Back to Blog
claude-code weekly-update auto-memory remote-control simplify batch anthropic

What's Up Claude Code: Week of February 24th

Olivier Legris ·

TL;DR

The theme this week: Claude Code is learning to remember and to ship for you. Auto-memory means agents build institutional knowledge across sessions. /simplify and /batch automate the two biggest time sinks — shepherding PRs to production and parallelizing code migrations. Remote Control goes wide, letting external tools drive Claude Code sessions programmatically.

Meanwhile, the stat that turned heads: 4% of all GitHub public commits are now authored by Claude Code. At current trajectory, SemiAnalysis believes it will cross 10% within months. The community is responding with infrastructure to match — from 13-agent marketing teams with peer review gates, to phone-calling skills, to self-healing production stores that fix 3am crashes without human intervention.

And the elephant in the room: Anthropic publicly refused to deploy in the Pentagon’s classified network, while OpenAI signed the deal. The “Cancel ChatGPT” movement went mainstream. Anthropic held the line. The community noticed.

Thirteen releases this week (v2.1.51 → v2.1.63), continuing the memory leak purge from last week.

1. Auto-Memory

Claude Code now automatically saves useful context to memory (v2.1.59). No more losing institutional knowledge between sessions — the agent builds a persistent memory of your codebase, preferences, and patterns.

Manage it with /memory. This is a quiet but profound change: the difference between an assistant that forgets everything between conversations and one that accumulates understanding over time.

2. /simplify and /batch — Production Automation Skills

Boris Cherny announced two new built-in Skills shipping in v2.1.63:

/simplify — Shepherds a PR to production. Reviews code, runs tests, fixes CI failures, handles merge conflicts, and lands the PR. Automates the tedious back-and-forth that eats hours after the actual coding is done.

/batch — Interactively plans code migrations, then executes them in parallel using dozens of agents. Each agent runs in its own git worktree with full isolation, tests its work, and puts up a PR.

Usage: /batch migrate src/ from Solid to React

These two together automate the bookends of development: getting code written (batch) and getting it shipped (simplify). Boris says he’s been using both daily.

3. Remote Control Goes Wide

claude remote-control (v2.1.51) expands to all users. External builds can now drive Claude Code sessions programmatically via local environment serving. Combined with the --sdk-url WebSocket flag from last week, this is the foundation for integrating Claude Code into any workflow — CI pipelines, custom IDEs, orchestration tools.

Also new: HTTP hooks (v2.1.63) can POST JSON to URLs instead of running shell commands, and hooks are now routed through the sandbox network proxy for security. The platform is becoming genuinely extensible.

4. Security Hardening

Several important security fixes this week:

  • Hook commands could execute without workspace trust acceptance (v2.1.51) — statusLine and fileSuggestion hook commands bypassed trust checks in interactive mode. Fixed.
  • HTTP hooks could interpolate arbitrary env vars (v2.1.51) — Header values could leak environment variables. Now requires an explicit allowedEnvVars list.
  • Bash permission classifier hallucinations continue to be addressed — the “always allow” system now computes smarter per-subcommand prefixes for compound bash commands (v2.1.59).

5. The Memory Leak Purge Continues

Last week we covered ~15 memory leak fixes. This week adds more:

  • Completed teammate tasks retained all messages after compaction (v2.1.63)
  • MCP server fetch caches not cleared on disconnect (v2.1.63)
  • MCP OAuth token refresh race condition with multiple instances (v2.1.59)
  • Bridge polling loop listener leak (v2.1.63)
  • MCP OAuth flow cleanup listener leak (v2.1.63)
  • Interactive permission handler listener leak during auto-approvals (v2.1.63)
  • Git root detection cache unbounded growth (v2.1.63)
  • JSON parsing cache unbounded growth (v2.1.63)
  • Bash command prefix cache leak (v2.1.63)
  • Hooks config menu navigation leak (v2.1.63)

Plus: heavy progress message payloads are now stripped during compaction in multi-agent sessions (v2.1.63), and /clear properly resets cached skills (v2.1.63).

6. Anthropic vs The Pentagon

The biggest non-technical story of the week. Sam Altman announced that OpenAI reached an agreement to deploy models in the Pentagon’s classified network. Anthropic publicly refused, responding to Secretary of War Pete Hegseth’s comments.

The “Cancel ChatGPT” movement went mainstream. For the Claude Code community, this crystallized something many already felt: the company behind the tool matters as much as the tool itself.

Boris Cherny’s reaction on X: “Proud to work at Anthropic.”

7. 4% of GitHub and Climbing

SemiAnalysis reported that 4% of GitHub public commits are now authored by Claude Code, with the trajectory suggesting 10%+ within months. For context, this makes Claude Code the single largest non-human contributor to open source software.

This pairs with Google’s research on Deep Thinking Ratio (DTR) — their finding that longer chain-of-thought correlates negatively with accuracy (-0.54). The implication: the quality of reasoning matters more than the quantity. Early termination of low-quality reasoning paths saves ~50% compute while improving accuracy.

8. Quality of Life & Bug Fixes

New features:

  • /copy command — Interactive picker for code blocks, with “always copy full response” option (v2.1.59, enhanced v2.1.63)
  • /model picker now shows human-readable labels and upgrade hints (v2.1.51)
  • /simplify and /batch bundled slash commands (v2.1.63)
  • HTTP hooks — POST JSON to URLs instead of shell commands (v2.1.63)
  • ENABLE_CLAUDEAI_MCP_SERVERS=false to opt out of claude.ai MCP servers (v2.1.63)
  • Project configs & auto memory shared across worktrees of the same repo (v2.1.63)
  • Plugin-provided commands available immediately after install (no restart) (v2.1.63)
  • Manual URL paste fallback for MCP OAuth when localhost redirect fails (v2.1.63)
  • VSCode: session rename/remove in sessions list (v2.1.63)
  • Custom npm registries and version pinning for plugins (v2.1.51)
  • Tool results >50K chars now persisted to disk (was 100K) for better context management (v2.1.51)

Notable bug fixes:

  • --worktree sometimes ignored on first launch (v2.1.53)
  • UI flicker where input disappeared after submission (v2.1.53)
  • Bulk agent kill (ctrl+f) now sends single aggregate notification (v2.1.53)
  • Graceful shutdown leaving stale sessions with Remote Control (v2.1.53)
  • Panic on Windows (“switch on corrupted value”) (v2.1.53)
  • WASM interpreter crash on Linux x64 & Windows x64 (v2.1.53)
  • VSCode remote sessions not in conversation history (v2.1.63)
  • REPL bridge message ordering race condition (v2.1.63)
  • Prompt suggestion cache regression reducing hit rates (v2.1.62)
  • Config file corruption from concurrent writes on Windows (v2.1.61)

Built with Claude Code

Community highlights from this week’s #cool-links and Reddit:

13-Agent Claude Marketing Team — The architecture post that sparked the most discussion this week. 13 Claude agents running marketing for a product, all on OpenClaw + PocketBase. The key pattern: peer review gates where agents can’t skip review, and heartbeat priority is “review others’ work first, then pick up your own.” Goal-driven task generation replaces manual creation. Running on a Claude Max sub — zero API costs.

clawr.ing — Phone Calling for AI Agents — OpenClaw skill that gives your agent an actual phone number. “Call me when X happens” + voice interaction. One-liner setup, no Twilio account, works internationally. For when chat messages are too easy to ignore.

Bosun — TDD-First Agent Orchestrator — Multi-agent pipeline orchestrator for Claude Code with test-driven execution: plan → test → implement → verify → auto-fix. MCP integration ready.

Neo — Browser-to-CLI API Capture — Chrome extension + CLI that captures every API call your browser makes, then lets you replay/query them from terminal. Auto-generates schemas, redacts auth at capture time. Bridges browser and CLI workflows.

Review Tower — Gamified Code Review — Claude Code plugin that turns PR reviews into a game. Every PR = tower, each file = floor. Challenge Mode: review blind, compare findings vs AI (2x XP for matches).

Agentic Emerald — AI Game Master — Claude Opus as DM for Pokemon Emerald. Watches gameplay via emulator, narrates in real-time, edits game state based on playstyle (teach moves, reward stats, turn Pokémon shiny). Fun use of the agent loop.

Self-Healing Production Store — An AI-operated e-commerce store where payment flows break at 3am and nobody gets paged — agents detect anomalies, diagnose, patch, test, and redeploy. Real operational architecture, not theory.

Claude Code LSP Integration Guide — Code navigation goes from 30-60s to 50ms with LSP enabled. Setup is ~2 minutes. Covers 11 languages with plugin architecture and troubleshooting. Probably the single highest-ROI optimization you can make.

Community Events

The global expansion continues — 28 upcoming events across 6 continents:

Next week (Mar 1-7):

Coming up:

Want to host one in your city? Submit an event on Luma →