Claude Report — 2026-09-12

🚀 Anthropic Official

Claude Code v2.1.269 (Sep 11)

Adds claude plugin eval — a CLI for running a plugin's eval suite against Claude Code and producing scored JSON/HTML reports, with claude plugin eval init to scaffold a starter suite — plus /output-style [name] for listing/switching output styles. Also ships Bash-tool file-change diffs, OpenTelemetry metrics with repo attributes, workflow concurrency limits, prompt-cache stability fixes for resumed sessions, and terminal fixes for VSCode, kitty, st, rxvt-unicode, and WezTerm.

Claude Agent SDK TypeScript v0.3.269 (Sep 11)

Stamps user_message_uuid/resume_reason on a turn's first complete assistant message rather than only stream events, fixes result.permission_denials omitting Read/Edit/Write calls blocked by path-scoped deny rules, fixes delayed interrupts during slow MCP OAuth sign-in, fixes missing tool_use_id on auto-resumed background subagent notifications, and makes plan mode always route writes through canUseTool even with allowDangerouslySkipPermissions set.

anthropic-sdk-ruby v1.70.0 (Sep 10)

Ships the same Managed Agents "auto" permission mode and content_too_large web_fetch error code as the Sep 10 Python/TypeScript SDK wave, extending that release to the Ruby SDK.

Smart reports (beta) for Claude Enterprise (Sep 10)

Introduces a beta feature that analyzes team-wide Claude usage — work completed, cost, friction points, and recurring patterns worth turning into shared skills — for Owners/Admins on Enterprise plans; capped at 10 reports per org per month during beta and unavailable with CMEK/HIPAA/Access Transparency configurations.

anthropics/skills — claude-api skill update (Sep 10)

Anthropic engineering updates the official claude-api skill in the anthropics/skills repo to cover the Managed Agents "auto" permission policy and the new ant beta:sessions connect CLI command, keeping the skill in sync with the Sep 10 API release.

claude-agent-sdk-python remains stalled at v0.2.152 (Sep 2); a Sep 11 commit only bumps the bundled CLI version reference to 2.1.269, with no new package release in ten days.

🔌 Claude Code Plugins

Plugin tracking skipped this run — official marketplace install-count cache unavailable.

🛠️ Skills

tt-a1i/archify

Fixes a non-reproducible Windows zip build in its packaging pipeline (Sep 12, #385) and merges a viewer-modularization follow-up and doc tweaks (Sep 11, #386, #393).

K-Dense-AI/scientific-agent-skills

Merges the heaviest activity of any tracked skill repo this window: install-command fixes, literature/database API companion skills, a genomic-intelligence v1.1 reconciliation (Sep 11), and companion identifier APIs for ontology-term-resolution (Sep 12, #262).

anthropics/skills

See Anthropic Official above — updates the official claude-api skill for Managed Agents auto permissions (Sep 10).

affaan-m/ECC

Ships dozens of commits across the window on its Claude Code/Codex/Cursor "agent harness" system, including a Sep 12 memory fix distinguishing incomplete reads from missing records.

Unchanged in window: ayghri/i-have-adhd (no commits since Sep 10, 15:48 UTC), agentskillexchange/skills (only automated sync-bot commits since its Sep 11 ledger-tasks-yylo addition), alirezarezvani/claude-skills (no push since Aug 30).

🤖 Agents & Subagents

anthropics/claude-agent-sdk-typescript

See Anthropic Official above — v0.3.269 (Sep 11) fixes permission-denial reporting and plan-mode write routing.

nicobailon/pi-subagents

Adds agent output-schema defaults (#2188), exposes retained completion output paths (#2187), and removes consumed workflow result indexes (#2186), all merged Sep 11-12 without a new version tag yet past v0.67.0.

anthropics/claude-agent-sdk-python

Bumps its bundled CLI version reference to 2.1.269 (Sep 11) but ships no new package release — the tagged version stays at v0.2.152, now ten days stale.

Unchanged in window: anthropics/commerce-agents (no commits since Aug 31), VoltAgent/awesome-claude-code-subagents (no commits since Sep 7), AMAP-ML/LongHorizon-Harness (no push since Aug 20).

🔗 MCPs & Integrations

upstash/context7

Adds native OpenTelemetry observability/tracing instrumentation to the MCP server (Sep 11, #3041), followed by a same-day version-package release.

DeusData/codebase-memory-mcp

Fixes tree-sitter parse budgeting to use thread CPU time instead of wall clock (#2180) and patches a daemon race during mid-HELLO cold storms, part of a dense Sep 11-12 Windows/daemon reliability push.

ChromeDevTools/chrome-devtools-mcp

Merges a burst of PRs through Sep 11 including CSS-formatter features, new telemetry, a trace-engine memory-leak fix, and a console-history-preservation fix — a new addition to this tracking list given its scale and velocity.

Unchanged in window: github/github-mcp-server (no commits since v1.12.1, Sep 8), oraios/serena (no new commits on main since Sep 8), modelcontextprotocol/typescript-sdk (no commits after its already-reported Sep 11 fix), modelcontextprotocol/servers (no commits since Sep 3), punkpeye/awesome-mcp-servers (no changes since Sep 8).

💡 Community — Workflows & Ideas

Stroq — block commands injected through MCP tool output (Sep 11)

Ships a local firewall pairing a PostToolUse "Scan & Taint" hook (normalizes and scans MCP/tool output for ~599 injection patterns) with a PreToolUse "Classify & Deny" hook that blocks or flags the next action based on taint and provenance; cites the real-world "fake Sentry errors over MCP" attack as motivation.

agent-guardrails-kit — deny-lists don't hold, allow-lists do (Sep 11)

Documents eight ways to bypass a Claude Code deny-list via shell interpreter layers (perl -e, python -c, glob/variable expansion) and ships an inverted allow-list approach for protected paths with quote-aware re-lexing and self-protection of the guard config.

Sigma Operator Stack (SOS) — project state between agent sessions (Sep 11)

Keeps append-only local records binding accepted project state and verification checks with source fingerprinting, marking a check "stale" the moment its underlying source changes, instead of having agents re-infer context from chat history.

Oh My Subagents (OMS) — durable multi-agent delegation (Sep 11)

Moves parent/child subagent "Waves" into durable SQLite/Postgres state instead of chat-transcript reconstruction, so a controller can resume a wave after interruption without the parent polling.

KyttoMCP — MCP server manager (beta) (Sep 11)

Local control panel reading existing MCP configs across Claude Desktop, Claude Code, Cursor, VS Code, and Codex into one view; validates servers via on-demand handshakes, estimates per-server token/context cost, and supports named MCP-stack profiles.

📰 Quick Mentions

• claude plugin eval docs — Anthropic publishes the eval-authoring reference for the new Claude Code v2.1.269 CLI feature.
• pi-subagents #2187 — Exposes retained completion output paths for subagent workflows.
• scientific-agent-skills #263/#264 — Adds literature and database API companion skills.