Claude docs changes for March 31st, 2026 [diff]

Executive Summary

• Fullscreen rendering is now available as a research preview (CLAUDE_CODE_NO_FLICKER=1), eliminating flicker, keeping memory flat in long sessions, and adding mouse support for terminal UIs like VS Code's integrated terminal and iTerm2.
• Computer use from the CLI is now documented — macOS Pro/Max users can enable Claude to control native apps and the screen via the built-in computer-use MCP server.
• GitHub Enterprise Server support is now documented — Teams and Enterprise organizations can connect self-hosted GHES instances for web sessions, Code Review, and plugin marketplaces.
• 300k output tokens on Message Batches API — Opus 4.6 and Sonnet 4.6 now support up to 300k output tokens in batch requests using the output-300k-2026-03-24 beta header.
• 1M context window beta (context-1m-2025-08-07) is retiring on April 30, 2026 for Claude Sonnet 4.5 and Sonnet 4 — migrate to Sonnet 4.6 or Opus 4.6.

New Claude Code versions

2.1.88

New features

• Added CLAUDE_CODE_NO_FLICKER=1 environment variable to opt into flicker-free alt-screen rendering with virtualized scrollback
• Added PermissionDenied hook that fires after auto mode classifier denials — return {retry: true} to tell the model it can retry
• Added named subagents to @ mention typeahead suggestions

Existing feature improvements

• Thinking summaries are no longer generated by default in interactive sessions — set showThinkingSummaries: true in settings to restore
• Auto mode denied commands now show a notification and appear in /permissions → Recent tab
• /env now applies to PowerShell tool commands (previously only affected Bash)
• Improved PowerShell tool prompt with version-appropriate syntax guidance (5.1 vs 7+)
• /usage now hides redundant "Current week (Sonnet only)" bar for Pro and Enterprise plans
• Collapsed tool summary now shows "Listed N directories" for ls/tree/du instead of "Read N files"
• Image paste no longer inserts a trailing space
• Pasting !command into an empty prompt now enters bash mode, matching typed ! behavior

Major bug fixes

• Fixed prompt cache misses in long sessions caused by tool schema bytes changing mid-session
• Fixed nested CLAUDE.md files being re-injected dozens of times in long sessions that read many files
• Fixed Edit/Write tools doubling CRLF on Windows and stripping Markdown hard line breaks (two trailing spaces)
• Fixed StructuredOutput schema cache bug causing ~50% failure rate in workflows with multiple schemas
• Fixed memory leak where large JSON inputs were retained as LRU cache keys in long-running sessions
• Fixed a potential out-of-memory crash when the Edit tool was used on very large files (>1 GiB)
• Fixed --resume crash when transcript contains a tool result from an older CLI version or interrupted write
• Fixed LSP server zombie state after crash — server now restarts on next request instead of failing until session restart
• Fixed hooks if condition filtering not matching compound commands (ls && git push) or commands with env-var prefixes (FOO=bar git push)
• Fixed /stats losing historical data beyond 30 days when the stats cache format changes
• Fixed scrollback disappearing when scrolling up in long sessions
• Fixed voice mode failing to request microphone permission on macOS Apple Silicon
• Fixed PowerShell tool incorrectly reporting failures when commands like git push wrote progress to stderr on Windows PowerShell 5.1
• Fixed PreToolUse/PostToolUse hooks not providing file_path as an absolute path for Write/Edit/Read tools

Claude Code changes

New Documents

Computer use (CLI) [Source]

New document covering computer use from the Claude Code CLI. Computer use is a macOS research preview (Pro/Max plan, v2.1.85+) that lets Claude open apps, click through UIs, and take screenshots. Enable it by turning on the built-in computer-use MCP server via /mcp. Includes sections on per-app approval, safety guardrails (global Esc key, terminal excluded from screenshots, machine-wide lock), example workflows (native app validation, layout bug reproduction, iOS Simulator testing), and differences from the Desktop app.

Fullscreen rendering [Source]

New document covering the opt-in fullscreen rendering mode (enabled with CLAUDE_CODE_NO_FLICKER=1). Fullscreen rendering draws the UI on the terminal's alternate screen buffer, eliminates flicker, keeps memory flat in long conversations, and adds mouse support. Documents mouse controls (click, drag, scroll), keyboard shortcuts for scrolling, transcript mode search (Ctrl+o + /), tmux caveats, and the CLAUDE_CODE_DISABLE_MOUSE=1 option for those who prefer native terminal selection.

GitHub Enterprise Server [Source]

New document covering Claude Code integration with self-hosted GitHub Enterprise Server instances. Teams and Enterprise admins can connect a GHES instance once via a guided GitHub App manifest flow, after which developers can use claude --remote, Code Review, teleport sessions, and plugin marketplaces with GHES repositories. Documents admin setup steps, GitHub App permissions, manual setup fallback, network requirements (Anthropic IP allowlisting), developer workflow, GHES plugin marketplace configuration using full git URLs, and limitations (GitHub MCP server not supported; use gh CLI instead).

Changed documents

Agent teams [Source]

• Added new "Use subagent definitions for teammates" section: teammates can now be spawned using a named subagent type, inheriting its system prompt, tools, and model — enabling role reuse across delegated subagents and agent teams. [line ~216]
• Clarified that team config is auto-generated and should not be hand-edited; changes are overwritten on the next state update. [line ~203]

Chrome [Source]

• Added link to the new computer use doc in the "See also" section for tasks that require native app control beyond browser automation. [line ~210]

Claude Code on the web [Source]

• Added "From the terminal" setup section: run /web-setup to connect GitHub using local gh CLI credentials, creating a default cloud environment without the GitHub App (App still needed for Auto-fix). [line ~37]
• Added note that self-hosted GitHub Enterprise Server instances are supported for Teams and Enterprise plans. [line ~72]

CLI reference [Source]

• Clarified --allow-dangerously-skip-permissions: adds bypassPermissions to the Shift+Tab cycle without starting in it, allowing composition with other modes. [line ~84]
• --dangerously-skip-permissions clarified as equivalent to --permission-mode bypassPermissions. [line ~94]
• --enable-auto-mode now documents that it's available on Team, Enterprise, or API plans (not just Team). [line ~104]
• --permission-mode now lists all valid values: default, acceptEdits, plan, auto, dontAsk, bypassPermissions. [line ~105]

Code review [Source]

• Added new "Troubleshooting" section covering how to retrigger failed/timed-out reviews (use @claude review once, not the GitHub Re-run button) and where to find findings that don't appear as inline comments (check run Details, Files changed annotations, review body). [line ~144]
• Added note that annotations and the severity table are written independently of inline comments, so they remain available even if GitHub rejects a comment. [line ~136]

Desktop [Source]

• Auto mode is now available on Team, Enterprise, and API plans (previously only Team). [line ~173]
• Updated computer use enable steps to include explicit "Update the desktop app" and "Turn on the toggle" numbered steps with UI path correction (Settings > General under Desktop app). [line ~181]
• CLI vs Desktop comparison table updated: CLI now shows "Enable via /mcp" for computer use on macOS. [line ~218]

Environment variables [Source]

• Added CLAUDE_CODE_DISABLE_MOUSE: disables mouse tracking in fullscreen rendering while keeping flicker-free rendering active. [line ~247]
• Added CLAUDE_CODE_NO_FLICKER: enables fullscreen rendering research preview. [line ~255]
• Added CLAUDE_CODE_SCROLL_SPEED: sets mouse wheel scroll multiplier in fullscreen rendering (1–20, 3 recommended to match vim). [line ~262]

Hooks [Source]

• TaskCreated hook exit code 2 now says "Rolls back the task creation" (previously "Prevents the task from being created"). [line ~275]
• Debug section updated: simplified default debug output; added CLAUDE_CODE_DEBUG_LOG_LEVEL=verbose for granular hook matching details. [line ~284]

MCP [Source]

• The featured remote MCP server list was updated. Many servers were removed (Slack, Intercom, NetSuite, Databricks, n8n, ClickUp, PostHog, Hex, etc.) and many new servers were added (Amplitude, Blockscout, Egnyte, Coupler.io, Dice, Port IO, Aiwyn Tax, Clarify, Clarity AI, Day AI, bioRxiv, ChEMBL, Clinical Trials, CMS Coverage, ICD-10 Codes, NPI Registry, DevRev, Harmonic, Krisp, LunarCrush, Mem, Metaview, Midpage Legal Research, Sprouts, Gainsight/Staircase AI, CB Insights, and more). [line ~23]

Model config [Source]

• Clarified that model setting is initial selection only, not enforcement — users can still switch to Default in /model. [line ~75]
• Added documentation for ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_OPUS_MODEL, and ANTHROPIC_DEFAULT_HAIKU_MODEL environment variables that control what the Default option and model aliases resolve to — needed to fully pin model versions. [line ~80]

Network config [Source]

• Added note about allowlisting Anthropic API IP addresses for self-hosted GitHub Enterprise Server instances behind a firewall. [line ~77]

Permission modes [Source]

• Auto mode now available on Team, Enterprise, and API plans (previously only Team). [line ~20]
• Expanded auto mode description to document the layered prompt injection defense: a server-side probe scans incoming tool results, and the classifier is never shown tool results so injected instructions cannot influence approval decisions. [line ~137]
• bypassPermissions mode clarified: still prompts for writes to .git, .vscode, and .idea to prevent accidental corruption; .claude also prompts except for .claude/commands, .claude/agents, and .claude/skills. [line ~209]

Platforms [Source]

• CLI platform now lists computer use (Pro and Max, macOS) as a supported feature. [line ~8]

Sandboxing [Source]

• Added note that computer use runs on your actual desktop (not sandboxed) with per-app permission prompts as the gating mechanism. [line ~278]

Settings [Source]

• defaultMode setting now documents all valid values (default, acceptEdits, plan, auto, dontAsk, bypassPermissions) and notes that --permission-mode overrides it for a single session. [line ~228]

Sub-agents [Source]

• Added note that subagent definitions from any scope (project, user, plugin, CLI-defined) are available to agent teams when spawning teammates. [line ~182]

Terminal config [Source]

• Added "Reduce flicker and memory usage" section pointing to the new fullscreen rendering mode (CLAUDE_CODE_NO_FLICKER=1) for users experiencing flicker or scroll position jumps. [line ~56]

Voice dictation [Source]

• Added explicit disclosure that voice dictation streams audio to Anthropic's servers for transcription (not processed locally). [line ~6]

Web scheduled tasks [Source]

• Added minimum interval of 1 hour for scheduled tasks; cron expressions firing more frequently (e.g. */30 * * * *) are now explicitly rejected. [line ~100]

API changes

Changed documents

Batch processing [Source]

• Added "Extended output (beta)" section: the output-300k-2026-03-24 beta header raises the max_tokens cap to 300,000 on Message Batches API for Claude Opus 4.6 and Sonnet 4.6. Standard batch pricing (50% discount) applies. Not available on Bedrock, Vertex AI, or Foundry. [line ~408]

Claude in Microsoft Foundry [Source]

• Sonnet 4.5 no longer has a 1M token context window on Microsoft Foundry — it now has a 200k token context window. The context-1m-2025-08-07 beta is retired. [line ~214]

Claude on Amazon Bedrock [Source]

• Sonnet 4.5 and Sonnet 4 no longer have a 1M token context window on Bedrock — they now have a 200k token context window. The context-1m-2025-08-07 beta is retired. [line ~182]

Claude on Vertex AI [Source]

• Sonnet 4.5 and Sonnet 4 no longer have a 1M token context window on Vertex AI — they now have a 200k token context window. The context-1m-2025-08-07 beta is retired. [line ~115]

Context windows [Source]

• Simplified context window description: only Opus 4.6 and Sonnet 4.6 have 1M context; Sonnet 4.5, Sonnet 4, and other models now have a 200k context window. Beta header note removed. [line ~91]

Extended thinking [Source]

• Added Haiku 4.5 output limit (64k) and reference to 300k batch API beta for Opus 4.6 and Sonnet 4.6. [line ~86]

Models overview [Source]

• Added note that Opus 4.6 and Sonnet 4.6 support up to 300k output tokens on the Message Batches API using the output-300k-2026-03-24 beta header. [line ~39]

Pricing [Source]

• Removed the long context pricing section for Claude Sonnet 4.5 and Sonnet 4 (beta retiring April 30, 2026). Removed "long context pricing" from the list of pricing modifiers that stack with prompt caching multipliers. [line ~72]

Python Agent SDK [Source]

• context-1m-2025-08-07 beta is retiring April 30, 2026 — requests using it with Sonnet 4.5 or Sonnet 4 will return a 400 error. Recommends migrating to Sonnet 4.6 or Opus 4.6. [line ~1162]

Release notes [Source]

• Added March 30, 2026 release notes: 300k output tokens beta on Message Batches API, and retirement of context-1m-2025-08-07 beta for Sonnet 4.5/4 on April 30, 2026. [line ~6]

Remote MCP servers [Source]

• The featured remote MCP server list was updated with many additions and removals, consistent with the changes to the Claude Code MCP page (new servers include Ticket Tailor, Blockscout, Close, Egnyte, Coupler.io, Dice, Port IO, Aiwyn Tax, Circleback, Clarify, Clarity AI, Day AI, healthcare servers like ICD-10/ChEMBL/Clinical Trials/CMS Coverage/NPI Registry, DevRev, DirectBooker, Harmonic, Krisp, LunarCrush, Mem, Metaview, Midpage Legal Research, and more). [line ~20]

Service tiers [Source]

• Removed Priority Tier token burndown rates for long-context requests on Sonnet 4.5 and Sonnet 4 (which doubled input and increased output), consistent with the retirement of the 1M context beta. [line ~32]

TypeScript Agent SDK [Source]

• context-1m-2025-08-07 beta is retiring April 30, 2026 — requests using it with Sonnet 4.5 or Sonnet 4 will return a 400 error. Recommends migrating to Sonnet 4.6 or Opus 4.6. [line ~2119]

What's new in Claude 4.6 [Source]

• Updated "Higher output token limits" section (previously "128k output tokens") to include the new 300k output capability on Message Batches API for Opus 4.6 and Sonnet 4.6. [line ~79]