Claude Code v2.1.212: /fork Becomes Background Session Copy + Runaway Loop Guards
Anthropic shipped Claude Code v2.1.212 with a behavioral change to /fork and new safeguards against runaway search and delegation loops. Here are the details.
Key Changes
- /fork now copies your conversation into a new background session (appears as its own row in
claude agents). The old behavior — launching an in-session subagent — is now/subtask. - Auto-mode reset:
claude auto-mode resetrestores the default auto-mode configuration. Use--yesto skip the confirmation prompt. - WebSearch limit: New session-wide cap of 200 WebSearch tool calls (default). Tune via
CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION. - Subagent cap: Per-session limit of 200 subagent spawns. Override with
CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION. Running/clearresets the budget. - MCP auto-background: MCP tool calls exceeding 2 minutes automatically move to the background. Configure the threshold (in ms) or disable via
CLAUDE_CODE_MCP_AUTO_BACKGROUND_MS. - Improved
/resume: Typing/resumein the agent view opens a picker of past sessions — including deleted ones — and resumes your pick as a background session.
Notable Fixes
- Plan mode no longer auto-runs file-modifying bash commands (e.g.,
touch,rm) without permission prompt or SDKcanUseToolcallback. - Fixed worktree creation following a repository-committed symlink at
.claude/worktreesthat could create files outside the repository. - SIGTERM during a running Bash tool in print/SDK mode now aborts the turn, kills the process tree, and exits with code 143.
- Windows fix:
/backgroundandclaude --bgno longer fail when Group Policy blocks PowerShell 5.1; the daemon now prefers PowerShell 7. - Shell mode (
!) now executes commands containing file paths even while the path autocomplete popup is open. /ultrareviewnow accepts PR references like#123,PR 123, and pasted URLs; error hints name the command you typed. It also fetches branches from origin and suggests closest branch name on typos.- Hosted sessions no longer fail at startup when repository settings configure mTLS certs, extra CA bundles, or OAuth scopes — these transport settings are now ignored with a warning.
- Fixed diff previews losing line numbers and +/- markers in narrow layouts.
- Fixed false
"Command timed out"on exit code 143.
Full details in the release notes.
📖 Read the full source: GitHub Claude-Code
👀 See Also
Qwen3 27B Outperforms Gemma 4 26B in Real-World Tool-Calling for Local AI Video Pipeline
A local AI video pipeline experiment shows Qwen3 27B handling tool-calling cleanly while Gemma 4 26B got stuck in loops. Also covers Said Image Turbo for local image generation and OpenCode orchestration hitting 174K context.
Claude Code users hitting usage limits faster than expected, bugs suspected
Anthropic acknowledges Claude Code users are exhausting quotas 'way faster than expected,' with users reporting maxed-out limits within hours. Suspected bugs in prompt caching may be inflating costs by 10-20x, and downgrading to version 2.1.34 reportedly helps.
Anthropic Paying SpaceX $15B/Year for Compute Through 2029
SpaceX IPO filing reveals Anthropic pays $1.25B/month through May 2029 for compute. The deal supports AI training on Colossus 1 & 2 facilities.
PeerZero: AI Agents Conduct Peer Review with Credibility-Based Incentives
PeerZero is a platform where AI agents submit research papers, review each other's work, and stake credibility on being right through a bounty system. Agents earn or lose credibility scores based on review accuracy, with vindicated outlier mechanics that reward independent thinking and punish groupthink.