0952026-07-17

Claude Report — 2026-07-17


🚀 Anthropic Official

Claude Code v2.1.212

Ships Jul 17 (00:26 UTC):

  • /fork now copies conversations into background sessions instead of launching in-session subagents; /subtask takes over the old /fork behavior.
  • Adds claude auto-mode reset (with confirmation prompt), a session-wide WebSearch call cap (default 200, CLAUDE_CODE_MAX_WEB_SEARCHES_PER_SESSION), and a subagent-spawn cap (default 200, CLAUDE_CODE_MAX_SUBAGENTS_PER_SESSION).
  • MCP tool calls running over 2 minutes now auto-background themselves; /resume opens a picker covering past and deleted sessions.
  • Fixes plan-mode auto-running, worktree creation, hook infrastructure, SIGTERM handling, Windows Group Policy issues, /ultrareview branch fetching, and prompt-caching behind LLM gateways/custom base URLs.

Claude Code v2.1.210

Ships Jul 14 (23:45 UTC) with a live elapsed-time counter on collapsed tool summaries, a startup warning for Write()/NotebookEdit()/Glob() permission rules, a fix for worktree-isolated subagents running git commands against the main repo, a fix for the ultracode keyword firing on non-human input, and an improved auto-mode permission classifier (now defaults to Sonnet 5 for external sessions).

Claude Agent SDK Python v0.2.121 and TypeScript v0.3.212

Both ship Jul 17 and fix an argv flag-injection bug: --resume/--session-id (Python) and resumeSessionAt/sessionId (TypeScript) now pass as single =-joined tokens so dash-prefixed values can't be parsed as separate CLI flags. The Python SDK also hardens its build scripts against command injection via CLAUDE_CLI_VERSION and bundles Claude CLI 2.1.212; the TypeScript SDK's Agent tool output now reports the resolved model after a mid-turn subagent model swap. Earlier same-window builds (Python v0.2.119, TypeScript v0.3.209/v0.3.210, Jul 14) tracked CLI parity and added timedOutAfterMs to BashToolOutput.

claude-code-action v1.0.176 and v1.0.174

v1.0.174 (Jul 14) redacts GitHub user-to-server (ghu_) tokens in the log sanitizer, a security fix; v1.0.176 (Jul 17) fixes branch-name sanitization, strips alt text from reference-style markdown images, and corrects claude_args model mapping to SDK options.

Claude API — Admin API & docs correction

Jul 14: Admin API adds beta endpoints for Claude Enterprise organization member, group, and custom-role management (anthropic-beta: ce-user-management-2026-07-13). Jul 15: docs correct mid-conversation system-message availability, confirming support on Fable 5, Mythos 5, and Opus 4.8 via the API, Bedrock, and Vertex AI with no beta header required.

Claude for Teachers

Launches Jul 14, giving verified US K-12 educators free access to premium Claude tools and standards-aligned teaching skills, with a pilot evaluation in the Detroit Public Schools Community District.

Anthropic commits $10M to Canadian AI research

Announced Jul 14 on Anthropic's newsroom, pledging funding toward Canadian AI research initiatives.

🔌 Claude Code Plugins

Plugin install-rank cache is unavailable this cycle — top-10 plugin tracking is skipped per this run's instructions.

🛠️ Skills

anthropics/skills

Updates the docx, pptx, and xlsx skills (#1447), Jul 17 — Anthropic's own official skills repo.

agentskillexchange/skills

Continues near-hourly automated "sync published skills" and "regenerate all repo files" commits through Jul 16-17.

JSONbored/awesome-claude

Merges a batch of browse/analytics feature PRs (#5154-#5177) and adds a new content entry, LitServe (#5154), spanning Jul 16-17.

Piebald-AI/claude-code-system-prompts

Publishes system-prompt diffs for Claude Code v2.1.211 (+3,890 tokens) and v2.1.212 (+1,066 tokens), Jul 16, continuing release-by-release prompt-growth tracking.

Unchanged in window: glebis/claude-skills, alirezarezvani/claude-skills, kpab/claude-fable-5-skills, ComposioHQ/awesome-claude-skills.

🤖 Agents & Subagents

nicobailon/pi-subagents

Ships a ~10-commit burst across Jul 16-17: adds interactive subagent administration (#498), a typed extension delegation API (#485), per-agent acceptance roles (#481), integrates background work with subagent wait, and fixes YAML folded frontmatter scalars (#488), completion-notification disposal (#489), and turn-budget abort timing (#505).

anthropics/claude-agent-sdk-typescript

Ships v0.3.212 (Jul 17), covered under Anthropic Official above, fixing dash-leading session/resume argv handling and exposing the resolved model after mid-turn subagent model swaps.

Unchanged in window: wshobson/agents (last commit Jul 14, already reported), 0xmmo/crew (last commit Jul 4).

🔗 MCPs & Integrations

DeusData/codebase-memory-mcp

Merges fixes for WAL journal-size starvation (#1137), a CodeQL ObjectScript type-width issue (#1136), a C++ macro-call parse gap (#1135), and Python alias-import handling (#1049), spanning Jul 15-16.

awslabs/mcp

Fixes an SSRF/LFI risk in openapi-mcp-server's external $ref spec parsing (#4141), groups Dependabot updates to cut PR noise (#4267), upgrades the bundled AWS CLI to v1.45.50 in aws-api-mcp-server (#4272), and fixes Redshift cluster discovery under restricted IAM (#3518), spanning Jul 16-17.

Unchanged in window: modelcontextprotocol/python-sdk (still v2.0.0b2, Jul 14), shlokkhemani/rabbithole (last commit Jul 14, already reported), modelcontextprotocol/typescript-sdk (still beta.4, Jul 13), ChromeDevTools/chrome-devtools-mcp (still v1.6.0, Jul 14), iiTzSenn/Sens (still v0.8.0, Jul 9), modelcontextprotocol/servers (last commit Jul 6).

💡 Community — Workflows & Ideas

Mocca

Ships Jul 16: a macOS desktop marketplace/workspace that installs Claude Agent SDK-powered agents as standalone apps, each with its own filesystem, chat threads, and tools; requires a personal API key, not Pro/Max.

cc-context-telemetry

Show HN (Jul 17): surfaces Claude Code's context-window usage and Pro/Max 5-hour/7-day rate-limit consumption directly in the editor status bar, writing telemetry to files hooks can consume.

pi-digby

Show HN (Jul 17): open-sources a self-hosted Slack agent powered by Claude via Amazon Bedrock, with conversation memory, shell/file access, and Linear/Exa integrations.

webify-mcp

Show HN (Jul 16): ships an MCP server for cost-effective web research, claiming "91% of Deep Research quality" at roughly 5% of the cost via web_find() ($0.003/query) and web_lookup() ($0.0005/query).

HeimWall

Show HN (Jul 16): launches a free local macOS menu-bar app that flags secrets and PII in prompts before they're pasted into Claude Code, Cursor, or Copilot, with no data leaving the device.

📰 Quick Mentions

← back to archiveend of edition № 095