problem · Revision 2
Finding: In an external report, langchain-core 1.6.9 and current master are said to undercount tool calls in count_tokens_approximately when an AIMessage with tool_calls has list content rather than string content. This is the author's report, not a KFA reproduction (https://gith
problem · Revision 1
Finding: In an external report on cloudflare/workers-sdk, `wrangler preview` reportedly fails after a successful asset upload with Cloudflare API error 10021, 'module with name _headers must have content type text/plain'. The report says this happens for SvelteKit apps using adap
problem · Revision 2
Finding: In an external report on run-llama/llama_index issue 23086, llama-index-vector-stores-lancedb 0.6.0 (also reported on main) builds delete and filter predicates with double-quoted IDs, and with lancedb 0.38.0 or later these fail with a schema error beginning "ValueError:
problem · Revision 1
Finding: In openai-agents-python 0.23.1 and main @ 71c2da4, per the reporter, compaction on a RedisSession can expose an empty or partial history to concurrent readers and can leave the session permanently empty if the process is killed mid-compaction. This is
problem · Revision 1
Finding: In openai-agents-python, per an external report, when a sandbox call fails and the manifest has configured mount authority, redact_mount_error_data raises a generic RuntimeError but first clears the original exception graph in place. This includes exceptions reachable th
problem · Revision 1
Finding: In an external report (not reproduced by KFA), a mustache-format ChatPromptTemplate with nested variables such as {{user.name}} and {{user.company}} renders correctly from nested dict input, but its generated input schema types the top-level variable as a string. As a
problem · Revision 2
Finding: An external report against langchain-core 1.6.9 installed from master (Python 3.11.15, tenacity 9.1.4) says RunnableLambda(...).with_retry(stop_after_attempt=2, wait_exponential_jitter=False) can return the wrong item in batch() and abatch() when one input recovers on re
problem · Revision 2
Finding: In an external report (https://github.com/cloudflare/workers-sdk/issues/16165), a Workers AI chat request to @cf/qwen/qwen3.8-27b with response_format json_schema and reasoning_effort "low" timed out with error code 3046 after about 120 seconds on one specific ~3,500-tok
problem · Revision 1
Finding: In an external report, claude-agent-sdk 0.2.160 on Python 3.11.16 with typing_extensions 4.16.0 silently registers a @tool with an empty input schema when input_schema is a class built on typing_extensions.TypedDict. No error or warning is raised (https://github.com/anth
problem · Revision 1
Finding: In an external report, a CodeActAgent handing off to a FunctionAgent inside an AgentWorkflow stored an assistant message without its tool-call block. The following tool result kept its tool_call_id, so the receiving agent saw an orphan tool response. The
problem · Revision 2
Finding: In an external report on cloudflare/agents (issue #2503, https://github.com/cloudflare/agents/issues/2503), a Think agent using @cloudflare/think 0.20.0, agents 0.26.0, ai 7.0.127, @ai-sdk/react 4.0.130 and @ai-sdk/openai 4.0.83 shows the resumed reply twice in the clien
problem · Revision 2
Finding: In a single external report (not reproduced by KFA), a Think messenger using chatSdkMessenger on @cloudflare/think 0.18.0, agents 0.23.0 and chat 4.40.0 (burst strategy) can leave an inbound message undispatched after its isolate dies mid-dispatch, for example during a
problem · Revision 1
Finding: In an external report on openai-agents 0.23.1 (Python 3.12.3, Linux), RunState.from_json can resume the wrong agent when two agents share a name and each has its own hooks object with mutable public state. If that state differs between to_json()
problem · Revision 1
Finding: In an external report against llama-index-core 0.14.25 (Python 3.12.3, Linux), HierarchicalNodeParser reportedly returns the input Document unsplit through its async entry points (aget_nodes_from_documents, acall, IngestionPipeline.arun), while the sync paths build the h
problem · Revision 1
Finding: In openai-agents 0.23.1 (Python 3.12.3, Linux), a user reports that OpenAIResponsesCompactionSession can compact a function_call that is still waiting for tool approval. After approval and resume, the function_call_output is stored without its function_call, and later ru
problem · Revision 1
Finding: In an externally reported regression, a Worker started with createTestHarness() hangs the Node.js process on the first RPC call that creates a Workflow (env.MY_WORKFLOW.create() inside a WorkerEntrypoint method called via getWorker().getExport()). The reporter says this
problem · Revision 1
Finding: In an external report (not reproduced by KFA), llama-index-core 0.14.25 on Python 3.12.3/Linux raises a pydantic ValidationError from LLM.predict_and_call and apredict_and_call when the LLM is not a function-calling model and no react_chat_formatter is passed. The error
problem · Revision 1
Finding: In the reported setup (@cloudflare/ai-chat@0.9.0, agents@0.17.0, ai@6.0.213), a single assistant turn that first calls a tool that settles and later calls a needsApproval tool puts both parts in one assistant message. Approving the second tool does not fire the
problem · Revision 2
Finding: In a single-reporter external report, a local Miniflare Workflow running sequential, immediately successful step.do() calls (ten-minute timeout, retries off) failed after 5,000 completed steps with a QuotaExceededError about 10,000 active timeouts. Versions reported: min
problem · Revision 2
Finding: In an external report on langgraph 1.2.12 (and main at 9a0394d88, per the reporter), with Python 3.13.11 on Linux and InMemorySaver, a graph compiled with interrupt_after does not stream the checkpoint saved at the pause on stream_mode="checkpoints" (or as