Back to Blog
claude-code weekly-update voice-mode loop auto-mode opus-4.6 community

What's Up Claude Code: Week of March 3rd

Olivier Legris ·

TL;DR

The theme this week: Claude Code is becoming ambient. Voice mode means you don’t need a keyboard. /loop means it runs without you asking. Auto Mode (March 12) means it stops asking you for permission. Stack these together and the direction is unmistakable: Claude Code is transitioning from a tool you use to a system that operates.

The infrastructure is straining under the weight. Anthropic publicly acknowledged that Claude and Claude Code traffic “grew faster than expected” — enough to stress their scaling systems mid-week. Forbes ran a piece about Claude “struggling to cope with the ChatGPT exodus.” Whether you see this as growing pains or validation depends on your patience for rate limits.

Five releases this week (v2.1.66 → v2.1.71), headlined by voice, scheduling, and a massive round of cross-platform fixes.

1. Voice Mode

Push-to-talk voice input is rolling out in Claude Code, live for ~5% of users now and ramping through the coming weeks.

Hold space, talk, release. The transcript streams in at your cursor position. You can type half a prompt, voice the messy middle, then finish typing. It’s composable — voice and keyboard are interleaved, not separate modes.

New in v2.1.69: STT support expanded from 10 to 20 languages — adding Russian, Polish, Turkish, Dutch, Ukrainian, Greek, Czech, Danish, Swedish, and Norwegian. The keybinding is now customizable via voice:pushToTalk in keybindings.json.

v2.1.71 fixed two startup issues: a 5–8 second freeze caused by CoreAudio initialization blocking the main thread after system wake, and a freeze when multiple claude.ai proxy connectors refreshed expired OAuth tokens simultaneously.

SuperWhisper users are already mourning their subscriptions.

2. /loop — Cron Inside Your Session

v2.1.71 adds /loop — run a prompt or slash command on a recurring interval right inside a Claude Code session.

/loop 5m check the deploy
/loop 1h run the test suite and report failures

This is the lightweight counterpart to the full scheduled tasks system. Where scheduled tasks are fire-and-forget background jobs, /loop keeps the task inside your current session with full context. Think of it as watch for AI tasks — continuous monitoring without leaving your conversation.

3. Auto Mode — Previewed for March 12

Anthropic announced a research preview of Auto Mode, expected to roll out no earlier than March 12.

The idea: Claude automatically handles permission prompts during coding. No more approving every file edit, shell command, or network request. You set boundaries, Claude operates within them.

This replaces the controversial --dangerously-skip-permissions flag with something more nuanced — scoped autonomy rather than blanket permission. The community response was immediate: 700+ upvotes and the most-discussed announcement of the week.

Pair this with /loop and scheduled tasks, and you get persistent autonomous operation: Claude runs on a schedule, handles its own permissions, and reports results. The “coding assistant” framing is starting to feel inadequate.

4. Opus 4.6 Defaults to Medium Effort

v2.1.68 changes the default effort level for Opus 4.6 to medium for Max and Team subscribers. Medium effort hits the sweet spot for most tasks — fast enough to not feel sluggish, thorough enough to not miss things.

The ultrathink keyword is back (v2.1.68) — type it to bump effort to high for the next turn when you need deep reasoning.

Also in this release: Opus 4 and 4.1 are removed from Claude Code on the first-party API. Users with legacy model strings pinned are automatically migrated to Opus 4.6.

5. Traffic Surge

Anthropic acknowledged that Claude and Claude Code usage spiked so hard this week it was “genuinely hard to forecast.” They’re actively scaling infrastructure.

Forbes covered the story under the headline “Claude Struggles to Cope with ChatGPT Exodus” — framing it as users migrating from OpenAI following the Pentagon deal backlash from last week.

The 2,400-upvote Reddit thread became a mix of frustration (rate limits during the spike) and flex (Claude is growing because it’s better). Whether the “ChatGPT exodus” narrative holds up or it’s just organic growth hitting a tipping point, the demand curve is clearly accelerating.

6. New Skills & Platform Features

/claude-api skill (v2.1.69) — Built-in skill for building applications with the Claude API and Anthropic SDK. Teach Claude Code how to use its own API.

/reload-plugins (v2.1.69) — Activate pending plugin changes without restarting. No more quitting and relaunching to pick up a new plugin.

/remote-control naming (v2.1.69) — Custom session titles (/remote-control My Project or --name "My Project") visible in claude.ai/code.

${CLAUDE_SKILL_DIR} (v2.1.69) — Skills can now reference their own directory, making them portable across machines.

InstructionsLoaded hook (v2.1.69) — Fires when CLAUDE.md or rules files load. Useful for plugins that need to react to project configuration.

includeGitInstructions setting (v2.1.69) — Strip built-in commit/PR workflow instructions from Claude’s system prompt. For teams with their own git conventions.

sandbox.enableWeakerNetworkIsolation (v2.1.69) — macOS-only escape hatch for Go programs (gh, gcloud, terraform) that need to verify TLS certs through a custom MITM proxy.

7. Bug Fixes & Stability

The theme this week: cross-platform polish and long-session stability.

Startup fixes:

  • 5–8s freeze with voice mode after system wake (v2.1.71)
  • UI freeze when multiple OAuth tokens refresh simultaneously (v2.1.71)
  • Chrome extension auto-detection permanently stuck on “not installed” (v2.1.71)
  • Extra VS Code windows opening at startup on Windows (v2.1.70)

Input/output:

  • stdin freeze in long-running sessions where keystrokes stop processing (v2.1.71)
  • Enter inserting newline instead of submitting over slow SSH (v2.1.70)
  • Clipboard corrupting CJK/emoji on Windows/WSL — now uses PowerShell Set-Clipboard (v2.1.70)

Agent & session stability:

  • Forked conversations (/fork) sharing the same plan file, causing overwrites (v2.1.71)
  • Background agent completion notifications missing output file path (v2.1.71)
  • Duplicate output in Bash tool error messages (v2.1.71)
  • Plugin installations lost when running multiple instances (v2.1.71)

API & proxy:

  • 400 errors with ANTHROPIC_BASE_URL and third-party gateways (v2.1.70)
  • “Model does not support effort parameter” with custom Bedrock inference profiles (v2.1.70)
  • Empty model responses after ToolSearch (v2.1.70)
  • Prompt-cache bust when MCP server with instructions connects after first turn (v2.1.70)
  • claude.ai connectors failing to reconnect after OAuth refresh (v2.1.71)

Built with Claude Code

Community highlights from this week:

SwiftUI Agent Skill by Paul Hudson (@twostraws) — 1,300+ stars in 4 days. An agent skill that teaches Claude Code (and Codex, and other AI tools) how to write proper SwiftUI. From the creator of Hacking with Swift. When someone with that much iOS authority packages their knowledge as an agent skill, it signals where the ecosystem is heading.

Claude-to-IM Skill — 630+ stars. Bridge Claude Code to messaging platforms — chat with your coding agent from Telegram, Discord, or any IM. A skill that turns Claude Code into a chatbot you can message from your phone.

Distill — Compress large CLI outputs into small, relevant answers for LLMs. Save tokens by distilling verbose command output before feeding it to Claude. Simple idea, 170 stars already.

Nemesis Auditor — “The Inescapable Auditor.” Iterative deep-logic security audit agent for Claude Code. Three passes: surface scan, logic analysis, adversarial testing. 157 stars.

Model Matchmaker — Local hook for Cursor and Claude Code that routes prompts to the right model tier automatically. Simple tasks get Haiku, complex ones get Opus. Saves money without thinking about it. 128 stars.

Bug Hunt — Adversarial bug hunting with 3 isolated agents: Hunter, Skeptic, and Verifier. Each challenges the others’ findings. 104 stars.

Claude Code Chief of Staff — Instruments Claude Code as a virtual personal chief of staff. Task management, scheduling, delegation. 101 stars.

Explore Claude Code — Interactive IDE-style documentation for learning Claude Code. The 1,000-upvote Reddit post describes it as “learning Claude Code by exploring it as it was designed.” 92 stars.

Community Events

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

This week (Mar 8-14):

Coming up:

Want to host one in your city? Join the Claude Community on Luma →