Claude docs changes for May 29th, 2026 [diff]

Executive Summary

• Claude Opus 4.8 launches as the new flagship model with adaptive-only thinking, xhigh/high effort defaults, fast mode at 2x the standard rate, and 1M-token context window
• Dynamic Workflows (research preview) let Claude orchestrate tens to hundreds of background subagents for large tasks; a new ultracode effort level uses xhigh reasoning + automatic workflow planning
• Mid-conversation system messages (Opus 4.8, Claude API only): insert {"role": "system"} messages mid-conversation to add instructions without invalidating the prompt-cache prefix
• New MessageDisplay hook in Claude Code lets hooks intercept and reformat assistant text as it renders on screen, without modifying the transcript
• Batch processing gains a documented table of unsupported parameters and new server-tool/web-search behavior notes

New Claude Code versions

2.1.153

New features

• Added skipLfs option to github/git plugin marketplace sources to skip Git LFS downloads during clone and update
• Status line commands now receive COLUMNS and LINES environment variables so scripts can size output to the terminal width
• /bg while Claude is responding now continues the response in the background session instead of dropping it
• /model now saves the selection as the default for new sessions; press s in the picker to switch for the current session only (replaces the d key)

Existing feature improvements

• claude agents: autocomplete in the dispatch input now suggests native slash commands and bundled skills
• claude agents: PR column now shows PR #N for a single PR or N PRs for multiple
• claude doctor now shows the result of the most recent update attempt
• Combined separate MCP authentication startup notifications into a single message
• macOS: background agents now appear as "Claude Code" in Privacy & Security across upgrades

Major bug fixes

• Fixed stateful MCP servers without the optional GET SSE stream reconnect-looping on tools/list (regression in v2.1.147)
• Fixed regression where a custom API gateway could receive the user's Anthropic OAuth credential instead of the gateway's own token
• Fixed subagent frontmatter MCP servers ignoring --strict-mcp-config, --bare, remote mode, enterprise managed MCP, and managed-settings allow/deny policies
• Fixed excessive memory usage (multiple GB) when resuming a session by transcript file path on machines with many stored sessions
• Fixed Agent tool with subagent_type: 'claude' silently discarding outputs written to gitignored paths via a temporary worktree

2.1.154

New features

• Claude Opus 4.8 is now available; defaults to high effort; use /effort xhigh for hardest tasks
• Dynamic workflows: ask Claude to write a workflow and it orchestrates tens to hundreds of background agents; use /workflows to view runs
• Fast mode on Opus 4.8 available at 2x the standard rate (vs 3x for 4.6/4.7) with 2.5x the speed
• claude agents: type ! <command> to run a shell command as a background session you can attach to and detach from; also available as claude --bg --exec '<command>'
• Plugins can now declare defaultEnabled: false in plugin.json or a marketplace entry; users opt in with claude plugin enable
• Stdio MCP server subprocesses now receive CLAUDE_CODE_SESSION_ID and CLAUDECODE=1 in their environment
• claude mcp list/get now show unapproved .mcp.json servers as ⏸ Pending approval instead of auto-approving when output is piped
• Streaming tool execution is now always enabled, including when telemetry is disabled or on Bedrock/Vertex/Foundry

Existing feature improvements

• The lean system prompt is now the default for all models except Haiku, Sonnet, and Opus 4.7 and earlier
• Claude now reserves multiple-choice questions for decisions it genuinely cannot make itself
• /simplify now runs a cleanup-only review (reuse, simplification, efficiency) instead of the full bug-hunting review
• /plugin Discover tab now pins plugins matching the current directory with a "suggested for this directory" annotation
• Renamed /effort slider labels from "Speed"/"Intelligence" to "Faster"/"Smarter"
• Claude in Chrome: pick which connected browser to use via /chrome or in-chat when multiple browsers are connected
• Deprecated CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE (will be removed 06/01)
• Improved auto-mode classifier's detection of data exfiltration, particularly bulk repository transfers

Major bug fixes

• Fixed rm -rf $HOME not being blocked as a dangerous path when HOME has a trailing slash
• Fixed background-agent completion notifications triggering premature "out of context" behavior on some 1M-context models
• Fixed orphaned claude --bg-pty-host processes spinning at 100% CPU after the daemon exits on macOS
• Fixed worktree.baseRef: "head" resolving to the main checkout's HEAD instead of the current worktree's HEAD when spawning subagents or calling EnterWorktree from inside a linked worktree
• Fixed a single invalid allowedMcpServers/deniedMcpServers entry in managed settings discarding all managed-settings policy

2.1.156

Major bug fixes

• Fixed an issue when using Opus 4.8 where thinking blocks were modified, leading to API errors

Claude Code changes

New Documents

Workflows [Source]

New document covering dynamic workflows (research preview, requires v2.1.154+). Workflows are JavaScript scripts Claude writes to orchestrate subagents at scale — dozens to hundreds of agents running in the background while the session stays responsive. Covers when to use workflows vs subagents/skills, running the bundled /deep-research command, having Claude write a custom workflow, the ultracode effort level that auto-plans workflows for every substantive task, and how to manage, pause, resume, and save runs.

Changed documents

Agent SDK / Hooks [Source]

• New MessageDisplay hook added to the SDK hooks table: fires when an assistant message with text completes, for redacting or reformatting displayed text without changing the transcript [line 116] [Source]

Agent SDK / Python [Source]

• MessageDisplay hook added to the list of events supported by the Python SDK
• New note: on Linux, sandbox failIfUnavailable defaults to False in the Python SDK (unlike TypeScript where it defaults to true); commands run unsandboxed with a warning when the sandbox can't start unless "failIfUnavailable": True is set explicitly

Agent SDK / TypeScript [Source]

• MessageDisplay hook event type and MessageDisplayHookInput interface added, with turn_id, message_id, index, final, and delta fields
• New failIfUnavailable field documented on SandboxSettings (default true in TypeScript): stops at startup when the sandbox can't start; set to false to fall back to unsandboxed execution with a warning

Agent View [Source]

• New feature: type ! <command> in the dispatch input to run a shell command as a background session; also available as claude --bg --exec '<command>'
• PR status indicator changed from a colored dot to labeled PR #N text (or N PRs for multiple)
• /bg while Claude is mid-response now continues the response in the background session
• /model picker: s key now switches session-only; d key removed

Agents [Source]

• "Worktrees" replaced with "Dynamic workflows" (workflows.md) as one of the four parallelism approaches in the comparison table

CLI Reference [Source]

• New --exec flag: runs a shell command as a PTY-backed background job instead of a Claude session; use with --bg
• New --prompt-suggestions flag: emits a prompt_suggestion message in --output-format stream-json --verbose mode (off by default)
• --fallback-model now also covers retired or unavailable models in addition to overloaded ones

Commands [Source]

• New /deep-research <question> command: a bundled workflow that fans out web searches across several angles and returns a cited report
• New /reload-skills command: re-scans skill directories mid-session without restarting
• New /simplify [target] command: cleanup-only review (reuse, simplification, efficiency); use /code-review for bug hunting
• New /workflows command: opens a progress view for running and completed workflows
• /code-review updated: now accepts ultra level and --fix flag; ultra is the preferred alias for /ultrareview
• /effort now includes ultracode as a new level
• /model behavior changed: Enter saves as default; s applies session-only

Environment Variables [Source]

• New CLAUDE_CODE_DISABLE_WORKFLOWS: disables dynamic workflows entirely
• New CLAUDE_CODE_PROPAGATE_TRACEPARENT: propagates W3C trace context to custom proxies
• New OTEL_METRICS_INCLUDE_ENTRYPOINT: includes session entrypoint in OpenTelemetry metrics attributes
• CLAUDECODE and CLAUDE_CODE_SESSION_ID now also set in stdio MCP server subprocesses
• CLAUDE_EFFORT now reports ultra value for ultracode sessions
• CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE documented as deprecated

Hooks [Source]

• New MessageDisplay hook event: fires while assistant message text streams, delivering text in batches via a delta field; returns displayContent to replace what's rendered on screen without modifying the transcript
• effort field in hook input now reports "ultra" as a value (for ultracode sessions)
• SessionStart hook can now return sessionTitle (to name the session automatically) and reloadSkills (to trigger a skill re-scan after the hook installs new skills)
• SessionStart input now includes session_title so hooks can avoid overwriting a title the user set explicitly
• ${CLAUDE_PROJECT_DIR} is now also set in the environment of stdio MCP servers and plugin LSP servers
• Updated hooks lifecycle diagram to include MessageDisplay as a display-only event

Interactive Mode [Source]

• /btw side-question overlay now supports f key to fork into a new full session, x to clear earlier exchanges, and Up/Down to scroll
• Prompt suggestions in print mode: off by default, enabled via --prompt-suggestions flag

Keybindings [Source]

• / in vim normal mode now opens reverse history search (same as Ctrl+R)
• modelPicker:setAsDefault (key d) replaced with modelPicker:thisSessionOnly (key s)

MCP [Source]

• Pending approval servers from .mcp.json now appear in claude mcp list as Pending approval; claude mcp get also shows rejected status

Model Configuration [Source]

• Claude Opus 4.8 is now the default opus alias on the Anthropic API (requires v2.1.154+); default effort is high
• New ultracode effort level: combines xhigh reasoning with automatic dynamic workflow orchestration; session-only, cannot be set via effortLevel or --effort in config
• /model now saves as default by default; s key applies to the current session only

Plugin Marketplaces [Source]

• New defaultEnabled field in marketplace entries (takes precedence over plugin.json): ships plugins disabled by default, requiring user opt-in
• claude plugin marketplace remove now supports --scope option to remove from a specific settings scope without uninstalling
• claude plugin marketplace list --json now includes source-specific fields in output

Plugins Reference [Source]

• New defaultEnabled field in plugin.json: ships a plugin disabled by default; users opt in with /plugin or claude plugin enable (requires v2.1.154+)
• MessageDisplay hook event type added to plugin hook events

Setup [Source]

• claude doctor now shows the result of the most recent update attempt and lists fixes when the npm global directory isn't writable

Skills [Source]

• New disallowed-tools frontmatter field: removes specific tools from Claude's available pool while a skill is active (useful for background loops to prevent AskUserQuestion)
• ${CLAUDE_EFFORT} substitution now also reports the ultra value (for ultracode)
• allowed-tools now accepts comma-separated strings in addition to space-separated

Status Line [Source]

• COLUMNS and LINES environment variables are now set before status line scripts run, enabling terminal-width-aware output (requires v2.1.153+)
• effort.level field now reports ultra value for ultracode sessions

Sub-agents [Source]

• MCP restrictions (--strict-mcp-config, --bare, enterprise managed MCP, allowedMcpServers/deniedMcpServers) now apply to MCP servers declared in subagent frontmatter as of v2.1.153

Voice Dictation [Source]

• Voice dictation is now unavailable when HIPAA compliance is enabled for the organization

API changes

New Documents

Build with Claude / Mid-Conversation Effort Example [Source]

New guide showing how to build an orchestration mode using mid-conversation system messages. Demonstrates a full agentic loop that toggles xhigh effort and automatic workflow use via system messages appended mid-conversation, with parallel subagent fan-out and structured report_findings results. Available on Claude Opus 4.8 only.

Build with Claude / Mid-Conversation System Messages [Source]

New guide for the mid-conversation system message feature (Claude Opus 4.8, Claude API only). Explains how appending {"role": "system"} messages mid-conversation allows adding instructions without invalidating the prompt-cache prefix. Covers placement constraints, caching interaction, use cases (mode switches, per-turn context injections, policy changes), and limitations.

Managed Agents / Cloud Sandboxes Reference [Source]

New reference page documenting the specifications of cloud sandboxes for Managed Agents. Lists pre-installed languages (Python 3.12+, Node.js 20+, Go 1.22+, Rust, Java, Ruby, PHP, C/C++), databases (SQLite local, PostgreSQL/Redis clients), system utilities, and development tools.

Changed documents

Agents and Tools / Agent Skills / Quickstart [Source]

• Code examples updated to claude-opus-4-8

Agents and Tools / MCP Connector [Source]

• MCP connector now supports Message Batches API requests; MCP tool calls via the Batches API are priced the same as regular requests

Agents and Tools / Tool Use / Advisor Tool [Source]

• Claude Opus 4.8 added as a supported advisor and executor model
• New stop_reason field on BetaAdvisorResultBlock and BetaAdvisorRedactedResultBlock: reports the advisor sub-inference's stop reason (same values as top-level stop_reason); max_tokens indicates the advisor's output was truncated

Agents and Tools / Tool Use / Code Execution Tool [Source]

• Claude Opus 4.8 added to the supported model table for both code_execution_20250825 and code_execution_20260120 tool versions

Agents and Tools / Tool Use / Overview [Source]

• Tool use system prompt token counts updated for all models; Claude Opus 4.8 added (290 tokens for auto/none, 410 for any/tool)
• Token counts for existing models changed significantly (e.g., Opus 4.7 updated from 346/313 to 675/804)

Agents and Tools / Tool Use / Web Search Tool [Source]

• Claude Opus 4.8 added to the dynamic filtering support list for web_search_20260209
• New note: the Batches API throttles web search per organization for large batches; links to the Claude Console Limits page to view or request rate limit increases

API / Beta [Source]

• Two new beta header values added: thinking-token-count-2026-05-13 and mid-conversation-system-2026-04-07
• Claude Opus 4.8 ("claude-opus-4-8") added to the model enum with description "Frontier intelligence for long-running agents and coding"
• BetaAdvisorResultBlock and BetaAdvisorRedactedResultBlock now include stop_reason field

API / Messages [Source]

• New MidConversationSystemBlockParam type added: represents system instructions that appear mid-conversation instead of via the top-level system parameter; type is "mid_conv_system"
• New url_not_in_prior_context error code added to the web fetch error enum
• ContentBlockParam union expanded to include MidConversationSystemBlockParam (now 14 variants)

API / Messages / Create [Source]

• role field now accepts "system" in addition to "user" and "assistant" to support mid-conversation system messages
• MidConversationSystemBlockParam documented as a new content block type
• Claude Opus 4.8 added to the model enum

Build with Claude / Batch Processing [Source]

• New pricing row for Claude Opus 4.8: $2.50/MTok batch input, $12.50/MTok batch output
• New table documenting unsupported parameters in batch requests (e.g., stream, speed, store, cache_hint, max_tokens: 0, research_preview_2026_02) with reasons for each
• New "Server tools and the agentic loop" section: all server tools work in batches; the batch loop runs more iterations per turn before returning pause_turn; web_search is throttled per organization in batch mode
• Extended output beta (output-300k-2026-03-24) now includes Claude Opus 4.8

Build with Claude / Context Windows [Source]

• Claude Opus 4.8 added to the 1M-token context window group; note added that on Microsoft Foundry, Opus 4.8 has a 200k-token context window
• Compaction beta support list updated to include Claude Opus 4.8

Build with Claude / Effort [Source]

• Claude Opus 4.8 added to supported models for max and xhigh effort levels
• New "Recommended effort levels for Claude Opus 4.8" section: recommends xhigh for coding/agentic use, high as default; links to the new orchestration mode guide
• New note clarifying that Claude Code's ultracode mode is not an API effort level — it pairs xhigh effort with multi-agent permissions via mid-conversation system messages

Build with Claude / Extended Thinking [Source]

• Claude Opus 4.8 added: manual extended thinking (budget_tokens) not supported (returns 400); adaptive thinking only, with display defaulting to "omitted"
• New usage.output_tokens_details.thinking_tokens field documented for tracking how many billed output tokens were spent on reasoning
• Model comparison table expanded with Claude Opus 4.8 column

Build with Claude / Fast Mode [Source]

• Claude Opus 4.8 added to supported fast mode models at $10/$50 per MTok (vs $30/$150 for 4.6/4.7) — 2x the standard rate
• Fast mode for Opus 4.6 is now deprecated; will be removed approximately 30 days after the Opus 4.8 launch
• Fast mode for Opus 4.8 is Claude API only (not available on Bedrock, Vertex, or Foundry)

Build with Claude / Prompt Caching [Source]

• New pricing row for Claude Opus 4.8: $5/MTok input, $0.50/MTok cache hits, $25/MTok output; added to 1,024-token minimum cacheable prompt length group
• New note: on Opus 4.8, you can append a {"role": "system"} message mid-conversation without invalidating cached prefixes

Build with Claude / Structured Outputs [Source]

• Claude Opus 4.8 added to the supported model list for structured outputs on the Claude API, Amazon Bedrock, and Vertex AI

Build with Claude / Task Budgets [Source]

• Task budgets now listed as available in beta on Claude Opus 4.8 (previously only Opus 4.7)
• Code examples updated to use the streaming pattern (client.beta.messages.stream)

Build with Claude / Working with Messages [Source]

• New note: temperature, top_p, and top_k are not supported on Claude Opus 4.7 and later (including 4.8); setting them returns a 400 error
• New note: refusal responses (stop_reason: "refusal") on Opus 4.7+ include a stop_details object identifying the policy category
• New "System role in messages" section documenting mid-conversation system messages for Opus 4.8

Get Started [Source]

• Prerequisites section now explicitly lists having an API key
• Quickstart steps expanded to include more language tabs (C#, Go, PHP, Ruby)
• Step 2 now includes creating a virtualenv before installing the SDK
• Example output simplified to show a condensed response format

Intro [Source]

• Featured "latest model" updated to Claude Opus 4.8, described as Anthropic's most capable model for complex reasoning and agentic coding

Managed Agents / Agent Setup [Source]

• Note added: fast mode for Opus 4.6 is deprecated as of the Opus 4.8 launch and will be removed approximately 30 days later

Managed Agents / Environments [Source]

• Terminology updated throughout: "cloud container" renamed to "cloud sandbox"; broken link updated from cloud-containers.md to cloud-sandboxes-reference.md

Managed Agents / Overview [Source]

• "Container" terminology replaced with "sandbox" throughout; code examples updated to claude-opus-4-8