problem · Revision 1
Finding: In an external report, Claude Code 2.1.296 on Windows 10 Pro 19045 (npm install) behaves as follows. After /remote-control is used in one session, a detached daemon reportedly keeps adopting and resuming sessions the user exited or killed, including
problem · Revision 1
Finding: In an external GitHub report, Output.array({ element }) in ai@7.0.137 (also reported on upstream main at ba05943) generates a response-format schema that mishandles a root-recursive Zod element, because the root self-reference $ref "#" ends up pointing at the array
problem · Revision 1
Finding: In a single external report, Claude Code CLI 2.1.296 (native install) on Fedora 44 with GNOME, gnome-keyring and GDM automatic login (login keyring locked at session start) showed a keyring unlock dialog at startup and a second when the
problem · Revision 1
Finding: In an external report on anthropics/claude-code issue #100924 (https://github.com/anthropics/claude-code/issues/100924), a plugin Pane in the Claude desktop app (Code mode, desktop 2.31226.1, engine 2.1.295, macOS 15.7.3) redrawing about every 213 ms reportedly stopped B
problem · Revision 1
Finding: In an external report on Windows desktop Codex (Microsoft Store OpenAI.Codex 26.930.3930.0, observed 2026-10-09), starting a new local Work chat in an existing project intermittently fails with the UI error "Could not use this project for a local chat"
problem · Revision 1
Finding: Multiple users report that Codex fails OAuth login to the official Meta Ads MCP endpoint https://mcp.facebook.com/ads. Dynamic client registration returns HTTP 400 invalid_client_metadata ("Dynamic registration is not available for this client") before any browser consen
problem · Revision 1
Finding: One user reports that on Windows 11 Pro ARM64 with Codex App 26.1002.7124.0 and CLI 0.161.0, native Windows sandbox setup fails in both elevated and unelevated modes. The error is "windows sandbox failed: helper_unknown_error: setup refresh had errors" (source:
problem · Revision 1
Finding: In an external report on openai/codex issue 52688, paths containing a tab, LF or CR are silently altered when exec_command resolves workdir or apply_patch resolves file paths, so Codex uses a different path than the model named. This is
problem · Revision 1
Finding: In one external report, a plugin pane Input in the Claude desktop app (version 2.31226.1, engine Claude Code 2.1.295, macOS 15.7.3, Code mode) does not receive keyboard input when it appears, even though the Input has autoFocus and a
problem · Revision 1
Finding: In an external report against Claude Code 2.1.296 (also seen on 2.1.292 and 2.1.293) on macOS, auto mode sessions started with --remote-control inside tmux can be blocked by a terminal-only "Teach auto mode about your environment?" offer. Remote Control
problem · Revision 1
Finding: In an external report on openai/codex issue 50868, a ChatGPT Plus user on macOS 15.7.7 (Apple Silicon) says that selecting "Use a reset" in the /usage > Redeem usage limit reset picker of the CLI 0.158.0-alpha.2.1 bundled with Desktop
problem · Revision 1
Finding: In an external report on Claude Code 2.1.292 (native claude.exe, Windows 11 ARM64, interactive ConPTY session), removing the directory named by CLAUDE_CONFIG_DIR while the process runs reportedly makes it spin at about 1.1 CPU cores indefinitely, with no error
problem · Revision 1
Finding: In an external report, Codex CLI 0.162.1 running on a remote Linux machine over SSH shows a browser-launch error when the user Cmd-clicks a URL printed in Codex output, instead of the local terminal opening the URL (https://github.com/openai/codex/issues/52766). Reported
problem · Revision 1
Finding: According to an external report on BerriAI/litellm issue #45333 (no KFA reproduction), for Gemini 3 or newer models the Vertex/Gemini config in map_openai_params adds temperature=1.0 whenever the caller did not set a temperature, and additional_drop_params=["temperature"
problem · Revision 1
Finding: In an external report against LiteLLM v1.106.0 (https://github.com/BerriAI/litellm/issues/45561), sync Responses API streaming via litellm.responses(..., stream=True) in a plain for loop reportedly never closes the reasoning output item for bridged reasoning models. The
problem · Revision 1
Finding: In a reporter's LiteLLM v1.100.0 setup, Databricks models served through the chat -> Responses -> chat bridge (model_info.mode set to responses, or a Databricks model name without gpt, per the reporter) reportedly get plain-string user content rewritten as a
problem · Revision 1
Finding: In an external report on LiteLLM main (commits 4417bf08, a 1.106.0 dev tree, and 63a4f3f2 dated 2026-10-09), non-streaming /v1/messages calls routed to a native anthropic/... deployment with deployment-level tpm set reportedly do not accumulate token usage for that deplo
problem · Revision 1
Finding: In one external report, ChatGPT Desktop for Windows 26.1007.21434 on Windows build 22631 with WSL2 fails to run Codex chat commands (exec_command) for a project opened through a \\wsl$ UNC path, when Agent environment and Terminal are both set
problem · Revision 1
Finding: In an external report on openai/codex issue 52842, the Windows desktop app 26.1007.21434 (build 13901, prod, Windows 11 10.0.26200 x64) showed roughly 1.742 GiB of existing chat transcripts as Unclassified chats on the Storage page. This was observed for
problem · Revision 1
Finding: In a reporter-stated environment (CrewAI source main commit d38ceafb68c6ad33196c9ac052b9b67e74dab37c, Python 3.12.15), crewai_tools JSONLoader reportedly treats a valid UTF-8 JSON file with a leading byte-order mark as malformed. The result has parse_error set in metadat