Claude Test gains a spec author agent, a draft-writing tool and a stricter explorer

What

Claude Test is a plugin bundled with Claude Code for browser-based testing, started with /claude-test. This release changes it in several ways:

• A new agent, 'Claude Test's background spec author', writes draft test specs into proposed/. It may only use Read, Glob, ToolSearch and claude_test_write.
• claude_test_write is shown as 'Write Claude Test draft', and mcp__plugin_claude-test_browser__claude_test_write is on a pre-approved list, so it runs without asking. A renderer for draft results is also added.
• The explorer agent's instructions now say it has only Read, Grep and Glob, must not read .env files, keys, credentials or ~/.ssh, and must stay inside the project folder. Before, those credential rules sat in an author section, with no separate author agent.
• The run skill now fills in a second sandbox value and a list of background commands, alongside the sandbox value and user settings path it used before.
• New messages tell you to type /claude-test when the browser helper is not ready, such as 'Claude Test did not run this request. Type /claude-test to run the tests.'
• A new helper checks the macOS sandbox network setting allowLocalBinding and reports 'allowed', 'refused' or 'refused-fixed'. It sits beside a ${CLAUDE_TEST_LISTEN} placeholder that nothing else in the code uses yet.

The ToolSearch tool's own description is unchanged in this release.

Why

Claude Test can now draft test specs in the background, and its explorer is kept away from secrets and from files outside your project.