Version 2.1.92

A comprehensive multi-step wizard that walks you through configuring AWS Bedrock authentication, region selection, credential verification, and model pinning — all from within the CLI.

/setup-bedrock

• Supports four authentication methods: AWS SSO profile, Bedrock API key (bearer token), access key + secret, or existing environment credentials
• Auto-discovers AWS profiles from ~/.aws/config and ~/.aws/credentials
• Verifies credentials by calling STS GetCallerIdentity and listing Bedrock inference profiles
• Model pinning step tests each candidate (Sonnet, Opus, Haiku) with a one-token request to verify availability
• Saves configuration to ~/.claude/settings.json under env:
• Handles SSO session expiry, invalid credentials, and missing permissions with specific guidance
• Only visible when CLAUDE_CODE_USE_BEDROCK=1 is already set

Bedrock wizard UI components (search for "Set up AWS Bedrock", "Bedrock configuration saved to ~/.claude/settings.json", "How do you authenticate to AWS?")

Resume a Claude Code session that was started on claude.ai/code directly in your local CLI.

/teleport
# or use the alias:
/tp

• Loads a remote session's conversation history into your local Claude Code instance
• Requires remote sessions to be enabled (allow_remote_sessions policy)
• Only available when authenticated
• Shows "Session resumed successfully" on completion or "Teleport cancelled" on abort

Teleport command registration (search for "Resume a Claude Code session from claude.ai", "Teleport cancelled")

Set a session-only prompt-based Stop hook with a quick one-liner, allowing you to define a stopping condition that Claude evaluates after each turn.

/stop Has Claude completed all requested tasks?

• Creates a prompt-type hook on the "Stop" event for the current session only
• If a Stop hook already exists, shows the current condition and lets you edit, update, or delete it
• Submitting an empty prompt clears the hook
• Tab switches focus between the input and a "Delete this hook" option when editing an existing hook
• The hook condition is evaluated by Claude after each turn — if met, Claude stops

Stop hook command and UI (search for "Set a session-only Stop hook with a quick prompt", "Stop hook set", "Stop hook cleared", "tengu_stop_hook_command")

The /advisor command is now a full interactive dialog with rich descriptions, replacing the previous text-based interface.

/advisor           # Opens the advisor selection dialog
/advisor opus      # Set advisor directly
/advisor off       # Disable advisor

• New dialog title "Advisor Tool" with explanatory text about how the advisor works
• Explains: "When Claude needs stronger judgment — a complex decision, an ambiguous failure, a problem it's circling without progress — it escalates to the advisor model for guidance, then resumes"
• Highlights the Sonnet + Opus pairing: "near-Opus performance with reduced token usage"
• Shows a warning if the current main model doesn't support the advisor
• Valid advisor options are "opus" and "sonnet"
• Description changed from "Configure the advisor model" to "Configure the Advisor Tool to consult a stronger model for guidance at key moments during a task"
• Advisor enablement is no longer gated by canUserConfigure from tengu_sage_compass; now uses a simpler nb() check

Advisor dialog redesign (search for "Configure the Advisor Tool to consult a stronger model for guidance at key moments during a task", "Advisor Tool", "When Claude needs stronger judgment")

The /release-notes command now shows an interactive version picker instead of dumping all notes at once.

/release-notes

• Presents a scrollable list of versions (up to 10 visible at once)
• "Show all" option to display all versions' notes combined
• Select a specific version to view only its notes
• Versions are sorted by semver (newest first)

Release notes interactive picker (search for "Select a version to view its notes.")

A new notification appears when your context window grows large, showing approximate uncached token count and suggesting /clear.

• Triggers when uncached tokens exceed ~50,000
• Only shows if more than a certain time has passed since session start
• Format: ~Nk uncached · /clear to start fresh
• Helps users understand when conversation context is getting expensive

Context notification function (search for "k uncached", "/clear to start fresh")

The Write tool now advises: "For appending to an existing file, prefer shell redirection via Bash (e.g. echo "..." >> file) over rewriting the whole file." The old mode:'append' parameter has been removed from the Write tool schema.

Write tool description change (search for "For appending to an existing file, prefer shell redirection")

A new formatChord system renders keyboard shortcuts with platform-aware formatting. Supports three display styles: default (e.g., ctrl+A), compact (e.g., ^a), and symbol (e.g., ⌃A). Mac-specific glyphs (⌘, ⌥) are used when the platform is macOS. Chord sequences and arrow key groups are formatted intelligently.

Keyboard shortcut formatting (search for "Enter", "Esc", "⏎", "⌃" in the new key map definition at ~line 316314)

The OAuth flow now automatically appends offline_access to the authorization scope when the server supports it, enabling longer-lived refresh tokens.

OAuth scope handling (search for "Appended offline_access to authorization scope")

Organizations can now require that remote managed settings are freshly fetched before the CLI starts. If the fetch fails, the CLI exits with a clear message instead of proceeding with stale settings.

Managed settings enforcement (search for "Your organization requires remote managed settings to load", "When set in managed settings, the CLI blocks startup")

Scheduled tasks in proactive mode now validate that the cron expression doesn't fire more frequently than the configured minTickIntervalMinutes, preventing overly aggressive scheduling.

Cron validation (search for "Cron period too short")

When the auto-compact threshold is set by a server experiment (not user config), a notification now shows the threshold and how to override it: compacted at N · override with CLAUDE_CODE_AUTO_COMPACT_WINDOW=1000000.

Auto-compact notification (search for "override with CLAUDE_CODE_AUTO_COMPACT_WINDOW")

The memory index now handles both private and team memories under a single path structure. Private memories use file.md paths and team memories use team/file.md. The old separate "shared team memory" context tag has been removed.

Memory path description (search for "indexes both private and team memories")

A new sandbox setting lets users specify additional XPC/Mach service names to allow, supporting tools like 1Password CLI, Playwright, or other system services. Supports trailing-wildcard matching (e.g., "2BUA8C4S2C.com.1password.*").

Mach service setting (search for "macOS only: Additional XPC/Mach service names")

The CLAUDE_CODE_MAX_RETRIES environment variable is now validated to ensure it parses to a finite integer ≥ 0, preventing invalid values from causing unexpected behavior.

Retry validation (search for "Number.isFinite" near the CLAUDE_CODE_MAX_RETRIES handler)

Worktree cleanup now writes a CLAUDE_BASE marker to track the baseline commit, enabling accurate detection of commits ahead and uncommitted changes when deciding whether to auto-clean a worktree.

Worktree baseline tracking (search for "CLAUDE_BASE", "commitsAhead=")

The --remote-control-session-name-prefix flag now defaults to the hostname and is documented as configurable.

Remote control prefix (search for "Prefix for auto-generated Remote Control session names")

• Fixed compact streaming error messages to include hasStartedStreaming state for better debugging (search for "Compact streaming failed. hasStartedStreaming=")
• Hook condition evaluation prompt now includes structured response format guidance, requiring JSON with ok and reason fields (search for "You are evaluating a hook condition in Claude Code")
• Worktree removal hooks now report specific failure messages distinguishing between agent worktrees and regular worktrees (search for "WorktreeRemove hook did not remove")
• Team memory sync now gracefully handles inaccessible team directories instead of failing (search for "team-memory-sync: team dir inaccessible")
• Voice recording cancellation now explicitly discards without submitting (search for "cancelRecording: discarding without submit")

When identical tool results appear multiple times in a conversation, subsequent occurrences are replaced with a short reference to the original, saving context window space.

Feature-flagged behind tengu_onyx_basin_m1k (default: false)

• Hashes tool result content and tracks seen results with short IDs (r1, r2, etc.)
• Duplicate results are replaced with: <identical to result [rN] from your {tool} call earlier — refer to that output>
• Only applies to results longer than 256 characters and shorter than a configurable max
• Tracks telemetry via tengu_tool_result_dedup

Tool result deduplication (gated by tengu_onyx_basin_m1k, search for "result-id:", "identical to result")

Tool results will include [rerun: bN] footers, and future Bash calls can use {rerun: 'bN'} to exactly replay a prior command without retyping it.

Feature-flagged behind tengu_velvet_anchor (default: false)

• Each Bash command gets assigned a short alias (b1, b2, etc.)
• The rerun field and command field are mutually exclusive
• Error message for invalid alias: Unknown rerun alias 'bN'. Valid aliases this session: ...

Rerun alias system (gated by tengu_velvet_anchor, search for "Unknown rerun alias", "rerun")

A hands-free voice interaction mode with hold-to-talk and tap-to-toggle modes.

Feature-gated behind VOICE_HANDSFREE; internal-only annotation states "Hidden from public SDK types until external launch."

• Three modes: hold (hold to talk, default), tap (tap to start/stop+submit), and off
• Setting for "Submit the prompt when hold-to-talk is released (hold mode only)"
• Described as "Voice handsfree settings; behavior gated at read sites by feature(VOICE_HANDSFREE)"

Voice handsfree settings (search for "VOICE_HANDSFREE", "'hold' (default): hold to talk")

A setting to automatically activate autonomous background operation at launch for entitled users.

New setting infrastructure added; activation depends on entitlement.

• When true, proactive mode activates automatically at launch (if entitled)
• When false or null, the user must opt in via /proactive or --proactive
• Described as "Autonomous background operation configuration"

Proactive auto-activate setting (search for "Autonomous background operation configuration", "When true, autonomous background operation is activated automatically at launch")

The /tag command for toggling searchable tags on sessions has been entirely removed. All related UI (tag confirmation dialog, tag display in resume, tag-based search prioritization) has been removed.

The /vim command for toggling between Vim and Normal editing modes has been removed.

The Sleep tool (for waiting/resting) has been removed from the tool set.

The old approach of shipping pre-generated BPF filter files (unix-block.bpf) for seccomp sandboxing has been replaced with a multicall binary approach using /proc/self/exe and ARGV0 dispatch. The vendor/seccomp/ directory and architecture-specific filter lookup code have been removed.

The infrastructure for auto-improving skill definition files based on user preferences has been removed (the "You are editing a skill definition file" prompt and related code).

The --session-timeout CLI flag has been removed.