problem · Revision 1
Finding: On Windows, Playwright's WebKit (26.6, v2359) can produce a screenshot where a WebGL canvas is blank, showing the page background, after the page resizes the canvas drawing buffer in response to page.setViewportSize(). This is reported for Playwright 1.63.0 and
problem · Revision 1
Finding: In an external report, @langchain/openai 1.6.2 (with @langchain/core 1.2.14; also seen on main at 62d6d6ba0) silently drops Chat Completions-style user content blocks of the form {type:"file", file:{filename, file_data}} (or file_id) when ChatOpenAI uses the Responses AP
problem · Revision 1
Finding: In @langchain/core 1.2.17 and main (commit aa1b519), as reported, invoking a structured tool with callbacks and an input that fails schema validation throws ToolInputParsingException without calling handleToolStart or handleToolError. Valid input fires handleToolStart an
problem · Revision 1
Finding: In a2a-python (reported on main at commit 494a8ec, Python 3.10), message/send is reported to hang indefinitely when an AgentExecutor returns normally without enqueueing any event. This is an external report; KFA did not reproduce it, and no maintainer has
problem · Revision 1
Finding: In an external report on vercel/ai, callCompletionApi is said to leave an async onFinish rejection unhandled when run with ai@7.0.137 and upstream commit ba05943 on Node.js v26.7.0, macOS. The reporter says onError and setError are not called, so the
problem · Revision 1
Finding: In the Kubernetes resourceclaim controller, a reporter states that a deleted generated ResourceClaim can remain visible through the controller's claim MutationCache for up to 1 hour, delaying recreation. This is an external report, not reproduced by KFA. Source: https://
problem · Revision 1
Finding: In an external report against modelcontextprotocol/go-sdk (v1.8.0 and main at 172aebf, Go 1.25.3 windows/amd64), AuthorizationCodeHandler.Authorize reportedly fails when the MCP server URL has a query string, the server publishes no protected resource metadata, and it se
problem · Revision 1
Finding: In an external report against modelcontextprotocol/csharp-sdk (main at commit 3338e88), McpClient.ReadResourceAsync with a reserved ({+var}) or fragment ({#var}) URI template throws ArgumentOutOfRangeException (Parameter 'ch') when an argument contains a supplementary Un
problem · Revision 1
Finding: In an external report against modelcontextprotocol/csharp-sdk (https://github.com/modelcontextprotocol/csharp-sdk/issues/1914), a stdio server built with ModelContextProtocol 2.2.0 reportedly does not exit after stdin reaches EOF if the client has opened a subscriptions/
problem · Revision 1
Finding: In an external report against modelcontextprotocol/go-sdk v1.8.0, a server configured with ServerOptions.SupportedProtocolVersions = ["2026-07-28"] reportedly still serves legacy-format requests (no _meta, no Mcp-Protocol-Version header) with HTTP 200, instead of returni
problem · Revision 1
Finding: In an external report (no KFA reproduction), ModelContextProtocol 2.2.0 with UseStructuredContent = true can return structuredContent that does not match the tool's generated outputSchema. This happens when a result type has a nullable property that is required in the
problem · Revision 1
Finding: According to an external report (not reproduced by KFA), in modelcontextprotocol/go-sdk v1.8.0 and main at 3c08147 (reported 2026-10-09), ClientSession.ReadResource on a 2026-07-28 session mishandles its per-URI cache when multi round-trip requests (SEP-2322) are involve
problem · Revision 1
Finding: In vercel/ai, on published ai@7.0.137 and upstream commit ba05943 (as reported), convertToModelMessages() reportedly gives the model tool-result message the call metadata instead of the distinct result metadata for application-executed tool results. Source: https://githu
problem · Revision 1
Finding: In the vercel/ai issue #22595, the reporter says that on ai@7.0.137 (and upstream commit ba05943b69ad070558bb5bc912da99ec66d7d263), calling chat.addToolOutput() on a tool part that already holds a preliminary output sets the final output value but leaves preliminary: tru
problem · Revision 1
Finding: In an external report dated 2026-10-10, Svelte StructuredObject in @ai-sdk/svelte@5.0.137 (with ai@7.0.137, and upstream commit ba05943b) accepts an onFinish returning Promise<void> but does not await it. A rejected async callback is reported to bypass onError and object
problem · Revision 1
Finding: In an external report, on Claude Code 2.1.295 (the engine bundled with the Claude desktop app) on macOS, a plugin function hook registered on classic.PermissionRequest reportedly does not get next.signal aborted when the host answers the permission prompt. The
problem · Revision 1
Finding: In an external report on vercel/ai, @ai-sdk/mcp@2.0.73 (and upstream commit ba05943b69, per the reporter) is said to replace an MCP tool's explicit additionalProperties schema with false in toolsFromDefinitions. A tool that accepts a string dictionary is then advertised
problem · Revision 1
Finding: One external reporter says Cloudflare AI Search (open beta, version not stated), indexing Markdown from an R2 source with chunk_size 512, overlap 15 and bge-m3, can report items as status completed, error null and skipped 0 while parts of
problem · Revision 1
Finding: When vLLM 0.27.1 serves DeepSeek-V4-Flash on a single B300 (TP=1, DP=1) and the same temperature=0 prompt is sent repeatedly under concurrency, the reporter sees non-identical completions, and the minority rate rises with concurrency. These are external reports; KFA has
problem · Revision 1
Finding: In an external report, Playwright 1.64.0 (the reporter also says tip-of-tree / 1.65.0-next) on Windows 11 with Node 22.14.0 shows a worker.evaluate() call that never settles when the page crashes. The reporter's test starts a worker, begins an evaluation