Claude Code 24 fixes, Codex CLI voice mode, Cursor quiet week

What shipped

From 11 to 18 September, Anthropic pushed extensive updates to its terminal coding agent and OpenAI released new CLI features, while the AI-first IDE stayed the same. The changes focus on reliability, plugin testing, and new interaction modes for developers. These releases continue the shift toward more controllable agent workflows.

Cursor

Cursor: The AI-first IDE released no new features or fixes from 11 to 18 September, leaving users with the prior version for agentic coding tasks.

Claude Code

Anthropic released the bulk of the week's changes through Claude Code, issuing 24 updates centered on memory handling, plugin testing, and gateway stability. These fixes target daily reliability issues for terminal users and VS Code workflows. The volume of releases shows continued investment in production-grade agent controls.

• •Memory warnings Claude Code now shows critical memory alerts with recovery steps and adds configurable startup waits for MCP servers plus OpenTelemetry tracing for requests.
• •Skill syncing Background syncing of skills from claude.ai accounts arrived along with a send-now shortcut and login confirmations for the apps gateway.
• •Proxy fixes Requests no longer fail with 400 errors when ANTHROPIC_BASE_URL points to a proxy, correcting a regression from version 2.1.275.
• •Gateway headers Custom headers for LLM gateways, remote session forking, and MCP disconnect notifications were added with fixes for Bash permissions and prompt caching.
• •MCP reliability Critical memory warnings, startup controls, and tracing improvements arrived with fixes for transcript loops and tool timeouts.
• •Plugin eval The claude plugin eval command now runs test suites, while /output-style switching and Bash file diffs in results were introduced.
• •Git permissions A regression causing read-only git commands to request permission after long sessions was fixed in version 2.1.269.
• •Fast mode Remote sessions gained a fast mode, mouse controls in config, and per-command sandbox allowlists plus subagent options.
• •Reliability pass Targeted bug fixes improved session execution and overall stability across developer workflows.
• •Self-hosted cleanup The --remove-session-state flag now deletes per-session directories on self-hosted runners after each run ends.
• •Pricing sync Custom gateway.yaml pricing now reaches signed-in clients so /cost reports match actual spend meters.
• •OTel tags Repository attributes can now tag OpenTelemetry metrics and events with vcs.ref.head details when detailed logging is enabled.
• •Message recall Queued messages in cloud web sessions can be retracted before Claude reads them by pressing Escape or Up.
• •Hooks dialog A new Hooks dialog in the VS Code command menu lets users add, edit, and remove lifecycle hooks across config levels.
• •Agent Map VS Code now shows an Agent Map in the footer for viewing per-agent cards, transcripts, and stopping subagents.
• •Bash diffs File edit diffs from Bash commands now appear in tool results when the bashEditDiffEnabled setting is active.
• •Output styles The /output-style command lists and switches transcript formats for remote, cloud, and headless sessions.
• •Eval command claude plugin eval runs a plugin suite and produces scored JSON and HTML reports for reproducibility.
• •Git prompt fix Read-only git commands no longer trigger permission prompts after extended session runtimes.
• •Gateway sync Pricing sync, CIDR warnings, and endpoint error fixes were added for self-hosted setups.
• •VS Code tools Plugin eval, output-style switching, and an agent map with hooks dialog reached VS Code users.
• •Permission regression Read-only git commands stopped asking for permission after long sessions, fixing the 2.1.269 issue.
• •Runner options Version 2.1.268 added session cleanup flags, JSON plugin output, and fixes for idle CPU and endpoint errors.
• •Release 2.1.269 Plugin evaluation, output styles, agent maps, and prompt cache fixes shipped with message retraction support.

OpenAI Codex

OpenAI shipped ten updates to Codex CLI and related desktop tools, introducing voice conversations, new model support, and Windows daemon improvements. The changes expand interaction options and prepare for the GPT-5.5 retirement. Builders gain more ways to run agents locally or in mixed environments.

• •Voice mode Codex CLI 0.155.0 added experimental /voice conversations, live reasoning summaries in the TUI, and Touch ID checks for MCP requests.
• •GPT-6-Astra Version 0.154.0 introduced GPT-6-Astra model support, Git worktree management, and a shared Windows daemon.
• •Cygwin builds CI archives for native Windows voice releases were published with pinned source and binary package maps.
• •Python SDK Codex CLI Python SDK 0.154.0 added max and ultra reasoning levels plus ExternalMessage support for turn handling.
• •Quick chats ChatGPT desktop 26.908 brought floating Pets controls for quick chats on macOS and Windows plus Appshots on Windows.
• •Model retirement GPT-5.5 will leave ChatGPT and Codex on 14 October 2026, requiring migration to GPT-5.6 Sol for desktop users.
• •Worktree support Codex CLI 0.154.0 added GPT-6-Astra, experimental worktrees, inline questions, and Vim replace mode.
• •Voice packaging Internal build archives for Windows voice runtime were released with source mappings and license notices.
• •SDK updates The Python SDK gained reasoning-effort values, turn-level tier selection, and typed notification handling.
• •Desktop chats Floating Pets controls now let users start quick chats and monitor threads while working in other apps on Windows.

What this means for you

For Vibe Builders: You can test Claude Code's new plugin eval command and output-style switching to run repeatable agent checks without writing code. Codex CLI voice conversations and Pets quick chats let you interact with agents while staying in other apps. Cursor staying the same means your current setup remains stable for shipping small projects this week.

For Non-techies: Claude Code memory warnings and gateway pricing sync reduce surprise costs and crashes during daily agent use. Codex voice mode and Appshots make it easier to capture context from your desktop without extra steps. No Cursor changes keep your existing workflow intact while these tools move toward more reliable day-to-day actions.

For Developers: Claude Code's OpenTelemetry tagging, self-hosted runner cleanup, and Agent Map give clearer observability and subagent control in production pipelines. Codex CLI worktree support and reasoning-effort settings let you benchmark new models against current stacks. Watch the GPT-5.5 retirement date and test the Python SDK hooks before integrating into existing CI flows.

What to watch next

Track Claude Code MCP stability after the memory warnings and plugin eval adoption. Watch Codex CLI for Windows daemon updates and the October GPT-5.5 cutoff. Check whether Cursor issues a catch-up release next week.