Claude Report — 2026-07-09
- Claude Code ships v2.1.205 (Jul 8): transcript-tamper guard, full `/doctor` checkup mode, NTFS worktree-deletion fix.
- Agent SDKs mirror with Python v0.2.114 and TypeScript v0.3.205, the latter adding typed interrupt receipts.
- Anthropic publishes "The Making of Claude Code" feature and an Alberta government cybersecurity case study (Jul 6).
- MCP projects iiTzSenn/Sens and shlokkhemani/rabbithole each ship four-plus releases/phases in 48 hours.
- <a href="https://github.com/Nanako0129/pilotfish">pilotfish</a> (212 stars, Jul 8) debuts a Fable-5-planner-plus-cheap-workers delegation pattern for Claude Code.
🚀 Anthropic Official
Claude Code v2.1.205
Ships Jul 8 with a broad fix/feature batch: blocks auto mode from tampering with session transcript files, makes /doctor a full setup checkup (aliased /checkup), fixes Windows worktree removal deleting files outside the worktree via NTFS junctions, fixes --json-schema silently producing unstructured output on invalid schemas, makes auto mode ask before running rm -rf on an unresolved variable, cuts auto-updater peak memory by roughly 400MB via disk-streamed downloads, and reserves the "Claude Browser" MCP server name ahead of a Claude Desktop pane rename.
Claude Agent SDK Python v0.2.114 and TypeScript v0.3.205
Both ship Jul 8, tracking Claude Code v2.1.205. Python is a pure parity bump; TypeScript adds a still_queued field to interrupt-control responses, a typed receipt from Query.interrupt(), an interrupt_receipt_v1 capability flag, and structured name/body fields on peer-message session events.
The Making of Claude Code
Publishes Jul 6 as a feature piece tracing Claude Code's evolution from an internal CLI tool to Anthropic's flagship coding agent.
Government of Alberta uses Claude to find and fix cybersecurity vulnerabilities
Details Jul 6 how Alberta's government deploys Claude across its systems to identify and remediate security vulnerabilities.
🔌 Claude Code Plugins
Top-10 install-ranked plugin cache is unavailable again this cycle; items below come from general search and continuity with prior reports, not verified install rankings.
claude-helm
Ships four more releases (v2.13.0 Jul 8 through v2.16.0 Jul 9), adding test-scanning features and refactoring branch-creation logic.
toolshed (doc-lifecycle plugin)
Merges PRs #38-41 on Jul 8 (0.9.0 → 0.9.1): pins the marketplace via local checkout instead of a URL ref, adds version-agnostic pins with workflow-file fallback, ships a deterministic no-model-call version-bump lane, and grants actions:read to fix a CI 403.
compact-plus
Ships v1.0.4 (Jul 8), fixing the session-ID acquisition path so the /compact-plus skill works on marketplace-only installs, plus documentation corrections.
Unchanged in window: claude-mem, claude-hot-limit (repository unverifiable this cycle), claude-code-hooks-daemon.
🛠️ Skills
alirezarezvani/claude-skills
Opens two draft PRs Jul 8: #907 vendoring Microsoft's SkillOpt-Sleep into an "engineering" skill, and #906 fixing convert.sh tool-sync scanning and shadowing built-ins in si plugin status/review.
Piebald-AI/claude-code-system-prompts
Logs the v2.1.205 system-prompt diff (Jul 8, +23,674 tokens), the largest single-release jump tracked so far.
agentskillexchange/skills
New multi-contributor skill registry surfaces this cycle: ships three Claude Code agent skills (self-push, five-element-spec, multi-lens-review, PR #13) and RouterBase agent skills (PR #12) on Jul 9, plus SEO and x402-agent-tools skills on Jul 7.
JSONbored/awesome-claude
Refactors its JSON-LD builder and adds new agent/MCP entries (Marvin, Griptape, Tigris MCP Server) on Jul 9.
Unchanged in window: glebis/claude-skills, anthropics/skills, ComposioHQ/awesome-claude-skills, kpab/claude-fable-5-skills.
🤖 Agents & Subagents
anthropics/claude-agent-sdk-typescript
Ships v0.3.205 (Jul 8), the same release covered under Anthropic Official above, adding typed interrupt receipts and structured peer-message fields.
Unchanged in window: nicobailon/pi-subagents (still v0.34.0, Jul 7), 0xmmo/crew (no verifiable new activity).
🔗 MCPs & Integrations
iiTzSenn/Sens
Jumps from v0.2.0 to four new releases in two days — v0.4.0 and v0.5.0 (Jul 8), v0.6.0 and v0.7.0 (Jul 9) — shipping modular working-rule init, a shareable dashboard with QR codes, and a redesigned terminal UI.
shlokkhemani/rabbithole
Runs six build phases across Jul 8-9: converts its frontend to real ES modules, extracts a pure core module, stands up a browser-only web host with a BYOK IndexedDB store, adds PDF/URL/arXiv web ingestion, and ships a Pages deploy separating the marketing site from the app.
DeusData/codebase-memory-mcp
Extends its security-hardening batch into provenance work on Jul 8-9: merges SLSA build attestations (PR #970) and a "coverage-not-indexed" feature reporting deliberately-ignored files as a by-design coverage class (PR #974).
ChromeDevTools/chrome-devtools-mcp
Merges unreleased Jul 8-9 commits: a per-page collector refactor, heap-snapshot aggregate filtering, a page-selection fallback fix, and an --allow-unrestricted-paths config flag.
awslabs/mcp
Merges three Jul 8 fixes without a packaged release yet: a case-insensitive IAM managed-policy denylist, Aurora DSQL read-only session-state hardening, and a CloudWatch PromQL docs cleanup.
Unchanged in window: modelcontextprotocol/typescript-sdk (last release Jul 2), modelcontextprotocol/python-sdk (last release Jun 30).
💡 Community — Workflows & Ideas
pilotfish
New multi-tier delegation layer (created Jul 8, 212 stars) has a frontier model (Fable 5) plan and verify via role-based policy files while cheaper models execute subagent work, catching errors with fresh-context verification instead of trusting one big model end to end.
ditto
New tool (created Jul 8, 49 stars) mines local Claude Code/Codex/Cursor logs, redacts secrets, splits the corpus across multiple agents that independently extract behavioral patterns, and ranks traits by cross-agent consensus into a you.md profile read before every future task.
skill-extractor
New tool (created Jul 7, 8 stars) extracts reusable skills from agent transcripts with "outcome-weighted scoring," discounting techniques from failed sessions while still counting guardrails learned from failure.
cc-vision-hook
New hook toolkit (created Jul 8, 7 stars) routes pasted images to an external vision model via the Vercel AI SDK and injects text descriptions as additionalContext for vision-blind Claude models, with content-hash caching.
premium-router
New Claude Code skill pattern (created Jul 8, 8 stars) has Fable 5 handle judgment/classification design while Haiku 4.5 workers execute at volume, claiming roughly 10x lower cost on large corpora.
📰 Quick Mentions
- claude-antigravity-agents — Claude Code skill (98 stars, Jul 7) delegating work to Google Antigravity CLI sub-agents for parallel cross-model execution.
- Konnect — native KiCad 10 plugin (v0.1.1, Jul 6) exposing 171 schematic/layout/manufacturing tools via MCP for AI-assisted PCB design.
- obsidian-multivault-mcp — new multi-vault, multi-user MCP server (Jul 7, 29 stars) letting Claude read/write Obsidian vaults on a VPS over SSH or OAuth HTTP.
- rms-memory-mcp — new local-first, cross-IDE memory MCP server (v1.0.0 Jul 7, v1.0.1 Jul 8) using LanceDB semantic search with zero cloud dependency.
- dittu-mcp — new MCP server (Jul 9) that turns a phone's camera and Bluetooth into a remote mouse/keyboard for driving a desktop with zero target-machine install.
- parable — new unofficial "pseudo-Fable-mode" skill set (Jul 8, 14 stars) imitating Fable-style planning/verification discipline for Claude Opus.
- openwiki-for-claude-code — new tool (Jul 7, 9 stars) generating and maintaining OpenWiki-format repo docs from inside Claude Code.
- quickai — patches to v0.3.0 (Jul 8) and v0.4.0 (Jul 9), continuing the transcript-cost-profiler covered yesterday.
- ccmux and LianYaoHu — incremental patches (v1.0.2 Jul 8, and v0.1.3 Jul 9 respectively) to yesterday's tmux-pane tracker and sandboxed agent runner.