1252026-08-29

Claude Report — 2026-08-29


🚀 Anthropic Official

Claude Code v2.1.251 (Aug 28, 18:19 UTC)

Adds PreModelSwitch/PostModelSwitch hooks to block, confirm, or annotate model switches; streams foreground subagent tool calls live to Remote Control clients; adds a spend-limit bar to /usage and a per-session prompt-cache line (hit ratio + token counts) to /cost.

  • Fixes a symlink-swap security vulnerability in file tools that allowed path escape outside approved locations, plugin commands that could point outside the plugin directory, stuck "text content blocks must be non-empty" errors after thinking-only turns, Opus 5 requests failing at effort xhigh/max with thinking disabled, TUI lag with many parallel subagents, and agent-team message delivery bugs. Also trims ~5MB off the native binary.

Claude Agent SDK TypeScript v0.3.251 / Python v0.2.148 (Aug 28, ~18:20-18:34 UTC)

Both bump their bundled Claude CLI to v2.1.251 for parity with the release above; no independent SDK-level API changes reported.

Claude for Teachers, now available for schools and districts (Aug 28)

Extends the free educator offering to a centrally managed Enterprise org for whole schools and districts, under a single K-12 terms set with SSO, role-based access controls, and domain claiming; launches a companion pilot in Detroit Public Schools Community District studying educator wellbeing and practice.

Unchanged in window: API changelog (last entry Aug 27 — files/skills GA, personal/service-account API keys), anthropic-sdk-python (still v1.2.0, Aug 27), anthropic-sdk-typescript (still v0.122.0, Aug 27), Model Hardware Standard and Expanding support for scientists posts (both Aug 27, no new anthropic.com/news entries since).

🔌 Claude Code Plugins

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

🛠️ Skills

agentskillexchange/skills

Runs another automated catalog sync ("regenerate all repo files from live data") as recently as Aug 29, 07:27 UTC, continuing its near-daily cadence.

tt-a1i/archify

Merges a fix for site-language persistence across navigation (PR #159) plus star-history chart/CI cleanup (PRs #155-156), Aug 28.

Piebald-AI/claude-code-system-prompts

Publishes token-count diffs for the Claude Code v2.1.250 and v2.1.251 system prompts, Aug 28.

alirezarezvani/claude-skills

Merges PR #996 plus an automated codex-skills symlink sync and a new book-writing skill (PR #993), Aug 26.

Unchanged in window: anthropics/skills (last push Aug 21).

🤖 Agents & Subagents

nicobailon/pi-subagents

Merges PRs rendering parallel workflow groups as cards, hiding git-evidence subprocesses on Windows, and defaulting non-trivial requests to async lanes — all Aug 29.

anthropics/claude-agent-sdk-typescript

See Anthropic Official above — reaches v0.3.251 (Aug 28).

anthropics/claude-agent-sdk-python

See Anthropic Official above — reaches v0.2.148 (Aug 28).

Unchanged in window: AMAP-ML/LongHorizon-Harness (last push Aug 20), VoltAgent/awesome-claude-code-subagents (last push Aug 12).

🔗 MCPs & Integrations

DeusData/codebase-memory-mcp

Bounds search_code scan deadlines, adds index-time importance scoring via weighted-degree, and documents a runtime trace observation model, across PRs merged through Aug 29, 09:32 UTC.

modelcontextprotocol/python-sdk

Expires idle Streamable HTTP sessions by default and caps concurrent sessions (PR #3395); resolves tool output-schema refs within the schema document only (PR #3394), both Aug 28.

github/github-mcp-server

Exposes Copilot review-thread resolution reasons via a new merged PR, Aug 27.

modelcontextprotocol/typescript-sdk

Mirrors params.taskId into the Mcp-Name header for tasks requests per SEP-2663, Aug 28.

Unchanged in window: oraios/serena (last push Aug 20).

💡 Community — Workflows & Ideas

planning-with-files (Aug 28-29)

Persistent file-based planning system for coding agents — re-injects task_plan.md/findings.md/progress.md each turn so plans survive /clear, crashes, and context compaction; ships bug fixes (side-effect-free init help, npm/shell-script fixes) across Aug 28-29, installable via the Claude Code plugin marketplace.

The Load-Bearing Vocabulary of Claude (Aug 27)

Analyzes roughly 50M words across 1,000 daily GitHub PRs to isolate a Claude-generated-content cluster (about 40% of sampled PRs) and its distinctive tics ("load-bearing," "plainly," "survived"); auto-updates daily via GitHub Actions. Topped HN at 663 points.

Conduct (Aug 28)

Open-source control plane for AI agent governance: a policy engine with signed configs and hash-chained audit trails, plus an SDK-agnostic LLM proxy that can block, warn, audit, or inject instructions into LLM, shell-tool, and MCP calls, with 20+ compliance packs.

Talos (Aug 28)

"Brakes-first" agent (v0.15.1-alpha) built around a deterministic permission kernel that gates every tool call, binding approval to exact arguments with one-shot 30-second expiry, across 23 declared tools and four control channels (terminal, Telegram, email, WhatsApp).

Twing (Aug 28)

Coordination layer for parallel coding agents that gives real-time visibility into who's editing what and blocks conflicting concurrent edits on shared codebases.

📰 Quick Mentions

  • pi-black (Aug 28) — unofficial patch letting the Pi agent tool authenticate via Claude Max/Pro OAuth instead of API billing; explicitly unaffiliated with Anthropic.
  • claude-cookbooks (Aug 28) — removes unsupported cost/latency ratio claims from the CMA_plan_big_execute_small managed-agents cookbook.
← back to archiveend of edition № 125