Claude Code v2.1.252 Fixes Bash Errors, Remote Control Stalls
Anthropic shipped Claude Code v2.1.252 with four targeted fixes addressing nagging issues reported by developers. The release is a patch, not a feature drop, but the fixes tackle real workflow friction.
What’s fixed
- Bash command failures on macOS: Some Macs hit the error
task output swap refused (tasks dir moved or linked)when running Bash commands. This release resolves that, so shell tools should work reliably again. - “Always allow” not persisting: In projects without a
.claude/settings.local.jsonfile, the always allow setting wasn't being saved. v2.1.252 fixes that so permissions stick even when the file doesn't exist yet. - Remote Control session stalls: Sessions hosted by Claude Desktop or VS Code could hang for minutes after a tool finished if the connection to claude.ai was degraded. The stall is now eliminated, keeping the flow moving even on flaky networks.
- Oversized background notifications: When a background task failed with very large output (for example, git errors on a full disk), the resulting notification could push the conversation over the API request size limit. The fix trims large failures so the conversation stays within limits.
If you've been hitting any of these — especially the Bash error on Mac or the annoying stalls in Remote Control — upgrade to v2.1.252. The changelog is straightforward, and the fixes look solid.
📖 Read the full source: GitHub Claude-Code
👀 See Also
Glomz Octagon: Multi-Agent Code Reviews – 179 Agents, 1,333 Reviews, and the Network Effect
Glomz.com ran an experiment where 179 AI agents registered, submitted 433 code submissions, and generated 1,333 reviews in an "Octagon" arena. The 'review cascade' network effect is real — submissions with 3-5 initial reviews attracted more agents, with top submission getting 21 reviews.
Anthropic's circuit-tracing research reveals Claude 3.5 Haiku's internal mechanisms
Anthropic published circuit-tracing research on a simplified Claude 3.5 Haiku, revealing six specific behaviors including its default "I don't know" state, backward poem writing, and dual-path math processing.
Tokenmaxxing Is the New Stopwatch: Why Your AI Policy Needs to Be Coherent
Brian Meeker argues against vanity metrics like tokenmaxxing and shares his team's four-point AI policy: no mandate, understand generated code, survive without AI tools, care about teammates and customers.
Gemma 4 vs Qwen 3.5 Blind Evaluation Results with Claude Opus as Judge
A 30-question blind evaluation compared Gemma 4 31B, Gemma 4 26B-A4B, and Qwen 3.5 27B using Claude Opus 4.6 as judge. Qwen 3.5 27B won 46.7% of matchups but had lower average scores due to three zero-scoring responses.