Claude Code v2.1.142: New claude agents flags, Opus 4.7 default, and bug fixes
Anthropic released Claude Code v2.1.142 with significant updates to the claude agents subsystem, a default model change in fast mode, and a long list of bug fixes.
New claude agents flags
Eight new flags are available to configure dispatched background sessions:
--add-dir— additional directories--settings— settings file path--mcp-config— MCP configuration--plugin-dir— plugin directory--permission-mode— permission mode for the session--model— model override--effort— effort level--dangerously-skip-permissions— skip permission prompts
Previously these were only configurable via environment variables or the daemon configuration file; now they can be passed directly when starting a background agent.
Fast mode defaults to Opus 4.7
Fast mode now uses Opus 4.7 by default, replacing Opus 4.6. If you need to pin fast mode to Opus 4.6, set the environment variable:
CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE=1Plugin improvements
Plugins with a root-level SKILL.md and no skills/ subdirectory are now correctly surfaced as a skill. The /plugin details pane and claude plugin details command now display the LSP servers a plugin provides.
Notable bug fixes
- MCP_TOOL_TIMEOUT — fixed so it actually raises the per-request fetch timeout for remote HTTP and SSE MCP servers. Previously, tool calls were capped at 60 seconds regardless of the configured value.
- macOS sleep/wake — background sessions no longer disappear after sleep/wake. The daemon now detects clock jumps instead of treating them as elapsed idle time.
- Daemon upgrade — the daemon now exits cleanly after the binary is upgraded (e.g.,
brew upgrade), preventing dispatched agents from crash-looping on the deleted binary path. - Windows network drives — background agents no longer deadlock when working directories are on network drives; Ctrl+C now works during startup.
- Editor detection —
claude agents"v to open in editor" now uses your shell's$EDITOR/$VISUALinstead of the daemon's default editor. - Terminal color bleed — fixed background-color bleed when attaching to a session from Apple Terminal or other 256-color-only terminals.
- Permissions persistence —
claude --bg --dangerously-skip-permissionsnow persists across retire/wake cycles. - Plugin cache — no longer deletes the active plugin version directory when installation metadata is missing.
Other improvements
- Reactive compaction — the first summarize attempt now seeds from the original request's overflow size, avoiding a wasted near-full-context retry.
- Hook configuration — configuring a prompt- or agent-type hook for
SessionStart/Setup/SubagentStartnow shows a clear error suggesting to use a command-type hook instead. - Removed stale
/model claude-sonnet-4-20250514suggestion from Usage Policy refusal messages.
📖 Read the full source: GitHub Claude-Code
👀 See Also
MuninnDB adds Dream Engine for LLM memory consolidation with vault isolation
MuninnDB, a Go-based cognitive memory database, now includes a Dream Engine that performs LLM-driven memory consolidation between sessions using deduplication thresholds and semantic review. The system features vault trust tiers for data isolation and runs locally with Ollama.
Agent Memory Protocol (AMP): Open Spec for Interoperable AI Agent Memory on Top of MCP
AMP defines a standard interface for persistent memory in MCP-compatible agents with six core verbs: encode, recall, forget, consolidate, pin, and stats. Includes compliance test suite and reference implementation.
How Clawdbot Coordinates 6 AI Agents with a Production-Stable Work Queue
Clawdbot's team built a work queue system to coordinate 6 AI agents (design, code, marketing, ops) for their AI-operated store. The system features atomic task claiming, a state machine, retry logic with backoff, task chains, heartbeat tracking, and a daemon orchestrator.
Claude-ETA Plugin Adds Task Timing and Repair Loop Detection to Claude Code
Claude-ETA is a Claude Code plugin that times tasks, learns your actual velocity, and feeds real data back into Claude before it responds. It also detects repair loops by fingerprinting error content and intervenes after three identical failures.