Jonathon's AI Wiki

Claude Code Week 25 Release Digest (June 6-11, 2026)

9 min read

Source: raw/anthropic-watch-claude-code-tag-v2-1-166.md through -v2-1-173.md (CC v2.1.166–v2.1.173, June 6–11 2026, anthropic-ecosystem-watch stubs); ai-research/claude-code-changelog-w25-2026-06-10.md (authoritative GitHub CHANGELOG, pulled to de-truncate the watcher stubs); platform additions via raw/x-account-claudedevs-2064756984617021807.md (Apple Foundation Models) + raw/x-account-claudedevs-2064428351029449214.md (SDK refusal-fallback middleware), WWDC context via raw/WWDC_Isn_t_About_Siri._It_s_Jensen_Huang_s_Problem..md. SDK / cookbook / skills releases of the same week are folded into the SDK-releases article and Managed Agents. 2026-07-24 correction pass: raw/anthropic-watch-skills-commit-1f630fdf9259cec4a14913127dfd7c3b69ef72eb.md (anthropics/skills PR #1463, 2026-07-22) supersedes one claim in this digest — see the [!contradiction] callout below.

Week 25 is a Managed-Agents-enablement and Fable-5-access week, wrapped around a batch of Claude Code developer-experience hardening. The throughline runs across all four Anthropic feeds at once: CC v2.1.170 ships Fable 5 access; the Python/TS SDKs (v0.109 / v0.104) add Managed Agents deployments + environment-variable credentials and a frontier_llm refusal category; the cookbook adds a scheduled Managed-Agent example; and the claude-api skill documents both. The CC releases themselves are a DX/safety batch — fallbackModel, --safe-mode, /cd, disableBundledSkills, a new post-session hook, and hardened cross-session messaging — with v2.1.167/168 bug-fix-only.

Headline changes (Claude Code)

v2.1.166 (June 6) — fallbackModel, glob deny rules, cross-session auth hardening

  • fallbackModel setting (+ --fallback-model for interactive sessions). Configure up to three fallback models tried in order when the primary is overloaded or unavailable. Claude Code also now retries a turn once on the fallback model when the API rejects an unexpected non-retryable error (auth / rate-limit / request-size / transport errors still surface immediately). Resilience for long unattended runs — see dynamic workflows.
  • Glob patterns in deny-rule tool-name position. "*" denies all tools; allow rules reject non-MCP globs; unknown tool names in deny rules warn at startup. A real tightening of the permission grammar — see CLI Reference.
  • Hardened cross-session messaging (security). Messages relayed via SendMessage from other Claude sessions no longer carry user authority — receivers refuse relayed permission requests, and auto mode blocks them. Closes a privilege-relay path in multi-agent fan-outs (subagents / agent teams).
  • Thinking-disable controls. MAX_THINKING_TOKENS=0, --thinking disabled, and the per-model toggle now actually disable thinking on models that think by default via the Claude API (3P providers unchanged). Note this does not apply to Fable 5, whose adaptive thinking is always on.
  • Plus: claude update announces the target version before downloading; claude agents URL-filter; fixes for unprocessable-image token waste, worker-registration stalls, JetBrains 2026.1+ terminal flicker, and ${VAR} MCP-policy predicates.

v2.1.167 (June 6) & v2.1.168 (June 6) — bug fixes only

Both releases are “bug fixes and reliability improvements” with no documented user-facing feature changes.

v2.1.169 (June 8) — the DX release: —safe-mode, /cd, disableBundledSkills, post-session hook

  • --safe-mode (+ CLAUDE_CODE_SAFE_MODE). Start Claude Code with all customizations disabled — CLAUDE.md, plugins, skills, hooks, MCP servers — for clean-room troubleshooting. The fastest way to answer “is it my config or Claude Code?”
  • /cd command. Move a session to a new working directory without breaking the prompt cache mid-session — previously a directory change meant a cache-cold restart.
  • disableBundledSkills setting (+ env var). Hide bundled skills, workflows, and built-in slash commands from the model — a leaner tool surface for users who want only their own.
  • New post-session lifecycle hook (self-hosted runner). Runs after the session ends and before the workspace is deleted, so you can snapshot uncommitted work or export logs; the child-process SIGTERM→SIGKILL window is now configurable (default 5s). A net-new hook event — see Hooks.
  • /workflows opens immediately even while a turn is in progress; TaskCreate auto-repairs malformed inputs and includes the schema in validation errors.
  • claude agents --json now includes blocked + just-dispatched sessions, adds --all (completed sessions) and new id / state fields — better scripted-dashboard surface.
  • Hardening: untrusted project settings can no longer set OTEL client-cert paths without trust confirmation; invalid managed-settings entries now apply remaining valid policies instead of silently dropping the payload; the “CLAUDE.md too long” threshold scales with the model’s context window. Plus a long list of Windows / remote / macOS fixes and the MCP-policy enforcement gaps (reconnect / IDE-typed / --mcp-config / cold-start).

v2.1.170 (June 9) — Claude Fable 5 access

  • Update to v2.1.170 for Claude Fable 5 — the Mythos-class model made safe for general use. This is the CC-side gate for the day’s model launch.
  • Fixed sessions not saving transcripts (and not appearing in --resume) when launched from the VS Code integrated terminal or any shell that inherited Claude Code environment variables.

v2.1.172 (June 10) — nested sub-agents (5 levels deep), Bedrock region autodetect

  • Sub-agents can now spawn their own sub-agents, up to 5 levels deep — deeper nested delegation, the runtime substrate for dynamic workflows (which reached GA the same day).
  • Amazon Bedrock reads the AWS region from ~/.aws config when AWS_REGION isn’t set (matching AWS SDK precedence); /status shows where the region came from.
  • Plugin marketplace search bar when browsing a marketplace’s plugins in /plugin; model attribute added to the claude_code.lines_of_code.count OTEL metric.
  • Notable fixes: 1M-context sessions without usage credits getting permanently stuck now auto-compact back under the standard limit; background agents no longer read another directory’s project settings (.mcp.json approvals/trust) when dispatched onto a pre-warmed worker; availableModels restrictions now apply to subagent model overrides and the agent-dispatch model picker.

v2.1.173 (June 11) — [1m] suffix normalization for Fable 5

  • Fable 5 model names with a [1m] suffix are now normalized automatically — Fable 5 includes 1M context by default, so the suffix is stripped rather than treated as a distinct model. (If your configs carry claude-fable-5[1m] from the Opus-era pattern, they keep working — but the suffix is now redundant on Fable.)
  • Fixed a spurious “sandbox dependencies missing” startup warning on Windows when sandbox was enabled in settings.

Ecosystem — the Managed Agents deployments throughline

The same week’s SDK / cookbook / skills commits are the real story under the CC point-releases:

  • SDKs (Python v0.109.0 / TS v0.104.0): add support for Managed Agents deployments and environment-variable credentials. Point releases (v0.109.1 / v0.104.1) add a frontier_llm refusal category — the SDK-side hook for Fable 5’s topic-gated fallback. Folded into the SDK-releases article.
  • Refusal-fallback middleware across five SDKs (June 9). Python, TypeScript, Go, Java, and C# SDKs gained client-side middleware that detects a refusal, retries on Opus 4.8, and maintains context — completing the frontier_llm category above with an actual retry path for API consumers without server-side fallback (@ClaudeDevs; raw/x-account-claudedevs-2064428351029449214.md).
  • Apple Foundation Models support (June 10). Apple developers can now call Claude through Apple’s Foundation Models framework for multi-step reasoning, code generation, and longer-context tasks beyond the on-device models (@ClaudeDevs; raw/x-account-claudedevs-2064756984617021807.md). Lands the same week WWDC confirmed Apple opening its AI stack outward (Gemini alliance, private-cloud-compute expansion — raw/WWDC_Isn_t_About_Siri._It_s_Jensen_Huang_s_Problem..md).
  • Cookbook: a Sentry-triage scheduled Managed Agent example (provisions an agent with a sentry-cli vault credential, schedules it via the deployments API, writes a triage report each run; requires anthropic>=0.109.0), plus an async multi-agent orchestration cookbook (a fixed N-agent team with peer messaging through a shared hub, and a lead agent that spawns/monitors/dismisses async subagents on the Messages API). See Managed Agents and Agent Teams.
  • claude-api skill: new Managed Agents scheduled-deployments docs (cron schedules, deployment runs, pause/auto-pause), vault env-var credentials (secrets substituted at egress with networking allowlists), a system.message event for mid-session system-prompt updates (Opus 4.8 only), and a full Fable 5 / Mythos 5 model-table + migration section.

system.message: model coverage and prompt semantics

Existing claim: (from whats-new-2026-w25.md) — “a system.message event for mid-session system-prompt updates (Opus 4.8 only)” New source says: (from raw/anthropic-watch-skills-commit-1f630fdf9259cec4a14913127dfd7c3b69ef72eb.md) — “system.message works on four models, checks only the primary model, appends system context instead of replacing the prompt, and is accepted during a requires_action idle when it trails a tool result in the same request.” Status: resolved (2026-07-24) — upstream self-correction; the new value is authoritative.

Two independent errors in the W25 line: the model scope (four models, not Opus 4.8 alone) and the semantics (appends system context rather than updating/replacing the prompt). The 2026-07-22 claude-api skill sync (anthropics/skills PR #1463) files both under “Corrections to behavior the skill already documented” — the W25 claim traced to the skill’s own earlier text. Resolved rather than left open because the authority that published the original claim published the correction; the callout is retained so a reader holding the “Opus 4.8 only / replaces the prompt” version sees why it changed. The four models are still not named at the source — that gap is tracked in Managed Agents § Open Questions. Current corrected behavior is recorded there.

Key Takeaways

  • Managed Agents got scheduled deployments + vault credentials this week across SDK, cookbook, and skill — the most consequential W25 change for anyone running unattended agents. See Managed Agents.
  • fallbackModel (up to 3 models) is the resilience primitive for long runs — pair it with auto mode and dynamic workflows.
  • --safe-mode is the new first move when debugging a misbehaving Claude Code: disable all customization, then bisect.
  • /cd preserves the prompt cache across a working-directory change — a quiet but real cost win on long sessions.
  • New post-session hook fires after the session ends and before workspace deletion — the place to snapshot uncommitted work on self-hosted runners. See Hooks.
  • Cross-session SendMessage no longer carries user authority (v2.1.166) — a multi-agent privilege-relay hardening; relayed permission requests are refused and auto mode blocks them.
  • CC v2.1.170 is the Fable 5 access gate; the SDK frontier_llm refusal category is its server-side fallback hook.
  • v2.1.167 / v2.1.168 are bug-fix-only; the substance is in v2.1.166, v2.1.169, and v2.1.170.

Graph View

Backlinks