Source Intelligence
Follow Discord

DisclaimerUnofficial, and not affiliated with Anthropic. Nearly all of this is read straight out of what ships: npm bundles, captured prompts, published docs. Anthropic's own notes go in verbatim, marked as theirs. The rest is my reading, and every entry carries the strings behind it. If one looks wrong, vote it down and say why.

← All releases Home ←olderv2.1.81 v2.1.84newer→

Claude Code v2.1.83

32 entries read diff v2.1.81 → v2.1.83 Markdown

This release introduces the Advisor Tool — a server-side reviewer model you can configure with /advisor — along with claude-cli:// deep link support for opening Claude Code from external URLs, two new hook types (CwdChanged and FileChanged) for filesystem-aware automation, and the rebranded Ultrareview cloud code review. Enterprise users gain XAA IdP authentication for MCP servers and a managed-settings.d drop-in directory. Internally, Segment.io analytics has been completely removed.

Find
Pick an entry · j / k steps through
22 entries

Changesopen

Advisor Tool#

Feature flag
tengu_sage_compass Not enough to say

Nothing here resolved what this flag was doing on this version, so nothing here should be read as on or off.

This account: no value returned · anonymous baseline: no value returned · compiled default in v2.1.83: no gate table built for this version

These values were read against a different version of Claude Code, so treat them as the nearest reading available instead of one taken on this release.

Read once, for one account on one subscription tier, against v2.1.83. It isn't a statement about your account. What a flag value here can and cannot tell you

What

A server-side tool backed by a stronger reviewer model that automatically sees your entire conversation history. When invoked, the advisor reviews your work and provides feedback before you commit to an approach.

Usage
# Enable the advisor with a model
/advisor opus

# Check current advisor status
/advisor

# Disable the advisor
/advisor unset

# Via CLI flag
claude --advisor opus
Details
  • The advisor tool takes no parameters — it automatically receives the full conversation context
  • Designed to be called before substantive work (writing code, committing to an interpretation), when stuck, or when you believe a task is complete
  • Can be disabled via the CLAUDE_CODE_DISABLE_ADVISOR_TOOL environment variable
  • Only available when the server enables this capability; the /advisor command appears when user configuration is permitted
  • Supports Opus 4.6 and Sonnet 4.6 class models as advisor models
Evidence

Advisor tool feature (search for "advisor-tool-2026-03-01", "tengu_sage_compass", "/advisor")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

CwdChanged and FileChanged Hooks#

What

Two new hook event types that trigger automation when the working directory changes or when specific files are modified.

Usage in .claude/hooks.json:

{
  "hooks": {
    "CwdChanged": [{
      "command": "my-script.sh",
      "timeout": 10000
    }],
    "FileChanged": [{
      "command": "on-file-change.sh",
      "matcher": "src/**/*.ts",
      "timeout": 10000
    }]
  }
}
Details
  • CwdChanged: Fires after the working directory changes. Input JSON includes old_cwd and new_cwd. Hook output can include watchPaths to dynamically register additional file watchers via hookSpecificOutput
  • FileChanged: Fires when a watched file changes, is added, or is removed. Input JSON includes file_path and event (one of change, add, unlink). The matcher field specifies file glob patterns to watch
  • Both hook types set CLAUDE_ENV_FILE — write bash exports there to inject environment variables into subsequent BashTool commands
  • File watching uses chokidar under the hood for cross-platform filesystem monitoring
Evidence

New hook types (search for "CwdChanged", "FileChanged", "cwdchanged-hook-", "filechanged-hook-")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Ultrareview (Cloud Code Review)#

What

A rebranded and enhanced cloud-based code review feature. A cloud agent analyzes your branch for ~5–15 minutes, finding and verifying bugs, then delivers findings as a task notification.

Details
  • Launched in the background — users can continue working while the review runs
  • Review output arrives as a task notification when complete; press shift+↓ to view results
  • Requires a git repository with a GitHub remote and OAuth login
  • Replaces the previous "Remote review" terminology
Evidence

Cloud review feature (search for "ultrareview", "~5–15 min", "Cloud agent finds and verifies bugs")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

XAA IdP Authentication for MCP Servers#

What

Enterprise-grade OIDC authentication for MCP servers using the SEP-990 standard. Configure once and all XAA-enabled MCP servers authenticate silently using cached identity tokens.

Usage
# One-time setup: configure IdP connection
claude mcp xaa setup

# Login (opens browser for OIDC flow)
claude mcp xaa login

# Check current connection
claude mcp xaa status

# Clear configuration
claude mcp xaa clear
Details
  • Performs full OIDC discovery from the IdP issuer URL
  • Browser-based authorization with a local callback server on loopback
  • Caches id_token with automatic expiration; use --force to re-login
  • Supports --client-id, --issuer, --callback-port, and --id-token flags
  • Client secret can be read from the MCP_XAA_IDP_CLIENT_SECRET environment variable
  • Settings stored include issuer, clientId, callbackPort, and optionally clientSecret (in secure storage)
Evidence

XAA authentication system (search for "XAA IdP", "claude mcp xaa", "SEP-990")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Enterprise Managed Settings Drop-in Directory#

What

Enterprise administrators can now distribute policy settings via a managed-settings.d/ directory alongside the existing managed-settings.json file.

Details
  • JSON files placed in the managed-settings.d/ directory are merged into the enterprise policy configuration
  • Files are sorted alphabetically and merged in order, allowing layered configuration
  • Only .json files are loaded; hidden files (starting with .) are ignored
  • The health check display now distinguishes between "Enterprise managed settings (file + drop-ins)" when both sources exist
Evidence

Drop-in settings directory (search for "managed-settings.d", "Enterprise managed settings (drop-ins)")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Companion#

What

A small animated ASCII character that sits beside the user's input box and occasionally comments in a speech bubble. Each user gets a unique procedurally-generated companion based on their account.

Details
  • Companions are generated from a seeded random number generator keyed to the user's account UUID
  • Each companion has a species (from a pool of 18 creatures), rarity tier (common through legendary), eye style, optional hat, stats, and a chance to be shiny
  • Activated by server-side configuration (requires the companion config to be set)
  • Can be muted via companionMuted state
Evidence

Companion system (search for "companion_intro", "friend-2026-401", "companionMuted")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

/version Command#

What

A new slash command that prints the version of the currently running session, as distinct from what the auto-updater may have downloaded.

Usage
/version
Evidence

Version command (search for "Print the version this session is running (not what autoupdate downloaded)")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Ultraplan Enhanced with Web-Based Refinement#

Ultraplan now supports an interactive refinement workflow where plans can be generated and refined in Claude Code on the web (claude.ai/code), then brought back to the local CLI for execution. Users see a "No, refine with Ultraplan on Claude Code on the web" option during plan approval, and approved plans are automatically applied locally.

Evidence

Enhanced ultraplan flow (search for "Ultraplan approved in browser", "Refine local plan", "No, refine with Ultraplan on Claude Code on the web")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

MCP Registry for Official Server Validation#

Claude Code now fetches the official Anthropic MCP server registry at startup to validate MCP server URLs. This allows trusted servers to receive enhanced metadata forwarding (server name and tool name in API requests), improving the quality of tool-use responses.

Evidence

MCP registry validation (search for "mcp-registry", "https://api.anthropic.com/mcp-registry/v0/servers")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Plugin Marketplace GCS Distribution#

Official marketplace plugins are now distributed via Google Cloud Storage (downloads.claude.ai) with ZIP-based atomic updates and SHA-based cache validation, providing faster and more reliable plugin installation compared to the previous git-only approach. Falls back to git if GCS is unavailable.

Evidence

GCS plugin distribution (search for "downloads.claude.ai", "Official marketplace GCS")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

useAutoModeDuringPlan Setting#

A new boolean setting controls whether plan mode uses auto mode semantics (automatically accepting edits) when auto mode is available. Defaults to true. Can be configured at the user, local, flag, or policy level.

Evidence

Plan mode setting (search for "useAutoModeDuringPlan", "Whether plan mode uses auto mode semantics")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

Default Teammate Model Setting#

A new teammateDefaultModel setting allows users to configure the default model used when spawning teammate agents. Accessible through the /config settings UI under "Default teammate model".

Evidence

Teammate model config (search for "TeammateModel", "Default teammate model", "teammateDefaultModel")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

ctrl+x ctrl+k Keybinding for Killing Background Agents#

A new chord keybinding ctrl+x ctrl+k kills all running background agents. This supplements the existing ctrl+f toggle (press twice to stop agents).

Evidence

Kill agents keybinding (search for "ctrl+x ctrl+k", "chat:killAgents")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

sandbox.failIfUnavailable Setting#

A new boolean setting that causes Claude Code to exit with an error at startup if sandbox.enabled is true but the sandbox cannot start (missing dependencies, unsupported platform, or platform not in enabledPlatforms). When false (the default), a warning is logged but execution continues.

Evidence

Strict sandbox setting (search for "failIfUnavailable", "sandbox required but unavailable")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

shift+tab Plan Approval Feedback#

In plan mode, shift+tab now allows approving a plan while simultaneously providing feedback text. The input placeholder shows "shift+tab to approve with this feedback", streamlining the plan refinement workflow.

Evidence

Plan feedback approval (search for "shift+tab to approve with this feedback")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Worktree Name Validation Supports Path Segments#

Worktree names now support /-separated path segments (e.g., feature/my-branch). Each segment is independently validated for allowed characters (letters, digits, dots, underscores, dashes). The 64-character maximum length still applies to the full name.

Evidence

Path segment validation (search for "each \"/\"-separated segment must be non-empty")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

Improved Windows Terminal Detection#

Claude Code now detects and supports a broader range of terminal emulators on all platforms for launching new terminal sessions (used by deep links and remote sessions). On Windows, it prioritizes Windows Terminal, then PowerShell 7+, PowerShell 5, and Command Prompt. On macOS and Linux, it detects Ghostty, iTerm2, Alacritty, Kitty, WezTerm, and various others with proper --working-directory flag support.

Evidence

Terminal detection (search for "Windows Terminal", "com.mitchellh.ghostty", "com.googlecode.iterm2")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

MCP Tool Search Hints via anthropic/searchHint#

MCP servers can now provide search hints through the anthropic/searchHint metadata field on tool definitions. These hints help Claude find and select the right tools more effectively by providing natural-language descriptions of when to use each tool.

Evidence

MCP search hints (search for "anthropic/searchHint")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST Environment Variable#

A new environment variable for SDK hosts (desktop apps, IDE extensions) to signal that the API provider is managed by the host application, preventing users from accidentally overriding the provider configuration.

Evidence

Host-managed provider (search for "CLAUDE_CODE_PROVIDER_MANAGED_BY_HOST")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Memory System Prompt Refinements#

The memory guidance system prompt has been significantly rewritten for clarity:

  • MEMORY.md entries are now recommended to be under ~200 characters (up from ~150) with detail moved to topic files
  • Byte-size truncation added alongside line-count truncation for oversized MEMORY.md files
  • New instruction: "If the user says to ignore or not use memory: proceed as if MEMORY.md were empty"
  • HTML comment stripping added when processing MEMORY.md content
Evidence

Memory prompt updates (search for "Keep index entries to one line under ~200 chars")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Plugin Sensitive Options with Secure Storage#

Plugin configuration now distinguishes between non-sensitive values (saved to settings.json) and sensitive values (saved to macOS keychain or .credentials.json). Sensitive fields are scrubbed from settings.json after migration to secure storage.

Evidence

Plugin secrets handling (search for "sensitive)", "to secure storage", "saveMcpServerUserConfig: scrubbed")

Strings lifted out of the shipped bundle, so the claim above can be checked against them.

Related

Other releases about the same thing. Found by shared names or similar wording; neither means one caused the other.

1 entry

Bug Fixesopen

#

  • Fixed atomic file write cleanup: the temp file is now cleaned up even if it doesn't exist at the expected path (search for "Cleaning up temp file")
  • Improved plugin hooks reload: now tracks plugin-affecting settings changes rather than just enabledPlugins to avoid unnecessary reloads (search for "plugin-affecting settings change")
  • Fixed deep link input validation to reject control characters and enforce length limits, preventing potential injection issues (search for "Deep link cwd contains disallowed control characters")
  • Fixed MCP lazy dedup to properly suppress duplicate servers from plugins that overlap with claude.ai connectors (search for "[MCP] Lazy dedup: suppressing")
1 entry

Internal Changesopen

#

  • Segment.io analytics completely removed: The @segment/analytics-node library and all associated code (event factory, priority queue, emitter, stats tracking, UUID generation) have been removed. Telemetry is now handled through a lighter-weight internal system (search for "Segment.io" in old version)
  • Bootstrap API replaces clientData: The /api/oauth/claude_cli/client_data endpoint has been replaced by /api/claude_cli/bootstrap, which also returns additional model options alongside client data (search for "/api/claude_cli/bootstrap")
  • Improved debug repainting: New CLAUDE_CODE_DEBUG_REPAINTS environment variable enables detailed React component owner chain tracking for diagnosing UI repaint issues (search for "CLAUDE_CODE_DEBUG_REPAINTS")
1 entry

Notesopen

#

  • The Advisor Tool, Deep Link Protocol, and Companion features are controlled by server-side configuration and feature flags. They may not be immediately available to all users.
  • The managed-settings.d directory enables enterprise administrators to manage settings across multiple JSON files. Existing managed-settings.json files continue to work unchanged.
  • Ultrareview replaces the "Remote review" terminology from previous versions. The functionality is enhanced with cloud agent-based bug detection.
Verbatim
Official · Anthropic

Anthropic’s official release notes

Published verbatim by Anthropic for v2.1.83. Text is unmodified from the upstream changelog. Everything else on this page came out of the bundle instead, which is why the two lists don't match.

  • Added managed-settings.d/ drop-in directory alongside managed-settings.json, letting separate teams deploy independent policy fragments that merge alphabetically
  • Added CwdChanged and FileChanged hook events for reactive environment management (e.g., direnv)
  • Added sandbox.failIfUnavailable setting to exit with an error when sandbox is enabled but cannot start, instead of running unsandboxed
  • Added disableDeepLinkRegistration setting to prevent claude-cli:// protocol handler registration
  • Added CLAUDE_CODE_SUBPROCESS_ENV_SCRUB=1 to strip Anthropic and cloud provider credentials from subprocess environments (Bash tool, hooks, MCP stdio servers)
  • Added transcript search — press / in transcript mode (Ctrl+O) to search, n/N to step through matches
  • Added Ctrl+X Ctrl+E as an alias for opening the external editor (readline-native binding; Ctrl+G still works)
  • Pasted images now insert an [Image #N] chip at the cursor so you can reference them positionally in your prompt
  • Agents can now declare initialPrompt in frontmatter to auto-submit a first turn
  • chat:killAgents and chat:fastMode are now rebindable via ~/.claude/keybindings.json
  • Fixed mouse tracking escape sequences leaking to shell prompt after exit
  • Fixed Claude Code hanging on exit on macOS
  • Fixed screen flashing blank after being idle for a few seconds
  • Fixed a hang when diffing very large files with few common lines — diffs now time out after 5 seconds and fall back gracefully
  • Fixed a 1–8 second UI freeze on startup when voice input was enabled, caused by eagerly loading the native audio module
  • Fixed a startup regression where Claude Code would wait ~3s for claude.ai MCP config fetch before proceeding
  • Fixed --mcp-config CLI flag bypassing allowedMcpServers/deniedMcpServers managed policy enforcement
  • Fixed claude.ai MCP connectors (Slack, Gmail, etc.) not being available in single-turn --print mode
  • Fixed caffeinate process not properly terminating when Claude Code exits, preventing Mac from sleeping
  • Fixed bash mode not activating when tab-accepting !-prefixed command suggestions
  • Fixed stale slash command selection showing wrong highlighted command after navigating suggestions
  • Fixed /config menu showing both the search cursor and list selection at the same time
  • Fixed background subagents becoming invisible after context compaction, which could cause duplicate agents to be spawned
  • Fixed background agent tasks staying stuck in "running" state when git or API calls hang during cleanup
  • Fixed --channels showing "Channels are not currently available" on first launch after upgrade
  • Fixed uninstalled plugin hooks continuing to fire until the next session
  • Fixed queued commands flickering during streaming responses
  • Fixed slash commands being sent to the model as text when submitted while a message is processing
  • Fixed scrollback jumping when collapsed read/search groups finish after scrolling offscreen
  • Fixed scrollback jumping to top when the model starts or stops thinking
  • Fixed SDK session history loss on resume caused by hook progress/attachment messages forking the parentUuid chain
  • Fixed copy-on-select not firing when you release the mouse outside the terminal window
  • Fixed ghost characters appearing in height-constrained lists when items overflow
  • Fixed Ctrl+B interfering with readline backward-char at an idle prompt — it now only fires when a foreground task can be backgrounded
  • Fixed tool result files never being cleaned up, ignoring the cleanupPeriodDays setting
  • Fixed space key being swallowed for up to 3 seconds after releasing voice hold-to-talk
  • Fixed ALSA library errors corrupting the terminal UI when using voice mode on Linux without audio hardware (Docker, headless, WSL1)
  • Fixed voice mode SoX detection on Termux/Android where spawning which is kernel-restricted
  • Fixed Remote Control sessions showing as Idle in the web session list while actively running
  • Fixed footer navigation selecting an invisible Remote Control pill in config-driven mode
  • Fixed memory leak in remote sessions where tool use IDs accumulate indefinitely
  • Improved Bedrock SDK cold-start latency by overlapping profile fetch with other boot work
  • Improved --resume memory usage and startup latency on large sessions
  • Improved plugin startup — commands, skills, and agents now load from disk cache without re-fetching
  • Improved Remote Control session titles: AI-generated titles now appear within seconds of the first message
  • Improved WebFetch to identify as Claude-User so site operators can recognize and allowlist Claude Code traffic via robots.txt
  • Reduced WebFetch peak memory usage for large pages
  • Reduced scrollback resets in long sessions from once per turn to once per ~50 messages
  • Faster claude -p startup with unauthenticated HTTP/SSE MCP servers (~600ms saved)
  • Bash ghost-text suggestions now include just-submitted commands immediately
  • Increased non-streaming fallback token cap (21k → 64k) and timeout (120s → 300s local) so fallback requests are less likely to be truncated
  • Interrupting a prompt before any response now automatically restores your input so you can edit and resubmit
  • /status now works while Claude is responding, instead of being queued until the turn finishes
  • Plugin MCP servers that duplicate an org-managed connector are now suppressed instead of running a second connection
  • Linux: respect XDG_DATA_HOME when registering the claude-cli:// protocol handler
  • Changed "stop all background agents" keybinding from Ctrl+F to Ctrl+X Ctrl+K to stop shadowing readline forward-char
  • Deprecated TaskOutput tool in favor of using Read on the background task's output file path
  • Added CLAUDE_CODE_DISABLE_NONSTREAMING_FALLBACK env var to disable the non-streaming fallback when streaming fails
  • Plugin options (manifest.userConfig) now available externally — plugins can prompt for configuration at enable time, with sensitive: true values stored in keychain (macOS) or protected credentials file (other platforms)
  • Claude can now reference the on-disk path of clipboard-pasted images for file operations
  • Ctrl+L now clears the screen and forces a full redraw — use this to recover when Cmd+K leaves the UI partially blank. Use Ctrl+U or double-Esc to clear prompt input.
  • --bare -p (SDK pattern) is ~14% faster to the API request
  • Memory: MEMORY.md index now truncates at 25KB as well as 200 lines
  • Disabled AskUserQuestion and plan-mode tools when --channels is active
  • Fixed API 400 error when a pasted image was queued during a failing tool call
  • Fixed MCP tool calls hanging indefinitely when an SSE connection drops mid-call and exhausts its reconnection attempts
  • Fixed Remote Control session titles showing raw XML when a background agent completed before the first user message
  • Fixed remote sessions forgetting conversation history after a container restart due to progress-message gaps in the resumed transcript chain
  • Fixed remote sessions requiring re-login on transient auth errors instead of retrying automatically
  • Fixed rg ... | wc -l and similar piped commands hanging and returning 0 in sandbox mode on Linux
  • Fixed voice input hold-to-talk not activating when a CJK IME inserts a full-width space
  • Fixed --worktree hanging silently when the worktree name contained a forward slash
  • [VSCode] Spinner now turns red with "Not responding" when the backend hasn't responded for 60 seconds
  • [VSCode] Fixed session history not loading correctly when reopening a session via URL or after restart
  • [VSCode] Added Esc-twice (or /rewind) to open a keyboard-navigable rewind picker
  • [VSCode] Fixed "Fork conversation from here" and rewind actions failing silently after the session cache goes stale
Read it on GitHub →
System prompt

The system prompt was not captured for this release, so this page cannot say whether it moved.

Which releases were captured →