problem · Revision 1
Finding: In an external report, Crew.test() on crewAI main at commit 140bfcb (reported 2026-10-09; Windows 11, Python 3.13, uv venv) can raise IndexError ("list index out of range") after all iterations complete when a ConditionalTask is skipped in at least
problem · Revision 1
Finding: In an external report on crewAI 1.15.22 (main at commit 4ed2abc), the default executor (experimental AgentExecutor) reportedly emits ToolUsageFinishedEvent with from_cache=False even when a native function-calling tool call was answered from cache, so a cached run and a
problem · Revision 1
Finding: In an external report, LiteLLM SDK non-streaming acompletion (mistral/mistral-small-latest, versions 1.96.2 and 1.104.2) returns usage with prompt_tokens=0 and completion_tokens=0 when the provider response omits usage or sets it to null. This looks identical to a provid
problem · Revision 1
Finding: When google-adk at main commit 3c31fa510 (Linux, Python 3.11.16, mocked gemini-2.5-flash) handles a failed model call, the reporter says the invocation/invoke_workflow, invoke_agent, call_llm and generate_content spans end with ERROR status but carry no error.type attrib
problem · Revision 1
Finding: In reports on vercel/ai issue 13588, streamText in an Expo React Native client (expo 54, react-native 0.81.5, ai ^6.0.116) fails on a successful HTTP 200 SSE response with AI_APICallError 'Failed to process successful response'. The same code works in
problem · Revision 1
Finding: In google-gemini/gemini-cli, per an external report against main commit c9096a847, parseCustomHeaders splits the GEMINI_CLI_CUSTOM_HEADERS value on any comma followed by text and a colon, so a single header whose value contains such a pattern (for example JSON metadata o
problem · Revision 1
Finding: In langgraph-checkpoint-sqlite 3.1.1 (per a commenter's reproduction), SqliteStore reportedly ignores the documented `fields` key in its index configuration, falls back to ['$'], and embeds the whole serialized item rather than only the listed fields. Sources: https://gi
problem · Revision 1
Finding: In external reports on google-gemini/gemini-cli, the built-in GoogleSearch (google_web_search) tool can leave the CLI in a Thinking state for a very long time with no upper bound on tool execution. The original report cites CLI 0.61.0 on win32 (no
problem · Revision 1
Finding: In an external report on langchain-ai/langgraph issue 9001 (https://github.com/langchain-ai/langgraph/issues/9001), a thread checkpointed with a state key later removed from the schema, and resumed on a graph compiled with interrupt_before, is reported to pause before th
problem · Revision 1
Finding: In an external report for Gemini CLI 0.61.0 (win32, Node v24.14.1, gemini-api-key auth, gemini-3.5-flash), a project over its monthly spending cap saw repeated generic high-demand messages instead of the API error (https://github.com/google-gemini/gemini-cli/issues/29504
problem · Revision 1
Finding: In vLLM main at commit 5281e49 (per the reporter), DotsToolParser in vllm/tool_parsers/dots_tool_parser.py is reported to behave incorrectly in two ways during tool call extraction. This is an external report, not reproduced by KFA, and the cause is not maintainer-confir
problem · Revision 1
Finding: In an external report against LiteLLM v1.104.0, a proxied Vertex AI Gemini streamGenerateContent (alt=sse) call that gets HTTP 200 and then loses the connection before the first SSE event is not retried by the router; router_settings.num_retries is not applied.
problem · Revision 1
Finding: In vLLM main c23ca06b6d (0.31.1rc1.dev150, xgrammar 0.2.8), serving a Hermes-style model with --tool-call-parser hermes --enable-auto-tool-choice, a request with tool_choice="required" reportedly returns only one tool call when several are needed, whereas tool_choice="au
problem · Revision 1
Finding: In an external report (not reproduced by KFA), crewai deploy create/push on crewAI 1.15.23 (main), Windows 11, Python 3.13, venv, reads .env differently from crewai run, so a deployment can get different environment values than a local run. Source:
problem · Revision 1
Finding: In an external report against crewAI 1.15.22 (main @ 4ed2abc7), a failed non-streaming OpenAICompletion call/acall is said to emit LLMCallFailedEvent twice, for both chat completions and the Responses API (https://github.com/crewAIInc/crewAI/issues/7783). This is an exte
problem · Revision 1
Finding: In vLLM source main at commit ad73a4740dd780c5620099261738a30b979b262c, a reporter says different benchmark sweep combinations can map to the same result directory, and the cached-result branch then reuses the earlier experiment's measurements for the later one (https://
problem · Revision 1
Finding: In an external report, crewai 1.15.22 on Python 3.13 allegedly lets a Flow subclass that has only @listen methods and no @start() entrypoint run kickoff() without error. The reporter says it prints a completion message, returns None, leaves state
problem · Revision 1
Finding: In an external report on vLLM 0.30.0, serving a Qwen3_5ForConditionalGeneration (hybrid GDN) model with pipeline-parallel-size 2 and dflash speculative decoding fails at startup with a ValueError saying the model does not support dflash with pipeline parallelism. The rep
problem · Revision 1
Finding: In an external report on vLLM (reporter cites v0.30.0 and says the same code is on main), ECExampleConnector.has_cache_item() is said to create an empty per-hash directory in shared EC storage whenever a lookup misses. Source: https://github.com/vllm-project/vllm/issues/
problem · Revision 1
Finding: In google-adk, closing the merged live event stream returned by _merge_live_event_streams (src/google/adk/live/_runner_utils.py) can hang forever if aclose() is called while the internal one-slot merged queue holds an event. The reporter saw this on main at commit 4d0664