Skip to content

Claude Apex V7.0.0 — Three-Layer Auto-Routing

Latest
Latest

Choose a tag to compare

View all tags
@YousefNabil-SOC YousefNabil-SOC released this 23 Apr 19:48
· 8 commits to master since this release
v7.0.0
68bcfc9

v7.0.0 (2026-04-23) — Three-Layer Auto-Routing

Headline

The three-layer auto-routing system is the core V7 innovation: CARL (JIT rules) + CAPABILITY-REGISTRY (task → stack) + COMMAND-REGISTRY (intent → commands). Users never need to type slash commands — the system reads natural language and activates the right tools automatically.

New Capabilities

Three-Layer Auto-Routing

  • CARL now injects rules at UserPromptSubmit via carl-hook.py (UTF-8 fixed on all 4 file opens)
  • CAPABILITY-REGISTRY.md rewritten to route 30+ task patterns to skill+MCP+agent+CLI combos
  • COMMAND-REGISTRY.md added — 182 slash commands indexed with auto-invocation routing

7 New Custom Skills

  • premium-web-design — luxury web design skill with animation patterns, reference analyses, and GSAP/Framer Motion tool guides (205-line SKILL.md + patterns/ + references/ + tools/)
  • 21st-dev-magic — React+TS+Tailwind component generation via 21st.dev Magic MCP
  • instagram-access — authenticated Instagram scraping via instaloader (generic session file)
  • graphify — knowledge-graph skill; saves ~90% of file-reading tokens
  • graphic-design-studio — code-based graphic design (HTML/CSS/SVG/Pillow pipelines)
  • fireworks-tech-graph — technical documentation graphing
  • impeccable — design polish + micro-interaction refinement

3 New MCP Servers in default settings

  • @21st-dev/magic — premium React component generation
  • (context7, playwright, github, exa-web-search already present)

2 New Hooks

  • session-start-check.sh — reads ONLY essential files (CLAUDE.md, PRIMER.md, CAPABILITY-REGISTRY.md, AGENTS.md) + memory markers. Optimized against the prior "read everything" approach.
  • project-auto-graph.sh — on SessionStart, detects new project folders, queues them for user-opt-in graphify build (never auto-runs to protect token budget).

CARL Expanded

  • 7 → 9 domains: added BROWSER and PROJECT-NAVIGATION
  • 33 → 40 rules
  • All rules sanitized to generic, reusable templates (zero personal profile data, zero client data)

New Config Files

  • config/env.template — API key env var names with placeholders
  • config/command-registry.md — 182-command index with auto-invocation routing table
  • config/agents.md — agent divisions (Engineering, Business, Documentation, Cybersecurity, RuFlo, OMC)
  • config/auto-activation-matrix.md — task→tool activation mapping

Beginner Documentation (10 new guides)
Zero-prerequisite tutorials for users who have never used Claude Code or a terminal:

  • docs/00-START-HERE.md — "What is this?" for absolute beginners
  • docs/01-WHAT-IS-CLAUDE-CODE.md — defines Claude, Anthropic, CLI, terminal
  • docs/02-INSTALL-FROM-ZERO.md — every step from brand-new computer to running Apex
  • docs/03-FIRST-TIME-USING.md — first 10 commands with expected output
  • docs/04-WHAT-ARE-SKILLS.md — analogy: skills = recipe books Claude reads on demand
  • docs/05-WHAT-ARE-AGENTS.md — analogy: agents = specialist coworkers
  • docs/06-WHAT-ARE-MCP-SERVERS.md — analogy: MCP = superpowers plugged in
  • docs/07-WHAT-ARE-HOOKS.md — analogy: hooks = automatic reflexes
  • docs/08-WHAT-IS-CARL.md — analogy: CARL = librarian
  • docs/09-GLOSSARY.md — every term defined in one sentence
  • docs/10-TROUBLESHOOT-FOR-BEGINNERS.md — "It didn't work — now what?"

Fixed

  • carl-hook.py — UTF-8 encoding fix on all 4 open() calls (was corrupting non-Latin file content)
  • task-complete-sound.sh — 60-second cooldown added (prevents sound spam)
  • session-start-check.sh — reads only 4 essential files + memory markers (was reading everything, slow at session start)
  • effortLevel default lowered to high (never xhigh or max — self-healing enforces this)
  • Self-healing rules added to CLAUDE.md template
  • Post-compact recovery script simplified and sanitized (was referencing personal projects)

Changed

  • Version badges: 6.0 → 7.0 everywhere
  • Skill count: 1,308 (V6 claim) → 1,276+ verified on disk
  • MCP count: 12 → 15 (includes all optional)
  • Plugin count: 19 → 20 (added claude-md-management)
  • Slash commands: 107 → 182 (added 75 via user-created + plugin expansions)
  • Agent count: 108 remains stable (25 custom + 51 RuFlo + 19 OMC + 13 plugin)
  • Commands switch-project.md and templates.md rewritten as generic templates users edit for their own projects
  • README rewritten with V7 numbers, three-layer routing diagram, and "Start Here" beginner pointer
  • Install scripts (install.sh, install.ps1, install-interactive.sh) updated to install all 7 new skills + 2 new hooks + env.template
  • verify.sh updated with V7 expected counts (9 CARL domains, 40 rules, new skill checks)

Security

  • Full grep sweep before push: zero API keys, zero personal names, zero client/project data leaked
  • All API key references use ${VAR_NAME} placeholders
  • .env.template documents every required env var
  • .gitignore enforces .env exclusion

Attribution

  • Third-party credit preserved for OMC (Yeachan-Heo), PAUL (ChristopherKahler), SEED (ChristopherKahler), Claude Peers (louislva), Bun runtime
  • Skills from everything-claude-code plugin remain installed via /plugin install (not vendored in this repo)

Breaking Changes

  • ~/.claude/settings.json hooks section now requires UserPromptSubmit and SessionStart-chain entries for full auto-routing. The installer adds these non-destructively, but if you have custom hooks in those events, review the merge.
  • CARL domain count changed from 7 to 9 — any custom rules referencing old domain IDs must migrate.
Assets 2
Loading