fix(notebook): stop streaming jumpiness — disable smooth markdown + stable adapter#695
Open
fix(notebook): stop streaming jumpiness — disable smooth markdown + stable adapter#695
Conversation
Fixes the admin gate ("Kein Zugriff" page) caused by the legacy /status
handler stripping fields from the user response. The hand-picked literal
omitted is_admin and ~30 other UserProfile fields, so the frontend's
useAuthStore could never see is_admin=true even when the DB row had it.
Replaces the manual pick with toBetterAuthUser() — the same null-strip +
Zod-parse boundary already used by authMiddleware. The new ts-rest
contract enforces the response shape end-to-end, making this regression
class impossible to reintroduce.
- packages/contracts: authStatusContract + authStatusResponseSchema
reusing userProfileSchema.nullable()
- apps/api: authStatusContractRouter mounted in routes.ts before the
legacy authRouter; legacy /status removed from authCore.ts
- packages/shared: authStatus added to the typed ContractsClient
- middleware: toBetterAuthUser exported for reuse
…page Remove the two cards from the StatisticsSection grid so the startpage focuses on document count and date range. Collapse the grid from four columns to two so the remaining cards fill the row at all breakpoints, and shrink the matching skeleton placeholders.
Regolo's gemma4-31b endpoint hangs upstream — every notebook chat with
the default Gemma model just spun forever because the AI SDK has no
built-in first-token deadline. This change:
- Adds litellmFetchWithThinkingDisabled (sibling to regoloThinkingFetch)
injecting Ollama's `think: false` so LiteLLM-served gemma streams
content instead of burning its entire token budget on `reasoning`.
- Re-routes the user-facing "Gemma 4" model from Regolo → LiteLLM. Old
`gemma-regolo` ID is aliased server-side and migrated client-side
(chatStore v6) to the new `gemma-litellm` ID.
- Adds Qwen 3.6 27B as a selectable model (already in the existing
Regolo reasoning-stream allowlist, so no extra wiring).
- Introduces a 20s first-token deadline + single-step cross-provider
fallback (gemma-litellm ↔ gpt-oss-regolo) in responseStreamingService.
Qwen entries intentionally have no `fallback` field — the
Chinese-only-when-selected firewall (informed-consent boundary,
documented in ModelConfig).
- Fixes pre-existing bug: getModel('litellm', modelId) ignored the
modelId arg and always used LITELLM_DEFAULT_MODEL.
Fallback is silent end-user-side: server emits a `fallback` SSE event,
both runtime adapters log it to the browser console, no UI banner.
Implementation notes:
- streamAndAccumulate / streamAndAccumulateWithReasoning now have a
shared `wrapWithCompatCatch` factory and an `*OrThrow` internal layer
used by streamWithFallback. Existing chat router callers see the same
null-on-failure shape, plus the new deadline + empty-completion
safety nets for free.
- Single shared deadline across initial-probe iterations (was
accidentally giving 40s grace via per-call setTimeout).
- Reasoning streamer split into Phase-1 (race vs deadline until first
text) + Phase-2 (drain without race) — eliminates wasted Promise.race
microtask hops on every reasoning chunk after first content.
- Uses native AbortSignal.any() (Node 20.3+) instead of a hand-rolled
composeAbortSignals helper.
… fallback" This reverts commit 5ba1b6d.
…+ stable adapter Two unrelated causes worked together to make notebook chat answers visibly jump up and down during streaming: 1. assistant-ui's MarkdownTextPrimitive defaults to smooth=true, which re-buffers SSE chunks and reveals them character-by-character via requestAnimationFrame. With notebook's citation-dense answers (placeholder <sup> badges via processChildren(..., true)), this caused continuous line-wrap recalculation — observable as +400/-500 pixel oscillations in mid-stream. General chat is fine because its answers are shorter and less citation-dense. 2. NotebookChatProvider's getConfig was wrapped in useCallback with seven deps. Any upstream prop identity churn (e.g. config.collections rebuilt by getNotebookConfig on every render) recreated the adapter, which forced useLocalRuntime to reinitialize and reset scroll/streaming state. Visible as the "[Notebook] ⚠ Adapter RECREATED" warning firing before the first request. Fix: - New per-thread MarkdownStreamingContext. Default smooth=true preserves the typewriter feel for general chat. NotebookChatProvider sets smooth=false. CitationMarkdownText reads from context. - All getConfig inputs moved behind refs (same pattern already used for threadId/getFilters/onComplete). getConfig itself is now a stable useCallback([]) that snapshots refs at request time, so the adapter is created exactly once per provider mount.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Notebook chat answers visibly jump up and down during streaming on long, citation-dense responses. Two unrelated causes were compounding:
MarkdownTextPrimitivesmooth-streaming. assistant-ui's `smooth=true` default re-buffers SSE chunks and reveals them character-by-character via `requestAnimationFrame`. With notebook's citation-dense answers (placeholder `` badges via `processChildren(text, citationMap, true)`), this caused continuous line-wrap recalculation — observable as +400/-500 pixel oscillations mid-stream. Diagnosed via a temporary `ResizeObserver`-based viewport jump monitor: identical `+1256 / −1252` cancellation pairs landed on every `useSmooth` animation frame.General chat is unaffected — its answers are shorter, less citation-dense, and its provider doesn't churn the adapter the same way.
Fix
Files
Test plan