Claude Report — 2026-08-26

🚀 Anthropic Official

Funding better evaluations of AI's impact on wellbeing (Aug 25)

Anthropic announces research grants aimed at building better evaluations of how AI systems affect user wellbeing.

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

Large fix/feature batch: 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. Fixes fullscreen transcript corruption on resize, erratic scrolling, and slowdowns on long base64 lines; background sessions failing to open (45s timeout, EACCES); /fork from forked/backgrounded sessions starting empty; MCP tool-call interruption and empty-schema argument serialization bugs; several plugin-system bugs (duplicate cache dirs, doubled name prefixes, /reload-plugins reporting 0 skills, BOM-encoded plugin.json); /stats heatmap day-offset; and headless session cleanup on Windows/macOS.

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

Ships a single fix for a startup crash on Linux distributions shipping glibc 2.44 (Arch Linux, CachyOS, Fedora Rawhide).

Claude Code v2.1.241 (Aug 23, 00:52 UTC)

Ships generic "bug fixes and reliability improvements" with no itemized notes.

Claude Agent SDK TypeScript v0.3.246 (Aug 25, 22:31 UTC)

Adds an optional user_message_uuid linking replies/errors to the triggering user message, a modelUsage[*].costBasis field ('list'|'managed'|'unknown'), managedSettings.modelPricing support for host-managed billing, and a perTaskStopAffordance option scoping interrupt() behavior for background agents/workflows.

Claude Agent SDK TypeScript v0.3.245 / v0.3.241 (Aug 25 / Aug 23)

Parity releases syncing with Claude Code v2.1.245 and v2.1.241 respectively; no independent feature changes.

Claude Agent SDK Python v0.2.144 (Aug 25, 22:47 UTC)

Moves off the v0.2.143 plateau that had persisted since Aug 20; bumps the bundled Claude CLI to version 2.1.246.

No update in window: API release notes (last entry Aug 20 — Python SDK v1.0.0), anthropic-sdk-python (last v1.0.0, Aug 20), anthropic-sdk-typescript (last releases Aug 19), anthropics/skills (last commit Aug 21), anthropics/claude-cookbooks (last commit Aug 14).

🔌 Claude Code Plugins

Nenhum dos top 10 teve atualização concreta na janela. The only in-window commit found (claude-security, Aug 24) was already reported in yesterday's edition with no further evolution since, so it's not re-listed.

Unchanged in window: code-review, feature-dev, pr-review-toolkit, plugin-dev, skill-creator, mcp-server-dev, frontend-design, security-guidance, github (external), claude-security.

🛠️ Skills

agentskillexchange/skills

Syncs seven new published skills into its catalog on Aug 26 (discord-bot-manager, hol-guard, langchain-mcp-server, mongodb-mcp-server, notion-workspace-agent, openai-mcp-server, postgresql-slow-query-analyzer) and regenerates all catalog/index files.

tt-a1i/archify

Fixes invalid XML in generated diagrams by emitting boolean-style SVG data attributes (data-node-label, data-detail-anchor) with explicit empty values instead of bare attributes (Aug 26).

Piebald-AI/claude-code-system-prompts

Documents Claude Code v2.1.246 in its changelog, covering the new "Artifact workshop" skill, new SDK fields, and the syncClaudeAiPlugins setting (Aug 26).

Unchanged in window: alirezarezvani/claude-skills (LinkedIn plugin merge already reported Aug 25, no new evolution), anthropics/skills.

🤖 Agents & Subagents

nicobailon/pi-subagents

Tags three point releases in-window (v0.55.0 Aug 23, v0.56.0 Aug 23, v0.57.0 Aug 26), hardening the external-CLI adapter work with a one-use workflow child permit, resumed-workflow artifact preservation, symlinked agent discovery, and a preflight check that fails fast on missing/invalid subagent working directories.

anthropics/claude-agent-sdk-typescript

See Anthropic Official above — ships v0.3.246 (Aug 25) with user_message_uuid, costBasis, and perTaskStopAffordance.

anthropics/claude-agent-sdk-python

See Anthropic Official above — moves to v0.2.144 (Aug 25), ending its two-week plateau at v0.2.143.

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

🔗 MCPs & Integrations

modelcontextprotocol/python-sdk

Makes importing the deprecated mcp.server.fastmcp raise ModuleNotFoundError pointing to the migration guide (Aug 25); also fixes recursive tool return types to get an object-rooted output schema, logs handler exceptions by kind, and accepts boolean sub-schemas in tool properties (Aug 24).

github/github-mcp-server

Fixes reviewer OAuth scope validation to defer to the normal handler on missing/malformed repo arguments (Aug 25), on top of Aug 24 fixes preserving CORS across OAuth routes and terminating unknown resource-metadata routes.

modelcontextprotocol/typescript-sdk

Adds a 4MiB default cap on HTTP request bodies (413 on overflow) and a 100-message limit on JSON-RPC batches, and reorders middleware to validate Host/Origin before body parsing (Aug 25).

Unchanged in window: modelcontextprotocol.io roadmap, oraios/serena, DeusData/codebase-memory-mcp.

💡 Community — Workflows & Ideas

claude-adversarial-review (Aug 24)

Runs parallel reviewer agents (correctness/security/conventions) on an uncommitted diff, then pits skeptic agents against each finding to try to refute it before surfacing it — a concrete pattern for cutting AI-review false positives.

fireworks-open-eli5 (Aug 24)

Builds evidence-aware interactive visual explainers for Codex and Claude Code output, the highest-traction new community repo of the window (156 stars).

Claude-AD (Aug 24)

Encodes Active Directory pentest methodology as Claude Code skills/agents/slash-commands, orchestrating netexec, impacket, certipy, bloodyAD, and BloodHound CE with OPSEC telemetry and compliance mapping.

claude-kit (YesWeHack) (Aug 25)

Ships a bug-bounty-focused Claude Code plugin with three commands (/ywh:write, /ywh:triage, /ywh:gotchas) that block the model from asserting findings without real URLs/payloads/observed behavior.

Code-With-Claude transcripts (Aug 23)

Publishes full transcripts of all 19 talks from Anthropic's 2026 "Code with Claude" SF conference, produced via a Deepgram + ElevenLabs + Gemini correction pipeline that the author reports cut domain-term error rate from ~23% to 0.1%.

📰 Quick Mentions

• mcpd (Aug 25) — turns an arbitrary sandbox/container into a full MCP server on the fly instead of hand-writing tool bindings.
• Henka (Aug 23) — multi-tenant MCP server purpose-built for structured, semantics-aware code refactorings.
• Rust Analyzer MCP Server crate (Aug 25) — wraps rust-analyzer using a std-thread bridge instead of async to avoid tokio-runtime friction.
• MulmoTerminal (Aug 25) — dashboard for running many concurrent Claude Code sessions and surfacing which ones are blocked on human input.
• Editorial diagram types for Claude Code (Aug 25) — a skill defining magazine-style diagram output patterns as an alternative to default boxes-and-arrows.
• open-pstack (Aug 24) — brings pstack-style process inspection to Claude Code and Codex sessions.
• telegram-notifier (Aug 24) — universal Telegram notifier plugin/skill working across Claude Code, Codex, Cursor, and Windsurf.
• swissdevjobs-cli (Aug 23) — zero-dependency CLI/MCP server/Claude Code plugin for searching salary-transparent tech jobs.
• big-picture (Aug 23) — Claude Code plugin that leads output with plain-language decisions, hiding implementation detail behind a toggle.
• ai-native-sdlc (Aug 23) — reusable skill implementing Anthropic's published AI-native SDLC with human approval gates at each phase.