Auto mode is now available on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry via the new CLAUDE_CODE_ENABLE_AUTO_MODE env var (previously Anthropic API only)
The workflow trigger keyword changed from workflow to ultracode; natural-language requests like "use a workflow" also work
Advisor tool gains a max_tokens parameter to cap advisor output per call, with a new stop_reason field on results
claude-opus-4-8 is now listed in all API beta endpoint model enums; mid-conversation-system-2026-04-07 beta header removed
Protected paths list significantly expanded to cover more config dirs and dotfiles (.config/git, .devcontainer, .yarn, .mvn, and many shell/toolchain configs)
OTEL_RESOURCE_ATTRIBUTES values are now included as labels on metric datapoints, enabling slicing of usage metrics by custom dimensions like team or repo
/mcp now collapses claude.ai connectors you've never signed in to behind a "Show unused connectors" row
Existing feature improvements
claude agents rows now show done/total before the detail when work is fanned out; peek shows the longest-running item
Parallel tool calls: a failed Bash command no longer cancels other calls in the same batch — each tool returns its own result independently
Fullscreen mode: clipboard now uses wl-copy/xclip/xsel on Linux when available, copies to both clipboard and PRIMARY selection for middle-click paste, and the "hold {key} for native selection" hint now shows the correct key per terminal
Improved terminal rendering performance by stabilizing the layout engine's JIT compilation profile
Improved rendering performance for large file writes
[VSCode] Added a tip suggesting disabling terminal GPU acceleration (or running /terminal-setup) to fix garbled glyphs
Major bug fixes
Fixed forceLoginOrgUUID/forceLoginMethod managed-settings policies blocking third-party provider sessions (Bedrock, Vertex, Foundry, Mantle) alongside the org pin (regression in 2.1.146)
Fixed background subagent output corrupting claude -p stdout when using --output-format text or json
Fixed /autofix-pr reporting "cannot run on the default branch" when the session is inside a git worktree or another repository
Fixed --resume picker not showing sessions from the current directory when it isn't a git worktree (e.g., jj workspaces)
Fixed Windows hooks that invoke bash explicitly (e.g., /usr/bin/bash script.sh) failing with "command not found" or "cannot execute binary file"
Fixed OpenTelemetry log events being silently dropped when emitted before telemetry initialization completed
Fixed claude mcp list/get/add printing secrets to the terminal: ${VAR} references are no longer expanded, and credential headers and URL secrets are redacted
Fixed Workflow agents spawned with isolation: "worktree" in background sessions being blocked from editing files inside their own worktree
Fixed background sessions dispatched from claude agents booting on a stale model from the daemon's environment instead of the model in settings.json
Fixed completed subagents getting stuck showing as running when an error occurs while finalizing their result
Fixed EADDRINUSE errors from tools that bind Unix sockets under $TMPDIR when CLAUDE_CODE_TMPDIR is set to a deep path
Matchers now have documented comparison rules: a matcher with only letters, digits, _, and | is compared as an exact string with | as an OR separator; anything else is evaluated as a regex. * or empty string matches all occurrences. Updated the "Filter with regex matchers" section to "Filter with multi-tool matchers" to reflect this. [line 158] [Source]
Clarification that attached sessions always render in fullscreen mode, with notes on scrolling (PgUp/PgDn, mouse wheel) and why the tui setting does not apply. [line 154] [Source]
Updated availability note: auto mode on Bedrock, Vertex AI, and Foundry now requires CLAUDE_CODE_ENABLE_AUTO_MODE rather than being fully unavailable. [line 11] [Source]
Auto mode availability updated: for Enterprise deployments routing Desktop to Google Cloud Vertex AI, auto mode is off until CLAUDE_CODE_ENABLE_AUTO_MODE is set, and only Opus 4.7/4.8 are supported there. [line 74] [Source]
New CLAUDE_CODE_ENABLE_AUTO_MODE env var: enables auto mode on Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry (only Opus 4.7/4.8 supported on these providers). [line 172] [Source]
CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN clarification: does not apply to background sessions opened from agent view, which always use fullscreen rendering. [line 150] [Source]
CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE marked as removed in v2.1.160 and now a no-op; guidance updated to use /model + /fast on instead. [line 207] [Source]
"There's an issue with the selected model" error now has surface-specific guidance: interactive CLI runs /model, non-interactive mode uses --model flag or ANTHROPIC_MODEL, and Agent SDK surfaces a structured model_not_found error for programmatic handling. [line 518] [Source]
Added note that background sessions opened from agent view or claude attach always use fullscreen rendering, and tui/CLAUDE_CODE_DISABLE_ALTERNATE_SCREEN do not apply to them. [line 148] [Source]
Auto mode definition generalized: removed "Anthropic API only" qualifier now that it's available on other providers with an opt-in flag. [line 40] [Source]
Auto mode availability expanded: now available on Bedrock, Vertex AI, and Foundry for Opus 4.7/4.8 by setting CLAUDE_CODE_ENABLE_AUTO_MODE=1. New "Enable auto mode on Bedrock, Vertex AI, or Foundry" section with configuration examples. [lines 149-176] [Source]
tui setting description updated: background sessions opened from agent view always use fullscreen regardless of this setting. [line 244] [Source]
workflowKeywordTriggerEnabled updated: now controls the ultracode keyword (not workflow), and its display name in /config is now "Ultracode keyword trigger". [line 247] [Source]
grep, egrep, and fgrep (on a single file with no pipes or redirects) now satisfy the read-before-edit requirement in addition to cat, head, tail, sed. Also clarified that the recognized-for-deny-rules set differs from the read-before-edit list. [lines 129-130] [Source]
Workflow trigger keyword changed from workflow to ultracode. Natural-language requests ("use a workflow", "run a workflow") also now trigger workflows. The /config toggle renamed to "Ultracode keyword trigger". [lines 111-119] [Source]
New max_tokens parameter on the advisor tool definition to cap the advisor's total output (thinking + text) per call. Minimum 1024. Setting it above the advisor model's own output cap returns a 400 error. [line 94] [Source]
New stop_reason field on advisor_result and advisor_redacted_result variants, present when max_tokens is set on the tool definition. Carries the sub-call's stop reason ("end_turn" or "max_tokens"). [lines 143-145] [Source]
New "Capping advisor output" section with benchmark data: max_tokens: 2048 reduces mean advisor output ~7x with ~0% truncation; max_tokens: 1024 reduces ~10x but truncates ~10% of calls. [lines 418-465] [Source]
Added a security warning about prompt injection via tool results: content from tool results (web pages, uploads, third-party APIs) should be treated as untrusted, and instructions from those sources may attempt to redirect Claude. Recommends keeping untrusted content inside tool_result blocks rather than system prompts or plain user text. [line 61] [Source]
New "Claude flags tool results as prompt injection" troubleshooting section: when Claude refuses to act on a tool result, the fix is to move instructions out of the tool_result block into a subsequent user turn or a mid-conversation system message. [lines 44-47] [Source]
New "When Claude searches" section documenting when Claude triggers a web search (recent events, current data, explicit requests) vs. answers directly (stable facts, creative tasks, analysis of provided content). [lines 22-38] [Source]
Practical max_uses guidance added: simple queries typically use 1–3 searches; max_uses: 3 bounds cost for latency-sensitive lookups; 15–20 or omit for research agents. [line 132] [Source]
mid-conversation-system-2026-04-07 beta header removed from the AnthropicBeta enum (now "23 more" instead of "24 more"). [line 13] [Source]
BetaOutputTokensDetails object now formally documented as a named type with a detailed description of the thinking_tokens field, including how it's computed and how it relates to output_tokens. [lines 12589-12602] [Source]
output_tokens_details fields across all usage objects now reference BetaOutputTokensDetails by type link instead of inline object { thinking_tokens }. [line 9212] [Source]
claude-opus-4-8 added to BetaManagedAgentsModel enum as the new first/top model option across all agent create/update/list/retrieve endpoints. [line 71] [Source]
mid-conversation-system-2026-04-07 beta header removed from the beta enum across all beta API endpoint reference files.