Severity
Reports at this severity, newest stored first.
SearchParams aren't updated on client-side navigation when experimental cacheComponents and Draft Mode are both enabled; the page always shows the initial category until a hard reload. This is a caching key omission in Next.js App Router.
AsyncBaseTracer in langchain-core fails to forward the `name` argument to `_create_tool_run` and the `response` argument to `_errored_llm_run`, causing async tracers (including `with_alisteners`) to record incorrect run names and lose partial LLM outputs on streaming errors, unlike their sync counterparts.
The Supabase pricing page intermittently displays a client-side exception on initial load, but reloading the page resolves the issue. This suggests a hydration mismatch or client-side rendering inconsistency.
In Next.js 16.3.x with output: 'standalone', next-i18next.config.js is not copied into .next/standalone, causing 500 errors because next-i18next cannot locate its config at runtime. Regression from 16.2.12.
Next.js cacheComponents bfcache ordering causes React to reorder DOM siblings, re-parenting iframe subtrees and reloading them from the second visit onward.
ToolMessage content that contains standard LangChain content blocks (like images) is forwarded to the OpenRouter API without conversion, leading to a client-side ValidationError. This makes agent threads unrecoverable until history is trimmed.
Newly created hosted preview branches intermittently terminate active database sessions with FATAL SQLSTATE 57P01 after the branch reports ACTIVE_HEALTHY and after initial migrations/smoke queries succeed. The issue appears tied to preview-branch provisioning or the shared session pooler rather than application code.
Reading content_blocks on a legacy v0 multimodal block with an id raises TypeError due to duplicate id keyword argument in _convert_legacy_v0_content_block_to_v1.
When an invalid structured-output tool call and an ordinary tool call appear in the same AIMessage, the agent terminates after the first model call without feeding the validation error back to the model. As a result, no retry happens and structured_response is None.
Next.js 16.3.5 production build using Turbopack panics with a TurbopackInternalError: 'expected chunkable module for async reference' when the bknd package is used. The webpack build succeeds, indicating a Turbopack-specific issue with async module chunking.
When after() is called with a promise that never settles, the bookkeeping .then() callback registers in the caller's async context, pinning the entire request-scoped context (including RSC payloads) for the lifetime of the server process, leading to unbounded memory growth.
After an ISR revalidation turns a route that was prerendered as a fully static redirect() into a partially prerendered page, client-side navigations fail: RSC requests receive the static shell only (or 404), never resume, causing React error 412 and a full page reload. The build-time prefetchDataRoute pins the route to serve prebuilt static RSC content, ignoring the new postponed cache entry.
A scoped/fine-grained personal access token created with the Full access preset cannot reveal project API keys through the Management API, causing CLI commands like `supabase link` to fail with 403. Classic PAT works fine. This indicates a permission evaluation bug for scoped tokens with Full access.
Setting max_retries=0 in ChatOpenRouter does not disable retries; the omission of retry_config causes OpenRouter SDK to apply its default backoff policy, leading to unexpected repeated requests and potentially long hangs on 5xx responses.
In Next.js 16.3.0 with Rspack enabled (NEXT_RSPACK=true) on Windows using pnpm, direct dependencies listed in transpilePackages are incorrectly externalized during development, leading to runtime errors such as 'No matching bindings found'. The issue is specific to the dev server and pages router; webpack dev and rspack production build work correctly.
RemoteGraph's invoke/ainvoke returns message dictionaries instead of BaseMessage instances, causing CompiledSubAgent to produce empty ToolMessage content when used as a subagent. This breaks the parent agent's ability to read subagent results.
Next.js Streaming Metadata renders a <div hidden> as the first child of <body>. React 19 hydration matches only by tag name and claims any prepended <div> (e.g. consent manager or chat widget) as the metadata wrapper. Hydration then fails with Minified React error #418, and React's clearContainerSparingly removes the third-party DOM node.
langchain-openrouter streaming paths raise plain ValueError with the numeric error code interpolated into prose, discarding structured error classification and the is_retryable attribute. Callers cannot distinguish retryable 502/429 from non-retryable 401/403. Non-streaming invoke path also fails with a misleading no-choices ValueError because the error field is dropped during SDK parsing.
In Next.js 16.3.3 with cacheComponents enabled and a client component in the layout, the first request to a route with exactly 8 URL segments causes a RangeError (maximum call stack size exceeded) in React Flight's requireModule/getComponentNameFromType. The error disappears after any shallower route compiles or when using Node 22. This only affects dev mode (Turbopack or webpack).
Streaming CommaSeparatedListOutputParser produces incorrect fields when a quoted CSV field containing a comma is split across chunk boundaries. Non-streaming parsing is correct, but transform/atransform silently corrupts data.
Filtering with the `is` operator and an uppercase value like 'NULL', 'NOT NULL', 'TRUE', or 'FALSE' generates invalid SQL because the value is compared case-sensitively and then treated as a literal string instead of a SQL keyword.
When `convert_to_openai_tool` receives a dict that is already in OpenAI tool format (e.g. `{"type": "function", "function": {...}}`) along with `strict=True`, it returns the dict unchanged without adding `strict` or setting `additionalProperties: false`. This silently disables Structured Outputs for `bind_tools` and `with_structured_output`, potentially causing model outputs that do not match the JSON schema.
The upgrade codemod modifies eslint from a caret range to an exact latest major (10.10.0) even when target Next.js peer dependencies allow ESLint 9, causing peer dependency conflict with eslint-plugin-import and aborting the upgrade.
A client disconnecting during /_next/image optimization causes the response generator promise to never settle, permanently poisoning the cache key for all subsequent requests on self-hosted Next.js (next start/standalone).
aggregate_lines_to_chunks() repeatedly appends to a growing string, causing quadratic time complexity when many adjacent lines share the same metadata. The issue is observable with large inputs (e.g., 16,000 paragraphs take ~5.8 seconds) and becomes progressively worse.
The middleware replaces a trailing string in system message content with a cache_control-only dict, losing the original text.
Setting a database password containing extended ASCII characters (e.g., è, Ö, Æ) is accepted by the Supabase UI, but subsequent dashboard connections to the PostgreSQL database fail with authentication errors, making the database inaccessible through the dashboard.
Next.js dev server calculates the V8 max-old-space-size based on host total memory (os.totalmem()), ignoring container memory constraints. In a Docker container with a memory limit, this sets an excessively high heap limit, potentially leading to out-of-memory (OOM) kills or performance degradation. Node.js already provides process.constrainedMemory() which respects cgroup limits; Next.js should use it when available.
Next.js SWC styled-components transform hoists the css prop to the styled component declaration, causing a ReferenceError due to temporal dead zone when the css prop references a const declared later in the module.
langchain-openai 1.6.4 rejects shell, programmatic_tool_calling, and local_shell tool types in bind_tools, and the Responses API conversion layer does not handle shell_call, shell_call_output, local_shell_call, local_shell_call_output, program, program_output items, nor preserves allowed_callers and caller fields, breaking GPT-6 programmatic tool calling end-to-end.