This Week in Coding Agents: October 19, 2026

Week 43, 2026 · Published October 19, 2026

Claude Code	2.1.303
Codex CLI	0.163.0
Gemini CLI	0.64.0

The short version

Claude Code’s changelog lists 2.1.303 as the top release. npm latest is 2.1.303 (October 18). 2.1.302 changes claude -p and Agent SDK sessions with no permission mode to start in default (Manual) on third-party providers, with telemetry off, or on a fresh install. --permission-mode auto keeps auto. Idle compaction landed in 2.1.299: a 200K+ token terminal conversation left idle about 54 minutes is compacted before the one-hour prompt cache expires.
Codex CLI 0.163.0 (October 15, tag rust-v0.163.0) is the current stable. Daybreak is on by default for eligible accounts. Last week it was opt-in. You can send a follow-up while Codex is still answering, and /config now searches settings across categories. The releases list shows 0.164.0-alpha.4 as a prerelease above that stable tag.
Cursor has no version number. Cloud Agent Canvases and Environment Improvements (October 16): cloud agents can build interactive canvases and share a finished one by link. The Environments API is in beta. The same post says environments hold up to 300 repositories.
GitHub Copilot has no version number. Gemini 4 Argon is generally available to Pro+, Max, Business, and Enterprise. Three models leave on November 16: Claude Opus 4.8 (annual individual plans keep it), Claude Opus 5, and GPT-5.6 Terra.
Gemini CLI 0.64.0 shipped October 13. The notes I read are mostly fixes. PR #29528 stops headless mode from reporting an untrusted folder as trusted.

Claude Code

2.1.299 (October 14) through 2.1.303 (October 18). I opened the live CHANGELOG.md. The first heading is 2.1.303. There is no 2.1.297 or 2.1.298 heading. npm publish times match those dates.

2.1.303 (October 18). The changelog line is “Bug fixes and reliability improvements”. I am not filling that in from memory.

2.1.302 (October 16). CLAUDE_CODE_AUTO_MODE_MODEL picks the model for the client-side auto mode classifier. It does not change server-side auto mode. /config gets an Idle compaction row; notices after an idle compaction say how to turn it off. /design-sync and /design-login warn that standalone Claude Design (claude.ai/design) closes by December 14, 2026. claude -p and Agent SDK sessions with no permission mode start in default on third-party providers, with telemetry off, or on a fresh install. --permission-mode auto keeps auto. Fork subagents (subagent_type: "fork") are on by default in claude -p, the Agent SDK, the desktop app, and VS Code, as they already are in the terminal. CLAUDE_CODE_FORK_SUBAGENT=0 turns them off; claude mcp serve never offers them. Desktop and VS Code always run the main conversation’s subagents in the background; a subagent’s own Agent calls in -p, the SDK, desktop, and VS Code now run in the foreground by default. At most 12 prompt or agent hooks of one event run at a time. Image sources over 128 MB are refused instead of being loaded into memory. MCP servers that negotiate protocol 2026-07-28 but omit resultType, ttlMs, or cacheScope no longer show empty tool lists. Windows: writes on exFAT and FAT32 are no longer refused or prompted for every file.

2.1.301 (October 15). CLAUDE_CODE_RETRY_WATCHDOG_MAX_USAGE_LIMIT_WAIT_MS stops unattended retry mode (CLAUDE_CODE_RETRY_WATCHDOG) from waiting out a usage limit that resets too far away. The fable (and best) alias retries once on the previous Fable model when the newest one is not available to your organization. OpenTelemetry tool_decision events gain classifier_verdict and permission_mode, and classifier_category when OTEL_LOG_TOOL_DETAILS=1. Switching to a smaller context window now summarizes with the previous model first. Workflow agents no longer skip a step and still report success after reading the harness note that their task text carries no user authority.

2.1.300 (October 14). /advisor lands on Claude Platform on AWS. No advisor runs until you pick one with /advisor, advisorModel, or --advisor. CLAUDE_CODE_DISABLE_ADVISOR_TOOL=1 turns it off. --sandbox and --no-sandbox on claude remote-control now exit with an error (they did nothing). Set sandbox.enabled in settings instead. Large requests to api.anthropic.com use gzip even with telemetry off; CLAUDE_CODE_GZIP_REQUEST_BODIES=0 turns that off. updatedInput from PreToolUse and PermissionRequest hooks is validated like Claude’s own input: unknown keys drop, and input the tool rejects denies the call. PermissionRequest setMode updates in Cowork cloud sessions are ignored. Session transcripts also record each API request’s system prompt, tools, and parameters.

2.1.299 (October 14). Idle compaction in the terminal: a 200K+ token conversation left idle about 54 minutes is compacted before the one-hour prompt cache expires. CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS changes the 200K minimum. "idleCompaction": false in settings.json turns it off. Both need v2.1.290 or later. CLAUDE_CODE_DISABLE_FILE_MENTIONS sends Claude the paths of @ file mentions instead of the file contents. MCP allow rules now reach only the server they name: mcp__a no longer allows a server named a__b or a_. Write mcp__<server>__* with the whole normalized name. Agent tools: lists follow the same rule. UserPromptSubmit hooks run before a skill or custom slash command expands, so a hook that blocks the prompt also stops its embedded shell. Plugin hooks that cannot be handed an action, or cannot hand back an answer, now refuse the action instead of letting it through. Empty .mcp.json is no longer reported as invalid. A sandboxed shell can no longer write into a --plugin-dir folder whose code Claude Code reloads and runs outside the sandbox.

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 and still say to set permissions.defaultMode if you do not want the built-in start mode.

Codex CLI

0.163.0 (October 15). I opened the rust-v0.163.0 release page. The releases list shows 0.164.0-alpha.4 as a prerelease above the current stable tag rust-v0.163.0. src/data/tool-versions.json records latest as 0.163.0.

0.163.0 (October 15). /config searches preferences across categories, applies supported settings immediately, and marks which changes need a restart. You can send a follow-up while Codex is still responding; interrupting a turn also stops active Code Mode cells and nested tool calls. /keymap gets a leader key (default Ctrl+X); shortcut hints stay open until you choose or cancel, and voice mute moves to F9. /daemon can restart a managed background server after a warning, then reconnects. Independent read-only tools can run in parallel, including skill, memory, and conversation-history lookups. Daybreak is enabled by default for eligible accounts. codex exec can also select it with an OpenAI API key when the server offers access.

Resuming a saved thread restores its permissions and network approvals, still honoring explicit overrides and current managed policy. The network proxy checks permission before a DNS lookup, rejects duplicate Host headers, and stops credential aliases from bypassing inspection hooks. /side returns you to an existing side conversation after you switch threads. Codex no longer bundles a patched zsh; it uses the standard shell executor, and configs that still set the retired shell flags get a migration notice. Bundled OpenAI Docs guidance now covers migrations to the GPT-6 family.

Cursor

No version number. I opened the changelog, both posts from this week, the Canvases docs, and the Cloud Agents API page.

Cloud Agent Canvases and Environment Improvements (October 16). Cloud agents can build interactive canvases as part of their work. Share a finished canvas by sending its link. The Canvases docs say to use Publish on the canvas toolbar, then copy the link. Anyone on your team opens it read-only in the browser. Shared canvases need a paid plan (Premium and higher, Teams, Enterprise, and legacy Pro, Pro+, and Ultra), team membership, and a privacy mode that allows data storage. Legacy Privacy Mode blocks sharing. Free accounts cannot create shares. Team admins can turn shared canvases off under team settings > Shared Canvases.

The Environments API is in beta: create and manage the cloud environments agents run in. The changelog says environments now support up to 300 repositories and that large environments load 5x faster. The API page I opened lists a 500 repository maximum on POST /v1/environments (repos array) and on launching an agent with a repos list.

Origin App Marketplace (October 13). Origin lists more than 20 apps across deployments (Cloudflare, Netlify, Render, Spacelift, Vercel), CI (Buildkite, CircleCI, CloudBees, Depot, EngFlow, Namespace, RWX), code review (CodeRabbit, Greptile, Tenki), security (Artemis Security, Semgrep, Socket Security, Upwind), monitoring (Sentry), and project management (Linear). The marketplace is open to public submissions: create an app in your namespace, build against the Origin API (app JWT plus installation access tokens, webhooks, check runs, PR comments, reviews), make it public, then submit a slug, tags, and a demo video of install, uninstall, and core flows.

GitHub Copilot

No version number. I opened each changelog URL below. The Copilot label index lists the same seven posts for this week.

Multi-repository support for GitHub in Slack and Teams (October 15). @GitHub in Slack and Teams can work across multiple repositories in one conversation: create issues, open PRs, list your open PRs, or investigate using frontend and backend repos without switching threads. You can mention @GitHub anywhere in the message, not only at the start. Slack shared messages and threads can be used as context.

Upcoming deprecation of selected models (October 15). On November 16, 2026, these leave every Copilot surface (chat, inline edits, ask and agent modes, completions): Claude Opus 4.8, Claude Opus 5, and GPT-5.6 Terra. Suggested replacements: Claude Opus 5.5, GPT-6.1 Sol or GPT-6 Luna. Claude Opus 4.8 stays for individual Copilot subscribers on annual plans. Under default model enablement, the replacements are on for Business and Enterprise unless an admin turned off the global default or disabled that model.

Agentic autofix for code scanning alerts, general availability (October 15). GitHub Code Security and GitHub Advanced Security on GitHub Enterprise Cloud, including data residency. Assign a default-branch alert to Copilot; it explores, proposes a fix, reruns CodeQL, iterates, then opens a draft PR. You can assign up to 25 alerts together, or set assignees to ["copilot-swe-agent[bot]"] on the Update a Code Scanning Alert REST API. Alerts that exist only on PRs or other branches cannot be assigned. An enterprise admin must enable the “Agentic autofix” Advanced Security policy, then an org or repo admin turns it on in Code Security settings. It consumes AI Credits from the Code Security budget and GitHub Actions minutes.

HydraFusion updates and HydraFusion Max (October 14). Still research preview in VS Code, the Copilot app, and Copilot CLI. The pool now includes Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna. Cascade can add a judging step before repairing. Critique inspects the workspace with tools and runs parallel critic models. HydraFusion Max uses the same architecture with higher reasoning effort. Available on Pro, Pro+, Business, and Enterprise with experimental or preview features enabled, subject to model policy.

Gemini 4 Argon (October 13). Generally available to Pro+, Max, Business, and Enterprise. Picker coverage: VS Code, Visual Studio, Copilot CLI, coding agent, the Copilot app, JetBrains, Xcode, Eclipse. Provider list pricing under usage-based billing. Rollout is gradual. Business and Enterprise admins manage it through the model policy; new models stay enabled unless the global default is off or this model is disabled.

Enterprise managed settings enforcement (October 13). forceRemoteSettingsRefresh: true makes Copilot CLI, VS Code, and the Copilot app fetch current server-managed settings at startup instead of using a cached policy. If the client cannot confirm current policy, it fails closed. The sandbox key now enforces a minimum client version: pinning an older client that includes this key forces an update before Copilot runs.

Client telemetry user attribution (October 12). telemetry.capture.identity adds user.name (GitHub login), process.user.name, and host.name to OpenTelemetry data. Identity capture is off by default. Five content classes are also off by default and independent: prompts, responses, toolArguments, toolOutput, policyDetail. Fine-grained capture is in Copilot CLI, the Copilot app, and VS Code Agent Host. Managed values win over local settings, SDK config, and environment variables. captureContent remains a shorthand for the five content classes and never turns on identity.

Gemini CLI

0.64.0 (October 13). I opened the v0.64.0 release. The releases list shows that tag as the current stable release. Nightly v0.66.0-* tags and v0.65.0-preview.0 sit above it.

The notes I read are mostly fixes: atomic settings persistence with backup recovery, Ctrl+C reaching the cancellation handler during an active operation, @file:line resolution, Windows ConPTY IME cursor forwarding, and Enter/Space confirming selection lists. PR #29528 makes headless folder trust report the resolved state (true, false, or undefined) instead of always telling the parent the folder is trusted. PR #29499 serializes file tool operations per path and writes files atomically, so parallel sub-agents stop losing updates on the same file. PR #29450 migrates a2a-server settings from flat V1 keys to nested V2 in memory without rewriting settings.json on disk.

What this means for your setup

If your claude -p or Agent SDK jobs on Bedrock, Vertex, Foundry, a gateway, telemetry-off, or a fresh install relied on last week’s auto start, pass --permission-mode auto after 2.1.302. With no flag they start in default (Manual). Keep --permission-mode dontAsk on CI. Permissions and unattended runs already says -p starts in Manual and already passes dontAsk. Unattended agents in CI already passes dontAsk.
Idle compaction will rewrite a long terminal session left sitting near the one-hour cache boundary. Turn it off with "idleCompaction": false in settings.json or the /config row if you do not want that. CLAUDE_CODE_IDLE_COMPACT_MIN_TOKENS raises the 200K floor.
CLAUDE_CODE_DISABLE_FILE_MENTIONS is the switch if @ mentions are blowing the context window. Claude still gets the paths and can Read them.
Fork subagents are now on in -p and the SDK. Set CLAUDE_CODE_FORK_SUBAGENT=0 if a CI job should not spawn them.
claude remote-control --sandbox and --no-sandbox now exit. Put sandbox.enabled in settings.
MCP allow rules of the form mcp__a no longer cover a server whose normalized name starts with a__ or a_. Rewrite them with the full server name.
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, New repo agent setup, and Context is a budget still describe a required @AGENTS.md bridge. A CLAUDE.local.md you add for personal notes will hide AGENTS.md unless you set claude-md-and-agents-md.
Codex Daybreak is on by default for eligible accounts in 0.163.0. If you turned it off last week because cli_daybreak had to be explicit, check /config and your feature flags. codex exec can select it with an API key when the server offers access.
Cursor canvases from cloud agents: Publish from the toolbar, send the link. Team admins who do not want shareable canvases turn them off in team settings > Shared Canvases. The Environments API is the way to create those environments from a script; the changelog’s 300-repo cap is the product claim I read, while POST /v1/environments currently documents a 500 max.
Copilot Business and Enterprise admins: enable Gemini 4 Argon if your model policy does not pick up new models. Move workflows off Opus 4.8, Opus 5, and GPT-5.6 Terra before November 16. Turn on agentic autofix only after the enterprise Advanced Security policy is set; it bills Code Security AI Credits. Set forceRemoteSettingsRefresh if a cached policy is not acceptable. Leave telemetry.capture.identity and the content classes off unless an audit destination is approved.
Multi-model routing still prices Haiku 4.5 at $1/$5 and Sonnet 5, and does not name Haiku 5.5, Sonnet 5.5, or Gemini 4 Argon.
Gemini CLI 0.64.0 headless runs now honor an untrusted folder instead of pretending it is trusted. Atomic writes are the default for the file tools.

Pages we updated this week

None this week.