Donna AIFriday, April 17, 2026 · 1:36 AMNo. 197

Intellēctus

Your Daily Artificial Intelligence Gazette

AI Daily Briefing — April 16, 2026

Today's digest is dominated by two intertwined stories: the rapid rollout of Claude Opus 4.7 and the immediate hotfixes that followed, plus a growing community alarm over GitKraken silently injecting monitoring hooks into Claude Code's global settings. Buckle up — it's been a busy 24 hours in the agentic coding world.

Model Updates

Opus 4.7 ships — then immediately gets a rate-limit fix. Anthropic dropped Claude Opus 4.7 today, and within hours users on subscriptions were hitting unexpectedly high rate-limit consumption on long-context requests. Anthropic confirmed the bug and pushed a fix: rate limits on Claude subscriptions weren't properly adjusted for extended-context usage in Opus 4.7. Community testing via the Pelican Test benchmark showed notable improvements over prior Opus generations. Separately, a Reddit community member running for the Best of AI Webbys Award noted the recognition as a milestone for a four-month-old startup — worth a glance if you track who's building in the Claude ecosystem.

Claude Code Developer Corner

v2.1.112: Auto-mode availability fix for Opus 4.7. The headline patch in Claude Code v2.1.112 resolves a critical regression: users in auto mode were seeing a "claude-opus-4-7 is temporarily unavailable" error even when the model was live. This blocked agentic workflows that depend on auto model selection — if you experienced mysterious unavailability errors today, updating to v2.1.112 is the fix. No other breaking changes are noted in this release.

Opus 4.7 is now the default model in Claude Code. With the update, Claude Code routes auto-mode sessions to Opus 4.7. Some users are already flagging sticker shock on token budgets since there's currently no in-tool rollback to Opus 4.6. Plan your usage accordingly if you're on a capped subscription. There's also community chatter about adaptive thinking controls — specifically, whether CLAUDE_CODE_DISABLE_ADAPTIVE_THINKING=1 is being honored in 4.7, with at least one report that it may be ignored. Worth monitoring.

🚨 Security alert: GitKraken injecting hooks into Claude Code global settings. A widely circulated warning (posted repeatedly by @user95373) claims GitKraken is silently injecting hooks into Claude Code's global settings file to monitor all events — including prompts and configuration changes — and transmitting them to GitKraken's servers, even when AI features are disabled. This is a significant supply-chain trust concern: Claude Code hooks run with local filesystem access, and a third-party tool modifying your global hook configuration without explicit consent is a serious boundary violation. Recommended action: inspect your Claude Code global settings (~/.claude/settings.json or equivalent) for unexpected hook entries, especially any pointing to GitKraken endpoints, and audit before your next session.

Claude Code Routines: agentic automation triggered by GitHub events. Multiple developers are highlighting that Claude Code Routines — which fire automated coding tasks on GitHub events — represent a qualitative shift from suggestion-based tools like Copilot or Cursor. Unlike autocomplete, Routines can execute a full workflow: click your app, run tests, push a PR — all without human touch. Combined with computer use capabilities (CLI control, test runners, dev tooling), this makes Claude Code a genuine background agent rather than a chat interface. One developer reports running PR reviews, changelog drafts, and test reports automatically every morning before opening a laptop.

Playwright MCP + /loop = multi-feature shipping days. Developers pairing Claude Code's /loop command with the Playwright MCP server for browser-driven tasks are reporting they can ship multiple features per day across parallel tabs. This combo lets Claude autonomously navigate web UIs, run integration tests, and iterate — no human in the loop required for each cycle.

Salesforce ships 60+ MCP tools for Claude Code. Salesforce released a suite of 60+ MCP tools giving Claude Code (and Cursor, Codex, Windsurf) live read/write access to any Salesforce org. That means 150,000 Salesforce developers can now build agents that query records, trigger flows, and execute business logic without writing custom API integrations. This is one of the most substantial enterprise MCP drops to date.

claude-mem: session memory plugin gaining traction. The open-source claude-mem plugin — which captures a full Claude Code session, compresses it with AI, and reinjects it as context in future sessions — hit 2,300 stars within 24 hours of launch. For developers hitting context-window resets between work sessions, this is a practical workaround worth evaluating.

Thinking levels are configurable in Claude Code but not Claude.ai. Researcher Ethan Mollick flagged a UX gap: thinking levels can be set in Claude Code but not in Claude.ai (Claude Cowork). With adaptive thinking off, there's no thinking at all — and the controls are poorly documented. If you need extended reasoning outside of a coding context, CLI is currently your only lever.

Worth Watching

  • Andrej Karpathy's CLAUDE.md file is apparently the anchor of a trending GitHub repo with 48,965 stars (+7,939 in one day) compiling his LLM coding observations into a single prompt file. If you haven't seen it, it's effectively a free Claude Code course in one markdown file.
  • Graphify for Claude Code — a new tool that builds knowledge graphs from your codebase with the goal of reducing token consumption and improving agent performance on large repos. Early-stage but worth watching for teams with sprawling monorepos.
  • LLM-Wiki MCP server — AtomicMem's llmwiki ingests URLs, compiles them into a structured wiki, and exposes ingest, compile, and query as native MCP tool calls — effectively giving any Claude Code agent a persistent external memory layer with zero custom skill code.
  • Desktop app performance complaints are getting louder. Multiple developers today called out the Claude Code desktop app as slow on macOS and frequently unavailable. The CLI remains the more reliable path for serious workloads.

Sources

  • Opus 4.7 Pelican Test — https://reddit.com/r/ClaudeAI/comments/1snakvf/opus_47_pelican_test/
  • Best of AI Webbys Award — https://i.redd.it/iqw5k88hylvg1.png
  • @trq212: Rate limit bug fix announcement — https://x.com/trq212/status/2044869038455894151
  • @trq212: Quick fix for higher rate limit usage in Opus 4.7 — https://x.com/trq212/status/2044869569555411383
  • Claude Code v2.1.112 Release — https://github.com/anthropics/claude-code/releases/tag/v2.1.112
  • Claude Code CHANGELOG v2.1.112 — https://github.com/anthropics/claude-code/blob/main/CHANGELOG.md#21112
  • @user95373: GitKraken hook injection warning — https://x.com/user95373/status/2044869527012364791
  • @surti_sneh65996: Opus 4.7 as new default, no rollback — https://x.com/surti_sneh65996/status/2044868207882715287
  • @RobEvanSanders: DISABLE_ADAPTIVE_THINKING may be ignored in 4.7 — https://x.com/RobEvanSanders/status/2044868248924295427
  • @dorlando1997: Claude Code Routines overview — https://x.com/dorlando1997/status/2044868531741966801
  • @coo_pr_notes: Morning routine automation with Routines — https://x.com/coo_pr_notes/status/2044867839434396068
  • @NerdyDudeofEAH: /loop + Playwright MCP — https://x.com/NerdyDudeofEAH/status/2044868434467631164
  • @newclawtimes: Salesforce 60+ MCP tools — https://x.com/newclawtimes/status/2044867941330727315
  • @dorlando1997: claude-mem plugin — https://x.com/dorlando1997/status/2044868555188146247
  • @emollick: Thinking levels in Claude Code vs Claude.ai — https://x.com/emollick/status/2044868563626824062
  • @DC__64: Thinking toggle in Claude Code — https://x.com/DC__64/status/2044867968044003802
  • @seelffff: Andrej Karpathy CLAUDE.md repo — https://x.com/seelffff/status/2044868944830644317
  • @thegb: Graphify knowledge graph tool — https://x.com/thegb/status/2044868765117210964
  • @Greenie989: LLM-Wiki MCP server — https://x.com/Greenie989/status/2044868473889992924
  • @TheTokenSensei: Claude Code macOS desktop slow — https://x.com/TheTokenSensei/status/2044868051032547423
  • @soumitrashukla9: Claude Code desktop frequently down — https://x.com/soumitrashukla9/status/2044869218160538022
  • @Ozzzagen: Auto mode Opus 4.7 unavailability bug — https://x.com/Ozzzagen/status/2044868982495179060
  • @FlowKitsAI: Claude Code Routines GitHub event triggers — https://x.com/FlowKitsAI/status/2044868845484269854