Claude Code 2.1.260: Diff Panel, /reload-plugins, Fable 5.1 Fixes
Anthropic's Claude Code CLI updates to v2.1.260, adding a visual diff panel, expanding headless session commands, and fixing a batch of permission, caching, and model-switching bugs that have been tripping up developers.
New Features and Commands
- Diff panel: Opens beside the conversation in fullscreen mode, showing uncommitted changes as Claude edits. Toggle it with
/diff. - /reload-plugins: Now available in headless sessions (via
-pand Agent SDK), so it appears in Claude Code Desktop and SDK command lists. - Text form of /advisor:
/advisor,/advisor <model>,/advisor off—for desktop app, Remote Control, and other headless sessions. - OIDC scope_on_refresh: Added to the Claude apps gateway for identity providers that return an
id_tokenon refresh only when requested again. - New desktop policy keys: Support for
userPluginMarketplacesEnabledanduserPluginUploadsEnabled.
Notable Fixes
- Permission paths with parentheses: Fixed rules being dropped, which left read-only folders writable.
- Invalid regex in file permission rules: A single rule with an uncompilable pattern now guards the literal path instead of crashing all edits.
- Bash auto-approval bypass: zsh commands hiding command substitution in
REPORTTIME,REPORTMEMORY, orDIRSTACKSIZEassignments now prompt for approval. - Fable 5.1 context window: Agents pinned via
ANTHROPIC_DEFAULT_FABLE_MODELsometimes ignored the[1m]tag and ran with a 200K context window—now fixed. - Fable 5.1 prompt caching: Context after tool results was re-sent uncached; caching now covers it.
- Model switching: Stuck after a plugin hook load failure or when an org-managed plugin's marketplace couldn't load; now re-checks each time and names the cause.
- SDK MCP servers: Desktop connectors sometimes missing from first turn; they now appear immediately.
- /rewind reliability: No more false success reports and stale file-read tracking after rewinds.
- Terminal rendering: Flags and joined emoji no longer split across lines or leave stale text.
Also fixed: Bedrock's local issuer certificate error when root CA is only in OS store, permissions.blockReadsOutsideWorkingDirectories hiding git config on macOS, managed settings not loading for Enterprise/Team users with leftover API keys, and Remote Control accepting invalid model picks.
📖 Read the full source: GitHub Claude-Code
👀 See Also

LeMario: Training a JEPA World Model on Super Mario Bros — Technical Walkthrough and Postmortem
Benjamin Bai reproduces LeWorldModel's JEPA architecture from scratch and trains it on Super Mario Bros. The model predicts five-step futures but fails at long-horizon planning — a detailed technical postmortem.

Claude Code CC 2.1.124 and 2.1.126: File Modification Budget Exceeded Reminder, Harness Instructions Update, REPL Awaits Clarification, and Malware Analysis Reminder Removed
CC 2.1.124 adds a system reminder for file changes omitted due to budget limits, updates harness instructions with explicit insertion points, and clarifies REPL auto-await behavior. CC 2.1.126 removes the malware analysis post-read reminder.

The AI Dependency Trap: Why Over-Reliance on LLMs May Erode Core Skills
A contrarian take arguing that heavy reliance on AI chatbots will lead to atrophy of critical thinking, writing, research, and learning abilities.

Claude Code 2.1.76 adds MCP elicitation, worktree improvements, and fixes for context limits
Claude Code version 2.1.76 introduces MCP elicitation support for structured input during tasks, adds worktree.sparsePaths for large monorepos, and fixes 'Context limit reached' errors on 1M-context sessions. Version 2.1.75 made 1M context windows default for Opus 4.6 on Max, Team, and Enterprise plans.