Claude Code adds AGENTS.md fallback, aligning with the cross-tool config standard

Claude Code’s September 18 release adds support for AGENTS.md, the emerging vendor-neutral convention for project-level agent instructions. In any project that lacks a CLAUDE.md, the CLI now reads AGENTS.md instead, with the choice configurable under ‘Project instructions’ in /config. The feature isn’t yet available on the Bedrock, Vertex, or Foundry backends. The move is notable because it signals Anthropic converging on a shared standard rather than locking users into its own proprietary file, easing life for teams that juggle multiple AI coding tools.

The rest of the changelog is a heavy maintenance pass, dominated by crash and hang fixes across the CLI, SDK, and headless (claude -p) paths. Several address corrupted or malformed ~/.claude.json state — bad theme, customApiKeyResponses, and MCP placeholder values that previously killed startup — while others stop sessions from hanging silently after internal errors, now exiting with code 1 and a reported error. Tooling reliability also improved: Grep, Glob, Edit, and Write now surface clear errors for resource exhaustion, non-matching edits, and writes aimed at directories instead of failing quietly.

For enterprise and gateway operators, the release hardens the Claude apps gateway with a new egress-boundary proxy mode, static upstream headers, NO_PROXY handling for telemetry, and stricter marketplace-policy parsing so one malformed entry no longer disables the whole policy. Plugin management, auth (fixing spurious logouts when older builds run alongside newer ones), prompt-cache behavior after resume, and Windows-specific edge cases round out a broad stability-focused update.