Jonathon's AI Wiki

Claude Code Week 23 Release Digest (May 28-30, 2026)

12 min read

Source: raw/anthropic-watch-claude-code-tag-v2-1-154.md / -v2-1-156.md / -v2-1-157.md (CC v2.1.154/156/157), raw/anthropic-watch-anthropic-sdk-python-tag-v0-105-0.md / -v0-105-1.md / -v0-105-2.md (Python SDK v0.105.0-v0.105.2), raw/anthropic-watch-anthropic-sdk-typescript-tag-sdk-v0-100-0.md / -sdk-v0-100-1.md / -aws-sdk-v0-3-1.md / -foundry-sdk-v0-2-4.md (TS SDK v0.100.0-v0.100.1 + aws/foundry sub-packages), raw/anthropic-watch-anthropic-cookbook-commit-2ec3c494*.md / -31cfc00c*.md (Agent SDK hosting cookbook), raw/anthropic-watch-skills-commit-da20c925*.md (skills #1216 — Opus 4.8 migration guide); raw/x-account-claudecodelog-2060460301825290524.md (@ClaudeCodeLog v2.1.157 changelog summary — 2026-05-29)

Week 23 is the Claude Opus 4.8 launch week. The model lands inside Claude Code as v2.1.154 (May 28), which also flips on dynamic workflows (the ultracode orchestrate-tens-to-hundreds-of-agents primitive), drops a far cheaper Fast mode for Opus 4.8, and makes the lean system prompt the default for current-generation models. The Python (v0.105.0) and TypeScript (sdk-v0.100.0) SDKs ship API-side claude-opus-4-8 support the same day in Stainless lockstep, alongside two new API surfaces — mid-conversation system blocks and usage.output_tokens_details. The next 24 hours are cleanup: CC v2.1.156 patches an Opus-4.8 thinking-block API error, v2.1.157 lands a plugin-loading + worktree-management batch, and SDK point releases carry a streaming-compaction fix plus dependency plumbing. On the docs side, the cookbook ships a new Agent SDK self-hosting cookbook (Docker / Modal / Kubernetes) and anthropics/skills adds an Opus 4.8 migration guide to the claude-api skill.

v2.1.155 not captured

The ecosystem watcher captured v2.1.154, then jumped to v2.1.156 — v2.1.155 is absent from raw/ and its changelog is not in this digest. Confirm via the GitHub release page if the 155 detail matters.

Headline changes

Claude Code v2.1.154 (May 28) — Opus 4.8 + dynamic workflows

Source: raw/anthropic-watch-claude-code-tag-v2-1-154.md (tagged 2026-05-28T18:00:55Z, author ashwin-ant); the GitHub stub truncated mid-sentence at “Plugins can now …”, so the Additional v2.1.154 surfaces below are recovered from the live changelog (ai-research/claude-code-docs-changelog-2026-05-29.md). Note: that changelog renders the default effort as “xhigh”; the GitHub release note (authoritative) says high — high is correct.

The marquee release of the week — where Claude Opus 4.8 becomes the in-CLI default and dynamic workflows flip on:

  • Opus 4.8 is the new default model, defaulting to high effort. Use /effort xhigh for the hardest tasks. The reasoning-effort knob (low/medium/high/xhigh/max) is now the primary tuning surface, and high is the out-of-the-box setting.
  • Dynamic workflows (research preview). “Ask Claude to create a workflow and it orchestrates work across tens to hundreds of agents in the background. Run /workflows to view your runs.” CLI-side enablement of the feature the Dynamic Workflows announcement and the [[claude-ai/claude-code-workflows-tool-walkthrough|/workflows walkthrough]] document — dynamically-planned, self-verifying parallel fan-out.
  • Fast mode on Opus 4.8 — much cheaper. “Now available at a fraction of its previous cost: 2x the standard rate for 2.5x the speed.” Recalibrates the Fast-mode economics from the W22 baseline (where Fast had quietly defaulted to Opus 4.7).
  • Lean system prompt is now the default for all models except Haiku, Sonnet, and Opus 4.7-and-earlier. Current-generation models (Opus 4.8) boot with the lean prompt; older/smaller models keep the prior prompt.
  • Fewer forced multiple-choice prompts. “Claude now reserves the multiple-choice question prompt for decisions it genuinely cannot make itself, instead of asking when it already has enough context to proceed.” The inverse of the W22 (v2.1.146) fix that stopped auto-mode from over-suppressing AskUserQuestion — this one stops it from over-asking.
  • /simplify is cleanup-only again. “/simplify now runs a cleanup-only review (reuse, simplification, efficiency, altitude) and applies the fixes, instead of running the full /code-review --fix bug-hunting review.” Partially reverts the W22 (v2.1.152) merge that folded /simplify into /code-review --fix. Net state: /code-review = bug-hunting, /simplify = cleanup-only — distinct commands again. Update muscle memory.
  • /effort slider relabeled from “Speed”/“Intelligence” to “Faster”/“Smarter”.
  • claude agents: ! <command> runs a shell command as an attachable/detachable background session (also claude --bg --exec '<command>'). Extends the W22 --bg + /resume bg unification with an inline one-keystroke launcher.
  • claude agents: /logout now actually signs you out instead of being routed to a background session.
  • ←← to open the agents view now works on Bedrock, Vertex, Foundry, and with telemetry disabled.
  • Claude in Chrome: pick which connected browser to use via /chrome → "Select browser…", or in-chat when multiple browsers are connected.

Additional v2.1.154 surfaces (recovered from the live changelog — the GitHub stub cut off at “Plugins can now …“):

  • CLAUDE_CODE_SESSION_ID and CLAUDECODE=1 are now passed to stdio MCP servers — an MCP process can identify the invoking session for per-session telemetry / rate limits without out-of-band config.
  • Plugins can declare defaultEnabled: false (installed-but-disabled by default; users enable features via /plugin), and the /plugin Discover tab pins directory-relevant plugins (those matching the project’s language/stack surface first).
  • Streaming tool execution is always enabled, including on Bedrock / Vertex / Foundry (previously gated).
  • Improved auto-mode data-exfiltration detection — the classifier better catches commands that would move data outside the workspace.
  • /remote-control autocomplete shows “Disconnect Remote Control” when a session is already active.
  • Deprecated CLAUDE_CODE_OPUS_4_6_FAST_MODE_OVERRIDE env var is removed on 2026-06-01.

Claude Code v2.1.156 (May 29) — Opus 4.8 thinking-block hotfix

Source: raw/anthropic-watch-claude-code-tag-v2-1-156.md (tagged 2026-05-29T01:42:23Z). A single-line hotfix ~7 hours after v2.1.154:

  • Fixed an issue when using Opus 4.8 where thinking blocks were modified, leading to API errors. A day-one Opus-4.8 regression — anyone who picked up v2.1.154 and saw thinking-related API errors should jump straight to v2.1.156. Pairs with the same-day SDK fix in TS sdk-v0.100.1 (carry encrypted_content on beta compaction blocks) — both are thinking/reasoning-content correctness fixes against the new model.

Claude Code v2.1.157 (May 29) — plugin auto-loading + worktree management

Source: raw/anthropic-watch-claude-code-tag-v2-1-157.md (tagged 2026-05-29T20:20:38Z). Note: truncated mid-sentence at “Fixed copy-on-select in claude agents not …” — the final entry is incomplete.

New surfaces:

  • Plugins in .claude/skills directories are now auto-loaded — no marketplace required. A meaningful step in the skills-as-first-class-primitive arc (W22’s CLAUDE_CODE_SYNC_SKILLS, bundled claude-api skill): drop a plugin into .claude/skills and it loads without registering a marketplace.
  • claude plugin init <name> scaffolds a new plugin in .claude/skills — a create-and-it-just-works local plugin loop.
  • Autocomplete for /plugin arguments — subcommands, installed plugin names, and plugins from known marketplaces.
  • Config setting to disable the “workflow” keyword trigger for dynamic workflows. ^[inferred — surfaced in the @ClaudeCodeLog v2.1.157 summary, not in the truncated GitHub stub] Opt out of having the word “workflow” in a prompt auto-launch the dynamic-workflows orchestrator, for teams that use the term in ordinary requests.
  • claude agents: the agent field in settings.json is now honored for dispatched sessions, with --agent <name> to override.
  • EnterWorktree can now switch between Claude-managed worktrees mid-session.
  • tool_decision telemetry events now include tool_parameters (bash commands, MCP/skill names) when OTEL_LOG_TOOL_DETAILS=1. Extends the W22 (v2.1.145) OTEL subagent-trace work with parameter-level detail behind an explicit opt-in.
  • Claude-managed worktrees are left unlocked when the agent finishes, so git worktree remove/prune can clean them up.

Fixes:

  • Unprocessable images (zero-byte, corrupt) attached via paste, MCP, or dialog no longer crash the request — they become a text placeholder. Sibling to the W22 (v2.1.144) image-extension-mismatch fix; the image-handling surface keeps hardening.
  • Sandbox network permission prompts no longer appear in auto and bypass-permissions mode when using the desktop app, IDE extensions, or SDK.
  • claude agents completed sessions now retire correctly when an idle subagent is parked or had leaked a backgrounded shell.
  • claude agents: Esc now cancels a slow “opening…” instead of leaving the list unresponsive.
  • Background agent worktrees under .claude/worktrees/ are no longer orphaned after the 30-day job-retention sweep.
  • Background sessions re-attached after sleep/wake now tell the model the correct date.
  • Copy-on-select in claude agents — (fix description truncated in the stub; confirm on the GitHub release page).

Claude Code v2.1.158 (May 30) — auto mode reaches the enterprise clouds

Source: raw/anthropic-watch-claude-code-tag-v2-1-158.md (tagged 2026-05-30T02:42:16Z, author ashwin-ant). A single-line release:

  • Auto mode is now available on Bedrock, Vertex, and Foundry for Opus 4.7 and Opus 4.8 — opt in with CLAUDE_CODE_ENABLE_AUTO_MODE=1. Closes the gap where auto mode (the permission-classifier that minimizes prompt interruptions, made opt-in-free for first-party endpoints in W23’s v2.1.154) was unavailable to teams running Claude Code against the three enterprise model-provider clouds. Pairs naturally with dynamic workflows, which Anthropic recommends running under auto mode to avoid per-action permission stops on large fan-outs.

Anthropic SDKs — Opus 4.8 support + new API surfaces (May 28-29)

Sources: raw/anthropic-watch-anthropic-sdk-python-tag-v0-105-{0,1,2}.md; raw/anthropic-watch-anthropic-sdk-typescript-tag-{sdk-v0-100-0,sdk-v0-100-1,aws-sdk-v0-3-1,foundry-sdk-v0-2-4}.md. All by stainless-app[bot]; .0 releases May 28, patch/sub-package releases May 29.

  • Python v0.105.0 / TS sdk-v0.100.0 (May 28) — same headline feature in lockstep: api: Add support for claude-opus-4-8, mid-conversation system blocks, and usage.output_tokens_details (Python f18b014, TS bb0bf27). Three things land together: (1) the claude-opus-4-8 model id as first-class; (2) mid-conversation system blocks (inject system-role content partway through a conversation — the API-side surface behind the CLI’s lean-prompt change); (3) usage.output_tokens_details for finer-grained output-token accounting. Python v0.105.0 adds one TS-absent extra — support custom file size caps (#1825) (7e5f944) — plus a docs chore and the cookbook example rename managed-agents private-sandbox-worker → self-hosted-sandbox-worker (#1822).
  • TS sdk-v0.100.1 (May 29) — patch. Bug fix: streaming: carry encrypted_content on beta compaction blocks (#1025) (eccddf3) — the TS port of the Python v0.104.1 fix from W22; the streaming accumulator now preserves encrypted_content on compaction blocks. Plus a standardwebhooks lockfile chore (5e9b523).
  • Python v0.105.1 (May 29) — chore-only. internal: use Trusted Publishing for PyPI releases (1d04fc5). OIDC-based PyPI publishing; no public-facing change.
  • Python v0.105.2 (May 29) — empty. Changelog body shows no entries. A bare version bump.
  • TS aws-sdk-v0.3.1 and foundry-sdk-v0.2.4 (May 29) — chore-only. Both ship the identical standardwebhooks lockfile chore (5e9b523), cut across all three TS sub-packages. Pure plumbing.

Net: anyone calling Opus 4.8 via a Python or TS client should bump to py v0.105.0+ / TS sdk-v0.100.0+ for first-class model support, mid-conversation system blocks, and usage.output_tokens_details. TS streaming users running context-compaction should take sdk-v0.100.1. The aws-sdk / foundry-sdk / py v0.105.1/.2 bumps are chore-only. See Anthropic SDK Releases — May 2026 for the full rollup.

Cookbook: Agent SDK self-hosting cookbook — Docker / Modal / Kubernetes (May 28)

Sources: raw/anthropic-watch-anthropic-cookbook-commit-2ec3c494*.md (feat, kevintang513) and -31cfc00c*.md (merge of PR #677).

  • feat(claude_agent_sdk): add hosting cookbook — Docker, Modal, and Kubernetes tiers (2ec3c494), merged via PR #677 as “Agent SDK cookbook 07: Hosting your agent.” A new entry in the claude_agent_sdk series covering how to host a Claude Agent SDK app across three deployment tiers — local containerization (Docker), serverless (Modal), and orchestrated clusters (Kubernetes). Distinct from the W22 CMA self-hosted-sandbox worker templates (Anthropic-orchestrated loop, customer-side tool execution); this is for the fully self-hosted Agent SDK (you run the whole loop). The two fill out the “where does my agent run” matrix. Follow-up (May 29-30): PR #678 plus several kevintang513 commits (address #677 review comments, ruff format + lint fixes) hardened this hosting cookbook — Content-Length body-size limits, GATEWAY_TENANTS parsing, k8s AGENT_IMAGE validation, kind-quickstart.sh auto-restart. Refinement of the same PR #677 surface, no new cookbook.

Commit-message-only source

Both cookbook stubs are GitHub commit notifications — only the commit/PR title was captured, not the notebook contents. Pull the cookbook for the implementation detail.

Skills: Opus 4.8 migration guide added to the claude-api skill (May 29)

Source: raw/anthropic-watch-skills-commit-da20c925*.md (commit da20c925, PR #1216, rlancemartin). Note: commit body truncated at “invalid vault credentials don’t block se…“.

PR #1216 — “Add Opus 4.8 migration guide and model updates to claude-api skill” — does two things:

  • (a) Opus 4.8 migration guide + model updates added to the [[claude-ai/managed-agents|claude-api]] skill — the new W23 content (migration guidance for moving claude-api-skill code onto claude-opus-4-8).
  • (b) Re-lands the W22 CMA self-hosted-sandbox doc surface in the same PR (the shared/managed-agents-self-hosted-sandboxes.md content, sessions.update(), >100K MCP-output file offload). This is the same content already documented in Week 22 Digest and the self-hosted-sandbox article — no new CMA surface here; #1216 bundles the prior skills doc work (PR #1164) with the new Opus 4.8 layer.

Key Takeaways

  • Opus 4.8 is the week’s headline, arriving across all three surfaces at once: the CLI (v2.1.154 default model, high effort), the API/SDKs (py v0.105.0 / TS sdk-v0.100.0 add the claude-opus-4-8 id), and the docs (claude-api skill migration guide via #1216). See Claude Opus 4.8.
  • Dynamic workflows ship in the CLI with Opus 4.8 (v2.1.154) — ultracode / “ask Claude to create a workflow,” orchestrating tens-to-hundreds of agents, /workflows to view runs. Full coverage in Dynamic Workflows and the [[claude-ai/claude-code-workflows-tool-walkthrough|/workflows walkthrough]].
  • Two new API surfaces land with Opus 4.8 support: mid-conversation system blocks and usage.output_tokens_details (both py v0.105.0 / TS sdk-v0.100.0). The mid-conversation system block is the API-side counterpart to the CLI’s lean-prompt-by-default change.
  • Day-one Opus 4.8 correctness fixes dominate May 29: CC v2.1.156 fixes thinking-block-mutation API errors, TS sdk-v0.100.1 ports the Python v0.104.1 encrypted_content-through-compaction fix. If you adopted v2.1.154 / the .0 SDKs on launch day, take the May 29 patches.
  • /simplify reverts to cleanup-only (v2.1.154), undoing the W22 merge that folded it into /code-review --fix. The two commands are distinct again.
  • Fast mode on Opus 4.8 is much cheaper — 2x standard rate for 2.5x speed.
  • Plugins auto-load from .claude/skills with no marketplace (v2.1.157), plus claude plugin init scaffolding — another step in skills/plugins becoming a first-class local primitive.
  • A new Agent SDK self-hosting cookbook (Docker / Modal / Kubernetes, PR #677) complements the W22 CMA self-hosted-sandbox templates.
  • Stainless lockstep holds: Python and TS shipped Opus 4.8 support the same day, and the May 29 patch wave hit both SDKs together.
  • Auto mode reaches enterprise clouds (v2.1.158, May 30): now available on Bedrock / Vertex / Foundry for Opus 4.7 + 4.8 via CLAUDE_CODE_ENABLE_AUTO_MODE=1 — the recommended pairing for dynamic workflows.
  • v2.1.155 was not captured by the ecosystem watcher — its changelog is not in this digest.
  • Claude Opus 4.8 — the model that launched this week; v2.1.154 makes it the in-CLI default and the SDKs add the API id
  • Dynamic Workflows in Claude Code — the orchestrate-tens-to-hundreds-of-agents primitive that flips on in v2.1.154 (ultracode)
  • [[claude-ai/claude-code-workflows-tool-walkthrough|Claude Code /workflows Walkthrough]] — the hands-on mechanics companion to dynamic workflows
  • Week 22 Digest — the immediately-preceding window; CMA self-hosted sandboxes, the /simplify↔/code-review merge that v2.1.154 partially reverts, and the Python v0.104.1 compaction fix that TS sdk-v0.100.1 now ports
  • Anthropic SDK Releases May 2026 — full May SDK rollup including these v0.105.x / sdk-v0.100.x releases
  • Managed Agents — the claude-api skill that PR #1216 adds the Opus 4.8 migration guide to
  • Managed Agents Self-Hosted Sandboxes — the CMA self-hosted surface PR #1216 re-lands alongside the Opus 4.8 docs
  • Claude Agent SDK — the SDK the new Docker/Modal/Kubernetes hosting cookbook targets
  • CLI Reference — /effort, /simplify, /code-review, /workflows, claude agents, /chrome surfaces touched this week

Graph View

Backlinks