Claude Code releases 2.1.84: PowerShell preview, task hooks, idle-return clearing

• Claude Code 2.1.84 adds an opt-in PowerShell tool for Windows automation, new TaskCreated and HTTP WorktreeCreate hooks, and model-override env vars for pinned third-party backends, according to the 2.1.84 changelog.
• Anthropic also shipped a cluster of runtime fixes in the release thread, including a partial-clone startup bug that could trigger mass blob downloads, MCP cache leak fixes, and prompt-cache improvements that now work with ToolSearch and configured MCP tools.
• The system prompt changed in ways that affect agent behavior: the prompt diff shows stronger guidance to batch git and PR tool calls in parallel, repo-qualified owner/repo#123 references, and the removal of an explicit top-level “Avoid over-engineering” rule.
• Separately, Anthropic published how Claude Code auto mode works in the engineering post, and Team users can now enable it with claude --enable-auto-mode, while an official plugin update says iMessage is now available as a channel.

The biggest functional addition is the changelog's opt-in PowerShell preview for Windows, which exposes PowerShell commands from the CLI for Windows-side automation and scripting; Anthropic also linked dedicated tool docs in the PowerShell reference. The same release adds TaskCreated hooks, HTTP WorktreeCreate hooks that can return hookSpecificOutput.worktreePath, and x-client-request-id request headers for timeout debugging.

For teams running pinned models on Bedrock, Vertex, or Foundry, the release notes add ANTHROPIC_DEFAULT_{OPUS,SONNET,HAIKU}_MODEL_SUPPORTS overrides plus _MODEL_NAME and _DESCRIPTION variables to control capability detection and the /model picker label. The CLI surface diff in the follow-up thread also lists new idle-threshold env vars, allowedChannelPlugins as a managed setting, and a VS Code rate-limit banner with usage percentage and reset time.

A lot of the practical work here is reliability. Anthropic says the changelog fixes a startup path where partial-clone repos could trigger “mass blob downloads,” resolves an MCP tool/resource cache leak on reconnect, and fixes subagents failing with API 400 when both outer and inner sessions used --json-schema. It also says startup is about 30 ms faster, claude "prompt" now renders before MCP servers finish connecting, /stats screenshots are 16× faster, and global system-prompt caching now works when ToolSearch is enabled.

The prompt changes are small on paper but meaningful in harness behavior. The diff summary says Claude no longer has an explicit top-level “Avoid over-engineering” rule, even though narrower instructions remain: no extra features, no unnecessary refactors, and no one-off abstractions. The same update standardizes GitHub references as owner/repo#123, matching a release note that bare #123 links no longer auto-render.

The stronger shift is around parallelism. Screenshots in

show commit and PR playbooks now explicitly marking which bash commands should run in parallel, with language that “multiple tools can run in one response” for “optimal performance.” At the same time, the prompt diff says Claude is no longer explicitly told to “speculatively” fan out extra Glob and Read calls, which should reduce some broad prefetch behavior.

That tradeoff is already visible in user reaction. One practitioner complained in a user report that lazy-loading tools “adds latency and error opportunity,” while Anthropic's own release notes focus on prompt-cache gains and faster startup. The net effect looks less like a capability jump than a recalibration of when Claude should parallelize aggressively and when it should keep context lean.

Anthropic's new engineering write-up explains auto mode as a middle layer between manual approvals and --dangerously-skip-permissions. According to the engineering post, the system uses a server-side prompt-injection probe on tool output plus a transcript classifier that decides whether each action should be allowed or denied before execution.

Auto mode is already rolling out beyond individuals. Anthropic engineer Cat Wu said in the Team rollout post that “almost everyone on our team uses this as a daily driver,” and that Team users can enable it with claude --enable-auto-mode and enter it with Shift+Tab. A practitioner reading the design details called out a “17% false-negative rate on real overeager actions” in a reaction post, alongside an architecture diagram showing that subagents run the same classifier pipeline recursively.

The channel surface is expanding too. An official plugin announcement says iMessage is now available as a channel, and that lands in the same release window as allowedChannelPlugins for team and enterprise admins in the 2.1.84 notes. Put together, 2.1.84 is not just a CLI patch: it widens where Claude can act, adds more hooks around task creation and worktrees, and adds more policy controls for how those actions are exposed.