$ the-wire · showcase
Codex reroutes untrusted context and adds device auth; Gemini CLI patches ship
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
Codex turned on tool-output formatting for untrusted context by default, added device authorization to app-server gateway OAuth, and Gemini CLI shipped a stack of operator-facing fixes around recursive shell sessions, aborted runs, and Seatbelt folder trust.
Enable untrusted context tool outputs by default openai/codex
Untrusted additional context such as browser information and goal objectives now reaches the model as function_call_output items in the agent namespace instead of as user messages, and the feature is stable and on by default. If your prompts or evals depend on the old message shape, setting features.untrusted_as_tool_output = false preserves it.
Add device authorization to app-server gateway OAuth openai/codex
The app-server gateway OAuth flow now advertises supportsDeviceAuth in account/gatewayOAuth/read and accepts deviceAuth: true on login, returning a verificationUrl, userCode and loginId before approval. Saved credentials are reused or refreshed where possible, in which case the response comes back with prompt: null and no second device authorization.
fix(cli): prevent recursive interactive gemini sessions in shell mode google-gemini/gemini-cli
Shell Mode and the [Shell] tool now refuse to spawn a nested interactive gemini session and reject management subcommands like gemini config set, so typing gemini at the ! prompt no longer recursively launches the CLI. Queued and restored LLM prompts are also isolated from Shell Mode.
fix(cli,core): kill child processes, record cancelled tools, and preserve signal exit code on headless abort google-gemini/gemini-cli
On SIGTERM, SIGINT or SIGHUP during a non-interactive -p run, cleanup now kills detached child shell process groups and PTYs so background commands stop when gemini exits. Tool calls are persisted as the model stream completes, and --resume synthesizes an interrupted functionResponse when no final result was recorded.
fix(cli): resolve folder trust before entering the macOS Seatbelt sandbox google-gemini/gemini-cli
Since the Seatbelt profiles began denying access to ~/.gemini/trustedFolders.json, every launch under the macOS sandbox re-showed the folder trust dialog and any choice failed with exit code 52. Folder trust now resolves before the sandbox is entered.
fix(core): send the reasoning disable to custom models with saved overrides cline/cline
Setting Reasoning Effort to None on a custom OpenAI Compatible model with saved overrides silently omitted reasoning_effort: "none" from the request; the disable is now sent. On gpt-5.6/gpt-6 chat-completions endpoints that omission affected every tool call.
fix(core): reject MessageBus.request() immediately when publish fails google-gemini/gemini-cli
MessageBus.request() rejects immediately when its publish step fails, with the original error, and releases its timer and response listener instead of waiting out the 60-second default and returning a generic timeout.