v2.1.166
read at source ↗ github.com
v2.1.166
Source: Anthropic Claude Code Date: 2026-06-06 URL: https://github.com/anthropics/claude-code/releases/tag/v2.1.166
Summary
v2.1.166 is the largest substantive Claude Code release in this batch. The two headline changes are: a fallbackModel setting (up to three fallback models tried in order when the primary is overloaded or unavailable, now applying to interactive sessions as well as headless) and a hardened cross-session authority model (messages relayed via SendMessage from other Claude sessions are stripped of user authority — receivers refuse relayed permission requests and auto mode blocks them). Additional changes include MAX_THINKING_TOKENS=0 and --thinking disabled now correctly suppressing thinking on API-side thinking-capable models, one-retry fallback on unexpected non-retryable errors, and roughly a dozen bug fixes covering terminal compatibility (JetBrains, Kitty protocol), process lifecycle, and managed settings enforcement.
Implications
- The agent-host ownership split. The relayed-
SendMessageauthority drop is a hard architectural choice: multi-agent chains cannot escalate permissions through message relay. This hardens the trust boundary between coordinating sessions and prevents a compromised or over-permissioned child agent from bootstrapping user-level authority in a parent. It’s the right call, and it will break naive orchestration patterns that assume session-to-session trust is transitive. - Fleet-correctness substrate.
fallbackModelbrings reliability-by-default to interactive sessions — not just CI/headless runs. The one-retry-on-unexpected-error addition is a complementary signal: Anthropic is building resilience into the client layer rather than requiring operators to handle it. As fleets scale, client-layer retry/fallback becomes load-balancing infrastructure. - The trust/provenance hardening arc. The managed-settings fixes (invalid entries silently disabling policy enforcement,
${VAR}references not matching) are quietly important for enterprise deployments. Silent policy bypass is a compliance failure mode; this release closes at least two such paths.