This Week in Coding Agents: September 14, 2026

Week 38, 2026 · Published

Claude Code2.1.270
Codex CLIpython-v0.154.0
Gemini CLI0.59.0

The short version

  • Claude Code’s changelog lists 2.1.270 as the top release (September 12). There is no 2.1.264 heading. 2.1.269 adds claude plugin eval and /output-style. 2.1.268 limits TodoWrite and the Task* tools to older model IDs unless you set CLAUDE_CODE_ENABLE_TODO_TOOLS=1.
  • Codex CLI 0.154.0 (September 9, tag rust-v0.154.0) adds experimental --worktree / /worktree and removes the deprecated codex mcp-server entry point. The fetch script recorded latest as python-v0.154.0 (September 10), which is the Python SDK tag, not a newer CLI.
  • Cursor has no version number. The September 10 changelog entry is Cursor Projects: a coordinator agent in the left nav, running on a cloud machine, in beta.
  • GitHub Copilot has no version number. Business and Enterprise can now set managed permissions for agent operations. MAI-Code-1-Flash dropped on September 10. Copilot CLI put Project HydraFusion in /experimental.
  • Gemini CLI 0.59.0 shipped on September 8. The notes I read are MCP OAuth SSRF and fail-closed workspace-trust fixes, not a new model.

Claude Code

2.1.265 (September 8) through 2.1.270 (September 12). I opened the live CHANGELOG.md. The first heading is 2.1.270. There is no 2.1.264 heading on that page.

2.1.270 (September 12). The only note on the page: read-only git commands in Bash no longer ask for permission after a session has been running for a while. The changelog calls that a regression in 2.1.269.

2.1.269 (September 11). claude plugin eval runs a plugin’s eval suite and writes a scored JSON plus HTML report (claude plugin eval --help). /output-style [name] lists and switches output styles, including Remote Control, cloud, and other headless sessions. The Bash tool can attach a diff of files it changed when bashEditDiffEnabled is on. New knobs: OTEL_METRICS_INCLUDE_REPOSITORY (adds vcs.* attributes; commit events also get vcs.ref.head.* when OTEL_LOG_TOOL_DETAILS is set), CLAUDE_CODE_GATEWAY_MODEL_DISCOVERY_TIMEOUT_MS (default 3s), CLAUDE_CODE_WORKFLOW_MAX_CONCURRENT_AGENTS (1 to 256). A spinner tip points at /focus. Remote and headless sessions no longer report “waiting for your input” while background agents are still running (CLAUDE_CODE_BG_TASKS_REPORT_RUNNING=0 restores the old status). A deny or ask rule that starts with ! now applies only inside the settings source that wrote it; a bare ! is ignored. /ultrareview --post posts the PR comment itself and prints the link. Cloud-synced skills are named anthropic-skills: plus the skill name. VS Code gets an agent map, a Hooks dialog, a Permission rules dialog, and live subagent progress in Focus view.

2.1.268 (September 10). A Claude apps gateway with pricing: in gateway.yaml pushes the same rates through managed settings so /cost and telemetry match the spend meter. gatewayInternalNetworks lets admins allow /login from the org’s own public IPv4 block. claude self-hosted-runner --remove-session-state (default off) deletes each session’s _sessions/ directories when the session ends. claude auth status --json includes configDirectory. claude plugin install, uninstall, update, enable, and disable accept --json. WebFetch now fails after 300 seconds (CLAUDE_CODE_WEBFETCH_DEADLINE_MS, 0 turns the deadline off). Third-party ANTHROPIC_BASE_URL endpoints no longer fail every turn with HTTP 400 from an Artifact-tool regex added in 2.1.265. /plugin install, enable, and disable take effect when you close the menu. TaskCreate, TaskGet, TaskUpdate, TaskList, and TodoWrite are offered only on Claude 3.x, Opus 4.0 through 4.7, Sonnet 4.0 through 4.6, and Haiku 4.5; set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 on other models. Plain WebFetch deny and ask rules no longer cover Artifact reads. Cowork sessions that skip approvals refuse Artifact reads of a local file outside the session folders or behind a symlink.

2.1.267 (September 9). maxEffortLevel (top-level or per model under modelSettings) caps effort on every provider, including Bedrock, Vertex, and Foundry. --system-prompt-snapshot off re-renders the system prompt on every request. effort: frontmatter on commands, skills, and subagents now applies on Opus 4.7, Opus 4.8, and Fable 5. /model no longer re-sends every tool definition. claude remote-control re-registers when its server credential expires instead of dropping every attached session. --use-anthropic-git-proxy is reported at registration and warns on each session that still clones through the legacy git proxy.

2.1.266 (September 8). Reverts a 2.1.265 regression: CLAUDE_CODE_USE_GATEWAY on its own no longer forces Cloud-gateway sign-in. It is ignored again unless ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN are both set. The note says no configuration change is needed.

2.1.265 (September 8). --plugin-dir can point at a folder of plugins; each child folder with a manifest loads, and children added or removed while running are picked up. Tool results saved to disk cap at 1 GB. /model opusplan[1m] is accepted again. MCP servers configured as http that only speak legacy HTTP+SSE now fall back to SSE. -p, Agent SDK, and cloud sessions keep a cd across user messages. VS Code archives sessions inactive for 14 days by default (“Archive inactive sessions”). Machines with forceLoginGatewayUrl in managed settings start as Claude apps gateway sessions.

Codex CLI

Python SDK 0.154.0 (September 10, tag python-v0.154.0) and CLI 0.154.0 (September 9, tag rust-v0.154.0). I opened both release pages. The releases list shows 0.155.0-alpha.4 as a prerelease. src/data/tool-versions.json records latest as python-v0.154.0 because that non-prerelease tag is newer than rust-v0.154.0.

Python SDK 0.154.0 (September 10). Install with pip install --upgrade openai-codex==0.154.0 (Python 3.10 or later). It bundles openai-codex-cli-bin==0.154.0. New: max and ultra reasoning-effort values; ExternalMessage on sync and async run() and turn() (tool-level authority, not user authorization); include_turns on resume/fork, turn_service_tier for one new turn, and source metadata. History selection changes the returned response, not model context. Migrations on that page: HookMetadata now wraps the handler in .root (hook.root.command, check hook.root.handler_type first); some notifications have typed payloads (read named fields, not .params); late-joining turn handles start from the attach point and can raise TransportClosedError after completion (thread.read(include_turns=True) for saved history). Custom codex_bin overrides need CLI 0.151.0 or newer for ExternalMessage and the new history options.

0.154.0 (September 9). GPT-6-Astra is in the model picker and Amazon Bedrock catalogs. Experimental worktrees: --worktree or /worktree creates an isolated checkout for a new or forked session, then you can browse and resume them. You can answer questions inline with suggested choices or custom text while Codex keeps working, without losing the main draft. Windows sessions can share a background Codex server, with daemon lifecycle commands and managed updates. Vim gets R replace mode with undo and dot-repeat. /copy can copy status output or individual session fields, and copied responses keep formatting in rich-text apps. Existing sessions pick up newly installed plugin tools and refresh skills and hooks after an external plugin upgrade or rollback. Startup no longer runs workspace-controlled helpers before trust is established. The deprecated codex mcp-server entry point is gone.

GitHub Copilot

No version number. I opened each changelog URL below.

Add VS Code Agents to Copilot usage metrics (September 11). Generally available fields on 1-day and 28-day enterprise and organization reports: daily_active_vscode_agent_users, totals_by_vscode_agent (session_count, total_user_messages). User reports add used_vscode_agent and the same per-user totals. These cover the dedicated VS Code Agents window only, not editor-window Agent Mode. Optional fields stay absent or null when the data is missing. Access: enterprise owners, billing managers, organization owners, and custom roles with View Copilot Metrics, with the Copilot usage metrics policy enabled.

Auto-resolution and analysis updates in Copilot code review (September 11). On rereview, Copilot resolves its own comments when a later commit addresses them. Applying a Copilot suggestion now gets a generated commit message instead of the generic autofill. Reviews use the Copilot SDK shell tools behind the agent firewall. Lite effort now runs an ensemble of agents and merges their findings. The page says those analysis changes do not change how you request a review.

Weekly releases, September 7 (September 10). Copilot app: bring Jira issues onto a shared canvas and carry that context into investigation, implementation, and PR prep. Copilot CLI: Project HydraFusion is in /experimental. You pick it like a model; it routes between local, cloud, and compound models. VS Code 1.137: recurring agent tasks (hourly, daily, weekly, or on demand) in public preview; voice mode experimental; issue and PR details in the Agents window experimental. JetBrains sandbox policies are in public preview (same story as the September 8 JetBrains entry).

MAI-Code-1-Flash deprecated (September 10). Dropped the same day across Chat, inline edits, ask/agent modes, and completions. Suggested replacement: MAI-Code-1.1-Flash. Enterprise admins may need to enable that model in policy. The page says you do not have to remove the old one.

Enterprise managed permissions for agent operations (September 9). Copilot Business and Enterprise. Block, require approval, or allow shell commands, file reads and edits, and network domains. Managed restrictions cannot be weakened by user or workspace settings, auto-approval, or saved approvals. Team-specific policies are supported. Generally available in the Copilot app, Copilot CLI, and VS Code sessions that use Agent Host.

Enterprise-managed sandbox in Copilot for JetBrains (September 8). Public preview: admins set sandbox enablement, filesystem and network access, proxy, developer-tool access, and macOS Keychain access. Managed restrictions lock the matching IDE controls. Sandbox settings appear only when the org enables the Editor Preview flag or sets a managed sandbox on/off. Also on that page: cross-file cursor jumps for next-edit suggestions; global files and folders in chat context; enterprise policy diagnostics; /ide in Copilot CLI to attach a terminal session to the JetBrains IDE (public preview); OpenTelemetry settings in “GitHub Copilot – Chat” generally available.

Cursor

No version number. I opened Cursor Projects (dated September 10, 2026 on the changelog index).

Projects is in the left-hand nav. A coordinator agent plans, delegates to implementers, and brings the work back for you to check. The coordinator does not write code itself. The page says it can run as many agents in parallel as the work needs, including “thousands of subagents.”

A Project runs on its own cloud computer. Closing your laptop does not stop it. When something must run on your machine, the coordinator starts a local agent there.

Each Project keeps a shared file set that syncs across the cloud and local machines its agents use. Agents add research, artifacts, and what they learn about the repo and how you like work done.

You can tell the coordinator to watch a Slack channel, run on a schedule, or follow your PRs. The example on the page: point Slack at a bug-report channel and it starts delegating when a bug arrives.

The page says Projects is in beta and rolling out to all users.

Gemini CLI

0.59.0 (September 8). I opened the v0.59.0 release. The releases list shows that tag as the current stable release. Nightly and v0.60.0-preview.0 tags sit above it.

Notes I saw: MCP OAuth metadata discovery and authentication block SSRF; workspace trust is fail-closed and mcpServers are filtered in restricted mode. No new model name appears on that page.

What this means for your setup

  • If you still start Codex with codex mcp-server, 0.154.0 removed that entry point. No learn or playbook page here names that command.
  • If you want isolated Codex checkouts, 0.154.0’s --worktree and /worktree are the new flags. They are marked experimental on the release page.
  • If you consume the Codex Python SDK, 0.154.0 moves hook fields under .root and can return partial events on a late-joined turn handle. Pin openai-codex==0.154.0 only after you change those call sites.
  • Claude Code users who still call TodoWrite or TaskCreate on Opus 4.8, Fable 5.1, or other IDs outside the 2.1.268 list need CLAUDE_CODE_ENABLE_TODO_TOOLS=1. Context is a budget does not name those tools, so I am not asking for a refresh from this.
  • Claude Code plugin authors can run claude plugin eval after 2.1.269. Headless and Remote Control sessions can switch styles with /output-style.
  • A deny or ask rule that starts with ! is now local to the settings file that wrote it (2.1.269). Permissions and unattended runs and Unattended agents in CI still match the --permission-prompts none, --bare, dontAsk, and codex exec --sandbox workspace-write --ask-for-approval never flags I opened this week. Those pages do not document ! rules.
  • Copilot Business and Enterprise admins: enable MAI-Code-1.1-Flash if anyone still picked MAI-Code-1-Flash. Turn on managed permissions if you want shell, file, and network rules that users cannot weaken. Leave HydraFusion in /experimental until you have measured the routing.
  • Cursor users who want a standing coordinator for a feature or migration should open Projects in the left nav. That is a new surface. It does not change a command on the Cursor playbook (that page is still a draft).

Pages we updated this week

None this week.

Sources