Moonshot routed requests through Claude

Highlight

New claude plugin eval command scores plugin behavior

You can now run claude plugin eval to execute a plugin's eval suite against Claude Code and get reproducible, scored results as JSON and an HTML report. If you build or maintain plugins, this gives you a repeatable way to catch regressions before shipping; see claude plugin eval --help for the suite format.

Worth Knowing

Switch output styles from anywhere with /output-style

The new /output-style [name] command lists and switches output styles from any session, including Remote Control, cloud, and other headless contexts. You no longer need a local terminal to change how Claude formats its replies mid-session.

Bash tool results now include file diffs

When the Bash tool edits files, its results now include a diff of what changed, controlled by the bashEditDiffEnabled setting. This makes it easier to review side effects of shell commands without a separate diff step.

New effort cap and system-prompt snapshot controls

You can now set maxEffortLevel at the top level or per model under modelSettings to cap effort across every provider, including Bedrock, Vertex, and Foundry, while still letting users pick lower levels. A new --system-prompt-snapshot off flag renders the system prompt fresh on each request, which is useful when you're iterating on prompt text.

Plugin changes apply without reloading

Installing, enabling, or disabling a plugin from /plugin now takes effect as soon as you close the menu. You no longer need to run /reload-plugins afterward, and in VS Code the Manage plugins dialog applies changes to open sessions without a restart.

VS Code gets agent map, hooks, and permission dialogs

The VS Code extension adds an agent map behind the footer's "N agents" pill with per-agent cards and read-only transcripts, plus new Hooks and Permission rules dialogs for viewing and editing entries across user, project, and local settings. Focus view also shows live progress rows for running subagents under each tool-call group.

Permission-rule scoping and symlink handling tightened

Deny and ask rules starting with ! now apply only within their own settings source, and several path-scoping gaps closed: rules on symlinked directories like /etc, /tmp, and /var now apply when a path is given by its real location, Bash(tee:*) allows no longer cover destinations outside your working directories, and deny rules apply through env -C and eval wrappers. Review any settings that relied on the old behavior.

Under the Hood

Prompt caching holds across model switches and resumes

A large batch of fixes stops prompt-cache misses when you switch models with /model, resume sessions where MCP connectors reconnect at different moments, spawn subagents, or use --system-prompt/--append-system-prompt. Extended thinking is also preserved when tool lists change mid-session on supported models.

Regression fixes for git prompts, third-party endpoints, and CPU pinning

Read-only git commands no longer start asking for Bash permission after a session has been running for a while, third-party endpoints via ANTHROPIC_BASE_URL no longer fail every turn with HTTP 400 (a regex in the Artifact tool schema those endpoints rejected), and a busy loop that pinned a CPU core in long-running idle sessions is gone.

From Anthropic

In the News