Claude Code Daily Briefing - 2026-06-06

Release Summary

Today (6/6) brought v2.1.166. Yesterday’s v2.1.165 was a quiet “bug fixes and reliability” release, but v2.1.166 lands three substantive changes at once: model fallback, deny-rule globs, and cross-session message security.

Full release notes

New Features & Practical Usage

Don’t stall when your primary model goes down — fallbackModel / --fallback-model (v2.1.166)

v2.1.166 adds a fallbackModel setting that lets you specify up to three fallback models, tried in order when the primary model is overloaded or unavailable. The --fallback-model flag, previously headless-only, now applies to interactive sessions too. On top of that, when the API returns an unexpected error that was previously treated as non-retryable, Claude Code retries the turn once on the fallback model — while auth, rate-limit, request-size, and transport errors still surface immediately as before.

# Fall back to Sonnet if the primary model is overloaded (now works in interactive sessions)
claude --fallback-model claude-sonnet-4-6

In a settings file you can list up to three models under fallbackModel, tried in order (follow the official settings docs for the exact schema once it’s updated). With back-to-back multi-model error incidents from June 2 through June 5, this is a practical availability guard for unattended pipelines and CI: one model wobbling no longer kills the job.

GitHub v2.1.166

Globs in deny rules — lock down every tool with a single "*" (v2.1.166)

As of v2.1.166 you can use glob patterns in the tool-name position of a deny rule, and "*" denies all tools. Conversely, allow rules reject non-MCP globs (to prevent over-broad grants), and unknown tool names in deny rules now warn at startup.

// .claude/settings.json — a deny-by-default skeleton: lock everything, then allow only what you need
{
  "permissions": {
    "deny": ["*"],
    "allow": ["Read", "Grep", "Glob"]
  }
}

When building a strict sandbox or a read-only exploration agent, instead of denying tools one by one you can express a whitelist in a single line: lock everything with "*", then open only what’s needed. (Check the official permissions docs for exact precedence and combination rules.)

GitHub v2.1.166

Developer Workflow Tips

Anthropic’s own playbook for writing skills — nine categories and the Gotchas section

Anthropic’s Claude Code team published the lessons it has learned running hundreds of skills internally (Thariq Shihipar, June 3). The actionable highlights:

• Skills aren’t “markdown files” — they’re folders, and can bundle scripts, assets, and data.
• Cataloging their internal skills, they found everything clusters into nine categories: library/API reference, product verification, data fetching/analysis, business-process/team automation, code scaffolding/templates, code quality/review, CI/CD and deployment, runbooks, and infrastructure operations.
• One skill, one category. The best skills fit cleanly into one; the ones that try to do too much straddle several and confuse the agent.
• The Gotchas section is the highest-signal content in any skill. Capture the failure points Claude repeatedly hits, and keep adding edge cases over time.
• For knowledge skills, don’t state the obvious. Skip what the model already knows and focus on information that pushes Claude out of its normal way of thinking.
• Progressive disclosure. Don’t front-load everything; point Claude to referenced files for specific situations so it reads them at the right time.

Even their marketplace is organic: no central committee — people drop skills into a sandbox folder on GitHub, and once one gains traction it’s promoted via PR.

Anthropic — Lessons from building Claude Code: How we use skills

Turn thinking off for bulk, mechanical work to cut cost and latency — MAX_THINKING_TOKENS=0 (v2.1.166)

As of v2.1.166, MAX_THINKING_TOKENS=0, --thinking disabled, and the per-model thinking toggle now actually disable thinking on models that think by default (via the Claude API; third-party providers are unchanged). For high-volume mechanical passes that need little reasoning — structured transforms, simple substitutions, formatting cleanup in subagents — thinking tokens tend to add cost and latency with no benefit, so turning them off trims per-call cost and response time.

# Disable thinking for this session (even on models that think by default)
export MAX_THINKING_TOKENS=0
claude

Conversely, keep thinking on for work where reasoning drives the outcome — architectural decisions, hard debugging. Deciding where to spend your reasoning budget by task type is one axis of cost governance heading into the June 15 Programmatic Usage Credits change.

GitHub v2.1.166

Security & Limitations

June 5 — elevated errors across many Claude models (6/5)

On June 5, Anthropic’s status page recorded an incident with elevated error rates across multiple Claude models. It started around 08:08 PT (15:08 UTC), with recovery times varying by model — Opus 4.6 recovered first (08:25 PT), Opus 4.5 last (10:29 PT) — before all models fully recovered. Coming after the June 2 outage and June 3 incidents, this is several days of choppy model availability, and a concrete reason to set up the fallbackModel above to hedge against your primary model wobbling.

Claude Status

v2.1.166 — relayed cross-session messages no longer carry user authority (6/6)

v2.1.166 hardens cross-session messaging. Messages relayed via SendMessage from other Claude sessions no longer carry user authority — receivers refuse relayed permission requests, and auto mode blocks them. This closes a path where one session could drive another like a tool to bypass permissions. Two related permission fixes are worth checking too:

• An invalid entry no longer silently disables the rest of your policies. Previously, a single invalid entry in managed settings could quietly disable enforcement of the remaining valid policies in that file; this is fixed.
• MCP-server predicates now match with ${VAR}. Managed-settings allowedMcpServers/deniedMcpServers predicates that used ${VAR} references failed to match; this is fixed.

GitHub v2.1.166

Community News

• “Running an AI-native engineering org” — how the Claude Code team actually works (6/3): Fiona Fung, Director of Engineering for Claude Code and Claude Cowork, shared how the team’s processes and structure changed once agentic coding became the default. With a six-month roadmap going stale by month three, they moved from long-range planning to “just-in-time” planning — prototyping fast, pulling internal users in early, and iterating on feedback. Code review now leans on Claude for style, linting, and bug detection, while humans focus on domain expertise, security-sensitive code, and product judgment. Managers start as ICs to live the engineering experience firsthand, and roles have blurred (PMs prototype code, engineers take on design). claude.com blog

• A salesperson who’d never opened a terminal built a Gmail auto-draft app, “CLAFTS” (6/5): Anthropic GTM PM Jared Sires had never used a terminal before joining, yet built CLAFTS (Claude Drafts) — a Gmail-embedded app of roughly 4,300 lines, almost all written by Claude Code. Facing a book of 600–700 accounts and answering emails until 9–10 p.m., he built an app that pulls context from Google Drive and other tools, references Anthropic’s public docs via web search, and drafts replies in his own writing style, waiting in his drafts folder for review. He now packages these tools as a Claude Cowork plugin for the whole sales team. A vivid example of non-engineers shipping internal tools through agentic coding. claude.com blog

Minor Changes Worth Knowing

• claude update announces the target version: instead of going silent, it tells you which version it’s about to download (v2.1.166)
• claude agents URL filter: typing a URL into the list filters to the session whose first prompt contained it (v2.1.166)
• Recurring “image could not be processed” error fixed: the repeating error and the extra token usage it caused when an unprocessable image was sent are both fixed (v2.1.166)
• JetBrains terminal flicker fixed: synchronized output is enabled on IntelliJ, PyCharm, WebStorm, etc. (2026.1+) to stop flickering (v2.1.166)
• Kitty keyboard protocol Shift+non-ASCII fixed: inputs like Shift+ä → Ä were being dropped in WezTerm, Ghostty, and kitty (v2.1.166)
• Worktree background-agent crash loop fixed: background agent sessions that entered a git worktree no longer crash-loop with “No conversation found” when reopened from claude agents (v2.1.166)

Recommended Reads

• “How to build an AI-native startup”: A six-step framework arguing that the edge comes from discipline, not a better model. Systematically map your workflows, build context systems, encode skills together with evals, and run weekly improvement loops, and a small team learns exponentially faster than a traditional org. It’s the same thesis as Anthropic’s skills post above — “encode your skills” — applied at the company level. cyberfund (X)

• “When AI Creates Itself: Our Progress Toward Recursive Self-Improvement” — Anthropic: Claude’s task-completion horizon is doubling roughly every four months, and if that trend continues alongside enough compute, AI could eventually design and build successor models on its own. Still, deciding which problems matter most remains a human comparative advantage, and meaningful safeguards would require verifiable international coordination among multiple frontier labs. A measured piece that asks how to think about the trajectory as we hand agents more and more. Anthropic

• “Did Claude increase bugs in rsync?”: A data-driven check on whether Claude-assisted rsync releases (v3.4.2, v3.4.3) were actually buggier than past releases. Using severity-weighted bugs per 10 commits, permutation and Fisher’s exact tests return p-values of 46% and 74% — “no statistical evidence that Claude releases are abnormally buggy.” In fact the pre-Claude v3.4.1 had the worst bug rate in the whole dataset. A calm rebuttal of vibes-based criticism of AI coding, with numbers. alexispurslane.github.io

Interesting Projects & Tools

• harness-starter-kit — a repo “harness” so coding agents stop repeating mistakes: An open-source starter kit that persists project rules and decision history inside the repo via AGENTS.md, decision logs, and failure-tracking files, so agents can reference past context across sessions. It ships workflow scripts and profile examples for Django, Next.js, and FastAPI. Where oh-my-harness leaned on enforcing rules, this one emphasizes accumulating the memory of decisions and failures. GitHub

• Agent memory architecture plugin — tackling memory noise and decay: An open-source plugin (Apache 2.0) that structures agents as repos with actual files rather than prompts, and manages memory through a ticket-based system with curator oversight to fight noise and decay. It’s portable across Claude, Codex, and Gemini runtimes with no vendor lock-in, and offers three modes: a self-evolving single agent, an orchestrated multi-agent team with a validation layer, and repackaging of existing workspaces. A head-on attempt at the long-session “memory” problem. GitHub