Claude Code 2.1.260 opens a live diff panel

Claude Code 2.1.260 adds a diff panel that opens beside the conversation in fullscreen mode and fills with uncommitted changes as Claude edits, toggled with /diff. The release notes on Github list it alongside a long run of permission, model and session fixes.

At a glance

• Permission rules for Edit, Write and Read whose paths contain parentheses were being dropped as invalid and ignored by the Bash sandbox, which left folders marked read-only writable.
• /cost and the status line's prompt_cache field now name a likely cause for prompt-cache misses, such as changed tool definitions or a system prompt edit, or idling past the TTL.
• The 2.1.259 change applying Read() deny rules to Bash arguments is reverted after it blocked npm run build under a Read(./**/build/**) rule in every mode and made cd && grep prompt in auto mode.

For anyone running Claude Code under a policy, the permission work matters more than the diff panel. Three separate defects quietly widened what an agent could touch: parenthesised paths discarded, one uncompilable pattern disabling every file rule, and zsh assignments hiding command substitution from the auto-approval check. Read together, they suggest that a deny rule is only as strong as the parser behind it, and that a silently ignored rule is more dangerous than a rejected one.

Three permission bugs left deny rules unenforced

Edit, Write and Read rules whose path contains parentheses were treated as invalid and skipped by the Bash sandbox, so folders declared read-only stayed writable. A single file rule with an uncompilable pattern, such as an unclosed bracket, made every file edit fail with Invalid regular expression; that deny rule now guards the literal path it spells.

Bash permission checks also auto-approved zsh commands that hid a command substitution inside a REPORTTIME, REPORTMEMORY or DIRSTACKSIZE assignment, and those now prompt for approval. Rules with text after the closing parenthesis, such as Bash(ls) x, never matched anything and are reported as invalid settings instead of being ignored.

Glob and Grep no longer probe the search path on disk before the permission check; a missing path is reported after permission is decided, as Read does. Commands typed at the ! bash-mode prompt run outside the sandbox even when sandbox.allowUnsandboxedCommands is set to false.

Claude Fable 5.1 now appears in the /model picker

The picker had omitted Fable 5.1 for organizations entitled to it, and the model was accepted only when typed as /model claude-fable-5-1. Prompt caching on it did not cover the context attached after tool results, so that context was re-sent as uncached input on every tool-call turn.

Agents pinned with model: fable ignored the [1m] tag on an ANTHROPIC_DEFAULT_FABLE_MODEL pin and ran silently with a 200K context window. Auto-compact now covers 1M-context models: Opus and Fable sessions compact shortly before the limit, and recovery compaction on very large contexts stops timing out at ten minutes.

Changing /effort mid-session on Fable 5.1 keeps the prompt cache intact, and model switching, which stayed blocked for the rest of a session after a plugin hook load failure or an unloadable organization-managed marketplace, is re-checked at each switch, with the refusal naming the cause.

Managed settings now load for Enterprise users with a leftover API key

Managed settings failed to load for claude.ai Enterprise and Team users who still carried an API key from an earlier /login. /status had listed a signed-in claude.ai account and a configured API key as if both were in effect; the credential not in use is now marked.

On Amazon Bedrock, model discovery, token counting and AWS SSO/STS credential calls failed with unable to get local issuer certificate when the corporate root CA lived only in the OS certificate store. Managed skillOverrides keyed on a bundled skill's alias, such as checkup for /doctor, now apply, and Skill(name) deny rules cover nested skills listed as <dir>:name.

Claude in Chrome follows the organization's admin setting, so when an admin turns it off, --chrome, /chrome and the browser tools are unavailable. The Claude apps gateway gained oidc.scope_on_refresh for identity providers that return an id_token on refresh only when openid is requested again.

Headless sessions and the VSCode pills

The release also carries /reload-plugins into headless sessions, where it now appears in the Claude Code Desktop and SDK command lists, and gives /advisor a text form for the desktop app, Remote Control and other -p or Agent SDK sessions: /advisor, /advisor <model>, /advisor off.

In VSCode, the footer model pill shows the selected effort level and the pills return to their earlier compact size, while the session list's status filter adds Open and Closed. The /ultrareview wait for cloud reviews rises to 45 minutes from 30.