v2.1.174

Source: Anthropic Claude Code Date: 2026-06-12 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.174

Summary

v2.1.174 is a dense bug-fix release across model-picker UX, provider credential plumbing, and fleet-level session isolation. The standout fixes address Bedrock GovCloud routing (wrong inference-profile prefix causing 400s), background sessions inheriting the wrong provider’s env vars, and /advisor pre-selecting a model blocked by the availableModels allowlist. Also included: skill hot-reload now diffs and re-announces only changed skills rather than blasting the full listing, and VSCode gains per-agent and per-skill usage attribution in the Account & usage dialog.

Implications

• Fleet-correctness. Several fixes target multi-session and background-daemon isolation — env vars leaking between sessions, pre-warmed workers failing auth after sitting idle, and Workflow subagents missing attribution headers. These are exactly the failure modes that surface at scale when the host is running dozens of concurrent background agents; the pattern indicates ongoing hardening work specific to fleet deployments.
• Coding-agent host war. The VSCode usage-attribution breakdown (cache misses, long context, subagents, per-skill/agent/plugin/MCP) gives the VSCode integration a reporting surface that rivals managed enterprise dashboards. This is Anthropic tightening the native IDE footprint rather than ceding it to third-party plugins.
• Trust/provenance hardening. The git commit co-author attribution bug (wrong model name on some models) and the skill hot-reload diff fix are both correctness fixes in the provenance layer — who or what produced an artifact, and what changed between versions. Small, but the direction is consistent.