// Written by: bifrost, openrouter requests, the raw OpenAI calls OpenInference records, the Langfuse SDK.
 // Handled: string and block content, tool_calls nested or flattened, tool messages, refusal,
 // reasoning_content, finish reasons.
 // Not yet: streaming delta chunks, legacy completions choices[].text, audio output.
@@ -88,6 +89,19 @@ func (f *formatting) openAIChatMessage(m map[string]any) OutputMessage {
 	return msg
 }
 
+// openAIChatSingle converts a message recorded outside a list: the first choice's message the Langfuse
+// OpenAI wrapper writes as the output, or the additional_kwargs of a LangChain AIMessage, {tool_calls}
+// without a role, which the Langfuse LangChain handler writes for a tool-calling turn:
+// https://github.com/langfuse/langfuse-python/blob/18baf80450dd5f9c1f85916eb2d1b3c72d2052e1/langfuse/openai.py#L446
+// https://github.com/langfuse/langfuse-python/blob/18baf80450dd5f9c1f85916eb2d1b3c72d2052e1/langfuse/langchain/CallbackHandler.py#L1558
+func (f *formatting) openAIChatSingle(m map[string]any) OutputMessage {
+	msg := f.openAIChatMessage(m)
+	if msg.Role.IsZero() {
+		msg.Role = RoleAssistant
+	}
+	return msg
+}
+
 // toolCallPart reads {id, function: {name, arguments}}; a gateway may flatten it to {id, name, arguments|args}.
 func toolCallPart(c map[string]any) Part {
 	call := ToolCallRequestPart{Type: PartTypeToolCall, ID: optionalString(c["id"]), Name: stringOf(c["name"])}

pkg/types/aiobservabilitytypes/message_format_openai_chat_test.go

@@ -205,6 +205,30 @@ func TestFormat_OpenAIChat(t *testing.T) {
 			]`,
 			wantOutput: `[]`,
 		},
+		{
+			// Langfuse SDK generation: the history as a list, the first choice's message as the output
+			name:       "SingleMessageOutput_AssistantMessage",
+			input:      `[{"role": "system", "content": "You are a support agent on a phone call."}, {"role": "user", "content": "Hello?"}]`,
+			output:     `{"role": "assistant", "content": "Hi, this is a courtesy call about your order. Do you have a minute?"}`,
+			wantInput:  `[{"role": "system", "parts": [{"type": "text", "content": "You are a support agent on a phone call."}]}, {"role": "user", "parts": [{"type": "text", "content": "Hello?"}]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "Hi, this is a courtesy call about your order. Do you have a minute?"}]}]`,
+		},
+		{
+			// a structured output that happens to carry a role key is not a message
+			name:       "ObjectWithRoleKeyButNoContent_KeptAsGeneric",
+			output:     `{"role": "admin", "permissions": ["read", "write"]}`,
+			warnings:   []string{"message format not recognised, kept as generic"},
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "", "parts": [{"type": "generic", "content": "{\"permissions\":[\"read\",\"write\"],\"role\":\"admin\"}"}]}]`,
+		},
+		{
+			// Langfuse LangChain handler: the additional_kwargs of the AIMessage, without a role, as the output
+			// of a tool-calling turn
+			name:       "RolelessToolCallsOutput_AssistantToolCallPart",
+			output:     `{"tool_calls": [{"id": "call_1", "type": "function", "function": {"name": "get_weather", "arguments": "{\"city\": \"Oslo\"}"}}], "refusal": null}`,
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "tool_call", "name": "get_weather", "id": "call_1", "arguments": {"city": "Oslo"}}]}]`,
+		},
 		{
 			name:      "ResponseWithScalarChoice_KeptAsGeneric",
 			output:    `{"choices":["oops",{"finish_reason":"stop","index":1,"message":{"role":"assistant","content":"hi"}}]}`,

pkg/types/aiobservabilitytypes/message_format_openai_responses.go

@@ -2,13 +2,15 @@ package aiobservabilitytypes
 
 import "strings"
 
-// Format: OpenAI Responses, a request {input, instructions} or a response {output, status} of items:
+// Format: OpenAI Responses, a request {input, instructions}, a response {output, status} or a bare list
+// of items:
 // https://github.com/openai/openai-openapi/blob/aeb6a041dc56c011ca9f12127ec898f911100aa0/openapi.yaml#L47597
 // https://github.com/openai/openai-openapi/blob/aeb6a041dc56c011ca9f12127ec898f911100aa0/openapi.yaml#L69691
 // https://github.com/openai/openai-openapi/blob/aeb6a041dc56c011ca9f12127ec898f911100aa0/openapi.yaml#L53834
 // https://github.com/openai/openai-openapi/blob/aeb6a041dc56c011ca9f12127ec898f911100aa0/openapi.yaml#L57596
-// Written by: gateways recording raw Responses bodies, and the raw calls OpenInference records for the
-// Responses API and the OpenAI Agents SDK when its flattened messages are absent.
+// Written by: gateways recording raw Responses bodies, the raw calls OpenInference records for the
+// Responses API and the OpenAI Agents SDK when its flattened messages are absent, and LiteLLM's
+// langfuse_otel callback, the function_call items of a tool-calling turn as a bare list.
 // Handled: message items with input_text, output_text, refusal, input_image and input_file content;
 // function_call and custom_tool_call with their outputs; reasoning summaries; compaction; every other
 // *_call item as a server tool call and *_output item as its response; the status and incomplete_details
@@ -32,6 +34,19 @@ func isResponsesRequest(m map[string]any) bool {
 	return isItem
 }
 
+// isResponsesItems looks for an item only the Responses API has, a tool call or its output, a reasoning
+// or compaction item, all without a role; a list of messages has none.
+func isResponsesItems(list []any) bool {
+	for _, item := range list {
+		m, _ := item.(map[string]any)
+		typ := stringOf(m["type"])
+		if m["role"] == nil && (typ == "reasoning" || typ == "compaction" || typ == "mcp_approval_response" || strings.HasSuffix(typ, "_call") || strings.HasSuffix(typ, "_output")) {
+			return true
+		}
+	}
+	return false
+}
+
 func isResponsesResponse(m map[string]any) bool {
 	if _, ok := m["output"].([]any); !ok {
 		return false

pkg/types/aiobservabilitytypes/message_format_openai_responses_test.go

@@ -73,6 +73,23 @@ func TestFormat_OpenAIResponses(t *testing.T) {
 			wantInput:  `[{"role": "system", "parts": [{"type": "text", "content": "Be brief."}]}, {"role": "user", "parts": [{"type": "text", "content": "Say hi"}]}]`,
 			wantOutput: `[]`,
 		},
+		{
+			// LiteLLM langfuse_otel: the function_call items of a tool-calling turn as a bare list
+			name:       "BareFunctionCallItems_OneAssistantMessage",
+			output:     `[{"id": "chatcmpl-1", "name": "set_stage", "call_id": "call_7", "type": "function_call", "arguments": {"fields": {"stage": "consent"}}}]`,
+			wantInput:  `[]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "tool_call", "name": "set_stage", "id": "call_7", "arguments": {"fields": {"stage": "consent"}}}]}]`,
+		},
+		{
+			name:  "BareItemsWithMessage_ToolRound",
+			input: `[{"type": "message", "role": "user", "content": [{"type": "input_text", "text": "Weather in Oslo?"}]}, {"type": "function_call", "call_id": "call_9", "name": "get_weather", "arguments": "{\"city\":\"Oslo\"}"}, {"type": "function_call_output", "call_id": "call_9", "output": "4C, cloudy"}]`,
+			wantInput: `[
+			  {"role": "user", "parts": [{"type": "text", "content": "Weather in Oslo?"}]},
+			  {"role": "assistant", "parts": [{"type": "tool_call", "name": "get_weather", "id": "call_9", "arguments": {"city": "Oslo"}}]},
+			  {"role": "tool", "parts": [{"type": "tool_call_response", "id": "call_9", "response": "4C, cloudy"}]}
+			]`,
+			wantOutput: `[]`,
+		},
 		{
 			// openinference capture, CreateEmbeddings span: an embeddings request also has input and model
 			name:       "EmbeddingsRequest_NotResponses_KeptAsGeneric",

pkg/types/aiobservabilitytypes/message_format_openinference.go

@@ -4,8 +4,9 @@ import "github.com/SigNoz/signoz/pkg/valuer"
 
 // Format: the flattened llm.input_messages.N.message.* and llm.output_messages.N.message.* keys, the
 // provider-neutral copy of the conversation the OpenInference instrumentations write on LLM spans. A side
-// without them falls back to the raw provider payload the mapper moved onto gen_ai.*.messages; pipecat
-// streams the reply into output.value alone.
+// without them, or whose buckets hold no message, falls back to the raw provider payload the mapper moved
+// onto gen_ai.*.messages; pipecat streams the reply into output.value alone, LiteLLM's OpenInference mode
+// leaves the output bucket empty on a tool-calling turn.
 // https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/semantic_conventions.md#L77
 // https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/tool_calling.md#L37
 // https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/spec/multimodal_attributes.md#L35
@@ -28,15 +29,14 @@ func (f *formatting) openInference() bool {
 		return false
 	}
 	inputLabel, outputLabel := FormatterOpenInference, FormatterOpenInference
-	if len(input) > 0 {
-		f.out.Input = inputMessages(openInferenceMessages(input))
+	if msgs := openInferenceMessages(input); len(msgs) > 0 {
+		f.out.Input = inputMessages(msgs)
 	} else if raw := attribute(f.attributes, GenAIInputMessages); raw != nil {
-		var msgs OutputMessages
 		msgs, inputLabel = f.side(decode(raw), RoleUser)
 		f.out.Input = inputMessages(msgs)
 	}
-	if len(output) > 0 {
-		f.out.Output = openInferenceMessages(output)
+	if msgs := openInferenceMessages(output); len(msgs) > 0 {
+		f.out.Output = msgs
 		if fr := finishReason(stringOf(attribute(f.attributes, openInferenceFinishReason))); fr != nil {
 			for i := range f.out.Output {
 				f.out.Output[i].FinishReason = fr

pkg/types/aiobservabilitytypes/message_format_openinference_test.go

@@ -160,16 +160,30 @@ func TestFormat_OpenInference(t *testing.T) {
 			// which the mapper moves onto gen_ai.output.messages
 			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-pipecat/src/openinference/instrumentation/pipecat/_attributes.py#L430
 			// https://github.com/Arize-ai/openinference/blob/12f2cc035edfb00b60b656d977f07abfb7a69900/python/instrumentation/openinference-instrumentation-pipecat/src/openinference/instrumentation/pipecat/_observer.py#L628
-			name: "OutputKeysAbsent_RawOutputFormatted_Warns",
+			name: "OutputKeysAbsent_RawOutputFormatted",
 			attributes: map[string]any{
 				"llm.input_messages.0.message.role":    "user",
 				"llm.input_messages.0.message.content": "hi",
 				GenAIOutputMessages:                    "Hello there.",
 			},
-			warnings:   []string{"bare assistant text, role assumed", "output formatted as text, input as openinference"},
+			warnings:   []string{"bare assistant text, role assumed"},
 			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "hi"}]}]`,
 			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "Hello there."}]}]`,
 		},
+		{
+			// LiteLLM's OpenInference mode leaves the output bucket empty on a tool-calling turn; the calls are
+			// the Langfuse output the mapper moved onto gen_ai.output.messages
+			name: "EmptyOutputBucket_RawOutputFormatted",
+			attributes: map[string]any{
+				"llm.input_messages.0.message.role":     "user",
+				"llm.input_messages.0.message.content":  "Yes, go ahead.",
+				"llm.output_messages.0.message.role":    "assistant",
+				"llm.output_messages.0.message.content": "",
+				GenAIOutputMessages:                     `[{"id": "chatcmpl-2", "name": "set_stage", "call_id": "call_8", "type": "function_call", "arguments": {"fields": {"stage": "email"}}}]`,
+			},
+			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "Yes, go ahead."}]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "tool_call", "name": "set_stage", "id": "call_8", "arguments": {"fields": {"stage": "email"}}}]}]`,
+		},
 		{
 			// message.role left unset by some instrumentors: tool_call_id, tool_calls or neither decide it,
 			// as the collector-contrib normalizer does

pkg/types/aiobservabilitytypes/message_format_test.go

@@ -122,6 +122,31 @@ So, the ball costs 5 cents.`,
 			wantInput:  `[]`,
 			wantOutput: `[]`,
 		},
+		{
+			// Langfuse SDK generation: the app passes the request through superjson, the output is the parsed
+			// structured output
+			name:       "SuperJSONEnvelope_Unwrapped_StructuredOutputGeneric",
+			input:      `{"json": {"model": "gpt-4o-mini", "messages": [{"role": "system", "content": "Extract the main article of the page as JSON."}, {"role": "user", "content": "# Council approves new bike lanes"}], "temperature": 0}}`,
+			output:     `{"hasArticle": true, "article": {"title": "Council approves new bike lanes"}}`,
+			warnings:   []string{"message format not recognised, kept as generic"},
+			wantInput:  `[{"role": "system", "parts": [{"type": "text", "content": "Extract the main article of the page as JSON."}]}, {"role": "user", "parts": [{"type": "text", "content": "# Council approves new bike lanes"}]}]`,
+			wantOutput: `[{"role": "", "parts": [{"type": "generic", "content": "{\"article\":{\"title\":\"Council approves new bike lanes\"},\"hasArticle\":true}"}]}]`,
+		},
+		{
+			name:       "SuperJSONEnvelopeWithMeta_Unwrapped",
+			input:      `{"json": {"messages": [{"role": "user", "content": "What time is it in Oslo?"}]}, "meta": {"values": {"messages.0.at": ["Date"]}}}`,
+			output:     `{"json": {"choices": [{"finish_reason": "stop", "message": {"role": "assistant", "content": "It is 14:05 in Oslo."}}]}, "meta": {"values": {}}}`,
+			wantInput:  `[{"role": "user", "parts": [{"type": "text", "content": "What time is it in Oslo?"}]}]`,
+			wantOutput: `[{"role": "assistant", "parts": [{"type": "text", "content": "It is 14:05 in Oslo."}], "finish_reason": "stop"}]`,
+		},
+		{
+			// a json key beside others is a payload of its own, not the envelope
+			name:       "JSONKeyBesideOthers_NotUnwrapped_Generic",
+			input:      `{"json": {"messages": []}, "schema": "v1"}`,
+			warnings:   []string{"message format not recognised, kept as generic"},
+			wantInput:  `[{"role": "", "parts": [{"type": "generic", "content": "{\"json\":{\"messages\":[]},\"schema\":\"v1\"}"}]}]`,
+			wantOutput: `[]`,
+		},
 		{
 			name:       "NoMessageAttributes_EmptyLists",
 			attributes: map[string]any{GenAIRequestModel: "gpt-4o", GenAIToolName: "get_weather"},