Version 2.1.90

A new /powerup command that launches a set of 10 interactive lessons teaching Claude Code features that "most people miss." Each power-up includes animated terminal demos, explanatory text, and can be marked as completed.

/powerup

• 10 lessons covering: @ file mentions, permission modes (shift+tab), undo/rewind, background tasks, CLAUDE.md memory, MCP tool integration, skills & hooks, subagents, remote control/teleport, and model/effort switching
• Progress is persisted — unlocked power-ups carry across sessions (stored via powerupsUnlocked in local settings)
• A celebration animation plays when all 10 are completed ("All powered up")
• Suggested to new users on the welcome screen when the terminal is tall enough (≥44 rows): "New here? Run /powerup to learn the features most people miss."
• Telemetry events: tengu_powerup_lesson_opened, tengu_powerup_lesson_completed

Power-up command and lesson system (search for "Discover Claude Code features through quick interactive lessons")

Claude Code now supports "Claude Platform on AWS" as a first-class API provider, alongside Bedrock, Vertex, and Foundry.

CLAUDE_CODE_USE_ANTHROPIC_AWS=1 claude

• Set CLAUDE_CODE_USE_ANTHROPIC_AWS=1 to enable the provider
• Requires ANTHROPIC_AWS_WORKSPACE_ID (or workspaceId in constructor)
• Optionally set ANTHROPIC_AWS_BASE_URL or let it derive from AWS_REGION / AWS_DEFAULT_REGION (pattern: https://aws-external-anthropic.{region}.api.aws)
• Supports AWS SigV4 request signing via @aws-sdk/credential-providers
• Supports explicit credentials via awsAccessKey + awsSecretAccessKey, AWS profile, or a custom providerChainResolver
• Set CLAUDE_CODE_SKIP_ANTHROPIC_AWS_AUTH=1 to skip authentication (for development/testing)
• Fast mode is not available on this provider (same as Bedrock, Vertex, and Foundry)

AnthropicAws client class (search for "Claude Platform on AWS" and "ANTHROPIC_AWS_WORKSPACE_ID")

A new /toggle-memory slash command that disables or re-enables automemory for the current session.

/toggle-memory

• When toggled off: "this conversation will not write or read new memories, and previously-loaded memory content should not be referenced"
• When toggled back on: "memory content may be referenced and new memories can be saved"
• Memory read/write operations are blocked while toggled off, returning guidance to run /toggle-memory to re-enable
• Fires tengu_memory_toggled telemetry with toggled_off boolean
• The command is isHidden: false so it appears in command listings, but isEnabled: () => false (the toggle itself is always available; the isEnabled likely gates something else)

Toggle-memory command definition (search for "Toggle automemory off/on for this session")

The sandbox proxy can now route outbound traffic through an upstream (parent) HTTP proxy, enabling use in corporate environments with mandatory egress proxies.

• Configured via the network.parentProxy setting, with sub-keys for http, https, and noProxy
• Falls back to standard HTTP_PROXY / HTTPS_PROXY / NO_PROXY environment variables when the setting is not explicitly configured
• NO_PROXY supports hostname suffixes, wildcard *, and CIDR ranges (both IPv4 and IPv6)
• Localhost and loopback addresses always bypass the parent proxy
• CONNECT tunneling through the parent proxy for HTTPS traffic, with proper authentication (Basic auth from proxy URL credentials)
• Hop-by-hop headers are stripped when forwarding through the proxy
• Logs configured proxy URLs at startup: Parent proxy configured: http=... https=...

Parent proxy routing logic (search for "Upstream HTTP proxy for outbound connections" and "Parent proxy configured")

• The two-step memory save process (write file, then update MEMORY.md index) has been simplified — the MEMORY.md index step has been removed, and memory is saved directly to individual files. The number of MEMORY.md references decreased from 10 to 6 between versions.
• Memory retrieval now prefixes content with "Retrieved for possible relevance — use only if it actually applies to what the user asked" to reduce false positive influence.
• Memory selection is now more conservative with user-profile and project-overview memories: the system explicitly avoids matching on surface keyword overlap (e.g., a profile mentioning "DB performance" won't be selected just because a question contains the word "performance").
• Memory extraction now skips runs where no user prose has been added since the last extraction, reducing unnecessary extraction cycles (search for "[extractMemories] skipping — no user prose since last extraction").
• The extraction prompt now includes "If nothing is worth saving, output only 'Nothing to save.' Do not explain why." for cleaner no-op results.

Memory selection system prompt (search for "Be especially conservative with user-profile") and extraction guard (search for "no user prose since last extraction")

When a PostToolUse hook modifies a file that Claude just edited (e.g., a formatter running after a write), Claude Code now automatically detects the change and re-syncs its internal read-file state. This prevents stale-file errors on the next Edit tool call for that file.

Hook file sync handler (search for "PostToolUse hook modified" and "re-synced readFileState")

Docker commands now reject --host, --context, --config, and --tls flags, as well as short flags -H and -c, which could be used to redirect Docker operations to a remote daemon and escape the sandbox.

Docker flag validation (search for startsWith("--host") near the docker command validator)

• Background job operator (&) detection: PowerShell commands using the & operator now trigger a security prompt with the message "Command uses the background job operator (&) which spawns a child PowerShell process" (search for "background job operator").
• Remove-Item -Recurse safety: Deleting the working directory (including .git and .claude) via Remove-Item -Recurse now requires manual approval (search for "Remove-Item -Recurse targeting").
• Action parameter validation: PowerShell common parameters like -ErrorAction, -WarningAction (and their aliases -ea, -wa, etc.) are now validated to ensure they use expected values (SilentlyContinue, Stop, Continue, Ignore), preventing abuse of these parameters.

PowerShell AST parser (search for "hasBackgroundJob" and "background job operator")

• Mount point cleanup is now deferred when multiple sandboxes are still active, using a reference counter to avoid premature cleanup (search for "Deferring mount point cleanup").
• Added --unshare-user flag in certain sandbox configurations for tighter process isolation.
• Write paths that get wiped by a denyRead tmpfs mount are now automatically re-bound (search for "Re-bound write path wiped by denyRead tmpfs").
• Deny path deduplication prevents redundant sandbox mount operations.

Sandbox refcount (search for "sandbox(es) still active") and re-bind logic (search for "Re-bound write path wiped by denyRead tmpfs")

The sandbox proxy now provides more specific error messages for connection failures:

• "CONNECT handshake timed out" — timeout during proxy CONNECT
• "Proxy closed during CONNECT handshake" — premature close
• "Proxy refused CONNECT: ..." — non-2xx response from proxy
• "CONNECT response header too large" — headers exceeding 16KB
• "Denying malformed host: ..." / "Rejecting malformed SOCKS host: ..." — input validation
• "Invalid destination host for CONNECT" / "Invalid destination port" — bad target validation
• "Invalid parent proxy URL, ignoring: ..." — misconfigured parent proxy URL
• "unsupported scheme or empty host" — bad proxy URL scheme

CONNECT tunnel error handling (search for "CONNECT handshake timed out" and "Proxy refused CONNECT")

Users who cannot upgrade their plan now see a clearer message: "Upgrade not currently available. For additional usage, run /extra-usage." instead of a generic error.

Upgrade handler (search for "Upgrade not currently available")

Debug logging now records auto-mode permission opt-in status, including checks across user, local, flag, and policy settings scopes, making it easier to diagnose permission configuration issues.

Auto-mode logging (search for "[auto-mode] hasAutoModeOptIn")

• Fixed network allowlist hostname matching to correctly reject IP addresses for wildcard domain patterns (e.g., *.github.com no longer incorrectly matches bare IP addresses) (search for "qI_" in the NC1 function)
• Removed the FetchHttpHandler polyfill and several duplicate URI encoding/query string builder modules that were dead code from the AWS SDK bundling (removed bc7, rn7, D9q, f9q, N9q, y9q, B9q functions)

A companion feature referenced as "buddy" that appears to be gated by configuration.

Dark-launched / configuration-gated

• The only visible artifact is the message: "buddy is unavailable on this configuration"
• Appears to be checked via Fd8() and returns null when unavailable
• No further implementation details are visible in this version

Buddy availability check (search for "buddy is unavailable on this configuration")

Infrastructure for synchronizing conversation transcripts to a server for session continuity across devices.

Feature-flagged / server-controlled

• Uploads main conversation events and subagent entries to a persistence backend
• Tracks events since last compaction, skips oversized entries
• Integrated with the bridge:repl system for session forwarding
• Log prefix [persistence-sync] for debugging
• CCR v2 internal event writer registered for transcript persistence

Persistence sync system (search for "[persistence-sync] Uploaded" and "CCR v2 internal event writer")

• The full HTML generation for "Claude Code Insights" reports was removed from the CLI bundle. If you used /insights to generate local HTML reports, this functionality may have moved to the web interface.
• The /resume command now filters out loop sessions and entrypoint-based sessions from the resume picker (search for "filtered from /resume").