Claude Report — 2026-10-19

🚀 Anthropic Official

Claude Code v2.1.302 (Oct 16)

• Adds a new Idle compaction toggle in /config, and CLAUDE_CODE_AUTO_MODE_MODEL to pick the model the client-side auto-mode classifier uses.
• Turns on fork subagents (subagent_type: "fork") by default in claude -p, the Agent SDK, desktop app and VS Code (already default in terminal); CLAUDE_CODE_FORK_SUBAGENT=0 disables it.
• Defaults claude -p/Agent SDK sessions with no permission mode to default mode on third-party providers (was auto); pass --permission-mode auto to keep the old behavior.
• Desktop app and VS Code now always run main-conversation subagents in the background, matching terminal behavior; a subagent's own nested Agent calls run in the foreground by default.
• Caps prompt/agent hooks at 12 concurrent runs per event; the rest queue.
• Reorganizes Claude Tag admin settings into four tabs (Bundles, Connectors, Skills and plugins, Channels) and groups Claude Tag channel-manager roles separately from org roles.
• Fixes plugin marketplaces re-cloning on every refresh and git plugins showing version "unknown" under git's reftable format; fixes MCP servers showing no tools (or every call failing) when negotiating protocol 2026-07-28 but omitting resultType/ttlMs/cacheScope; fixes cloud-session routines that failed partway not resuming correctly — part of 43 total bug fixes.

Claude Code v2.1.303 (Oct 18)

Ships bug fixes and reliability improvements with no further detail published.

Claude Agent SDK (TypeScript) v0.3.302 / v0.3.303 (Oct 16 / Oct 18)

Fixes structured-output results so API refusals return as errors instead of being delivered as structured_output; turns on fork subagents by default in SDK sessions; changes the default permission mode on third-party providers to default; v0.3.303 brings the SDK to parity with Claude Code v2.1.303.

Investing in the public benefit (Oct 18)

Anthropic commits, subject to stockholder approval, 5% of its shares to the new Anthropic Public Benefit Fund — a wholly-owned LLC, not a 501(c)(3) — with an additional 2% addable at board discretion for every 50% increase in company value above a baseline; combined with founder and employee equity pledges, roughly 20% of Anthropic's shares are now earmarked for public-good work, with none usable for lobbying or political giving.

Anthropic pitches Congress a draft frontier-AI safety bill (Oct 16)

Anthropic quietly circulates draft federal legislation, per Politico, that would let the government block a frontier model's release only over a foreseeable, imminent risk of $1B+ in damages or 50+ deaths/serious injuries, and only via three pathways (bioweapon, model escaping developer control, or cyberattack); independent evaluators registered with Commerce would have 72 hours to report an "imminent enumerated risk," only a DC federal judge could suspend a model on Commerce's request, and the bill would preempt some state AI-safety disclosure rules.

Anthropic opens Claude's values to outside input (Oct 18)

Anthropic convenes law professors and democracy experts on campus to debate who should set Claude's values, launching a consultation process with ethicists, community organizers, policy experts and users weighing in on contested behavior and cross-cultural values — while Anthropic retains final decision-making authority.

🔌 Claude Code Plugins

Plugin tracking skipped this run — marketplace install-count cache unavailable, per instructions.

🛠️ Skills

None of the top 5 updated in window.

Unchanged in window: addyosmani/agent-skills (last commit Oct 13), anthropics/skills (last commit Oct 9), kharmanskyi/open-steps (last commit Oct 5), K-Dense-AI/scientific-agent-skills (only an automated security-scan-report bot commit on Oct 19, no content change), cloudflare/security-audit-skill (last commit Sep 14).

🤖 Agents & Subagents

davila7/claude-code-templates (Oct 16-19)

Ships ultra-config — a centralized settings-profiles system with a designed /config form and per-model effort settings — adds real 404 pages, llms.txt, OpenAPI and trust pages to its public agent dashboard, enables Supabase row-level security on its download-tracking tables, and swaps its Bright Data sponsor section for Swiftproxy; this builds on this week's bluesky-feed mod and NVIDIA SkillSpector scoring badge.

nicobailon/pi-subagents v0.79.0 / v0.79.1 (Oct 17-18)

v0.79.0 lets a session inherit background runs from its predecessor in handoff scenarios and lets coordinators resume finished workers without restarting processes; v0.79.1 follows by withholding the bg_wait/subagent_supervisor tool definitions (~1,500 chars) until a session actually delegates, and stops embedded hosts from launching children with the wrong bundled Pi install.

wshobson/agents (Oct 17)

Fixes the documentation-generation skill to avoid overwriting an existing ADR README in its table-of-contents example.

Unchanged in window: VoltAgent/awesome-claude-code-subagents (last commit Oct 11), anthropics/commerce-agents (last commit Aug 31).

🔗 MCPs & Integrations

modelcontextprotocol/servers v2.0.0 (Oct 19)

Refactors every reference server (filesystem, memory, sequential-thinking, fetch, git, time) onto the modern-era MCP spec (2026-07-28) while still answering legacy clients: TypeScript servers migrate to SDK v2, Python servers move to the high-level MCPServer class with standardized error handling and structured output schemas, and the Node version floor rises to 24.

upstash/context7 context7-mcp@4.4.3 (Oct 19)

Updates MCP SDK dependencies so tool schemas convert lazily, cutting per-request server build costs, improves HTTP server shutdown to finish open connections gracefully, and adds clearer error logging for malformed client requests.

Unchanged in window: ChromeDevTools/chrome-devtools-mcp (last release Oct 15), DeusData/codebase-memory-mcp (last release Oct 16, already covered previously, no new evolution), oraios/serena (last commit Oct 15).

💡 Community — Workflows & Ideas

A hook that suggests a better way before Claude builds (Oct 17)

A 36-line pre-message hook plus one skill file and a ledger lets Claude propose a simpler/mechanical alternative before building what was literally asked for, logging which suggestions get accepted so future suggestions follow the user's actual preferences.

📰 Quick Mentions

• MCP reference servers hit v2.0.0 (Oct 19) — full modern-spec refactor, see MCPs section above for detail.
• Anthropic Public Benefit Fund announced (Oct 18) — 5% share commitment, see Official section above.