1082026-08-12

Claude Report — 2026-08-12


🚀 Anthropic Official

Compliance API now covers local Cowork and Claude Code sessions

Adds GET /v1/compliance/apps/sessions/local (and per-session/messages endpoints) so Enterprise compliance teams can pull consolidated transcripts of Cowork and Claude Code sessions that run on users' own machines, in beta.

Claude Code v2.1.228

Fixes interactive sessions that stopped redrawing after a rare layout error, Git/Git Bash not being found on Windows when launched from a parent folder, /tui reverting to an earlier model after /model changes, Remote Control leaking conversation history into a connected session, and cross-session messaging starting without an inbox; hardens skills synced from claude.ai against shadowing local commands.

Claude Sonnet 5 pricing made permanent

The introductory $2/$10 per Mtok rate is now standard; the scheduled September 1 increase to $3/$15 will not happen.

Claude Code v2.1.227

Fixes feature flags evaluating without subscription tier on expired-token session starts, Bash commands failing under claude-code-action with allowed_non_write_users, and /tui restoring rewound conversations; improves slash-command menu performance.

🔌 Claude Code Plugins

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

🛠️ Skills

agentskillexchange/skills

Adds a new FreeFM NetEase Private FM skill (Aug 12) on top of its continuous automated sync from the Agent Skill Exchange.

Piebald-AI/claude-code-system-prompts

Mirrors Claude Code's v2.1.227 (Aug 10) and v2.1.228 (Aug 11) system-prompt and changelog content within hours of each release.

Unchanged in window: anthropics/skills (last Aug 7), JSONbored/awesome-claude (no window activity found), alirezarezvani/claude-skills (last major update Aug 3).

🤖 Agents & Subagents

nicobailon/pi-subagents

Ships v0.46.0 (Aug 11) adding prompts.render and a TypeScript project-pane API, v0.47.0 (Aug 11) adding strict model-scope enforcement, and v0.47.1 (Aug 12) fixing artifact-cleanup retention and workflow recovery.

anthropics/claude-agent-sdk-typescript

Ships v0.3.227 (Aug 10) for parity with Claude Code v2.1.227, then v0.3.228 (Aug 11) carrying usage.output_tokens_details through agent tool-result outputs.

AMAP-ML/LongHorizon-Harness

Ships v0.1.4 (Aug 11) adding a browser-driven React/FastAPI Dashboard for starting, steering, and approving long-horizon agent runs.

Unchanged in window: Chachamaru127/claude-code-harness (last May 10), VoltAgent/awesome-claude-code-subagents, wshobson/agents, 0xmmo/crew.

🔗 MCPs & Integrations

DeusData/codebase-memory-mcp

Ships MCPB bundle packaging and publishes to the MCP Registry (Aug 11), then merges installer and permission fixes — a release-asset install bug and a private-directory DACL inheritance issue (Aug 12).

Unchanged in window: awslabs/mcp (last release Aug 11, already reported), github/github-mcp-server (last v1.9.0 Aug 10, already reported), oraios/serena (last v1.7.0 Aug 9, already reported), shlokkhemani/rabbithole, modelcontextprotocol/python-sdk, modelcontextprotocol/typescript-sdk.

💡 Community — Workflows & Ideas

How to organize Claude Code for product work

Recommends splitting a Claude Code workspace into stable-context, active-project, and dated-notes folders, with a two-minute filing habit that routes every correction into preferences, context, or skills so fixes compound instead of repeating (35 points, 26 comments).

Claude finds a new lowest-ever counterexample to Borsuk's Conjecture

An independent researcher has Claude identify a set of 321 points in ℝ⁶³ that can't be split into 64 smaller-diameter subsets, pushing the smallest known dimension where the conjecture fails down from 64 to 63.

Why does CLAUDE.md keep growing? Catastrophic remembering in agentic coding

A study of 247,694 instruction lifetimes across 1,867 repos finds agentic prompt files more than triple over their lifetime because deleting an instruction without its rationale risks a regression; annotating instructions with reasoning cut excess growth by 99.3% in tests.

Watermark backlash: a removal tool appears, and a rebuttal explains why most can't work

Days after Claude's content watermark shipped, a "remove the watermark" tool surfaces alongside a technical rebuttal arguing most such tools can't actually strip a signal embedded in word-choice rather than invisible characters.

Show HN: Parley — your coding agent can talk to a teammate's agent

Builds an append-only, audit-logged coordination layer so separate coding-agent sessions can message each other asynchronously and escalate blocking issues to humans, echoing Anthropic's own cross-session messaging feature with a cross-tool version (9 points, 5 comments).

📰 Quick Mentions

← back to archiveend of edition № 108