Claude Code v2.1.271: Fast Mode in Remote Sessions, Per-Command allowed_domains, and a Batch of Bash Permission Fixes
Claude Code v2.1.271 is out, and it's mostly a hardening release with a few genuinely new knobs. Here's what actually changed.
New features
- Fast mode in Remote sessions. Cloud and self-hosted runners now honor the host's fast-mode setting, or
/fasttyped inside the session — where your organization allows it. - Mouse support in
/config(fullscreen). Wheel scrolls the settings list, clicking a setting's value changes it, and the row under the pointer is highlighted. claude self-hosted-runner --drain-marker-file <path>. If that file exists at a SIGTERM drain, the runner reports its exit to the server as a host drain. Telemetry only.- Per-command
allowed_domains. In auto mode with sandboxing, Bash, PowerShell, and Monitor commands can have the hosts they need reviewed alongside the command and opened for that command alone; other hosts are refused. omitClaudeMdin agent frontmatter and--agentsJSON. Custom and plugin subagents can run without user, project, and local CLAUDE.md files. Managed policy files still load.--accept-command <sha256>forclaude plugin installandclaude plugin update, accepting exactly the command a prior--jsonrun displayed, instead of-y.- Pricing multipliers up to 10. The
modelPricingmanaged setting and the Claude apps gatewaypricingblock now accept a multiplier above 1 for marked-up internal chargeback rates.
Notable fixes
- Cached org policy no longer gets reused after switching accounts, orgs, or API keys, and refreshes when the credential changes mid-session instead of waiting for the hourly check.
- Tool and command lists now update when the org policy finishes loading after startup or changes mid-session.
- An unreadable or unparseable enterprise
managed-mcp.jsonis no longer ignored — it now keeps exclusive MCP control (user/project/plugin servers don't load) and warns at startup. - Org policy is no longer fetched through (and rejected by) third-party local proxies set via
ANTHROPIC_UNIX_SOCKET. - Cloud sessions no longer reject every subagent tool call (
updatedInput … failed schema validation) when a workflow or agent approval lands after the session's worker restarted. /fast offnow actually turns fast mode off when the org has it disabled, instead of answering "Fast mode unavailable".- Sessions started with
CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECKno longer re-send fast requests every turn after an API rejection. - Bash permission checks got several patches: files read by
fmt,column, and similar commands after unrecognized options; files expanded from wildcards in patterns or option values (e.g.grep -v dir/* file); shell variable declaration flags no longer misrepresent the command; twocds, a subshell, or acd + gitchain no longer skip the prompt underpermissions.blockReadsOutsideWorkingDirectories. - Stale
.git/config.lockon Linux no longer breaksgit checkout -b,git push -u, orgit configfor the rest of the session. - macOS settings watcher falls back to polling when the system file-event service is saturated.
- Resumed
claude -psessions whose tools all come from MCP servers no longer fail with "At least one tool must have defer_loading=false". - LLM gateways returning non-streaming replies as
text/plainno longer cause "API returned an empty or malformed response". - MCP
list_changednotifications in a tight loop no longer cause sustained high CPU and repeated tool-list requests. - MCP OAuth client registration handling fixed (denying consent forced a new one, another redirect URI's registration was reused, and a concurrent write could delete a valid one).
If you're on self-hosted runners or use org-managed policy files, the MCP and policy-cache fixes are the ones worth updating for.
📖 Read the full source: GitHub Claude-Code
👀 See Also

Medvi's $1.8B AI Company Claims Face Scrutiny Over Legal and Ethical Issues
Gary Marcus critiques the viral story of Medvi, a claimed $1.8B AI company built by one person in two months, highlighting class action lawsuits for spam violations and questions about revenue reporting and compliance.

Benchmark Results for Qwen3.5 Models with 2K to 400K Context on RTX 4090
A developer tested multiple Qwen3.5 model variants on an RTX 4090, measuring performance across context windows from 2,048 to 400,000 tokens. The benchmarks include time-to-first-token metrics and reveal issues with some models requiring KV offload testing.
OpenClaw v2026.11.1 Patch: Fixes Message Refusals, Doctor Exit Codes, and Stuck Gateway Writes
OpenClaw v2026.11.1 is a patch release with 83 pull requests and 67 contributors — no new features, just fixes for false message refusals, Doctor false-positive inspections, and Gateway settings writes hanging mid-apply.

Meta to capture employee mouse movements and keystrokes for AI training
Meta plans to start capturing employee mouse movements and keystrokes for AI training data, according to a Reuters report. The article has generated discussion on Hacker News with 33 points and 7 comments.