WHY THIS MATTERS
Anthropic released Claude Code 2.1.260 with a package of changes that combines greater autonomy for headless sessions and agents with important fixes to the Bash sandbox, file permissions, and integrations. For Brazilian teams, the most valuable gain is not a flashy feature, but fewer silent risks in automations that already run with limited supervision.<p><strong>CONFIRMED.</strong> Anthropic published Claude Code 2.1.260 on September 3, 2026, in a release focused on the two qualities that determine whether a development agent is useful in production: autonomy and control. The update adds ways to work with fewer interruptions, while fixing permission paths that could be ignored or applied incorrectly.</p><h2>More context, less interruption</h2><p>The release adds a diff panel beside the conversation in fullscreen mode, showing uncommitted changes as Claude edits a project. The <code>/diff</code> command makes inspection part of the workflow instead of forcing users to switch between terminal and editor. The <code>/cost</code> command and session status also now show likely causes for prompt cache misses, a small detail that helps explain why a task became slower or more expensive.</p><p>Headless sessions and the SDK gain <code>/reload-plugins</code> and a text form of <code>/advisor</code>. In practice, this brings automated workflows closer to the interactive experience and makes it easier to operate agents in pipelines, servers, and desktop tools without relying on visual menus.</p><h2>The important part is permissions</h2><p>Anthropic fixed Edit, Write, and Read rules whose paths contain parentheses. Those rules could be discarded as invalid, leaving folders that were supposed to be read-only writable through the Bash sandbox. The release also changes how incomplete regular expression patterns are handled, so one broken deny rule no longer causes every file edit to fail.</p><p>It also fixes zsh commands that concealed command substitutions inside assignments to variables such as REPORTTIME, REPORTMEMORY, and DIRSTACKSIZE. Instead of auto-approving them, Claude Code prompts for approval again. It is not a glamorous changelog item, but it matters when an agent can touch real repositories.</p><h2>What changes for Brazilian SMBs</h2><p>Agencies and small companies using Claude Code to build dashboards, integrations, or automations can gain speed without treating the risk bundle as inevitable. The editorial recommendation is to update a test environment first and watch three things: whether permission rules still protect sensitive directories, whether long-session costs remain explainable, and whether plugins and subagents resume work correctly.</p><p>My read: this is an update worth adopting, but not one to roll into production blindly. It does not create a new category of agent. It makes the existing path more reliable, and that is worth more than a pretty demo. Autonomy is useful only when its action perimeter remains legible.</p><h2>Sources</h2><p>Official source and independent corroboration: <a href="https://github.com/anthropics/claude-code/releases/tag/v2.1.260" rel="noopener noreferrer">Anthropic's v2.1.260 GitHub release</a> and <a href="https://www.reddit.com/r/ClaudeAI/comments/1qvgdc5/claude-code-v21262130-what-changed/" rel="noopener noreferrer">a technical community report from ClaudeAI</a>.</p>