Claude Report — 2026-10-20

🚀 Anthropic Official

Claude Code v2.1.304 (Oct 19)

76 changes: 42 bug fixes, 19 improvements, 8 new features, 3 security-related changes.

• Adds /auto-mode-setup, a guided setup that reads the project and recent sessions to teach auto mode the environment, now also available on Bedrock, Vertex, Foundry, Claude apps gateway and telemetry-disabled sessions.
• Adds /claude-api managed-agents-secure, which checks a Managed Agents agent's settings against Anthropic's published security guide without changing anything.
• Adds review levels to Code Review: organization owners choose Quick, Standard (default), or Deep, with per-repository overrides, plus a "review every repo in the org automatically" option.
• Adds a warning on Pro/Max plans when resuming an inactive conversation would burn a noticeable share of the 5-hour usage limit.
• Fixes claude --worktree silently deleting the worktree and its branch at exit when HEAD was detached or on another branch — a commit-loss risk.
• Fixes context_management: Extra inputs are not permitted errors behind LLM gateways by resending the request without that field.
• Fixes a "Context limit reached"/"Prompt is too long" error when stopping a turn mid-compaction.
• Fixes a failed mod reload silently dropping the still-loaded version's tools and commands.
• Improves responsiveness and throughput when many subagents run concurrently, and memory use in long/resumed conversations.
• [VSCode] Adds mod drawing in the chat panel for Bedrock/Vertex/Foundry/gateway/telemetry-disabled sessions; background subagents now collapse into one status line per Agent card.

Claude Agent SDK (TypeScript) v0.3.304 (Oct 19)

Brings the SDK to parity with Claude Code v2.1.304 — no SDK-specific changes called out beyond the version sync.

Anthropic in talks for a second $45B Nscale compute deal (Oct 19)

Nscale and Anthropic are reportedly negotiating roughly $45B for a second building at Nscale's Monarch, West Virginia campus — 460MW running Nvidia's Vera Rubin chips — on top of the $45B deal struck in August; Anthropic also holds right of first refusal on Monarch's planned third building, per Bloomberg.

🔌 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 20)

Points its Stripe MCP config at the official remote Stripe server instead of a third-party one, rewrites the stripe-integration skill with current Stripe guidance, and adds a confirm-before-write setting for Stripe MCP calls plus a mobile-friendly reading layout for its blog.

nicobailon/pi-subagents v0.79.2 (Oct 20)

Makes delegation survive restarts — a background child's completion notice now reaches the parent even if the session restarted first — fixes steering so a timed-out steer no longer blocks later steers on that run, stops children from inheriting a session's APPEND_SYSTEM.md, and fixes workflows launched from very long tool-call-id providers (e.g., LiteLLM with Gemini) failing at launch.

Unchanged in window: wshobson/agents (last commit Oct 17, already covered previously, no new evolution), VoltAgent/awesome-claude-code-subagents (last commit Oct 11), anthropics/commerce-agents (last commit Aug 31).

🔗 MCPs & Integrations

oraios/serena v2.1.0 (Oct 19)

Fixes a critical security issue where TypeScript projects auto-executed a repo-supplied tsserver.js (now pinned to Serena's managed TypeScript unless the project is explicitly trusted), restores Claude Desktop compatibility by renaming the session_id parameter back to session, and adds a native-LSP typescript_tsgo language server plus VB.NET support via Roslyn.

upstash/context7 context7-mcp@4.4.4 (Oct 19)

Fixes --port to reject non-integer or out-of-range values instead of crashing on retry, rejects --port/--api-key combined with incompatible transports, and narrows the server instructions so agents reach for Context7 specifically when porting code between libraries, platforms, or SQL dialects.

Unchanged in window: modelcontextprotocol/servers (v2.0.0 shipped Oct 19 but already covered previously, no new evolution), ChromeDevTools/chrome-devtools-mcp (last release Oct 15), DeusData/codebase-memory-mcp (last release Oct 16).

💡 Community — Workflows & Ideas

krites-dev/krites (Oct 19)

A Claude Code plugin that blocks session stop until configured test/check commands pass, protects named files from agent edits, and generates a cryptographically signed receipt for each verification run.

ehnruiz/claude-skills (Oct 19)

Ships 21+ skills, subagents (scout/implementer/reviewer) and six hooks built around proving a fix's test actually fails without the fix, routing work to the cheapest capable model, and handing off cleanly between sessions.

gojko/lights-on-software-factory (Oct 20)

A boilerplate for an unattended Claude Code "software factory" — sandboxed execution, a bash-command regulator, linter-based feedback, and a workflow orchestrator the author reports cuts token spend by about 80% versus a bare-bones Claude workflow.

Optimizing MCP tools so Claude Code, Codex and OpenCode actually find them (Oct 16-18)

A five-step framework (vocabulary alignment, front-loaded first-line impact, guided outputs, ~5,000-token result budgets, real-job test suites) addresses how Claude Code, Codex, Grok Build and OpenCode use different tool-discovery mechanisms (name-based, keyword filter, BM25, partial catalogs) that each need separate tuning.

Quorus-dev/Quorus (Oct 19)

A relay system that puts Claude Code and Codex in a shared "room" to split a task — one agent fixes code in its own git worktree while the other reviews before push — coordinated via message fan-out and a distributed lock.

📰 Quick Mentions

• Earshot (Oct 19) — talk to Claude Code hands-free via AirPods, nod to send.
• Craig (Oct 19) — a TUI for running and watching multiple parallel coding agents at once.
• Faztudo (Oct 20) — a local chat UI for coordinating several coding-agent CLIs from one place.
• Remakecraft (Oct 19) — a Claude Code plugin that rewrites an app from observed behavior rather than reading its source.
• Replit drops Anthropic as its default model provider for OpenAI (Oct 17) — per Replit's CEO, a competitive signal in the coding-agent market.