Claude Code v2.1.215 — An Anthropic-Run Google Cloud Backend, a No-Fallback Model Lock, and a Background-Session Cleanup#

Published on July 19, 2026

Part of the Claude Code Version Tracker series. | Official Env Vars | Official Changelog

Claude Code v2.1.215 adds five environment variables over the last analyzed build, v2.1.212 — a sixth first-party model backend (an Anthropic-operated inference plane on Google Cloud), a CLAUDE_CODE_NO_MODEL_FALLBACK lock that pins a session to exactly the model you chose, an OpenTelemetry content-size cap, and a CLAUDE_PID guard that stops the agent from killing its own process.[1] Alongside the binary additions, the changelog lands a large batch of permission-check hardening, a background-session lifecycle cleanup, reasoning effort in the subagent status line, and a behavior change to how /verify and /code-review run.

New Environment Variables#

VariableLikely Purpose
CLAUDE_CODE_USE_ANTHROPIC_GOOGLE_CLOUDA new provider-backend toggle sitting in the binary's exact provider family — beside CLAUDE_CODE_USE_BEDROCK, USE_VERTEX, USE_FOUNDRY, USE_ANTHROPIC_AWS, USE_MANTLE, and USE_GATEWAY — and read by the same guard that suppresses error reporting under a managed provider. It resolves a companion config set (ANTHROPIC_GOOGLE_CLOUD_PROJECT, _LOCATION, _WORKSPACE_ID, _BASE_URL) distinct from the existing Vertex path — pointing at an Anthropic-operated inference plane on Google Cloud rather than customer-owned Vertex. Not named in the changelog.
CLAUDE_CODE_SKIP_ANTHROPIC_GOOGLE_CLOUD_AUTHThe skip-auth companion to the toggle above, grouped in the auth-bypass list with SKIP_BEDROCK_AUTH, SKIP_VERTEX_AUTH, SKIP_FOUNDRY_AUTH, SKIP_ANTHROPIC_AWS_AUTH, and SKIP_MANTLE_AUTH. It tells Claude Code to skip the provider auth handshake when the host has already injected Google Cloud credentials — the pattern host-managed and CI deployments rely on. Undocumented.
CLAUDE_CODE_NO_MODEL_FALLBACKWhen set, the availability chain collapses to [primary]: buildAvailabilityFallbackChain() returns only the model you selected, and a tripwire error fires if any path still attempts a fallback pivot ("a model-fallback pivot was attempted while the no-fallback guarantee is active … this branch should be unreachable"). It has teeth — compaction refuses to run when it would require a model swap ("Compaction unavailable: CLAUDE_CODE_NO_MODEL_FALLBACK is set and model substitution is disabled"). Built for determinism and compliance: no silent downgrade to a cheaper model for background work. Undocumented.
CLAUDE_CODE_OTEL_CONTENT_MAX_LENGTHCaps the length of content attributes on OpenTelemetry spans and logs; the truncation function takes the minimum of this value (default 60 KB) and the standard OTEL attribute-length limits, appending a "[TRUNCATED — Content exceeds NKB limit]" marker. The one addition here the changelog names ("configure the 60 KB truncation limit") — a payload-size and privacy control for enterprise telemetry pipelines.
CLAUDE_PIDHolds the CLI's own process ID and powers a self-protection shim: the injected shell wrapper around pkill/pgrep refuses to signal any match equal to ${CLAUDE_PID}, so an agent-run pkill -f can't terminate Claude Code itself. It is also propagated to child sessions and hooks alongside CLAUDE_CODE_SESSION_ID and TRACEPARENT. The changelog records the resulting fix: "Fixed Bash tool killing the Claude session when a pkill -f pattern accidentally matched the CLI's own process."

Fixes and Improvements Worth Knowing#

Background sessions get cleaned up. A background session runs detached from your terminal — you start it with /background, claude --bg, or /fork, it works in its own process with a row in claude agents, and you can reopen it later. This release fixes several lifecycle bugs: idle sessions no longer keep a process alive forever, completed ones can now be removed via claude rm or the agent view, sessions started outside a git folder can be deleted, and reopening a stopped session restores its saved conversation.

Reasoning effort in the subagent status line. The subagentStatusLine payload now includes reasoning effort, so a custom status line can show each subagent's model and effort at a glance.

/verify and /code-review stop auto-running. Claude no longer runs these skills on its own — you invoke them when you want them. Fewer surprise runs and fewer surprise tokens.

Tighter permission checks. More Bash commands now prompt instead of auto-running: odd file-descriptor redirects, commands over 10,000 characters, and zsh subscripts inside [[ ]]. A rule like Edit(src/**) no longer approves writes to any src/ in the tree, a PowerShell 5.1 bypass is closed, and docker/Podman commands with daemon-redirect flags now require approval.

What These Tell Us#

The provider list is the headline. CLAUDE_CODE_USE_ANTHROPIC_GOOGLE_CLOUD is the sixth first-party backend and the second ANTHROPIC_-branded one after ANTHROPIC_AWS — a pattern of Anthropic running its own inference plane inside each hyperscaler (its own AWS, its own Google Cloud) rather than routing through the cloud's native model service (Bedrock, Vertex). The dedicated ANTHROPIC_GOOGLE_CLOUD_PROJECT/_LOCATION/_WORKSPACE_ID config and its own skip-auth flag mark this as a full enterprise path, not an experiment — and it lands with no changelog mention.

CLAUDE_CODE_NO_MODEL_FALLBACK and CLAUDE_PID are both about removing surprises. The first guarantees the model you selected is the model that ran — down to blocking a compaction that would otherwise substitute one — which matters for reproducibility and for policy regimes that forbid a silent downgrade. The second is a safety rail against the agent harming its own runtime; the pkill guard reads as a direct response to autonomous sessions issuing broad process kills. Read next to the background-session cleanup and the permission-hardening pass, the theme is containment: as more work runs unattended, the guardrails around it get thicker.

This analysis is conducted for educational and research purposes under fair use principles. All trademarks and software referenced belong to their respective owners. This content is not intended to infringe on any intellectual property rights, circumvent any protections, or encourage unauthorized access to proprietary systems.

Sources#

  1. Claude Code Official Changelog — v2.1.215 and v2.1.214 release notes

Let's Connect

BlueskyLinkedInThreads
← Back to all posts