Between 7 and 13 November, coding agents got new modes of operation, and some changes affect compatibility and automation. Imagine Semen, administrator of the fictional “Pigsty No. 42” lab: he’s updating agents and needs to check not just the new buttons, but also how the changes behave in CI scenarios.

The most notable new features are experimental Code Mode in Claude Code, the Qwen Code 1.0 preview with native packages, and model provider selection in Codex CLI. Meanwhile, a change in behavior opencode run may leave an automated process waiting indefinitely. Oink, release notes matter more than the urge to “update everything.”

Claude Code: JavaScript instead of separate tool calls

Claude Code 2.1.323 introduces experimental Code Mode. Instead of calling tools one at a time, Claude generates JavaScript that carries out those calls. The mode has to be enabled manually. Anthropic refers to documentation on configuration and the option to disable the feature for an organization, but the linked page is unavailable at the time of writing. So the exact rules and limitations are still unclear.

For Semen, this means the new mode shouldn’t be treated as a drop-in replacement for the usual way of interacting: first, he needs to figure out how it’s configured in his particular environment. This is an experimental feature, not grounds for assuming unknown properties. OpenCode is also developing Code Mode: version 2.2.6 introduced an experimental virtual machine for bytecode.

Another practical Claude Code change concerns third-party models. For an unknown model identifier, the program now assumes a context window of 1 million tokens instead of 200 thousand. If the connected model actually has a smaller window, the session may fail on every subsequent turn after the window fills up. Anthropic recommends setting CLAUDE_CODE_DISABLE_1M_CONTEXT=1 or configuring early compaction via /autocompact.

In the same version, progress summaries for background subagents can be disabled with the variable CLAUDE_CODE_DISABLE_AGENT_PROGRESS_SUMMARIES=1. It removes both the summaries themselves and the requests that generate them; according to the notes, the variable has been supported since version 2.1.318. Also, plugin.json adds the field engines["claude-code"], which defines a compatible version range: an incompatible plugin won’t be installed, updated, or loaded.

There are other integration details too: the first-version MCP client has been removed, and MCP_SDK_GENERATION=v1 runs the v2 client in compatibility mode. The subagentStatusLine setting has been removed, although old files with this key continue to load. In Claude Code 2.1.321, the default Fable model became Claude Fable 5.2; a saved Fable 5.1 selection carries over to fable for Pro and Max plans, and the previous model can be restored with the command /model claude-fable-5-1.

Qwen Code and Codex: native installation and provider selection

Alibaba’s Qwen Code 1.0 Preview brings together Qoder and Qwen Code agent capabilities for coding, debugging, implementation, and testing. It supports Qwen and Qoder accounts, multiple model providers, and compatible endpoints; session resumption and non-interactive script execution are also available.

An important installation change is native packages for macOS, Linux, and Windows. There’s no longer any need to install Node.js separately: a package for the required operating system and architecture is available among the release Assets. In preview.1, they fixed an issue that prevented Qwen accounts with a subscription from using subscription models after signing in, and added parallel tool calls for Qoder models. This is a preview version, so interfaces, options, and behavior may change; the stable 0.27.x branch continues to receive updates.

Codex CLI 0.168.0 shows model providers in the TUI, under Sessions → Model providers. There you can set the default provider for new conversations, but already open and resumed chats keep their provider. Additional features are enabled with the options config.toml: features.tui_workspace opens a side conversation alongside the main one, while tui.notification_method = "osc777" enables terminal notifications.

A separate option features.image_generation_animation is for creating and editing animated GIFs, with settings for frame count and duration. A message can also include an audio recording if the model supports audio. And if the configuration explicitly sets features.multi_agent, Codex warns about it and suggests agents.enabled to disable subagent tools.

OpenCode in CI: waiting without a timer

In OpenCode 2.2.6, the command opencode run waits for background commands, subagents, and image generation launched by the prompt to finish, then lets the model respond based on their results. There’s no built-in time limit. If the prompt starts a long-running process such as a development server or file watcher, execution won’t finish on its own. For CI, Semen needs an external timeout; termination via SIGTERM cancels the session and closes the process with code 143.

For running in a separate Git worktree, there are -w or --worktree=<name>; цей режим не поєднується з --continue and --session. The --archived option removes script sessions from the input and lists. MCP servers can be combined into a shared pool with pool: true, if the projects have the same settings: they then use a single connection. Version 2.2.7 added Auto mode for GitHub Copilot, which selects a model and falls back to the account’s default chat model if it fails.

Kimi Code, DeepSeek Harness, and other changes

Kimi Code 2.6.0 no longer prohibits multiple Edit calls for the same file or requires rereading the file before each edit. The full-screen transcript now groups runs of tool calls and reasoning into collapsed rows; the previous view can be restored via the options tool_call_density = "ungrouped" і thinking = "ungrouped" у tui.toml. The Server API changed how the agent profile is passed: it’s now specified in agent_config.profile when creating a session, and the old transcript endpoints have been replaced by a turns stream over WebSocket. Version 2.6.1 fixes slow loading of long sessions and hangs with empty responses during context compaction.

DeepSeek Harness—an open agent toolkit with a web interface and desktop application—remains in developer preview. 0.2.2-rc.2 added Git worktree change browsing by branch, commit, and file type, GitHub review support, and nested subagents. AgentTeam lets you create a separate chat with shared participants and a task panel. rc.3 introduced experimental chat import from the DeepSeek app and web version. Both builds are published on npm with the tag next, while latest points to rc.1.

Among other notable releases, Crush 0.101.0 lets you manage skills through the command palette, and in Yolo mode blocks dangerous commands without asking an extra question. Pi 1.5.1 added sign-in through pi auth login <provider> without starting a session and changed how requests are sent to OpenAI-compatible models: an empty "tools": [] field is no longer sent. If a proxy requires this field in the tool-call history, set compat.requiresToolsWithToolHistory: true.

Goose shows context window usage, and you can ask a completed subagent a follow-up question. Zed fixed auto-compaction in long turns, which might not get a chance to run again. Ollama 0.40.16 launches Ripple Preview with the command ollama run ripple-preview and claims roughly 27 GiB lower memory usage for Qwen3.8 Flash Next on Apple silicon. Gemini CLI 0.68.0 also added, among other things, a countdown after error 429.

What to check before updating

For Semen, the order is simple: check the risks first, then enable the new features. If Claude Code is using a third-party model, he needs to reconcile its actual context window with the new assumption or set CLAUDE_CODE_DISABLE_1M_CONTEXT=1. CI scenarios using opencode run should run with an external timeout when background tasks may take a long time.

Integrations with the Kimi Code Server API and the package @opencode/ai need interface-change checks, while a proxy in front of Pi needs its requirements for the toolsfield checked. Qwen Code 1.0 and DeepSeek Harness are labeled as preview or developer preview, so their settings may change. In this technical pigsty, new features are tempting, but check compatibility and the completion of automated tasks separately: that’s where the nastiest surprises hide, oink.