The week's story is competitive pressure showing up in two places at once.

OpenAI shipped GPT-5.5 on Wednesday with the same agentic-coding pitch as Claude. The API followed a day later: $5 per million input tokens, $30 per million output tokens, 1M-token context window. Anthropic spent the same week reversing its own Claude Code regression stack - a caching bug that dropped thinking history, a verbosity prompt change that hurt coding quality, a quietly-lowered default reasoning effort - all rolled back, xhigh restored as the coding default, usage limits credited for affected subscribers. Then on Friday night they shipped 2.1.120, which crashed claude --resume and claude --continue with a JS runtime error and reintroduced a macOS sandbox failure on resumed sessions. The npm latest tag stayed pinned to 2.1.119, no GitHub Release was cut, and ~25 bug reports landed within 48 hours. The model race is one story; the harness around the model is a different and increasingly louder one.

From us: we shipped Yaw Mode for Claude Code - an overlay that bundles rules and agents while preserving conversation history across sessions. Plus a public methodology for grading any MCP server against the spec in 30 seconds (88 criteria, A–F), a Tailscale audit prompt that resolves five common scenarios in one agent run, and a piece on the local-vs-remote MCP deployment tradeoff.

From the Yaw blog

Claude Code this week

  • The Apr 23 post-mortem - Anthropic publicly identifies the three changes behind the month-long quality dip (Mar 4 reasoning-effort downgrade, Mar 26 caching bug, Apr 16 verbosity prompt), confirms all reverted/fixed, resets subscriber usage limits
  • 2.1.120 (Apr 24) - --resume/--continue crash with UKH/FKH/g9H is not a function; macOS sandbox regression on resumed sessions. Anthropic posted a status-page advisory Apr 25 01:45 UTC: auto-update rolls clients back to 2.1.119 on next check; npm latest tag held at 2.1.119; no GitHub Release was cut for 2.1.120
  • 2.1.119 (Apr 23) - persistent /config writes to ~/.claude/settings.json; PowerShell commands auto-approvable; MCP servers connect in parallel
  • 2.1.118 (Apr 23) - /cost and /stats merged into /usage; vim visual/visual-line modes; hooks can invoke MCP tools directly
  • 2.1.117 (Apr 22) - /model selection persists across restarts; CLAUDE_CODE_FORK_SUBAGENT=1 enables forked subagents
  • 2.1.116 (Apr 20) - /resume 67% faster on large sessions; inline progress in the thinking spinner

The broader week