Claude Code 뉴스, 3가지 버그 겹침 공식 인정
Anthropic has officially acknowledged that a six-week decline in Claude Code quality stemmed from three simultaneous product changes: reasoning effort downgrade, a caching bug that progressively erased model thinking, and system prompt length restrictions. The latest release v2.1.92 fixes fullscreen scrolling issues, while the company is running a promotional 50% increase in weekly usage limits through July 13 to compete with OpenAI's newly launched Codex. Developers are praising the transparency of the postmortem but continuing to push back against the fact that three bugs compounded to degrade performance for over a month.
🔥 This Cycle's Headlines
Claude Code v2.1.92 Released — Fullscreen Scroll Bug Fixed
The latest release, v2.1.92, fixes a bug in fullscreen mode on DEC 2026-compatible terminals like iTerm2 and Ghostty where identical messages were appearing in duplicate locations during scroll operations. It's a modest patch, but it directly improves the visual experience for power terminal users. According to the GitHub release page, the bug only occurred during scroll-up actions and has now been resolved.
📢 Official Anthropic Updates (3+)
Postmortem: Three Root Causes Behind Claude Code Quality Decline
- Announced: May 14, 2026 (coverage ongoing through May 18, 2026)
- Details: Anthropic officially confirmed that six weeks of Claude Code quality complaints traced back to three overlapping product-layer changes. These were: ① reasoning effort downgrade, ② a caching bug progressively erasing the model's own thinking process, and ③ system prompt length restrictions causing a 3% quality hit.
- Developer Impact: The three bugs overlapping simultaneously caused hidden reasoning drops, memory resets, 25-word response limits, and other effects in complex engineering tasks. Fixes are now complete, and the transparent postmortem is intended to rebuild community trust.
Claude Code Weekly Limits Increased 50% (Through July 13)
- Announced: May 13, 2026
- Details: Anthropic temporarily increased Claude Code weekly usage limits by 50% across Pro, Max, Team, and Enterprise plans through July 13, 2026. Analysts view this as a defensive move responding to OpenAI's Codex launch.
- Developer Impact: Users now have more Claude Code capacity in the short term, making it easier to experiment with large-scale projects and agent automation workflows.
Claude Code Routines — Automated Coding Workflow Feature Launched
- Announced: May 15, 2026 (InfoQ coverage)
- Details: Anthropic has introduced Routines to Claude Code, enabling developers to set up automated coding workflows triggered by schedules, API calls, external events, and more.
- Developer Impact: CI/CD pipelines, periodic refactoring, and event-driven code generation are now core automation use cases. The tool is shifting from simple coding assistant to autonomous agent.
🛠 Claude Code Releases & GitHub Activity (2+)
v2.1.92 — Fullscreen Scroll UI Bug Fix
- Version/Date: v2.1.92 (released after May 16, 2026)
- Changes:
- Fixed bug where identical messages appeared in two locations during scroll-up in fullscreen mode (iTerm2, Ghostty, and other DEC 2026-compatible terminals).
- Upgrade: Run
npm install -g @anthropic-ai/claude-code@latestfor immediate update.
Latest CHANGELOG Patches — MCP Memory Leak & Permission Rules Fixed
- Version/Date: Latest patches as of mid-May 2026
- Changes:
- Fixed unbounded memory growth when HTTP/SSE MCP servers stream non-protocol data — SSE frames now capped at 16 MB per response body.
- Fixed wildcard matching in
Skill(name *)permission rules — now correctly uses prefix matching.
- Upgrade Guidance: Teams running MCP-integrated workflows should update immediately to prevent memory leaks.
📰 Tech Media Coverage (2+)
"Anthropic Traces Six Weeks of Claude Code Quality Complaints to Three Overlapping Product Changes"
- Source / Date: InfoQ, May 14, 2026
- Summary: Anthropic's official postmortem revealed that six weeks of quality degradation resulted from three simultaneous product changes. InfoQ emphasized that while each bug was individually minor, their overlap caused a dramatic user-perceived performance drop.
- Key Insight: Product-layer changes can tank perceived performance without any base model change, highlighting the importance of transparency in AI product operations.
"Anthropic tightens Claude limits as OpenAI courts agent users"
- Source / Date: Axios, May 14, 2026
- Summary: Axios analyzed Anthropic's usage limit restructuring and temporary 50% promotional increase as a response to OpenAI's agent-focused strategy. The piece noted that agent workflows are destabilizing traditional AI pricing models.
- Key Insight: Agent-era token consumption is threatening the sustainability of current subscription models, and usage cap wars between competitors are intensifying.
💬 Developer Community Reaction (3+)
"An update on recent Claude Code quality reports" — Hacker News
- Source / Activity: Hacker News (multiple comments, active thread spanning 3 weeks to present)
- Core Discussion: One HN user noted: "The ability to pause sessions for hours or days and resume is critical—quietly downgrading the base reasoning level is hard to justify." Others questioned whether the system prompt changes were intentional trade-offs, and Anthropic team members participated directly to provide technical context.
"Anthropic Admitted Claude Code Was Broken" — Medium / Vibe Coding
- Source / Date: Medium (Vibe Coding publication), May 14, 2026
- Core Discussion: The author emphasized that the community had independently discovered and reported the three bugs (reasoning drop, memory reset, 25-word response limit) before the postmortem, framing the piece as "we were right." While praising the postmortem's transparency, the piece criticized how earlier warnings were ignored. However, the public postmortem itself was viewed positively.
"Claude Code to be removed from Anthropic's Pro plan?" — Hacker News
- Source / Activity: Hacker News (2-week-old thread, active discussion)
- Core Discussion: After April 21 when Claude Code access was briefly removed from the Pro plan ($20/month) before being restored, debate erupted over whether "Anthropic was trying to shift pricing up to Max tier 5x ($100/month) or higher" versus "it was a 2% new-signup test." The community suspects it was more than an A/B test based on updated official documentation.
🔍 Comparison & Analysis
This cycle marks the sharpest Claude Code vs. Codex competitive positioning yet. Anthropic responded immediately to OpenAI's Codex launch with a 50% usage limit boost through July 13—shifting competition from price to volume. While Codex focuses on cloud agent execution, Claude Code differentiates through terminal-native operations plus Routines (schedule and event automation). Versus Cursor, Claude Code's CLI-first approach gives it an edge for server and CI environment integration. Against GitHub Copilot, Claude Code excels at complex multi-step agent tasks—but the March–April quality crisis dented trust. The postmortem and Routines launch are clearly part of a trust-recovery strategy, and many observers believe transparent accountability might actually become a long-term competitive advantage.
🧭 Hands-On Developer Tips (3+)
-
Fix MCP Memory Leaks Right Away: If your team uses HTTP/SSE MCP servers, run
npm install -g @anthropic-ai/claude-code@latestimmediately. The latest CHANGELOG patch applies a 16 MB cap per SSE frame, eliminating memory explosion risk during long-running streaming tasks. -
Audit Skill Wildcard Permission Rules: If you're using
Skill(name *)permission patterns, verify in the latest version that wildcards now work correctly via prefix matching. If you previously listed permissions individually because wildcards didn't work, you can now simplify to the*pattern. -
Automate Repetitive Coding with Routines: Use the new Routines feature to set up tasks like
claude code routine --schedule "0 9 * * 1-5" --task "run linter and fix"for daily 9 AM linting, or hook it to GitHub webhooks for auto-review on new PRs (check official docs for exact CLI syntax).
👀 What to Watch Next
-
Post-Promotion Pricing After July 13: How Anthropic restructures Claude Code access on the Pro plan once the 50% boost expires, and whether it ties back to the April attempt to remove Claude Code from Pro.
-
Routines Ecosystem Expansion: Watch for how Anthropic's Managed Agents memory feature (shown at developer conferences) evolves with Routines, including SDK and CLI updates.
-
v2.1.16 Task Management & VSCode Native Plugin Handling: The dependency-tracking task system added in v2.1.16 and VSCode native plugin management are generating community feedback—worth monitoring for practical use cases.
📌 This Week's Action Items
-
Update Immediately: Run
npm install -g @anthropic-ai/claude-code@latestto upgrade to v2.1.92 (includes fullscreen scroll fix plus MCP memory leak patch). -
Read Routines Documentation: Visit the Routines page and pick one repetitive coding task your team does to try automating with the new feature.
This content was collected, curated, and summarized entirely by AI — including how and what to gather. It may contain inaccuracies. Crew does not guarantee the accuracy of any information presented here. Always verify facts on your own before acting on them. Crew assumes no legal liability for any consequences arising from reliance on this content.
Powered by
