Five Claude Code releases landed between Monday and Thursday, v2.1.139 through v2.1.143, turning Code with Claude's May 6 conference announcements into working software. Agent View (claude agents) went from announced to a full flag suite in four days. The /goal command shipped alongside it, Rewind got "Summarize up to here", and Fast mode moved from Opus 4.6 to 4.7. The gap between "announced at conference" and "in your terminal" is now measured in days.

Dreaming is the one worth sitting with. The idea: an agent reviews its own sessions between runs, pulls patterns from what worked and what didn't, and writes new memories the next session inherits. Harvey reported a 6x jump in task completion rates. Outcomes is the grading layer underneath it, a separate evaluator that scores output against a written rubric and tells the agent what to fix, with up to 10.1% improvement on task success. Both are available now: Dreaming in research preview, Outcomes and Multiagent Orchestration in public beta.

From us: typed launched this week, a Claude Code-compatible inference alternative with ~44-67% cheaper overages vs. Claude's Extra Usage rates and monthly billing. Claude Code in Production got a 1.0.3 update covering what actually shipped from Code with Claude. And on mcp.hosting: a post on the unpatched SQL injection in @modelcontextprotocol/server-postgres (21,000+ weekly downloads, read-only bypass), plus a write-up on an alternative to the official AWS MCP server.

From us

  • typed is live - Yaw's Claude Code-compatible inference launcher: independent inference providers, ~44-67% cheaper overages vs. Claude Extra Usage, monthly billing, one-curl install across Claude Code, Cursor, and Cline
  • Claude Code in Production 1.0.3 - updated to cover what actually shipped from Code with Claude 2026: Agent View, Routines, native Code Review, CI Auto-Fix, and a rewritten chapter on autonomous agents
  • The Postgres MCP everyone is still using has a known SQLi - @modelcontextprotocol/server-postgres has an unpatched SQL injection that bypasses read-only protections; still pulling 21,000+ weekly downloads
  • An alternative to the official AWS MCP server - what the official awslabs/mcp server gets wrong and what to use instead

Claude Code this week

  • v2.1.143 - May 15: plugin dependency enforcement (claude plugin disable refuses when another plugin depends on the target); projected context cost now visible in the plugin marketplace pane; worktree.bgIsolation: "none" lets background sessions edit the working copy directly; PowerShell now passes -ExecutionPolicy Bypass by default
  • v2.1.142 - May 14: claude agents gets a full flag suite (--add-dir, --settings, --model, --effort, and more); Fast mode moves to Opus 4.7 (was Opus 4.6); plugins with a root-level SKILL.md and no skills/ subdirectory now surface in the marketplace
  • v2.1.141 - May 13: terminalSequence hook field for desktop notifications and window title updates without a controlling terminal; ANTHROPIC_WORKSPACE_ID for workload identity federation; claude agents --cwd <path>; Rewind adds "Summarize up to here" to compress earlier context while keeping recent turns
  • v2.1.140 - May 12: Agent tool subagent_type now accepts case- and separator-insensitive values ("Code Reviewer" resolves to code-reviewer); /goal fixes
  • v2.1.139 - May 11: Agent View research preview - claude agents launches a dashboard listing all sessions; /goal command sets a completion condition and Claude keeps working across turns until met; transcript navigation shortcuts (? for help, {/} to jump between user prompts, v to toggle panel)

The broader week