Claude Code v2.1.273 adds gateway diagnostic headers, fixes prompt-cache rewrite and permission bypasses
ashwin-ant · ghdev · 2026-09-16
anthropics/claude-code v2.1.273 ships notable fixes and features:
- Gateway diagnostics: new opt-in headers (CLAUDECODEGATEWAYHINTHEADERS=1) including request class, agent type, and compaction hints for LLM gateways.
- Reliability: notification when an MCP server disconnects and auto-reconnect gives up; session forking for claude --remote-control sessions started from the Claude app.
- Security: fixed Bash commands the permission checker can't fully analyze skipping prompts, a subshell hiding dangerous rm in bypass mode, and claude.ai-synced skills staying available after org-level Skills shutdown.
- Cost: /login, /upgrade, /extra-usage no longer discard prior thinking, avoiding a full prompt-cache rewrite on the next request.
- Also fixes 401/403 messaging on Bedrock/Vertex/Foundry, macOS screenshot reads, and shell-mode ! prefix handling.
More from coding & agent
- dspy-typesafeify adds typed inference to DSPy Signatures with one decorator — dbreunig · 2026-09-16
- Periodic Harness vs. vanilla Claude Code: a striking diff comparison — zainhas · 2026-09-16
- Dev runs coding agents inside tmux sessions viewed in his browser IDE — lucasmeijer · 2026-09-16
- Google's Artemis uses AI to navigate Android phones for testing, with a built-in MCP server — kevinkern · 2026-09-16
- Why isn't anyone vibe coding 'good enough' clones of every habit-protected high-margin app? — citrini · 2026-09-16
- Model upgrades are the best time to delete your agent workarounds — gethackteam · 2026-09-16