Claude Code 2.1.276 fixes the 2.1.275 proxy and gateway regression

Quick answer

Anthropic released Claude Code 2.1.276 on September 18 to fix a 2.1.275 regression that broke requests when ANTHROPIC_BASE_URL pointed to a proxy or gateway. This update summarizes official release notes, not an independently tested fix. Earlier account-sync and plugin changes remain covered below.

Download Chat AI Opens the official App Store or Google Play for your device.

Version 2.1.276 addresses failed requests through a gateway

Anthropic identifies an HTTP 400 error containing advisor_20260301 when ANTHROPIC_BASE_URL points to a proxy or gateway. The release notes attribute the regression to 2.1.275. Our suggested check: confirm the installed version and retry a non-sensitive request through the same approved gateway after upgrading. Do not bypass organizational routing controls to work around the error. This targeted fix does not establish that every gateway configuration is compatible.

Sources: Anthropic on GitHub

Account sync has separate opt-out controls

Set syncClaudeAiSkills or syncClaudeAiPlugins to false to disable the corresponding sync. Editing a synced skill locally does not save it back to the account. Our suggestion: review enabled skills before starting sensitive work and keep an independent copy of intentional edits.

Sources: Anthropic on GitHub

npm plugin installation skips package scripts

npm-sourced plugins are now integrity-checked and fetched with installation scripts disabled. The release also fixes credentials appearing in plugin and marketplace logs or messages. Neither change establishes that a plugin's runtime behavior is safe; review its source and requested access separately.

Sources: Anthropic on GitHub

Projects beta starts with a restricted rollout

Initial eligibility requires Pro or Max, Claude Code cloud-session use, and no existing web or desktop projects. Broader access on those plans is planned over the following week; Team, Enterprise, and the wider Claude experience follow later. Existing projects continue unchanged for now.

Sources: Anthropic

Parallel threads still need review and a usage budget

Each worker is a cloud session with its own repository copy and branch. Threads share memory; a library collects files and artifacts. Overlapping code can still produce merge conflicts, and parallel sessions can exhaust usage faster. Local execution is planned, not available in this launch. Our suggestion: start with a bounded task and review each diff before merging.

Sources: Anthropic

September 17: credential display and permission fixes

Anthropic says MCP errors and login-tool descriptions could display secrets resolved from configuration variables. Version 2.1.274 fixes that behavior. Certain special-shell-variable commands now require permission, and certain nested shell expansions are refused in worktree-isolated sessions.

Sources: Anthropic on GitHub

Recovery changes and practical upgrade checks

The release also addresses repeated tool-use-ID errors and goals lost when resuming compacted sessions. Our suggested check: use disposable credentials and a throwaway project to verify error redaction, permission prompts, and resume behavior after upgrading. This is guidance, not a completed test or a security guarantee.

Sources: Anthropic on GitHub

Claude Code 2.1.261 lets teams raise inline command and task output caps

The new bashOutputMaxChars and taskOutputMaxChars settings control how much Bash-command and background-task output Claude receives inline before the remainder is saved to a file. Each setting can be raised as high as 128,000 characters. A larger inline allowance can keep more output in the immediate model context, but it also consumes more context, so teams should increase it only when the added logs are useful for the task.

Sources: Anthropic on GitHub

/skill-doctor shows unused skills and their context cost

Claude Code 2.1.261 adds /skill-doctor, which reports loaded skills that were not used and the context they consume. This gives users a direct way to identify skills that add prompt overhead without helping the current workflow. The release notes do not prescribe a universal removal threshold, so decisions should account for how often a skill is needed across representative tasks rather than a single session.

Sources: Anthropic on GitHub

Policy diagnostics and subagent prompts become easier to manage

The /status and claude doctor outputs now include an Organization policy line that explains why policy loading failed, including cases where a proxy does not pass the endpoint through. A new --append-subagent-system-prompt-file option reads a subagent system prompt from a file when the prompt is too large for a practical command-line argument. These features improve diagnosis and configuration; they do not bypass an organization's managed policy.

Sources: Anthropic on GitHub

Remote sessions, proxies, and VS Code receive broad reliability fixes

Anthropic lists fixes for dropped context when resuming parallel tool calls, stale Remote Control permission states, sessions that appeared busy after stopping, early Stop signals that were ignored, missing connectors after a startup timeout, and high CPU use from failed agent wake-ups. The VS Code extension also gains add and remove controls in its MCP server dialog, clearer session-state indicators, and multiple fixes for restored, teleported, grouped, and archived sessions. These are targeted corrections rather than a guarantee that every network or IDE configuration is covered.

Sources: Anthropic on GitHub

Claude Code 2.1.260 puts uncommitted changes in a full-screen diff panel

The new diff panel gives terminal users a side-by-side view of uncommitted changes and can be toggled with /diff. Anthropic also updated /cost and the status-line prompt_cache field to explain likely reasons for prompt-cache misses. Together, these changes make review and cache troubleshooting visible inside the active Claude Code workflow rather than requiring a separate inspection step.

Sources: Anthropic on GitHub

The release fixes two Bash permission bypass paths

Anthropic says Claude Code could drop or ignore Edit, Write, and Read permission rules when a path contained parentheses, potentially leaving a folder writable even when a rule intended it to be read-only. Version 2.1.260 fixes that sandbox translation error. It also stops zsh REPORTTIME, REPORTMEMORY, and DIRSTACKSIZE assignments from being auto-approved when they hide command substitution. These are fixes to Claude Code's permission handling; they do not replace careful review of the rules a team deploys.

Sources: Anthropic on GitHub, Anthropic

Anthropic rolled back one over-broad denied-file check from 2.1.259

Version 2.1.259 expanded how Read() deny rules were applied to file-like arguments passed through Bash. In 2.1.260, Anthropic reverted that behavior because it could block unrelated commands: the release notes cite npm run build under a Read(./**/build/**) rule and unnecessary prompts for compound cd-and-grep commands. Teams that adopted 2.1.259 specifically for that broader Bash check should retest their policy expectations rather than assume every part of the earlier change remains active.

Sources: Anthropic on GitHub, Anthropic on GitHub, Anthropic

Headless and plugin workflows get new controls

The /reload-plugins command can now be used in headless sessions and appears in Desktop and Agent SDK command lists. Anthropic also added text-based /advisor forms for Desktop, Remote Control, and headless or SDK use, and removed the previous one-hour limit on background commands launched by subagents. Those changes are useful for long-running automation, but the 2.1.259 --permission-prompts none flag still denies any action that would otherwise require an interactive prompt; it does not silently approve tools.

Sources: Anthropic on GitHub, Anthropic on GitHub, Anthropic

Worktrees, sessions, and remote connections receive reliability fixes

The update repairs resume behavior when a worktree's Git metadata is lost, prevents team-agent transcripts from disappearing during retry waits, and fixes duplicate or phantom sessions. It also addresses a Chrome connection dropping during a task, URLs failing when installing plugin marketplaces, and output swapping between concurrent sessions in the same directory. These are targeted fixes listed by Anthropic, not a guarantee that every concurrent or remote workflow will behave identically across environments.

Sources: Anthropic on GitHub

Managed MCP and unattended controls from 2.1.259 still matter

The preceding 2.1.259 release introduced managedMcpServers for organization-provided HTTP and SSE servers and --permission-prompts none for fail-closed unattended sessions. It also changed allowedMcpServers to govern user-added servers while deniedMcpServers remains the way to block a literal managed server. Administrators upgrading through both versions should review those MCP lists and validate managed settings before relying on the newer 2.1.260 fixes.

Sources: Anthropic on GitHub, Anthropic, Anthropic

A focused upgrade check should cover output, skills, and permission edge cases

After upgrading, run /skill-doctor on representative work before pruning skills, and test bashOutputMaxChars and taskOutputMaxChars with a non-sensitive command whose output exceeds the configured cap. Check the Organization policy line if managed settings do not load. Then open /diff against a representative change and test a permission rule whose path contains parentheses. These checks are a practical reading of the release notes, not an Anthropic certification procedure.

Sources: Anthropic on GitHub, Anthropic on GitHub, Anthropic on GitHub, Anthropic

This is a Claude Code release, not a Chat AI model update

Version 2.1.261 updates Anthropic's Claude Code command-line, desktop, remote, and VS Code workflows. It does not announce a new Claude model or add a model to Chat AI. Chat AI availability is therefore marked not applicable for this article.

Sources: Anthropic on GitHub

Frequently asked questions

What readers usually ask

Which Claude Code version fixes the advisor_20260301 gateway error?

Anthropic lists the fix in 2.1.276. It addresses a regression introduced in 2.1.275 when ANTHROPIC_BASE_URL points to a proxy or gateway.

Does editing a synced skill update my Claude account?

No. Local edits are not saved back to the account.

What are the main changes in Claude Code 2.1.261?

The release adds configurable inline output caps, /skill-doctor, organization-policy diagnostics, a file-based subagent system-prompt option, and extensive remote-session and VS Code fixes.

What do bashOutputMaxChars and taskOutputMaxChars control?

They control how much Bash-command or background-task output Claude receives inline before additional output is saved to a file. Each can be raised to 128,000 characters.

What does /skill-doctor do?

It reports loaded skills that went unused and shows how much context they cost, helping users identify prompt overhead to review.

What permission issue does Claude Code 2.1.260 fix?

Anthropic says paths containing parentheses could cause Edit, Write, or Read rules to be dropped or ignored by the Bash sandbox. Version 2.1.260 fixes that translation problem.

Did version 2.1.260 keep every denied-file change from 2.1.259?

No. Anthropic reverted the application of Read() deny rules to Bash arguments because it produced false denials for commands such as npm run build and unnecessary prompts for compound commands.

Can headless sessions reload plugins now?

Yes. Claude Code 2.1.260 makes /reload-plugins available in headless sessions and exposes it in Desktop and Agent SDK command lists.

Does --permission-prompts none approve tools automatically?

No. The flag introduced in 2.1.259 automatically denies actions that would require a prompt. The active permission mode still decides actions that do not need a prompt.

Is Claude Code 2.1.261 available in Chat AI?

It is not a model or a Chat AI feature. It is a release of Anthropic's Claude Code tooling.

Evidence