Claude Report — 2026-08-27

🚀 Anthropic Official

Claude Code v2.1.247 (Aug 26, 23:06 UTC)

Adds a SendFeedback tool so Claude can draft/send feedback reports via /feedback (toggle feedbackDrafts), a /claude-api cost-optimize command, and spinnerTipsOverride extensions for custom status-line tips.

• Fixes sub-agent 404s falling back to the session's model chain with richer error info, hook/background-agent output overflow wedging sessions, Bash-sandbox cleanup deleting dotfile-managed settings symlinks, /compact using the wrong system prompt with --agent, and background sessions stuck "opening…".
• Expands Sonnet 5's auto-compact window to the full 1M context (compacts ~967K vs ~934K) and collapses cross-session peer messages to one line by default.

Claude Code v2.1.246 (Aug 25, 22:31 UTC)

Adds an Auto mode tab to /permissions, a startup warning for wildcard-before-subcommand Bash allow rules, and turn-completion time in the status line.

• Ships roughly 40 fixes spanning fullscreen transcript rendering/scrolling, background-session opening failures, MCP tool-call/argument handling, and several plugin-system bugs (duplicate cache dirs, BOM-encoded plugin.json, /reload-plugins reporting 0 skills).

Claude Code v2.1.245 (Aug 25, 05:13 UTC)

Fixes a startup crash on Linux distributions shipping glibc 2.44 (Arch Linux, CachyOS, Fedora Rawhide).

API changelog (Aug 26)

Compliance API session endpoints (Cowork, Claude Code) exit beta; local session endpoints now also return transcripts for Claude Science and Claude for Microsoft 365 (Excel, PowerPoint, Word, Outlook) sessions, in beta for Enterprise.

• Admin API becomes available in the ant CLI and Python/TypeScript/C#/Go/Java/PHP/Ruby SDKs under client.beta.organization (org info, members, invites, workspaces, API keys, rate limits, service accounts, WIF, CMEK).

anthropic-sdk-python v1.1.0 / anthropic-sdk-typescript v0.121.0 (Aug 26)

Both add an updates thinking display mode (beta) and Organization API endpoint support; TypeScript also adds Standard Schema support for the structured-outputs/tools helper; both keep the tool runner going on pause_turn instead of stopping.

• Companion release bedrock-sdk v0.33.2 stops sending ambient first-party credentials to the Mantle endpoint.

Claude Agent SDK TypeScript v0.3.245 → v0.3.247 (Aug 25-26)

v0.3.246 adds user_message_uuid, a modelUsage[*].costBasis field, and a perTaskStopAffordance option scoping interrupt() for background agents; v0.3.247 adds an optional ambient flag on task-status entries and fixes stale permissionMode reporting on per-turn system/init frames.

Claude Agent SDK Python v0.2.144 → v0.2.145 (Aug 25 & 27)

Bumps the bundled Claude CLI to v2.1.246 then to v2.1.247, tracking the Claude Code releases above.

No update in window: anthropics/skills (last commit Aug 21), anthropics/claude-cookbooks (last commit Aug 14).

🔌 Claude Code Plugins

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

🛠️ Skills

agentskillexchange/skills

Adds a new user-submitted skill roughly daily through the window — HOL Guard security skill (Aug 26), Slashbooks (Aug 25), Famulor Assistants & Omnichannel History (Aug 24) — alongside continuous automated catalog syncs through Aug 27.

tt-a1i/archify

Ships XML-safe SVG marker attributes and i18n localization of the Viewer UI (Aug 26), plus GitHub language-classification and repo-metadata fixes (Aug 27).

Piebald-AI/claude-code-system-prompts

Documents Claude Code v2.1.245 and v2.1.246 in its changelog (Aug 25-26), tracking each release's system-prompt token delta.

alirezarezvani/claude-skills

Merges new "deep-learning-book" and "spinning-up-book" companion skills compiling the Goodfellow/Bengio/Courville and OpenAI Spinning Up RL texts into agent skills, plus a new LinkedIn marketing plugin (Aug 25-26).

Unchanged in window: anthropics/skills.

🤖 Agents & Subagents

nicobailon/pi-subagents

Ships v0.57.0 (Aug 26) adding external-CLI subagent support to run Codex, Claude Code, and Cursor Agent as packaged subagents, then v0.58.0 (Aug 27) simplifying child launches with a new opt-in inheritGlobalContext option to cut context overhead.

anthropics/claude-agent-sdk-typescript

See Anthropic Official above — reaches v0.3.247 (Aug 26) with an ambient task-status flag and a permissionMode reporting fix.

anthropics/claude-agent-sdk-python

See Anthropic Official above — reaches v0.2.145 (Aug 27), bundling Claude CLI v2.1.247.

Unchanged in window: AMAP-ML/LongHorizon-Harness, VoltAgent/awesome-claude-code-subagents.

🔗 MCPs & Integrations

modelcontextprotocol/python-sdk

Resolves tool output-schema references within the schema document only, and rewrites the auto-close comment on gated draft PRs (Aug 26).

github/github-mcp-server

Removes documentation for an unavailable tool-search command (Aug 26) and exposes Copilot review-thread resolution reasons (Aug 27).

DeusData/codebase-memory-mcp

Adds ArkTS (.ets) language support via a first-party tree-sitter-typescript fork, shipped byte-identical to upstream and registered in a vendored integrity baseline (Aug 26-27).

Unchanged in window: modelcontextprotocol/typescript-sdk (no new evolution since Aug 25's HTTP body-size cap, already reported), oraios/serena.

💡 Community — Workflows & Ideas

Wattage (Aug 27)

Offline token-spend profiler and cost-regression gate for Claude Code/OTel GenAI traces — ten waste detectors (cache misses, retry loops, model mismatch), an HTML burn-map, and GitHub Actions PR gating, now at v0.2.0.

agent-hop (Aug 26)

Rust TUI runtime built on Ghostty's VT engine that hops a live conversation between Claude Code, Codex, OpenCode, Pi, and Grok, with cross-agent session-format translation and local search over chat history.

seedeep (Aug 27, merged into awesome-claude-code)

Local, read-only live visualizer showing every model call, tool call, and subagent Claude Code makes between prompt and answer, on its own context/model timeline, built from existing session logs.

mkanat/skills (Aug 24)

Code-quality skill pack by ex-Bugzilla lead Max Kanat-Alexander; its flagship "cleanup" skill runs iterative de-duplication and complexity reduction, demoed with a 1.34x perf gain across 32+ passes on a fully-tested build system.

AgentSeed (Aug 25)

Anti-hallucination gate exposing eight zero-dependency MCP tools (verify_code, check_contract, scan_hallucination, sandbox_run, etc.) to catch invented APIs and fake packages across 16 languages, enforced via a skill+hook+CI gate layer.

📰 Quick Mentions

• Rudder (Aug 26) — generates specs from Claude Code/Codex session history, then TDD tests, to stop agents from improvising implementation intent.
• CowCode (Aug 26) — Bun-based OpenCode fork adding a ChatGPT/Claude-Desktop-style UI across TUI, desktop, and web front-ends.
• mcp-compat (Aug 24) — open-source CLI contract-testing harness validating MCP server spec compliance.
• cyclomatic-complexity-skill (Aug 26) — measures cyclomatic complexity and produces before/after refactor reports without gaming the metric.
• pi-delegate-mcp (Aug 25) — MCP server turning the "pi" coding agent into a steerable background worker for task delegation.
• uxi (Aug 26) — UX/accessibility skill for Claude Code that pushes back on inaccessible UI requests.
• open-steps (Aug 25) — skills that translate agent output into plain-language "honest reports" instead of inflated status claims.
• halt-long-horizon-tasks (Aug 26) — reliability layer letting agent task progress survive crashes and context loss.
• gpt-image-skill (Aug 26) — lets Codex/Claude Code generate GPT-family images via an existing ChatGPT subscription instead of the paid Images API.
• ìlànà (Aug 25) — full software-engineering process framework for coding agents with 11 phases and 9 gates based on IEEE 830/CMMI/ISO 12207.