This Week in Coding Agents: November 2, 2026

Week 45, 2026 · Published November 2, 2026

Claude Code	2.1.316
Codex CLI	0.166.0
Gemini CLI	0.66.0

The short version

Claude Code’s changelog lists 2.1.316 as the top release. npm latest is 2.1.316. There are no 2.1.313 or 2.1.315 headings. 2.1.314 adds permissions.blockWritesOutsideWorkingDirectories. 2.1.316 adds CLAUDE_CODE_DISABLE_ULTRATHINK and CLAUDE_CODE_SUBAGENT_EFFORT. Fast mode on Opus 4.8 is deprecated as of October 30 and is removed on November 30 with no fallback.
Codex CLI 0.166.0 (October 29, tag rust-v0.166.0) is the current stable. GPT-6.1 Terra is in the bundled, Vertex AI, and Amazon Bedrock catalogs. codex queue --resume reloads an unloaded thread after you queue a message. 0.165.0 adds a built-in Vertex provider (model_provider = "gcp") and turns memory v2 on for fresh profiles. The releases list shows 0.167.0-alpha.3 as a prerelease above that stable tag.
Cursor has no version number and no changelog entry in this window. The newest post I opened is Agent SDK from October 20.
GitHub Copilot has no version number. Copilot Fusion is generally available in VS Code, the Copilot app, Copilot CLI, and JetBrains. GPT-6.1 Terra landed on October 26. Assisted permissions are generally available. The Copilot label index lists more October 26 to 28 posts than the fetch captured; I opened the extras that change models or permissions.
Gemini CLI 0.66.0 shipped October 27. PR #29774 adds a banner that the CLI is not optimized for Gemini 4. PR #29750 expands settings placeholders after .env loads.

Claude Code

2.1.311 (October 26) through 2.1.316 (October 31). I opened the live CHANGELOG.md. The first heading is 2.1.316. npm latest is 2.1.316. The changelog jumps 2.1.311 to 2.1.312 to 2.1.314 to 2.1.316.

2.1.316 (October 31). CLAUDE_CODE_DISABLE_ULTRATHINK turns the ultrathink keyword off. CLAUDE_CODE_SUBAGENT_EFFORT sets the effort subagents run at when nothing else names one, instead of inheriting the session’s. CLAUDE_CODE_DISABLE_LOOPED_CROSS_CERT_DROP keeps system cross-certificates that certify each other in a loop, for a private CA that needs one. The Claude apps gateway gains pooled group budgets (pooled_budget on the spend limits admin API), managed.policy_mode: layered so every matching policy applies and later ones merge over earlier ones, and guardrail.input_tagging on Bedrock upstreams (all_messages_and_system also tags the system prompt). Fast mode for Opus 4.8 is deprecated as of October 30, 2026 and is removed on November 30, 2026. After removal, a fast-mode request on Opus 4.8 returns an error and does not fall back to standard Opus 4.8. The changelog says to migrate to Opus 5.5 to keep the speedup. --restricted sessions no longer upload files from outside the working directories when a Read allow rule covered them. A managed autoMode.allow list without "$defaults" no longer gets the built-in exceptions back from a host --settings, a first /permissions rule, or /auto-mode-setup, and it no longer drops the built-in rule on scheduled-task prompts. Server-managed settings for Team and Enterprise logins apply without an approval prompt; a stored copy that does not verify is refused. sonnet, opus, and haiku on Vertex and Bedrock fall back to an older model in that family when the newest is not yet in the account or region. VS Code can sign in to a Claude apps gateway from the chat panel. In VS Code, ultrathink runs that turn at xhigh and shows in a rainbow, matching the terminal.

2.1.314 (October 30). permissions.blockWritesOutsideWorkingDirectories makes Write, Edit, and NotebookEdit refuse paths outside the working directories in every permission mode. Idle compaction lands in VS Code and Claude Desktop’s Code tab: a 200K+ conversation left idle is summarized, and the summary is used on the next message if the prompt cache has expired. Idle compaction also runs for subscribers with telemetry off, and on Bedrock, Vertex, Foundry, and LLM gateways once the one-hour prompt cache is on. claude -p that fails at startup now prints the error and exits 1 (it used to exit 0 with no output). With CI set, claude ultrareview and claude remote-control refuse an untrusted folder at once instead of asking the workspace trust question. /ultrareview now checks workspace trust before it runs git, as claude ultrareview does. Sandbox protections now apply to a worktree entered by claude -p --resume or --continue. Hooks that match Bash also run on shell commands started by the Monitor tool; those hooks can block or ask, not approve or rewrite. A settings file excluded with --setting-sources no longer still supplies auto mode rules and saved permission consents. FORCE_HYPERLINK=false (also no and off) turns hyperlinks off instead of on. sandbox.excludedCommands now means every command in a call must match an entry. An organization’s effort cap no longer fails every turn with “Invalid effort level”: Claude Code lowers the effort, retries, and says which level it uses. Microsoft Foundry falls back to an older model in the family when the resource has no deployment for the default Sonnet, Opus, or Haiku.

2.1.312 (October 27). oidc.user_id_claim lets the Claude apps gateway identify people by an id_token claim other than sub, such as Microsoft Entra ID’s oid. Bedrock, Vertex, Foundry, and Mantle now use the shorter system prompt for model names Claude Code does not recognize (custom deployment names, inference profile ARNs), even when they front Sonnet or Haiku. Set CLAUDE_CODE_SIMPLE_SYSTEM_PROMPT=0 to keep the full prompt. Automatic model switches after a flagged message also cover those unrecognized names; set CLAUDE_CODE_DISABLE_REFUSAL_FALLBACK=1 to turn them off. Usage limit messages now say “5-hour limit” instead of “session limit”. Prompt and agent hooks run at most 24 at a time per session, sub-agents included, plus at most 12 that have waited longer than their event’s timeout. A timeout under 1 second now counts as 1 second. A subagent’s autoCompactWindow can lower the window set by CLAUDE_CODE_AUTO_COMPACT_WINDOW; it still never raises that window. Claude Code stopped sending the claude_code.safety_event OpenTelemetry event that 2.1.307 added last week. Write now asks Claude to read a very large CLAUDE.md, rules file, or @-included file in a subdirectory all the way through, or to edit in place, instead of overwriting after seeing only the start. Per-directory and paths: skills are found after a Bash cd deeper into a git repository. An invalid command name no longer stops every command from loading, which could leave claude -p with no output.

2.1.311 (October 26). --hold-idle-file and --max-hold-min on self-hosted-runner keep an idle session while the host reports processes it left running. ultrathink in the interactive terminal now runs its turn at xhigh effort on Opus 5.5, Sonnet 5.5, Haiku 5.5, and Fable 5.1 through Anthropic’s API; typed while Claude is working, it raises the rest of that turn. The rainbow appears only where the keyword can raise effort, and not when the word is quoted or part of a file path. DISABLE_COMPACT in settings files and in a mod’s $.env.set can only turn compaction off, not back on. On Bedrock, Mantle, Vertex, and Foundry, Claude Code loads model settings from a signed file on downloads.claude.ai; set CLAUDE_CODE_MODEL_CATALOG=off to turn that off. Standard-context sessions that ran over 200K tokens now compact and continue instead of stopping with “Usage credits required for 1M context”. SubagentStop hooks run once when the subagent finishes, not about twice a minute while it runs. Claude Code no longer retries requests the API marked not to retry, which could hold a turn for minutes or, in unattended retry mode (CLAUDE_CODE_RETRY_WATCHDOG), indefinitely. VS Code sessions started while VS Code telemetry is off get DISABLE_TELEMETRY, so they send no usage metrics or error reports and fetch no feature flags. --output-format json and stream-json now write DEL, C1 control, and Unicode format characters inside strings as \uXXXX escapes; parsed values are unchanged.

I also opened the memory docs and the permissions page. Memory docs still say Claude Code reads AGENTS.md on its own when there is no CLAUDE.md or CLAUDE.local.md in the working directory or above it (v2.1.277+; default claude-md-or-agents-md). CLAUDE.local.md counts as a CLAUDE.md and blocks direct AGENTS.md load unless you set Project instructions to claude-md-and-agents-md. Permissions still label default as Manual.

Codex CLI

0.165.0 (October 27) through 0.166.0 (October 29). I opened the rust-v0.166.0 and rust-v0.165.0 release pages. The releases list shows 0.167.0-alpha.3 as a prerelease above the current stable tag rust-v0.166.0. src/data/tool-versions.json records latest as 0.166.0.

0.166.0 (October 29). Keyboard-driven transcript selection with word, line, and page motions, integrated search, and configurable shortcuts. Markdown renders in async questions, choices, answers, and transcript replies. GPT-6.1 Terra is in the bundled, Vertex AI, and Amazon Bedrock model catalogs. codex queue --resume resumes an unloaded thread after you queue a message. Safety-block explanations in the TUI now include selectable safer follow-up prompts. codex doctor --network-only runs configuration, connectivity, and WebSocket diagnostics. Acknowledged approval and permission settings survive resume and fork. Reasoning-token accounting is preserved across resume and fork so compaction does not fire early. Offline /quit and /exit recover from a TUI reconnect.

0.165.0 (October 27). A built-in Google Cloud Vertex AI provider, configured with model_provider = "gcp". GPT-6.1 Terra is in the model catalogs (also backported here). The session picker gains a repository filter, and the agent command center groups directories, which helps when you have linked worktrees. Experimental attached thread forks land on the app-server API. Reasoning effort changes during a conversation are on by default for supported models. Memory v2 is on by default for fresh profiles, and you can import external-agent memories into it. Fresh TUI sessions honor the server’s default approval reviewer when you did not pass an override. Queued agent messages no longer restart work after an interrupt or a switch to Plan mode. A provider “model not found” error is terminal; Codex stops retrying and keeps the message. Marketplace URL validation now rejects a backslash in HTTP(S) userinfo that could point Git at a different host than the one checked against allowed patterns.

GitHub Copilot

No version number. I opened each changelog URL below. The Copilot label index lists the ten October 28 posts the fetch captured, plus more October 26 to 28 posts. I opened the extras that change models, permissions, or auth.

GPT-6.1 Terra in GitHub Copilot (October 26). Generally available on Copilot Pro, Pro+, Max, Business, and Enterprise. The page calls it everyday agentic coding and terminal work at a lower cost than GPT-6.1 Sol, billed at provider list pricing. Rollout is gradual. Business and Enterprise admins manage access through model policy. Under default model enablement, new models are on unless an admin turned off the global default or disabled this model.

Assisted permissions generally available (October 26). In Copilot CLI, the Copilot app, and VS Code. A model classifies a permission request; clearly safe ones proceed, the rest still ask. CLI: /permissions, then Assisted. App: mode picker, Tool permissions, Assisted. VS Code: permissions picker, Assisted permissions. Enterprise managed settings can block Allow All while keeping Assisted, disable Assisted, and set the default mode for new sessions. If the default is unavailable or prohibited, the session starts with manual approval.

Server-to-server authentication (October 27). Public preview. A GitHub App with the read-only Copilot Requests organization permission can pass a short-lived installation access token to the Copilot SDK. Usage is billed to the organization that owns the installation. Model policies, budgets, and spending controls apply. AI usage CSV reports get an app_client_id field.

Route and orchestrate models with GitHub Copilot Fusion (October 28). Generally available in VS Code, the Copilot app, Copilot CLI, and JetBrains. Pick Copilot Fusion in the model picker, then an effort level: Standard, High, or Max. Fusion only uses models enabled in enterprise and organization model policy. Admins can disable a model for Fusion without disabling it for everyone, and can set Fusion at any effort as the default under enterprise managed settings. You are billed for each model Fusion uses, at that model’s rate. Paid plans get a 30% discount on those model costs for a limited time. Research preview auto, HydraFusion, and HydraFusion Max selections move to the matching Fusion effort.

Fusion local models (October 28). Experimental preview. Fusion Standard can choose between supported on-device models and cloud models in Copilot CLI, the Copilot app, and VS Code. Windows: managed download of MAI Code 1.1 Flash through Windows ML. CLI: experimental mode, then /model. App: Settings > Model providers, then Local model auto routing under Settings > Experiments. VS Code: chat.copilot.localModel.enabled. Bring-your-own MAI Code 1.1 Flash or Qwen3.8 Flash Next through a local OpenAI-compatible server also works on Windows, macOS, and Linux in the CLI. Local inference is not offline: routing decisions still run in the cloud.

Copilot app, Universe releases (October 28). Copilot Projects are in public preview: one home for work that spans repositories, with agents, context, and results together. Computer use is in public preview on macOS and Windows (Settings > Computer Use). Image generation and edit from chat is in the same post.

Generate and edit images (October 28). Public preview on Pro, Pro+, Business, and Enterprise. Your selected reasoning model calls a GPT Image 2.5 Flare tool, previews inline, and saves to the workspace. Billed on image tokens against AI credits, separate from the chat model. Org and enterprise owners can turn it off by disabling the image generation model in model policy. A visible “Made with AI” watermark is opt-in and off by default.

GitHub intelligence (October 28). Copilot Memory is generally available on paid plans across the app, CLI, VS Code, cloud agent, code review, Slack, Teams, and GitHub Mobile. Dreaming is public preview and opt-in: personal memory curation, repository customization (skills and instruction files as PRs), and code review customization. Dream runs consume AI credits. Code search can look past the checkout, including /research.

Shared Copilot Automations (October 28). Rolling out on Pro, Pro+, Max, Business, and Enterprise. Create an automation as Shared on github.com, the Copilot app, or VS Code. You cannot change the type later. Shared automations are visible to anyone with read access, editable with write access, billed to the repository owner, and attributed to Copilot. Organization repos need the “Allow use of Copilot cloud agent billed to the organization” policy.

Repository plugin defaults and enterprise plugin controls (October 28). Generally available in VS Code, Copilot CLI, and the Copilot app. A trusted repo can install and enable default plugins. Enterprises can require plugins, block plugins, and restrict install to approved marketplaces. Managed settings win over repo and personal choices.

GitHub Copilot in VS Code, Universe releases (October 28). Cloud agent sessions continue across VS Code, the Copilot app, GitHub Mobile, github.com, and Copilot CLI. Agent Host isolates sessions from a VS Code window. code agent host --github-environment lets GitHub Mobile steer an agent on your machine. The Copilot harness (Microsoft Copilot SDK) is the default for new sessions. Install Copilot CLI through the VS Code installer on Windows, when adding VS Code to PATH on macOS, or by running copilot on Linux. Computer use in VS Code: chat.agentHost.computerUse.enabled, then /computer on. Customization Marketplace is in preview. Connectors (Slack, Notion, Microsoft Learn) are in preview and skip a local MCP install.

Security extension for the GitHub Copilot app (October 28). Public preview. copilot plugin install github-security@awesome-copilot, or Customize in the app and search github-security. Browse and fix code scanning alerts, and run /security-review on local changes. GitHub Enterprise Server is not supported.

GitHub Enterprise now includes access to GitHub Copilot (October 28). Enterprises working with a GitHub or Microsoft account team can include Copilot with GitHub Enterprise Cloud for every licensed user. Self-serve is not open yet. Copilot Business and Copilot Enterprise remain for sale. Users without a GitHub Enterprise license get Copilot Standalone at $10 per user per month. There is no included usage allowance. Completions keep working with no budget; chat, code review, and Copilot in the cloud need budget.

GitHub Agentic Workflows generally available (October 28). Open source from GitHub Next. Markdown workflows compile to Actions YAML. Default engine is Copilot CLI and needs a Copilot license. gh aw forecast is no longer experimental. gh aw add-wizard adds templates.

Gemini CLI

0.66.0 (October 27). I opened the v0.66.0 release via gh release view. The releases list shows that tag as the current stable release. Nightly v0.67.0-* tags and v0.67.0-preview.0 sit above it.

PR #29774 adds a banner that Gemini CLI is not optimized for Gemini 4. PR #29750 resolves settings placeholders after .env files load, so a value such as "Authorization": "Bearer $GITHUB_MCP_PAT" expands instead of going to an MCP server as a literal. PR #29844 adds oauth.allowPrivateNetwork on an MCP server so OAuth can reach RFC 1918, RFC 6598, RFC 4193, .internal, and .local hosts when you set it. Off by default. PR #29832 stops forcing GIT_CONFIG_GLOBAL to /dev/null, so ~/.gitconfig (user.name, user.email, signing) applies again. System gitconfig stays disabled. PR #29760 sets GIT_ASKPASS, SSH_ASKPASS, GIT_SSH_COMMAND=ssh -o BatchMode=yes, GH_PROMPT_DISABLED, and GCM_INTERACTIVE=never so a background git fetch cannot freeze the TUI on a credential or SSH prompt. PR #29769 loads SKILL.md files with a single-line description, a BOM, or trailing space on the --- fence. PR #29773 accepts a "keys" array in keybindings.json so an override that unbinds several defaults is not discarded. PR #29757 checks that the active binary’s version actually moved after an auto-update, so a shadowed $PATH install no longer reports success and loops the update prompt.

What this means for your setup

Install Claude Code 2.1.316. If you pin @anthropic-ai/claude-code, bump the pin.
Set permissions.blockWritesOutsideWorkingDirectories if you want Write, Edit, and NotebookEdit refused outside the working directories in every permission mode, including bypass.
Set CLAUDE_CODE_DISABLE_ULTRATHINK=1 if you need the keyword off. Set CLAUDE_CODE_SUBAGENT_EFFORT if subagents should not inherit the session’s effort.
Leave Opus 4.8 fast mode. It is deprecated as of October 30 and is removed on November 30 with no fallback. Move that traffic to Opus 5.5 if you still want the speedup.
claude -p that fails at startup now exits 1. CI that treated a zero exit as “started” will start failing for real startup errors.
CI=true makes claude ultrareview and claude remote-control refuse an untrusted folder at once. /ultrareview now checks workspace trust the same way. Unattended agents in CI already passes dontAsk and does not mention this trust refuse.
If you started scraping claude_code.safety_event after 2.1.307, stop. 2.1.312 no longer sends it.
Idle compaction now runs in VS Code and the Desktop Code tab, and on telemetry-off, Bedrock, Vertex, Foundry, and gateway sessions once the one-hour prompt cache is on. The changelog line I opened for 2.1.290 still says "idleCompaction": false in settings.json turns it off.
Permissions and unattended runs already starts -p in Manual on 2.1.302+ and already passes dontAsk. It does not name permissions.blockWritesOutsideWorkingDirectories.
Claude Code already reads AGENTS.md when no CLAUDE.md or CLAUDE.local.md sits in the working directory or above it. One AGENTS.md, every tool and New repo agent setup already say that. Context is a budget still describes only CLAUDE.md load at session start. Instruction files is still a draft that only describes the bridge.
Codex 0.166.0: use codex queue --resume after a queued message on an unloaded thread. Use codex doctor --network-only when login or WebSocket fails. Fresh profiles get memory v2. Vertex users can set model_provider = "gcp" instead of a custom provider block. GPT-6.1 Terra is in the bundled catalog.
Copilot: pick Fusion in the model picker if you want routing instead of a named model. Enable GPT-6.1 Terra in model policy if the global default is off. Try /permissions Assisted in the CLI. The November 16 drop of Opus 4.8, Opus 5, and GPT-5.6 Terra is still ahead. Annual individual subscribers keep Opus 4.8. Teams users on the legacy GitHub app still need the consolidated app before November 16.
Multi-model routing still prices Haiku 4.5, Sonnet 5, Opus 5, GPT-5.6 Terra, GPT-5.6 Sol, and GPT-5.6 Luna, and still names GPT-5 Mini as the successor to o4-mini.
Gemini CLI 0.66.0: placeholders in settings.json now expand from .env. Set oauth.allowPrivateNetwork on an internal MCP server if OAuth discovery was failing on a private hostname. ~/.gitconfig applies again for user.name and user.email. A "keys" array in keybindings.json is valid.

Pages we updated this week

None this week.