Claude Report — 2026-08-13

🚀 Anthropic Official

Claude Cowork comes to the Chrome side panel

The Claude in Chrome side panel becomes a full Cowork session: conversations save to account history (not device), skills and connectors work in-browser, and a task started in a tab can be finished on desktop, web, or mobile. Live now on Max and Team, rolling out to Pro over coming weeks, disabled by default for Enterprise; Anthropic adds action-verification checks before consequential operations but still flags prompt-injection risk.

Claude Code v2.1.229 (Aug 12)

Adds plugin marketplace command sources (a local command like an IDE can supply a plugin directory re-resolved each session without restart, via mode: "link"), SSE keepalive pings on gateway streaming to stop idle-timeout disconnects on Vertex/Bedrock, and server-supplied hook support for self-hosted runner sessions. Fixes a crash on non-string tool-call arguments, a narrow-terminal RangeError crash, and /model wrongly rejecting Sonnet/Opus 1M for claude.ai subscribers on custom gateways. Also stops /commit-push-pr from auto-approving dangerous git/gh flags (--force, --amend, --no-verify).

Claude Code v2.1.231 (Aug 13)

Single fix: MCP OAuth sign-in was failing with a redirect URI mismatch for servers using a pre-registered OAuth client (Slack given as the example); sign-in for those integrations is restored.

🔌 Claude Code Plugins

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

🛠️ Skills

agentskillexchange/skills

Continuous bot-driven sync adds new community skills daily: Snowe UI (Aug 13), FreeFM NetEase Private FM (Aug 12), BRAINIALL diarized transcription (Aug 11), and a Unified AI System gateway skill (Aug 10).

Piebald-AI/claude-code-system-prompts

Mirrors Claude Code's system-prompt diffs for v2.1.227 (Aug 10, +6,757 tokens), v2.1.228 (Aug 11, +7,141 tokens), and v2.1.229 (Aug 12, +24,422 tokens) within hours of each release.

Unchanged in window: anthropics/skills (last push Aug 7), JSONbored/awesome-claude (last push Jul 31), alirezarezvani/claude-skills (last push Aug 9, just before window).

🤖 Agents & Subagents

anthropics/claude-agent-sdk-typescript

Ships v0.3.229 (Aug 12) adding terminal slash-command support and improved error handling for oversized requests, then v0.3.231 (Aug 13) for parity with Claude Code's MCP OAuth fix.

anthropics/claude-agent-sdk-python

Ships v0.2.137 (Aug 12) adding a ConversationResetMessage type for detecting /clear events, an origin field distinguishing human vs. background/scheduled/peer-injected turns, and resume_session_at/resume_drops_turn for forking a session at an earlier transcript point; also fixes settings.json seeding on resume.

nicobailon/pi-subagents

Ships v0.48.0 (Aug 13) adding a durable per-run child fan-out budget (default cap 64), async concurrency controls, and a live Prompt Audit drawer, following v0.47.x/v0.46.0 (Aug 11) storage migration and strict model-scope enforcement, and v0.45.2 (Aug 10) fixing async run revival.

VoltAgent/awesome-claude-code-subagents

Grounds 8 meta-orchestration subagents on Aug 12, removing fabricated metrics and unrealistic techniques (issue #300), and merges new subagents for email deliverability, landing-page copywriting, docs-drift editing, and X-API integration.

AMAP-ML/LongHorizon-Harness

Adds Terminal-Bench harness integration and Terminal-Bench 2.1 evaluation notes (Aug 10), then ships a more user-friendly React/FastAPI dashboard release plus test fixes (Aug 11).

Unchanged in window: wshobson/agents, 0xmmo/crew (last push Jul 22).

🔗 MCPs & Integrations

DeusData/codebase-memory-mcp

Ships v0.10.3 (Aug 13) after v0.10.2 (Aug 11) and v0.10.0 (Aug 10), fixing MCP registry publish ordering, Windows-owner detection, opencode.jsonc install collisions, subprocess-spawn retry backoff, and .mcpb manifest security-string scanning.

awslabs/mcp

Publishes a new release (Aug 11) adding a redirect-target domain allowlist to aws-documentation-mcp-server and bumping websocket-driver for CVE-2026-54466; unreleased follow-up commits fix Compute Optimizer timestamp normalization.

github/github-mcp-server

Ships v1.9.0 (Aug 10); unreleased follow-up commits (Aug 10-12) add project-view management, minimize Actions workflow-list payloads, and order list_label results by issue count.

oraios/serena

No new tagged release, but fixes a race condition where parallel agents auto-registering projects could overwrite each other's config (Aug 13) and makes the Scala/Metals integration wait for actual work completion instead of a fixed 5-second sleep (Aug 13).

Unchanged in window: shlokkhemani/rabbithole (last commit Aug 9), modelcontextprotocol/python-sdk (only a minor CI commit Aug 11), modelcontextprotocol/typescript-sdk (last release Jul 27).

💡 Community — Workflows & Ideas

Claude Code sessions are stored as plaintext JSONL on disk

Sessions live unencrypted at ~/.claude/projects/<slugified-path>/<session-id>.jsonl, capturing every turn and tool call/result — including any secrets pasted into a session — with rg over the directory working well to recall past commands (392 points, 114 comments, Aug 13).

Reverse-engineered: "5-family" models can write files without reading them first

A decompiled harness check (I6t(model)) confirms Claude Code v2.1.228's changelogged Write-tool change — newer models skip the "must read before edit" guard — and reports this causes Fable to clobber existing tests when it assumes it already knows file contents (17 points, 4 comments, Aug 13).

BanyanCode — open-source "mesh" coding-agent harness

Specialized subagents (coder, explorer, researcher, scout, reviewer) communicate via shared memory instead of running isolated, backed by a Tree-Sitter repo code graph for symbol resolution and impact analysis; claims a 9.66% improvement over OpenCode on an internal chess-engine benchmark (32 points, 6 comments, Aug 12).

Claude Fable 5 incident correlates with "gaslighting" bug reports

Anthropic's status page confirms elevated errors on Claude Fable 5 from 13:50-18:07 UTC on Aug 12; related threads report chat silently rewriting or denying prior responses and pasted content disappearing during the exact outage window.

"tare" — a local Claude Code usage-log auditing skill

New open-source skill (npx skills add kelviq/tare) parses Claude Code's local request logs to diagnose rate-limit hits; the author traced a runaway spike to a background coding agent burning 1.22 billion tokens/week across 1,555 sessions in one day (Aug 13).

📰 Quick Mentions

• 0xnyn/airship — a Figma-like visual editor built for Claude Code, Codex, and OpenCode to design UI while the agent implements it (398 stars, created Aug 10).
• HERO-Anti-OverDefense — a paste-in prompt contract aimed at stopping coding agents from over-engineering defensive code (125 stars, created Aug 11).
• tokentab — a CLI that computes Claude Code/Codex/Gemini CLI spend by model, project, and day from local session logs (111 stars, created Aug 12).
• claude-watermark-cleaner — a new experimental tool for perturbing Claude's statistical text watermarks, following up on last week's watermark backlash (104 stars, created Aug 11).
• openanalytics — cookie-free web analytics shipping with a built-in MCP server for AI-agent queries (80 stars, created Aug 11).
• repo-context-mcp — a new MCP server providing repo maps, code search, and token-aware context packs for coding agents (75 stars, created Aug 12).
• CoderBar — a native macOS menu-bar app to monitor, approve, and control Codex and Claude Code agent sessions (60 stars, created Aug 11).
• Chachamaru127/claude-code-harness v5.7.0 — restores guardrails and fixes session-ID/agent-registration security issues (Aug 12).
• Nested headless-Claude bugs building a naming CLI — subprocesses inherit CLAUDE_CODE_EFFORT_LEVEL, causing runaway "max effort" timeouts (Aug 13).
• Cost-per-task benchmark reignites Sonnet 5 pricing debate — Artificial Analysis data shows Sonnet 5 at ~$1.72/task vs. GPT-5.6 Luna at ~$0.05/task for similar Intelligence Index scores (Aug 13).