Version 2.1.119

A full-screen interactive TUI dashboard for managing all background agent sessions from a single place.

claude agents

• Displays all background sessions with live status: working, blocked, done, idle
• Jobs can be grouped by directory or state (toggle with Ctrl+S)
• Sessions can be pinned to the top of the list (Ctrl+T) for quick access; pins persist across sessions in .pins.json
• Sessions can be renamed inline (Ctrl+R)
• Sessions can be deleted (Ctrl+X to arm, press again to confirm; also stops running sessions)
• Sessions can be reordered within a group (Shift+Up/Down)
• A peek panel (Space) shows a split-view with reply input and session details
• An input field at the bottom lets you describe a new task to dispatch a new agent session — supports @template autocomplete and /skill slash commands
• PR status integration: for jobs monitoring PRs, shows CI pass/fail, review status, additions/deletions
• Quick-open numbered sessions with Cmd+1-9
• Help panel via ? shows all keybindings
• Empty state shows helpful suggestions: "paste a link, or 'review PR #123 for bugs' · 'fix the failing test' · 'babysit my PR until CI passes'"

Fleet view system (search for "claude agents", ".fleetview-heartbeat", "Agents here keep running")

A suite of new top-level CLI commands for managing background sessions without entering the interactive fleet view.

claude attach <id>    # Attach to a running background session
claude kill <id>      # Stop a background session
claude logs <id>      # Show recent output from a session
claude rm <id>        # Remove worktree and job state for a stopped session
claude respawn <id>   # Respawn an exited session

• After backgrounding a session, these commands are displayed as a quick reference
• claude kill retains the worktree (if used); prints "worktree retained at <path>" with a hint to use claude rm <id> for full cleanup
• claude rm removes both the worktree and job state files
• claude respawn --all respawns all exited sessions at once

Session management commands (search for "claude attach <id>", "claude kill <id>", "claude rm <id>", "claude logs <id>")

Install the Claude daemon as a persistent system service (launchd on macOS, systemd on Linux) so it runs automatically on boot and survives terminal closures.

claude daemon install     # Install as a system service
claude daemon uninstall   # Remove the system service
claude daemon start       # Start the daemon
claude daemon stop        # Stop the daemon
claude daemon restart     # Restart the daemon
claude daemon status      # Show daemon status and PID
claude daemon log         # View daemon logs

• On macOS, installs as a launchd agent (com.anthropic.claude-daemon)
• On Linux, installs as a systemd user unit with Restart=always and rate limiting (10 bursts per 60s)
• The daemon manages all background workers: scheduled tasks, remote control servers, assistants, and heartbeat workers
• Hot-reloads daemon.json configuration automatically (chokidar watcher with debounce)
• Centralized OAuth token management — proactively refreshes tokens, broadcasts to workers via IPC
• If headless auth fails, writes daemon-auth-status.json and sends a desktop notification prompting claude auth login
• Logs to ~/.claude/daemon.log
• Idle exit: daemon shuts down after a configurable period with no clients (unless installed as a service)

Daemon service management (search for "Install service", "Uninstall service", "com.anthropic.claude-daemon", "daemon.status.json")

A new JSON configuration file at ~/.claude/daemon.json that defines persistent background workers.

{
  "scheduled": [{
    "tasks": [{
      "id": "my-task",
      "cron": "*/30 * * * *",
      "prompt": "Check for new issues",
      "directory": "/path/to/project",
      "permissionMode": "dontAsk"
    }],
    "maxConcurrent": 1
  }],
  "remoteControl": [{
    "dir": "/path/to/project",
    "spawnMode": "worktree",
    "capacity": 32
  }]
}

• Worker kinds: scheduled (cron tasks), remoteControl (bridge to claude.ai), heartbeat (health monitoring)
• Configuration changes are hot-reloaded — the daemon diffs old vs new config and starts/stops/restarts only changed workers
• Unknown config keys are reported but don't block startup
• Interactive management via the /daemon slash command with tabbed UI for Scheduled Tasks, Remote Control, and Assistants
• Each tab supports add, edit, remove, enable/disable operations with inline forms

Daemon configuration system (search for "daemon.json", "config validation failed", "config reload failed")

A new slash command providing a tabbed interactive UI for managing daemon services.

/daemon

• Scheduled tab: lists all scheduled tasks with cron expression, directory, prompt, enabled/disabled status, permission mode, model, and timeout. Supports enable/disable, edit, and remove actions
• Remote Control tab: lists remote-control servers with directory, spawn mode, and running status
• Assistants tab: shows assistant entries with options to restart, clear conversation history, and uninstall
• New Scheduled Task form: accepts prompt, schedule (intervals like 5m, 2h, 1d or cron expressions with live human-readable preview), directory, auto-generated ID, permission mode, and model selection
• New Remote Control Server form: accepts directory (with trust check), name, and spawn mode (same-dir or worktree)
• New Assistant form: accepts directory, name, permission mode, and model
• Service management buttons: Install, Uninstall, Start, Stop, Restart

Daemon hub UI (search for "Claude Daemon", "New Scheduled Task", "New Remote Control Server")

An automated morning briefing that fires before the workday begins, pulling calendar events, overnight messages, and priority items.

• Fires at a random time between 7:00-8:59am daily (randomized to avoid thundering herd)
• Installed as a permanent scheduled task when an assistant is set up
• Reviews today's calendar events, identifies prep needs, and schedules dynamic pre-meeting check-ins
• Scans overnight inbox for unread email and chat mentions from important contacts
• Identifies the single most important priority for the day
• Outputs a concise brief: calendar summary, items needing attention, and top priority
• Time guard: skips if fired more than 2 hours after configured work start time
• Registered as the skill /morning-checkin at .claude/skills/morning-checkin/SKILL.md

Morning check-in routine (search for "morning-checkin", "/morning-checkin")

A periodic ambient awareness system that checks on tracked priorities every 2 hours during work hours.

• Fires every 2 hours via cron (0 /2 )
• Respects configurable quiet hours (default: 9am-5pm Mon-Fri) — silently exits outside that window
• Maintains a persistent priority list in .claude/catch-up-state.json tracking: priorities, lastSnapshot, lastRunAt, and noiseFloor
• On first run, bootstraps 2-3 priorities from CLAUDE.md and connected tools
• For each priority: checks if state has changed since last snapshot across source control, chat, email, calendar, docs, and issue tracking
• Triages findings into: assistant-can-act, user-should-act, FYI, or suppress
• When calendar events are found, dynamically schedules one-shot /pre-meeting-checkin tasks 2-15 minutes before each event
• If nothing is actionable, outputs "Nothing actionable." and the main agent won't relay it
• Learns over time: promotes new items, prunes resolved ones, demotes unchanged items

Catch-up routine (search for "catch-up-state.json", "/catch-up")

A dynamically-scheduled notification that fires 2-15 minutes before calendar events with meeting context and prep materials.

• Not on a fixed schedule — created as one-shot cron tasks by the morning check-in and catch-up routines
• Assembles: meeting document outline (if linked), recent chat/email context related to attendees, open questions from priorities, and notes from previous occurrences of recurring meetings
• Sends a notification directly to the user (runs in main context, not fork)
• Output format: meeting title, time to start, attendees, doc link, relevant context, and open items
• Offers to draft talking points or agenda but doesn't do so unasked

Pre-meeting check-in routine (search for "pre-meeting-checkin", "/pre-meeting-checkin")

New interactive UI flows for backgrounding, detaching, and managing the lifecycle of sessions.

• Background confirmation dialog: when backgrounding a session, shows "Background this session?" with options to keep running or detach
• Backgrounding… progress indicator while the session is being forked to the background daemon
• After backgrounding: shows helper commands (claude agents, claude attach, claude kill, claude logs)
• Detach prompt: "Detach (keep running)" option to leave a session while it continues in the background
• Warning for running tasks: "Background anyway (tasks will be abandoned)" when background tasks are active
• Nothing to background: shows "Nothing to background yet — send a message first" if no conversation exists
• Cannot background: shows "Cannot background — session persistence is disabled" when persistence is off
• Session status labels: "idle — waiting for trigger", "idle — attach to send a prompt", "actively progressing on the task", "stopped from session"
• Exit command description updated: "Exit the CLI (in a background session: detach or stop)"

Background session UI (search for "Background this session?", "Backgrounding", "Nothing to background yet", "Leave background session")

When setting up a daemon assistant, Claude Code scaffolds a complete directory structure with skill files, scheduled tasks, memory system, and personality templates.

• Creates .claude/agents/assistant.md — a voice-and-values template defining Claude's personality ("Warm, not performative / Smart, not showy / Direct, not blunt / Collaborative, not obedient")
• Creates CLAUDE.md at the project root — a user profile template with fields for name, timezone, work schedule, communication preferences, and catch-up hours
• Sets up skill directories for all four routines: dream, catch-up, morning-checkin, pre-meeting-checkin
• Pre-populates .claude/scheduled_tasks.json with three permanent tasks
• Initializes .claude/catch-up-state.json with empty state
• Memory directory established at ~/.claude/projects/<hash>/memory/

Assistant scaffolding (search for "assistant.md", "voice and values", "About The User")

Pin important sessions to the top of the fleet view for quick access.

In the claude agents view, press Ctrl+T on any session to pin/unpin it.

• Pinned sessions appear in a "Pinned" group at the top of the list
• Pins persist across sessions via .pins.json in the global config directory
• Auto-expands the pinned group when pinning a new session
• Available via Ctrl+T keybinding in the fleet view

Session pinning (search for ".pins.json", "pin to top", "ctrl+t to pin")

Configure Remote Control bridge servers as persistent daemon workers, making sessions accessible from claude.ai.

• Configured via daemon.json under the remoteControl key
• Each server maps a local directory to a Remote Control endpoint
• Spawn mode: same-dir (default) shares the directory, worktree gives each session its own git worktree
• Configurable capacity (default 32 concurrent sessions), permission mode, sandbox mode, and session timeout
• Sessions can be created on daemon start with createSessionOnStart: true
• Manages the Remote Control bridge headlessly with automatic OAuth token management

Remote control server configuration (search for "remote-control server", "New Remote Control Server", "spawnMode")

Several new user-configurable settings have been added.

• prUrlTemplate — URL template for PR links in the footer badge and inline messages. Supports placeholders: {host}, {owner}, {repo}, {number}, {url}. Example: "https://reviews.example.com/{owner}/{repo}/pull/{number}" (search for "URL template for PR links")
• theme — now documented with .describe("Color theme for the UI") (search for "Color theme for the UI")
• fileCheckpointingEnabled — now documented with .describe("Snapshot files before edits so /rewind can restore them") (search for "Snapshot files before edits")
• timestampMessages — stamps each assistant message with its arrival time, described as 'Show "Cooked for Nm Ns" after each assistant turn' (search for "Cooked for")
• todoTrackingEnabled — enables the todo/task tracking panel, described as "Enable the todo / task tracking panel" (search for "Enable the todo")
• notificationChannel — preferred OS notification channel (search for "Preferred OS notification channel")
• showFullToolOutput — shows full tool output instead of truncated summaries (search for "Show full tool output")
• keybindingMode — key binding mode for the prompt input (search for "Key binding mode")
• remoteControlAtStartup — start Remote Control bridge automatically each session (already existed, now with description)
• autoUploadSessions — mirror local sessions to claude.ai as view-only (already existed, now with description)

New settings (search for the .describe( strings listed above)

A new slash command with both interactive slider UI and text-mode for configuring the auto-compact window size.

/autocompact           # Opens interactive slider dialog
/autocompact 500k      # Sets window to 500,000 tokens
/autocompact auto      # Resets to automatic (recommended)

• Interactive mode: slider with up/down arrows, wrapping to "Auto" at boundaries
• Accepts token values in various formats: 500k, 200000, 200 (shorthand for 200,000), range 100k-1M
• Reset with: auto, reset, unset, or default
• Displays current window size, source (env var, settings, or auto), and whether it's capped by the model limit
• Shows warnings when: auto-compact is disabled, env var CLAUDE_CODE_AUTO_COMPACT_WINDOW takes precedence, or when overriding auto
• Strongly recommends the auto setting: "The auto setting picks a window tuned for your model and is strongly recommended for the best cost and performance."

Auto-compact configuration (search for "Auto-compact Window", "Configure the auto-compact window size", "CLAUDE_CODE_AUTO_COMPACT_WINDOW")

Claude Code now validates MCP tool names against the emerging MCP naming standard and warns about non-conforming names.

• Hard failures (tool rejected): empty names, names exceeding 128 characters, names with characters outside [A-Za-z0-9._-]
• Soft warnings (tool registered but warning logged): names with spaces, commas, leading/trailing dashes or dots
• Warning output references the MCP specification proposal: [SEP: Specify Format for Tool Names](https://github.com/modelcontextprotocol/modelcontextprotocol/issues/986)
• Registration proceeds for soft warnings: "Tool registration will proceed, but this may cause compatibility issues."

Tool name validation (search for "Tool name validation warning", "Tool name contains invalid characters", "Allowed characters are")

Background sessions and remote control servers can now be configured to use git worktree isolation, giving each session its own isolated copy of the repository.

• Configured via --spawn-mode worktree or in daemon.json spawnMode: "worktree"
• System prompt injected for worktree sessions: "This is a git worktree — an isolated copy of the repository"
• Agents in worktree mode are instructed to call EnterWorktree as their first action
• Session records now carry worktreePath, worktreeBranch, and worktreeHookBased fields
• After killing a worktree session, the worktree is retained with a hint to use claude rm for cleanup

Worktree spawn mode (search for "--spawn-mode", "spawnMode", "worktree retained at")

New CLI flag to override the default daemon configuration file path.

claude daemon --json-path=/path/to/custom-daemon.json run

• Default path remains ~/.claude/daemon.json
• Useful for running multiple daemon instances with different configurations
• Also exposed in service installation (ExecStart includes --json-path)

JSON path flag (search for "--json-path")

New CLI flag to control which settings sources are loaded.

claude --setting-sources=user,project,local

• Comma-separated list of setting sources: user, project, local
• Allows restricting which settings layers are applied for a session

Setting sources flag (search for "--setting-sources", "Comma-separated list of setting sources")

Added support for MCP completion/autocomplete of prompts and resource templates.

• New McpCompletable type registered as Symbol.for("mcp.completable")
• Validates completion request types: prompts must use ref/prompt, resource templates must use ref/resource
• Provides structured error messages: "Expected CompleteRequestPrompt, but got ..." / "Expected CompleteRequestResourceTemplate, but got ..."

MCP completions (search for "McpCompletable", "mcp.completable", "Expected CompleteRequestPrompt")

The dream routine has been upgraded from simple hardcoded prompts to a structured 4-phase SKILL.md with explicit protocol:

• Phase 1 (Preparation): Reviews logs/YYYY/MM/YYYY-MM-DD.md entries and session transcripts
• Phase 2 (Topics): Extracts events, lessons, and decisions into topic files (<topic-slug>.md)
• Phase 3 (Rules & Learnings): Records painful/inefficient experiences and user frustrations in learnings/<learning-slug>.md
• Phase 4 (Prioritization & Pruning): Keeps MEMORY.md under 200 lines, removes stale entries, adds newly important ones
• Now scheduled as a permanent task at a random time between 1:00-4:59am

Dream SKILL.md (search for "Nightly reflection and consolidation", "Prioritization and Pruning")

Usage limit messages have been updated to use "usage allocation" terminology for certain seat types:

• "Your usage allocation has been disabled by your admin" (new)
• "Now using your usage allocation" / "You're now using your usage allocation" (new)
• "Your seat type doesn't include usage" / "Your seat type doesn't include extra usage" (new)
• "Your org is out of usage · add funds to continue" / "· contact your admin" (new)
• "Your group's usage limit is set to $0" (new)
• Removed: "Request extra usage", "Switch to extra usage"

Updated usage strings (search for "usage allocation", "Your seat type doesn't include")

The stdout/stderr error handler now handles EIO in addition to EPIPE, and properly triggers a graceful exit callback when stdout encounters these errors.

Stream error handling (search for "EIO" in hJ8 function near line ~8728)

The Skill tool description has been updated to clarify skill parameter usage:

• Removed: "Set skill to the exact name of an available skill (no leading slash)"
• Added: "Only invoke a skill that appears in that list, or one the user explicitly typed as /<name> in their message. Never guess or invent a skill name from training data"
• Added the concept of plugin:skill fully qualified names for plugin-namespaced skills

Skill tool description (search for "Never guess or invent a skill name")

Slash commands now support explicit alias definitions with the description: "Alternate names that resolve to this command (e.g., /cost and /stats both resolve to /usage)".

Command aliases (search for "Alternate names that resolve to this command")

When a file has been modified on disk after Claude last read it, users now see a warning: "is now stale relative to disk — Read it again if you need current contents."

Stale file warning (search for "is now stale relative to disk")

Remote sessions now give a more specific error when trying to switch models: "Model picker shows local options in remote sessions — use /model <name> instead" and "Fast mode switching in remote sessions is coming soon — set at session creation for now".

Remote model switching (search for "remote-model-picker-unavailable", "Fast mode switching in remote sessions")

The Pro trial flow now includes:

• Trial start screen: "Your Pro plan includes N days of Claude Code"
• Interactive confirmation: "Press Enter to start your trial"
• Progress indicator: "Starting your trial..."
• Error handling: "Couldn't start your trial. Press Enter to continue."
• Trial duration display: "Trial: N day(s) left" in the status area
• Trial expiration route at /pro-trial-expired

Pro trial lifecycle (search for "Pro trial started", "Your Pro plan includes", "Starting your trial")

Several new options are available for the Agent SDK:

• askUserQuestion.previewFormat exposed via CLAUDE_CODE_QUESTION_PREVIEW_FORMAT env var
• canUseTool callback for dynamic tool permission decisions (mutually exclusive with permissionPromptToolName)
• getOAuthToken callback for custom OAuth token provision
• Input/output validation with structured error messages: "Input validation error: Invalid arguments for tool" / "Output validation error: Invalid structured content for tool"
• Tool annotations support via Zod schemas
• taskSupport for tools: "required" (needs task augmentation) vs "forbidden" (must use registerTool() instead)

SDK enhancements (search for "CLAUDE_CODE_QUESTION_PREVIEW_FORMAT", "canUseTool callback", "Output validation error")

Added support for Zod Mini (ZodMiniType, ZodMiniObject) alongside existing Zod schemas, with automatic detection and mixed-version handling. The system now throws a clear error on mixed versions: "Mixed Zod versions detected in object shape."

Zod Mini support (search for "ZodMiniType", "ZodMiniObject", "Mixed Zod versions")

Significantly expanded zodToJsonSchema converter supporting:

• jsonSchema2019-09 target with unevaluatedProperties support
• OpenAI-specific type handling (OpenAiAnyType) with reference strategies
• Date strategies: format:date-time, format:date, string, integer (unix-time)
• Email strategies: format:email, format:idn-email, pattern:zod
• Base64 strategy: contentEncoding:base64
• Warning messages for OpenAI compatibility: "Warning: OpenAI may not support records in schemas!" / "Warning: OpenAI may not support schemas with unions as roots!"

JSON schema generation (search for "jsonSchema2019-09", "OpenAiAnyType", "Warning: OpenAI may not support")

• Fixed handling of EPIPE and EIO errors on stdout/stderr — now properly destroys the stream and triggers a graceful exit callback instead of silently failing (search for "EIO")
• Improved handling of failed daemon connections: "Couldn't reach the daemon — it may be restarting. Press Enter to retry" instead of silent failure (search for "Couldn't reach the daemon")
• Added protection against recursive Zod schema references: "Recursive reference detected at" (search for "Recursive reference detected")
• Fixed regex pattern conversion failures with descriptive error: "Could not convert regex pattern at" (search for "Could not convert regex pattern")
• Added validation for --permission-mode flag values: "must be one of" (search for "--permission-mode must be one of")
• Added guard against duplicate task IDs: "task ids must be unique" (search for "task ids must be unique")
• Fixed transcript mirroring with bounded retry (3 attempts with short backoff; timeouts are not retried) to prevent silent data loss (search for "TranscriptMirrorBatcher")
• Added timeout handling for SessionStore operations: append(), load(), listSessions(), listSubkeys() all now have explicit timeout errors (search for "SessionStore.append() timed out")
• Fixed speculation pausing for backgrounded shells: "Speculation paused: backgrounded shell" (search for "Speculation paused: backgrounded shell")

The full HTML-based "Claude Code Insights" report generator — which produced an interactive dashboard with charts, project areas, friction analysis, usage patterns, and suggestions — has been removed from the CLI. The removal includes all HTML templates, CSS styles, JavaScript interactivity (timezone selector, collapsible sections, copy buttons), and the data rendering pipeline.

Removed Insights template (search for "Claude Code Insights" — reduced from 5 occurrences to 3 minor references)

The cG1 React component for /autofix-pr — which handled inline PR monitoring setup, Remote Control subscription, and session spawning — has been removed from the bundled source. (The broader autofix-pr infrastructure remains, suggesting this was refactored rather than removed entirely.)

Removed autofix-pr component (the cG1 function at old line ~472814 is gone; search for "tengu_autofix_pr" still returns hits in both versions)

The large inline /init command prompt (which guided the multi-phase CLAUDE.md setup flow) has been removed from the bundled source. The /init command likely still exists but its prompt content has been restructured or externalized.

Removed init prompt (the cN1 string at old line ~492860 covering Phases 1-8 is gone)

• The daemon installation (claude daemon install) creates a system service — users should run claude daemon uninstall before major upgrades if they experience issues with the service
• The daemon.json configuration format is new; there is no migration from previous daemon configurations since the daemon infrastructure was previously minimal
• Scheduled routines (morning-checkin, catch-up, pre-meeting-checkin) are installed when creating a daemon assistant via the /daemon command — they are not automatically enabled for existing setups
• The fleet view (claude agents) replaces the previous "List configured agents" description with "Manage background and configured agents", reflecting its expanded interactive capabilities