Version 2.1.76

A new built-in slash command to view and set the model effort level directly from the REPL, without navigating into /config.

/effort              # Show current effort level
/effort medium       # Set to medium
/effort max          # Maximum reasoning depth (Opus 4.6 only)
/effort auto         # Reset to automatic (default)

• Effort levels: low, medium, high, max, auto
• Named levels are persisted to user settings (effortLevel in settings.json)
• auto clears the setting, returning to the model default
• Each level has a descriptive summary (e.g. "Balanced approach with standard implementation and testing")

/effort command registration (search for name: "effort", description: "Set effort level for model usage")

A new hook event that fires after conversation compaction completes, letting you run shell commands or prompts in response to compaction.

• Hook event name: PostCompact
• Input includes compaction details and the summary text
• Allows post-processing after compaction (e.g., logging, notification, or custom follow-up actions)
• Hook output: exit code 0 shows stdout to user; other exit codes show stderr only

PostCompact hook event definition (search for hook_event_name: C.literal("PostCompact"), "Running PostCompact hooks…")

A new category of hooks that are registered internally by Claude Code, distinct from user-configured and plugin hooks.

• Shown separately in the hooks viewer as "Built-in Hooks" / "Built-in hooks (registered internally by Claude Code)"
• Cannot be modified by users—they are part of the Claude Code runtime
• Visible in the /hooks UI for transparency

Built-in hooks category (search for "Built-in hooks (registered internally by Claude Code)")

Built-in skills can now declare reference files that are extracted to a temporary directory at runtime, making them available for the skill prompt to reference.

• Skills define a files property in their registration, mapping relative paths to file content
• Files are extracted to a per-skill directory inside the Claude Code data directory
• Path traversal is blocked (bundled skill file path escapes skill dir: error)
• Files are written with O_NOFOLLOW to prevent symlink attacks
• The skill prompt is prefixed with Base directory for this skill: <path>

Bundled skill file extraction (search for "Bundled skill reference files are allowed for reading", "Failed to extract bundled skill '")

The /hooks command has been redesigned from a full hook management interface (add, edit, delete) to a read-only viewer. Users can browse hook events, matchers, and see hook details, but can no longer modify hooks through the UI.

What changed:

• Removed: "Add new hook", "Delete hook?", "Delete matcher?", "Save hook configuration", "Where should this hook be saved?" dialogs
• Added: "Hook details" detail view showing event, matcher, type, source, command/prompt, and status message
• Added: "This menu is read-only. To add or modify hooks, edit settings.json directly or ask Claude."
• Added: "To re-enable hooks, remove disableAllHooks from settings.json or ask Claude."
• Navigation: "Esc to go back" / "Esc to close" guides
• Description changed from "Manage hook configurations" to "View hook configurations"

Read-only hooks viewer (search for "View hook configurations", "This menu is read-only", "Hook details")

The --plugin-dir flag has been updated from accepting a single paths argument to being individually repeatable.

What changed:

• Old: --plugin-dir <paths...> (space-separated paths in one argument)
• New: --plugin-dir <path> (repeat the flag: --plugin-dir A --plugin-dir B)
• Session-only plugins are now labeled as "Session-only plugins (--plugin-dir):" in the UI
• The reserved marketplace name "inline" is used internally for --plugin-dir session plugins

Repeatable plugin-dir flag (search for "Load plugins from a directory for this session only (repeatable: --plugin-dir A --plugin-dir B)")

The preferred language setting description has been expanded to explicitly include voice dictation.

What changed:

• Old: Preferred language for Claude responses (e.g., "japanese", "spanish")
• New: Preferred language for Claude responses and voice dictation (e.g., "japanese", "spanish")
• New dictation language display: shows current dictation language with unsupported-language warning

Language setting update (search for "Preferred language for Claude responses and voice dictation", "Dictation language:")

New error messages provide better guidance when an ANTHROPIC_API_KEY belongs to a disabled organization.

• Subscription users: "Your ANTHROPIC_API_KEY belongs to a disabled organization · Unset the environment variable to use your subscription instead"
• Other users: "Your ANTHROPIC_API_KEY belongs to a disabled organization · Update or unset the environment variable"
• Organization verification errors now include actionable guidance about token scope and managed settings

Disabled org handling (search for "organization has been disabled", "Your ANTHROPIC_API_KEY belongs to a disabled organization")

The transcriptRetentionDays setting now has stricter behavior when set to 0.

What changed:

• Old: "Number of days to retain chat transcripts (0 to disable cleanup)" — 0 meant keep forever
• New: "Number of days to retain chat transcripts (default: 30). Setting to 0 disables session persistence entirely: no transcripts are written and existing transcripts are deleted at startup."
• This is a behavioral change: setting 0 previously preserved transcripts forever; now it deletes them

Transcript retention change (search for "disables session persistence entirely")

Agent worktrees that are no longer active are now automatically cleaned up.

• Worktrees with no uncommitted changes and no unpushed commits are removed
• The current session's worktree is never cleaned up
• Cleanup runs git worktree prune after removal
• Logged as: cleanupStaleAgentWorktrees: removed N stale worktree(s)

Worktree cleanup (search for "cleanupStaleAgentWorktrees")

Conversation compaction can now preserve a segment of recent messages instead of summarizing everything.

• Compacted boundaries include a preservedSegment with headUuid, anchorUuid, and tailUuid
• Loaders splice the preserved segment at the anchor point so resumed sessions include actual message content
• Metadata is serialized as preserved_segment in transcript files

Preserved segment in compaction (search for "preservedSegment", "Relink info for messagesToKeep")

The Read tool system prompt now encourages partial file reads when the needed section is known.

What changed:

• Added: "When you already know which part of the file you need, only read that part. This can be important for larger files."
• This supplements the existing guidance about offset/limit parameters

Read tool guidance (search for "When you already know which part of the file you need")

Full PDF reading (without specifying page ranges) now requires the Anthropic API and poppler-utils.

What changed:

• Old: "PDF reading is not supported. Install poppler-utils..."
• New: "Reading full PDFs is only supported with the Anthropic API. Use the pages parameter to read specific page ranges (e.g., pages: \"1-5\", maximum N pages per request). This requires poppler-utils: install with brew install poppler on macOS or apt-get install poppler-utils on Debian/Ubuntu."

PDF reading message (search for "Reading full PDFs is only supported with the Anthropic API")

Plugin dependencies can now be specified as bare names (without @marketplace) and are resolved against the declaring plugin's own marketplace.

• Dependency must be a plugin name, optionally qualified with @marketplace
• Bare names resolve against the parent plugin's marketplace source
• The "inline" marketplace name is reserved for --plugin-dir session plugins

Plugin dependency resolution (search for "Dependency must be a plugin name, optionally qualified with @marketplace")

A new reinitializeLspServerManager() function allows the LSP server manager to be restarted without restarting the entire session, useful when plugin configurations change.

LSP reinit (search for "[LSP MANAGER] reinitializeLspServerManager() called")

A dedicated McpToolCallError error type has been added for MCP (Model Context Protocol) tool call failures, providing better error categorization.

Error type (search for "McpToolCallError")

The ExitPlanMode tool description now explicitly states it's only for plan mode: "present plan for approval and start coding (plan mode only)". An error message was also added: "You are not in plan mode. This tool is only for exiting plan mode after writing a plan."

ExitPlanMode update (search for "present plan for approval and start coding (plan mode only)")

• Fixed autocompact circuit breaker to stop retrying after repeated failures: "autocompact: circuit breaker tripped after N consecutive failures — skipping future attempts this session" (search for "circuit breaker tripped")
• Prompt/memory count now only counts messages after the last compaction boundary, fixing inflated counts in long sessions (search for "compact_boundary" in UFY)
• Tool schema discovery warning now explains why typed parameters fail and how to fix it: "This tool's schema was not sent to the API — it was not in the discovered-tool set" (search for "This tool's schema was not sent to the API")

A background agent that automatically extracts and saves memories from conversations, running alongside the main conversation without user intervention.

Feature-flagged behind tengu_passport_quail (default: false)

• After each turn, a background agent analyzes recent messages and saves relevant information to memory files
• Uses the tengu_bramble_lintel flag to control how many turns to accumulate before running (default: 1)
• Supports both private (user-only) and team (shared) memory scopes
• Team memory format controlled by tengu_swinburne_dune flag (alternative two-step format)
• Tracks which files the conversation has already written to, skipping extraction if the user manually saved memories
• Coalesces rapid extractions: if one is already running, subsequent requests are stashed for a trailing run
• Sends notification: "Saved N memor(y|ies)" when memories are written
• Memory extraction agent is restricted to only Write, Read, and Edit tools within the memory directory
• System prompt: "A background agent automatically extracts and saves memories from this conversation. If the user asks you to remember or forget something, acknowledge it — the save happens automatically."

Memory extraction system (search for "tengu_passport_quail", "[extractMemories]", "extract_memories")

An in-session quality survey that can appear when a session is eligible.

Feature-flagged — controlled by a probability setting (suggested starting point: 0.05)

• Configuration: "Probability (0–1) that the session quality survey appears when eligible. 0.05 is a reasonable starting point."
• No evidence of the survey UI itself in the diff, suggesting infrastructure only

Survey config (search for "session quality survey")

Ability to cancel a pending async user message from the command queue by UUID.

Infrastructure added

• Operation: cancel_async_message drops a pending message that hasn't been dequeued yet
• Result includes cancelled boolean: false means the message was already dequeued or never enqueued
• Related to Remote Control bridge functionality

Async message cancellation (search for "cancel_async_message", "Drops a pending async user message")

A /bridge-kick debug command for injecting failure states into the Remote Control bridge for manual recovery testing.

Registered but explicitly disabled (isEnabled: () => !1)

• Subcommands: close, poll, register, reconnect-session, heartbeat, reconnect, status
• Only available when Remote Control is connected (USER_TYPE=ant)
• Intended for internal Anthropic debugging, not user-facing

Bridge debug command (search for "bridge-kick", "Inject bridge failure states")