Claude Code v2.1.129: Plugin URL flag, force sync output, and 20+ fixes

Claude Code v2.1.129 ships today with several new flags, env vars, and a heap of bug fixes — including a nasty /context command that was silently burning ~1.6k tokens per call. Here's what matters.
New flags and env vars
--plugin-url <url>— load a plugin .zip from a remote URL for the current session. No manual download required.CLAUDE_CODE_FORCE_SYNC_OUTPUT=1— forces synchronized output on terminals that auto-detection misses (e.g., Emacseat).CLAUDE_CODE_PACKAGE_MANAGER_AUTO_UPDATE— when set on Homebrew or WinGet installs, Claude Code runs the upgrade command in the background and prompts to restart. No more manual brew upgrade & restart dance.CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY=1— opt-in for/v1/modelsdiscovery in the/modelpicker. Was automatic in 2.1.126–2.1.128; now opt-in to avoid surprises with gateway deployments.
Plugin manifests: experimental namespace
themes and monitors in plugin manifests should now be declared under "experimental": { ... }. Top-level declarations still work but claude plugin validate will warn. Migrate your manifests to keep clean validation.
skillOverrides now works
The skillOverrides setting actually does something now:
off— hides the skill from both model and / command listuser-invocable-only— hides from model, still visible in /name-only— collapses the description
Behavior changes
- Ctrl+R history picker now searches across all projects by default (pre-2.1.124 behavior). Press Ctrl+S to narrow to current project or session.
- Third-party deployments (Bedrock, Vertex, Foundry, or ANTHROPIC_BASE_URL gateway) no longer show spinner tips pointing at first-party Anthropic surfaces.
claude_code.pull_request.countOTel metric now counts PRs/MRs created via MCP tools, not just shell commands.- Policy refusal error messages include the API Request ID for easier debugging.
Bug fixes that matter
- /context no longer dumps its rendered ASCII visualization grid into the conversation — was wasting ~1.6k tokens per call.
- 1-hour prompt cache TTL was silently downgraded to 5 minutes — now fixed.
- Cache-miss warning after /clear or compaction when changing /effort or /model — spurious warnings gone.
- External-editor handoff (Ctrl+G) blanking conversation history — fixed.
- Bash(mkdir *), Bash(touch *) allow rules not honored for in-project paths — now respected.
- deniedMcpServers patterns with
*://scheme wildcard not matching mixed-case hostnames — fixed. - OAuth refresh race after wake-from-sleep that could log out all running sessions — fixed.
- Server-managed settings policy not applying for users whose stored OAuth credentials lacked the
user:inferencescope — now enforced. - Agent panel below prompt hidden when subagents are running (regression in 2.1.122) — restored.
- /clear not resetting terminal tab title after a conversation — fixed.
- API errors with unrecognized 400 status codes showed raw JSON instead of the underlying error message — now properly parsed.
VSCode fix
/clear in VSCode extension now correctly clears both the conversation context and the displayed transcript.
Upgrade via brew upgrade claude-code or winget upgrade. Full changelog in the release.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Analysis of Claude Code's ~12K Token Forced System Prompt Reveals Priority Rules Overriding User Config
An analysis of Claude Code's injected ~12K token system prompt shows priority rules for song lyric bans, subagent delegation, and brevity that override user CLAUDE.md and memory files.

Qwen3.6 Plus Preview Available Free Through OpenRouter for OpenClaw
The Qwen3.6 Plus Preview model is now accessible at no cost via OpenRouter for OpenClaw users. Setup requires obtaining API keys from OpenRouter and configuring them in OpenClaw, with a prompt to set the thinking level during initial use.

Designing a Team of Agents: How Google Antigravity Structures Subagents for Autonomous Code Generation
Google Antigravity reveals its subagent architecture for autonomous coding: seven specialized agent types from the Sentinel (front-desk) to the Auditor (authenticity checker). Relevant for OpenClaw's subagent design.

Claude Code v2.1.181: /config Syntax, Sandbox Apple Events, Streaming Fixes
Claude Code v2.1.181 adds /config key=value syntax for inline settings, sandbox.allowAppleEvents on macOS, and CLAUDE_CLIENT_PRESENCE_FILE. It also upgrades Bun to 1.4, fixes prompt caching on custom API URLs, network drive write issues, and many startup regressions.