Full timeline of the March–April limits crisis, all changelog details, and what the new env vars actually control: turboai.dev/blog/claude-code-env-vars-v2-1-108

Workflow changes with real cost implications: `/model` warns before switching mid-conversation because the next response re-reads full history uncached — direct quota burn. `/recap` summarizes what happened while you were away instead of forcing you to scroll. The Skill tool now discovers built-in …

The `DISABLE_TELEMETRY` fix matters more than it looks. Users who set it were silently falling back to 5-minute cache TTL instead of the 1-hour TTL their plan includes. At $200/month on Max 20x, that's paying full price for 12x worse cache reuse — and no error, no warning, nothing. For months. Now …

Prompt cache TTL is now user-configurable. `ENABLE_PROMPT_CACHING_1H` opts into 1-hour TTL. `FORCE_PROMPT_CACHING_5M` locks to 5 minutes. A startup warning fires if caching is accidentally disabled. Rate limit errors (API throttle) and plan limit errors (quota burned) are now separate messages — yo…

Claude Code v2.1.108 is the release where six weeks of usage-limit story finally hits the codebase. Max 20x users watching $200/month quotas drain in 70 minutes. Privacy-conscious subscribers running `DISABLE_TELEMETRY` silently getting 12x worse cache economics for months. This update gives users …

Stability: queued message images no longer dropped. One-shot scheduled tasks stop re-firing. MCP tools available on first turn of headless sessions. Marketplace auto-update no longer breaks when plugin holds files open. 429 errors show clean messages, not raw JSON. Plan mode exit no longer suggests…

Terminal fixes: alt+enter newline works again (regression from 2.1.100). Leading whitespace preserved — ASCII art renders correctly. No garbled bash from clickable file links. No blank screen on prompt line wrap. Washed-out SSH colors fixed for Ghostty, Kitty, Alacritty, WezTerm, foot, rio, Contour.

QOL: WebFetch strips `<style>`/`<script>` so CSS-heavy pages don't exhaust content budget before reaching text. Long single-line writes (minified JSON) truncated in UI instead of paginating. MCP large-output truncation gives format-specific recipes — `jq` for JSON, computed Read chunks for text. Sk…

`ENABLE_AWAY_SUMMARY`: start `/loop`, walk away, get a debrief on return — what completed, what failed, no scrolling. `/loop` is now also `/proactive`. Background plugin monitors auto-arm at session start. PreCompact hooks let you block compaction (exit code 2 or `{"decision":"block"}`). Human-inte…

Stalled API streams no longer hang indefinitely. `ENABLE_BYTE_WATCHDOG` monitors byte flow below SSE — catches half-open TCP, proxy stalls, load balancer timeouts. After 5 min silence: abort, retry non-streaming. Network errors now show retry message immediately instead of silent spinner. Three-tie…

Worktree orchestration gets dedicated plumbing. `REPO_CHECKOUTS` controls where worktrees land on disk — fast SSD, outside iCloud sync. `BASE_REFS` overrides base ref detection so worktree agents diff against correct upstream. `RESUME_FROM_SESSION` chains sessions across context window boundaries. …

Claude Code v2.1.107 — worktrees become a real orchestration layer, stalled streams stop hanging your session, and away summaries debrief you when you return from `/loop`. 🧵

Full breakdown with complete changelog analysis: turboai.dev/blog/claude-code-env-vars-v2-1-104

Also from the changelog: /team-onboarding generates a ramp-up guide from your local Claude Code setup. Between proxy auto-trust and onboarding tooling, the focus this release is clearly on reducing setup friction for teams adopting Claude Code across an org.

CLAUDE_CODE_SDK_HAS_OAUTH_REFRESH tells the runtime that the host app (VS Code, JetBrains, custom integrations) already manages token refresh. Prevents both sides from trying to refresh the same token simultaneously.

Small flag, but it means fewer random auth failures in IDE extensions during long…

Claude Code now trusts the OS certificate store by default. Teams behind TLS inspection proxies (Zscaler, Netskope, Palo Alto) no longer need manual cert configuration to get Claude Code connected.

CLAUDE_CODE_CERT_STORE lets you revert to bundled-only certs if the OS store causes issues. Most tea…

Claude Code v2.1.104 shipped two env vars. Both target enterprise deployment — one for network trust, one for SDK auth coordination. 🧵

Full breakdown of all new variables, removals, and what the Bash overhaul means for auto-mode deployments:

turboai.dev/blog/claude-code-env-vars-v2-1-100

The Bash permission overhaul is the real story. 15+ fixes closing gaps where compound commands, backslash-escaped flags, and /dev/tcp redirects were auto-allowed. Simultaneously, safe patterns like LANG=C sort and cut -d / stopped triggering false prompts. Tightening real gaps while reducing false …

Perforce support landed. CLAUDE_CODE_PERFORCE_MODE makes Edit/Write fail on read-only files with a p4 edit hint. Claude Code has been git-only since launch. Perforce dominates game dev, semiconductor, and VFX pipelines — verticals with strict checkout-before-edit workflows. This reads like a specif…

CLAUDE_CODE_MAX_CONTEXT_TOKENS caps token consumption per session. Works with DISABLE_COMPACT — set both and sessions simply stop instead of summarizing. Enterprise teams get explicit cost control instead of hoping compaction keeps things reasonable.

Claude Code's latest release is one of the most security-significant in weeks. 30+ Bash permission fixes, Perforce VCS support, and a hard context budget knob for enterprise deployments. 🧵

Also: default effort level bumped to "high" for Team and Enterprise users. Anthropic wants the best-quality experience on managed plans, not a degraded one.

Full breakdown on the blog: turboai.dev/blog/claude-code-env-vars-v2-1-94

Smaller wins worth noting: --resume works across worktrees now. CJK text corruption in streaming is fixed. Write tool diffs compute 60% faster. Rate-limited subagents no longer look frozen.

TEAM_ONBOARDING marks a new product surface. Individual setup and team setup are different problems. When a new engineer joins a team already using Claude Code, they need to inherit permissions, MCP configs, and project conventions — not discover them from scratch. First dedicated variable for this.

MCP servers no longer inherit your full environment by default. MCP_ALLOWLIST_ENV lets you specify exactly which env vars each server can see. Every API key, every token — previously leaked into every subprocess. That changes now. Huge for orgs running third-party MCP servers.

Mantle auth is now live. Two new variables — USE_MANTLE and SKIP_MANTLE_AUTH — complete the Bedrock integration that started in v2.1.92 with the setup wizard. Auth proxy first, then runtime plumbing. A two-release rollout that suggests early Bedrock users validated it before the switch flipped.

Claude Code v2.1.94 dropped alongside the Mythos model preview, so it's easy to miss. Several interesting changes in this one. 🧵

Two more: CLAUDE_CODE_EXECPATH overrides binary path detection for symlinked or wrapped installs. CLAUDE_BASE appears to be a compact alternative to CLAUDE_CODE_API_BASE_URL for base configuration.

Full variable breakdown and what they signal about Anthropic's enterprise priorities: turboai.dev/bl…

CLAUDE_CODE_SKIP_FAST_MODE_ORG_CHECK disables org-level eligibility verification for /fast mode. The org check doesn't survive every corporate network topology — this flag closes the gap between "we have access" and "the check can reach the server to confirm it."