№ 1772026-10-21

Claude Report — 2026-10-21


🚀 Anthropic Official

Claude Code v2.1.305 (Oct 20)

  • Adds sandbox.allowNestedUserNamespaces so sandboxed Linux commands can create user namespaces.
  • Adds CLAUDE_CODE_RETRY_WATCHDOG_USAGE_LIMIT_RECHECK_MS so unattended retry mode notices a lifted usage limit before its scheduled reset.
  • Fixes requests failing on Vertex, Bedrock and other providers that reject tool search — Claude Code now retries without it.
  • Fixes repository access broken since v2.1.296 when git credential helpers read settings through GIT_CONFIG.
  • Fixes MCP servers from cloud sessions getting silently replaced by same-named servers from local settings.
  • Fixes Workflow resume re-running already-completed pipeline()/parallel() agents instead of reusing cached results.
  • Changes Code Review re-reviews to stop opening with "LGTM" while blocking findings are still open.

Claude Code v2.1.306 (Oct 21)

  • Fixes claude -p inside cloud sessions wrongly reporting "Not logged in" when started without the session's environment variables.
  • Fixes an API error getting masked as "prompt is too long" (triggering needless compaction) in sessions with many MCP tools.

Claude Code v2.1.303 (Oct 18)

Bug fixes and reliability improvements, no user-facing feature changes.

Claude Agent SDK (TypeScript) v0.3.305 (Oct 20) / v0.3.306 (Oct 21)

  • Adds recommendations and fix fields to get_usage's behavior windows.
  • Fixes servers registered via setMcpServers getting silently replaced later by a same-named server from the user's own settings.
  • Fixes startup({initializeTimeoutMs: null}) counting SessionStart/Setup hook time against its 60-second limit.
  • v0.3.306 is a parity-only bump matching Claude Code v2.1.306.

Claude Agent SDK (Python) v0.2.170 (Oct 21)

Fixes get_session_messages/get_session_messages_from_store dropping some tool calls and results from turns that ran tools in parallel — only one branch of the forked transcript was followed before; also switches PyPI release uploads to trusted publishing via GitHub OIDC. v0.2.169 (Oct 20) clarified that include_hook_events never suppresses SessionStart/Setup hook events.

anthropics/sandbox-runtime v0.0.81 (Oct 20)

Ships Windows-only sandboxing changes: a write-restricted token for sandboxed commands, a profile read-grant mechanism, and improved network loopback support.

Claude Developer Platform release notes (Oct 20)

  • Takes inference hooks out of beta for Claude Enterprise organizations.
  • Launches user profiles: platforms/gateways/resellers can tag each API request with an end customer, so a safety enforcement action hits one customer instead of the whole org; SDK and ant CLI support the new /v1/user_profiles endpoints.
  • Adds organization webhooks to the Admin API (beta, webhooks-2026-10-07 header) covering user-profile restriction/ban and un-ban events.

Anthropic discloses $80.7B H1 2026 net loss (Oct 21)

Reports an $80.7B H1 net loss, up from $13B in H1 2025, but $72B of it is an accounting charge from rising estimated value of convertible bonds — actual operating loss is about $8.1B on $4.6B of 2025 revenue; Q3 2026 revenue hit $16.8B (+45% QoQ, a slowdown from the prior quarter); infrastructure spend nearly tripled to $9B in H1. The disclosure lands as investors reportedly weigh a lower IPO valuation ($1.5T vs. the earlier $2T target) ahead of a planned pre-Thanksgiving listing.

🔌 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

nicobailon/pi-subagents v0.80.0 (Oct 21)

Moves to Pi 1.2.0 to make background work quieter — a background child's completion notice now starts the parent's turn itself when the parent was idle, cutting noisy transcript lines — and lets individual children in a workflow run with their own working directory or model, plus a way to disable slash commands while keeping the subagent tool itself.

Unchanged in window: davila7/claude-code-templates (only an automated data-refresh bot commit on Oct 21; last substantive change was Oct 20, already covered), wshobson/agents (last commit Oct 17, already covered previously), VoltAgent/awesome-claude-code-subagents (last commit Oct 11), anthropics/commerce-agents (last commit Aug 31).

🔗 MCPs & Integrations

None of the top 5 updated in window.

Unchanged in window: oraios/serena (v2.1.0 shipped Oct 19, already covered previously), upstash/context7 (v4.4.4 shipped Oct 19, already covered previously), modelcontextprotocol/servers (v2.0.0 shipped Oct 19, already covered previously), ChromeDevTools/chrome-devtools-mcp (last release Oct 15), DeusData/codebase-memory-mcp (last release Oct 16).

💡 Community — Workflows & Ideas

Gathin23/cleanyourtv (Oct 19)

A Claude-Code-built tool that debloats Android/Google TVs over Wi-Fi with no root — it disables manufacturer ad and tracking apps and installs a clean, Apple-TV-style launcher — went viral after TechCrunch coverage on Oct 19, now spanning 158 TV models across 52 brands and 15.1k stars.

Quorus-dev/Quorus (Oct 18-21, 8 releases)

Hardens its Claude Code/Codex shared-room relay across a 4-day release sprint (v0.5.1 → v0.6.5): masks participant keys and credential tokens in room posts, fixes a broken MCP command for agents joining via CLI, and raises the message size cap from 4 KB to the room's 50 KB limit with SHA-256 receipts for long fields.

📰 Quick Mentions

← back to archiveend of edition № 177