claude-cortex 2.5.0

Python CLI/TUI and intelligent framework for managing Claude and other AI agents

Cortex

This repository packages the Cortex (cortex) context management toolkit as a Claude Code plugin. It bundles the curated agents, commands, modes, rules, and supporting Python CLI + TUI so teams can install the complete experience through the plugin system or keep using the standalone cortex / cortex-ui scripts.

Note: The claude-ctx command has been deprecated but will remain available for a while.

📚 Docs: https://cortex.atlascrew.dev/

🎬 Presentations:

• Intro Overview
• Technical Deep Dive
• Executive Roadmap
• Feature Catalog

What's Inside

Core Assets (Loaded by Claude Code)

Directory	Purpose
agents/	Claude subagents with dependency metadata. Move to inactive/agents/ to park.
commands/	Slash command definitions (50+ curated commands)
skills/	Reusable skill modules (80+ skills) with activation triggers and dependencies
modes/	Behavioral context modules that toggle workflow defaults
rules/	Rule sets symlinked to ~/.claude/rules/cortex/ on launch
flags/	Modular token-efficient flag packs (22 categories, 3,380 tokens total)
hooks/	Automation hooks for pre/post tool execution and session events

Orchestration & Templates

Directory	Purpose
profiles/	Project type presets (frontend, backend, devops, etc.)
scenarios/	Multi-step workflow scenarios
workflows/	Complex workstream orchestration templates
prompts/	Prompt templates: guidelines, personas, and reusable snippets
templates/	Initialization templates for new projects
plugins/	Additional plugin modules (accessibility, frontend, UX)

Python Package (claude_ctx_py/)

Module	Purpose
cli.py	Main CLI entrypoint with all subcommands
launcher.py	cortex start implementation (plugin resolution, rules symlinking)
core/	Core domain logic (agents, skills, modes, rules, hooks, backups)
tui/	Textual-based terminal UI with 15+ specialized views
intelligence/	AI recommendation engine and pattern learning
memory/	Session memory and knowledge persistence
analytics.py	Usage analytics and skill metrics

Supporting Files

Path	Purpose
.claude-plugin/plugin.json	Plugin manifest for Claude Code marketplace
docs/	Documentation site (Jekyll-based)
tests/	Test suite with pytest
schema/	JSON schemas for validation
scripts/	Helper scripts for installation and maintenance

🆕 Latest Updates

• Hook logging – Hook failures now write to ~/.cortex/logs/hooks.log (override via CORTEX_HOOK_LOG_PATH).
• Extra plugin dirs – cortex-config.json supports extra_plugin_dirs to pass multiple --plugin-dir entries.
• Watch daemon mode – cortex ai watch --daemon runs watch mode in the background with status/stop controls.
• Template guardrails – The TUI detects missing templates/ files in the active CLAUDE_PLUGIN_ROOT and offers to initialize them or run the setup wizard.
• Multi-LLM consult skill – Ask Gemini, OpenAI (Codex), or Qwen for a second opinion; configure provider API keys via the TUI Command Palette -> "Configure LLM Providers".
• Asset Manager reliability – “Update All” and “Install All in Category” now behave consistently with clearer prompts.
• Flag toggles restored – Spacebar toggling works again in the Flag Explorer and Flag Manager, updating FLAGS.md immediately.

✅ Stability Update: AI + Context Management

We’ve fixed major issues across AI recommendations and context state tracking. Auto-activation and watch mode are more reliable, and context activation now uses .active-* state files with cortex doctor and cortex setup migrate to keep everything consistent.

🔥 New: Super Saiyan Mode

Universal visual excellence framework with platform detection:

• modes/Super_Saiyan.md – Core generic mode with auto-detection
• modes/supersaiyan/ – Platform-specific implementations (Web, TUI, CLI, Docs)
• claude_ctx_py/tui_supersaiyan.py – Enhanced Textual components
• examples/supersaiyan_demo.py – Interactive demo
• Three power levels: ⭐ Super Saiyan → ⚡ Kamehameha → 💥 Over 9000

Quick start:

python examples/supersaiyan_demo.py  # See it in action!

See Super Saiyan Integration Guide for details.

📦 New: Asset Manager

Install, diff, and update plugin assets directly from the TUI:

• Discover hooks, commands, agents, skills, modes, workflows, and flags
• Install/uninstall to any detected .claude directory
• Diff installed vs source, bulk install by category, update outdated assets

Quick start:

cortex tui
# Press 'A' for Asset Manager
# i=install, u=uninstall, d=diff, U=update all, I=bulk install, T=target dir

🌿 New: Worktree Manager

Manage git worktrees from the CLI or TUI.

CLI:

cortex worktree list
cortex worktree add my-branch --path ../worktrees/my-branch
cortex worktree remove my-branch
cortex worktree prune --dry-run
cortex worktree dir ../worktrees
cortex worktree dir --clear

TUI:

cortex tui
# Press 'C' for Worktrees
# Ctrl+N add, Ctrl+O open, Ctrl+W remove, Ctrl+K prune, Ctrl+B set base dir

🧭 New: Setup, Init & Migration

The installer and setup tooling have been overhauled to keep projects consistent across upgrades.

# Detect your project and recommend a profile
cortex init detect

# Apply a profile directly
cortex init profile backend

# Check init status
cortex init status

# Migrate CLAUDE.md comment activation → .active-* files
cortex setup migrate

🤖 New: AI Intelligence & Automation

Stay in Claude Code flow - Let AI manage the framework for you with context-aware intelligence, pattern learning, and auto-activation:

• Context Detection – Automatically analyzes changed files, detects auth/API/tests/frontend/backend
• Pattern Learning – Learns from successful sessions, recommends optimal agent combinations
• Workflow Prediction – Predicts agent sequences based on similar past work
• Auto-Activation – High-confidence agents activate automatically (≥80%)
• Watch Mode – Real-time monitoring with instant recommendations (no daemon required)
• TUI AI Assistant – Interactive view with recommendations and predictions (press 0)
• Skill Palette Shortcuts – Ctrl+P → type “Skill…” to run info, versions, deps, analytics, trending, or community install/validate/rate/search commands without leaving the TUI
• Multi-Reviewer Auto-Activation – Code changes can auto-activate multiple reviewers (quality, code, TS/React/UX, DB/SQL, performance, architecture)

Quick start:

# Get AI recommendations for current context
cortex ai recommend

# Auto-activate high-confidence agents
cortex ai auto-activate

# Start watch mode (real-time monitoring)
cortex ai watch

# Interactive TUI with AI assistant
cortex tui
# Press '0' for AI Assistant view
# Press 'A' to auto-activate recommendations

# Record successful sessions for learning
cortex ai record-success --outcome "feature complete"

Multi-review output example:

1. 🔵 quality-engineer [AUTO]
   Confidence: 85%
   Reason: Changes detected - quality review recommended

2. 🔵 code-reviewer [AUTO]
   Confidence: 75%
   Reason: Changes detected - code review recommended

3. 🔵 typescript-pro [AUTO]
   Confidence: 85%
   Reason: TypeScript changes detected - review recommended

4. 🔵 react-specialist [AUTO]
   Confidence: 80%
   Reason: React/UI component changes detected - review recommended

5. 🔵 ui-ux-designer [AUTO]
   Confidence: 80%
   Reason: User-facing UI changes detected - UX review recommended

Watch Mode Example:

══════════════════════════════════════════════════════════════════════
🤖 AI WATCH MODE - Real-time Intelligence
══════════════════════════════════════════════════════════════════════

[10:33:12] 🔍 Context detected: Backend, Auth
  3 files changed

  💡 Recommendations:
     🔵 quality-engineer [AUTO]
        85% - Changes detected - quality review recommended
     🔵 code-reviewer [AUTO]
        75% - Changes detected - code review recommended
     🔴 security-auditor [AUTO]
        95% - Auth code detected

[10:33:12] ⚡ Auto-activating 3 agents...
     ✓ quality-engineer
     ✓ code-reviewer
     ✓ security-auditor

See AI Intelligence Guide and Watch Mode Guide for complete documentation.

⭐ New: Skill Ratings & Auto-Feedback Loops

Phase 5 introduces a first-class feedback engine so skills can improve themselves:

• Ratings & Reviews – cortex skills rate <skill> stores star ratings, helpful/not-helpful votes, and optional text feedback in ~/.cortex/data/skill-ratings.db.
• Quality Metrics – cortex skills ratings <skill> shows averages, distributions, success correlation, and token efficiency; skills top-rated, skills export-ratings, and skills analytics expose the aggregate view.
• TUI Surfacing – The Skills table now includes a Rating column (press 5). Select a skill and press Ctrl+R to launch an inline rating dialog without leaving the terminal.
• Auto Prompts – Recent skill activations trigger modal prompts shortly after the TUI launches. The prompt explains why the skill was selected (usage count, task types, success rate) and offers to collect feedback on the spot. Dismiss once to snooze for 24 h; rating it clears future prompts until another burst of usage.
• Recommendation Feedback Loop – Ratings feed back into the AI recommender, so highly rated skills are prioritized and low-signal ones get demoted automatically (Feature 2 of the Phase 5 roadmap).

# Record a rating and optional review
cortex skills rate owasp-top-10 --stars 5 --review "Still the best security checklist"

# Inspect ratings/metrics
cortex skills ratings owasp-top-10
cortex skills top-rated --limit 5

# Export for analysis
cortex skills export-ratings --format csv > skill-feedback.csv

Within the TUI:

cortex tui
# Press 5 for Skills view, highlight a skill, press Ctrl+R to rate
# Auto prompts appear when the assistant detects a frequently used skill that lacks fresh feedback

See Phase 5 Roadmap for the broader Intelligence + Feedback plan.

🔌 New: MCP Server Management

Intelligent MCP server management - Observe, validate, and document your Model Context Protocol servers:

• Server Discovery – Automatically finds MCP servers from Claude Desktop config
• Configuration Validation – Diagnose issues and verify server setup
• Curated Documentation – Built-in guides for Context7, Serena, Sequential, Magic, and more
• Visual Dashboard – TUI view with server status, testing, and docs (press 7)
• Smart Recommendations – Integration with /tools:select for optimal MCP routing

Quick start:

# List all configured MCP servers
cortex mcp list

# Show server details and validation
cortex mcp show context7

# View curated documentation
cortex mcp docs serena

# Diagnose all servers
cortex mcp diagnose

# Generate config snippet
cortex mcp snippet playwright

TUI Interface:

cortex tui
# Press '7' for MCP Servers view
# t=test, d=docs, c=copy, v=validate

See MCP Management Guide for complete documentation.

⚙️ New: Token-Efficient Flag Management

Smart flag management - Control Claude's behavior flags with surgical precision and save tokens:

• Modular Flag Categories – 22 flag categories split into focused files (mode-activation, testing, debugging, etc.)
• Token Analytics – Real-time token counting shows savings per category (~100-250 tokens each)
• TUI Flag Manager – Visual interface for enabling/disabling flags (press Ctrl+G)
• Profile Integration – Flags auto-configure when switching profiles
• Config Auto-Update – Changes persist immediately to FLAGS.md

Flag Categories (3,380 tokens total):

Category	Tokens	Purpose
Mode Activation	120	Core behavioral flags (brainstorm, introspect, orchestrate)
MCP Servers	160	MCP server control (context7, sequential, magic, etc.)
Thinking Budget	140	Reasoning budget controls and cost-aware tuning
Analysis Depth	130	Thinking depth control (--think, --ultrathink)
Execution Control	150	Delegation, concurrency, iteration control
Visual Excellence	250	Super Saiyan, UI polish, design system
Output Optimization	120	Scope, focus, compression flags
Testing & Quality	170	TDD, coverage, mutation testing
Learning & Education	160	Educational modes, explanations
Cost Management	120	Budget limits, cost awareness
Refactoring Safety	140	Safe refactoring, behavior preservation
Domain Presets	150	Frontend, backend, fullstack presets
Debugging & Trace	110	Verbose logging, execution tracing
Interactive Control	130	Confirmation, pair programming modes
CI/CD	100	Headless, JSON output, automation
Auto-Escalation	180	Automatic reasoning depth adjustment
Performance Optimization	180	Profiling, benchmarking, scaling guidance
Security Hardening	190	Security-first workflows, threat modeling
Documentation Generation	170	Doc-driven workflows, reference output
Git Workflow	160	PR hygiene, commits, release steps
Migration & Upgrade	170	Version upgrades, compatibility guarantees
Database Operations	180	Schema changes, data safety, migrations

Quick start:

# Open Flag Manager in TUI
cortex tui
# Press Ctrl+G for Flag Manager
# Use ↑↓ to select, Space to toggle

# Apply profile with flags
cortex profile apply frontend
# Auto-enables: visual-excellence, testing-quality, debugging-trace
# Saves: ~1,120 tokens (52% savings!)

Example: Frontend Profile

Example configuration enables 6/22 categories (930 tokens):

• mode-activation, mcp-servers, analysis-depth
• execution-control, visual-excellence, output-optimization

When you switch to frontend profile:

• Auto-enables: testing-quality, domain-presets, debugging-trace
• Loads: 1,110 tokens (7 categories)
• Saves: 2,270 tokens (15 categories disabled)
• Savings: 67% reduction in flag overhead

All Profile Configurations:

Profile	Active Flags	Tokens Loaded	Tokens Saved	Savings
minimal	3 categories	430	2,950	87%
frontend	7 categories	1,110	2,270	67%
backend	7 categories	980	2,400	71%
devops	5 categories	640	2,740	81%
documentation	3 categories	430	2,950	87%
quality	7 categories	980	2,400	71%
full	22 categories	3,380	0	0%

Flag Manager Interface:

⚙️ Flag Manager

Status  Flag Category                    Tokens  File
──────  ────────────────────────────────  ──────  ─────────────────────────────
Summary 6/22 active                       930/3380 Saving 72% tokens (2450 tokens)
──────  ────────────────────────────────  ──────  ─────────────────────────────
✓ ON    ▸ Mode Activation Flags           120     mode-activation.md
✓ ON    MCP Server Flags                  160     mcp-servers.md
✗ OFF   Testing Quality Flags             170     testing-quality.md
✗ OFF   Learning Education Flags          160     learning-education.md
...

Controls: ↑↓ Select    Space Toggle    Changes saved to FLAGS.md

Location: Flag files live in ~/.cortex/flags/ and are referenced in ~/.cortex/FLAGS.md (which is included by CLAUDE.md).

Rules: Source rules live in ~/.cortex/rules/ and are symlinked into ~/.claude/rules/cortex/ on launch so Claude Code loads them. The launcher also adds rules/cortex/ to ~/.claude/.gitignore.

See Flag Management Guide for complete documentation.

The plugin manifest lives in .claude-plugin/plugin.json so Claude Code detects commands and agents automatically when the marketplace entry points to this repository.

Installation

Quick Start

# Install with pipx (recommended for CLI tools)
pipx install claude-cortex

# Or with uv
uv tool install claude-cortex

# Or with pip
pip install claude-cortex

# Initialize ~/.cortex with bundled assets (rules, flags, modes, etc.)
cortex install bootstrap

# Install shell completions and manpages
cortex install post

# Launch Claude Code with Cortex context
cortex start

Installation Options

pipx (recommended) – Isolated environment, no conflicts:

pipx install claude-cortex

uv – Fast, modern package manager:

uv tool install claude-cortex
# Or in a project: uv pip install claude-cortex

pip – Standard installation:

pip install claude-cortex

Development install – Editable mode for contributors:

git clone https://github.com/NickCrew/claude-cortex.git
cd claude-cortex
pipx install -e .
# Or: uv pip install -e .[dev]

Post-Install Setup

After installing the package, run these commands to complete setup:

1. Bootstrap ~/.cortex (required for first install):

cortex install bootstrap

This creates ~/.cortex/ and copies:

• rules/ – Rule markdown files
• flags/ – Flag category files (22 categories)
• modes/ – Behavioral mode files
• templates/ – Initialization templates
• cortex-config.json – Default configuration

Use --force to reinitialize: cortex install bootstrap --force

2. Install shell integrations (optional but recommended):

cortex install post

This installs:

• Shell completions (bash/zsh/fish)
• Manual pages (man cortex)
• Local architecture documentation

---

How cortex start Works

The cortex start command launches Claude Code with your Cortex configuration automatically applied. Here's what happens:

1. Configuration Loading

Reads ~/.cortex/cortex-config.json (created by cortex install bootstrap) plus FLAGS.md to determine what context to load.

2. Plugin Root Resolution

The launcher finds your Cortex assets in this order:

1. --plugin-root CLI argument (if provided)
2. CLAUDE_PLUGIN_ROOT or CORTEX_PLUGIN_ROOT environment variable
3. plugin_dir field in cortex-config.json
4. Bundled assets from the installed Python package
5. Claude's plugin registry (~/.claude/plugins/installed_plugins.json)

3. Rules Symlinking

Active rules from rules/ are symlinked into ~/.claude/rules/cortex/ so Claude Code automatically loads them. The launcher also adds rules/cortex/ to ~/.claude/.gitignore.

4. Claude Code Launch

Finally, cortex start executes Claude Code with:

• --plugin-dir pointing to the Cortex assets
• --settings if a settings.json path is configured
• Any extra --plugin-dir entries from extra_plugin_dirs
• Any persistent arguments from claude_args

Example Launch

# Basic launch
cortex start

# Override modes for this session only
cortex start --modes "Deep_Analysis,Quality_Focus"

# Override flags for this session only
cortex start --flags "security-hardening,testing-quality"

# Pass extra args to Claude (after --)
cortex start -- --model claude-sonnet-4-20250514

# Use a different Claude binary
cortex start --claude-bin /path/to/claude

Alias: cortex claude is equivalent to cortex start.

---

Directory Structure

What Gets Installed Where

Location	Purpose
~/.cortex/	Cortex home – configuration, rules, flags, modes, principles
~/.cortex/cortex-config.json	Launcher configuration (created on first cortex start)
~/.cortex/FLAGS.md	Active flag references (updated by TUI Flag Manager)
~/.cortex/PRINCIPLES.md	Generated principles (from principles/*.md)
~/.cortex/rules/	Rule markdown files
~/.cortex/flags/	Individual flag category files
~/.cortex/modes/	Behavioral mode files
~/.cortex/principles/	Principles snippet files
~/.claude/rules/cortex/	Symlinked rules – created by cortex start
~/.claude/settings.json	Claude Code settings (hooks, etc.)

Activation State Files

File	Purpose
.active-modes	List of currently active modes
.active-rules	List of currently active rules
.active-mcp	List of active MCP documentation
.active-principles	List of active principles snippets

---

Configuration: cortex-config.json

The launcher configuration file controls what context gets loaded when you run cortex start.

Location

• Default: ~/.cortex/cortex-config.json
• Created: Automatically on first cortex start

Full Example

{
  "plugin_dir": "~/Developer/claude-cortex",
  "plugin_id": "cortex",
  "extra_plugin_dirs": [
    "~/my-custom-plugin"
  ],
  "rules": [
    "workflow-rules",
    "quality-rules",
    "git-rules"
  ],
  "flags": [
    "mode-activation",
    "mcp-servers",
    "testing-quality"
  ],
  "modes": [
    "Deep_Analysis",
    "Quality_Focus"
  ],
  "principles": [
    "00-core-directive",
    "10-philosophy",
    "20-engineering-mindset"
  ],
  "settings_path": "~/.claude/settings.json",
  "claude_args": [
    "--model", "claude-sonnet-4-20250514",
    "--dangerously-skip-permissions"
  ],
  "watch": {
    "directories": ["~/projects/my-app"],
    "auto_activate": true,
    "threshold": 0.7,
    "interval": 2.0
  }
}

Field Reference

Field	Type	Default	Description
plugin_dir	string	auto-detected	Explicit path to Cortex assets. Overrides auto-discovery.
plugin_id	string	"cortex"	Plugin ID for registry lookup (if plugin_dir not set).
extra_plugin_dirs	string[]	[]	Additional plugin directories passed via --plugin-dir.
rules	string[]	all in rules/	Rule slugs (without .md) to symlink.
flags	string[]	from FLAGS.md	Fallback if FLAGS.md missing. Override with --flags.
modes	string[]	[]	Mode slugs to append to system prompt.
principles	string[]	all in principles/	Principles snippets to include.
settings_path	string	template if present	Path passed to Claude via --settings.
claude_args	string[]	[]	Args always passed to claude command.
watch	object	unset	Watch mode defaults (directories, auto-activate, etc.).

Notes

• Flags: Read from FLAGS.md when present; flags in config is a fallback.
• Rules: Specified by slug (filename without .md). Active rules are symlinked to ~/.claude/rules/cortex/.
• claude_args: Must be a JSON array of strings. Parsed with shell splitting if you accidentally provide a single string.

Overriding Per-Session

# Override modes for one session
cortex start --modes "Architect,Deep_Analysis"

# Override flags for one session
cortex start --flags "security-hardening,database-operations"

# Both
cortex start --modes "Quality_Focus" --flags "testing-quality"

---

Environment Variables

Variable	Purpose
CORTEX_ROOT	Override Cortex home directory (default: ~/.cortex)
CORTEX_SCOPE	Scope selection: project, global, or plugin
CLAUDE_PLUGIN_ROOT	Explicit plugin assets path
CORTEX_PLUGIN_ROOT	Alias for CLAUDE_PLUGIN_ROOT
CORTEX_HOOK_LOG_PATH	Override hook log location (default: ~/.cortex/logs/hooks.log)

Scope Selection

# Use project-local .claude/ directory
cortex --scope project status

# Use explicit plugin root
cortex --plugin-root /path/to/cortex status

# Set via environment
export CLAUDE_PLUGIN_ROOT="$HOME/Developer/claude-cortex"
cortex status

Shell completion

Shell completions are automatically installed when using the legacy installer (./scripts/deprecated/install.sh). For manual setup:

Automatic (recommended):

# Generate and install completions for your shell
cortex completion bash > ~/.bash_completion.d/cortex
cortex completion zsh > ~/.zsh/completions/_cortex
cortex completion fish > ~/.config/fish/completions/cortex.fish

# Show installation instructions
cortex completion bash --install

After adding or updating CLI subcommands (for example, setup migrate-commands), regenerate completions so the new options appear.

Using argcomplete (legacy method):

# Bash
register-python-argcomplete cortex > ~/.local/share/bash-completion/completions/cortex

# Zsh
register-python-argcomplete --shell zsh cortex > ~/.local/share/zsh/site-functions/_cortex

# Fish
register-python-argcomplete --shell fish cortex > ~/.config/fish/completions/cortex.fish

See Shell Completions Guide for detailed instructions.

Manual page (manpage)

A comprehensive manual page is available in docs/reference/cortex.1 and is automatically installed when using the legacy installer (./scripts/deprecated/install.sh).

View locally:

man docs/reference/cortex.1

Dedicated entries are also available for the TUI (man cortex-tui) and the workflow/scenario orchestration commands (man cortex-workflow).

Manual installation:

./scripts/deprecated/install-manpage.sh

After installation:

man cortex

The manpage documents all commands, subcommands, options, file locations, environment variables, and includes practical examples. It follows standard Unix manual page conventions and can be searched with / when viewing.

Advanced Features

For more advanced features, see the following guides:

• Warp AI & Terminal AI Integration
• Hooks and Auto-Suggestions

License & Attribution

This project is licensed under the MIT License - see the LICENSE file for details.

Credits

Cortex builds upon ideas and patterns from several excellent projects in the Claude Code ecosystem:

• obra/superpowers - Systematic debugging and quality gate patterns (MIT License)
• VoltAgent/awesome-claude-code-subagents - Specialized agent architecture and modular design (MIT License)
• SuperClaude-Org/SuperClaude_Framework - Behavioral modes, slash commands, and MCP integration patterns (MIT License)
• just-every/code - Multi-agent orchestration and reasoning control concepts (Apache-2.0 License)

See CREDITS.md for detailed attribution and a complete list of inspirations and dependencies.

Development notes

• Update the version in .claude-plugin/plugin.json whenever you publish a new release.
• Keep semantic changes to commands or agents alongside changelog entries in CLAUDE.md or RULES.md.
• Use claude plugin validate . to confirm the manifest structure prior to publishing.
• Run strict type checks before opening a PR.
  • python3 -m mypy --strict claude_ctx_py
  • The CI harness also drops the latest failure output in /tmp/mypy.log; keep that file around when iterating locally so you can jump directly to errors with your editor.
  • New modules should prefer TypedDict/Protocol over raw dict/Any, and Textual mixins need explicit state attributes so the UI keeps passing --strict.

For marketplace configuration examples, see ../claude-private-marketplace.

Preview the docs locally

The documentation site under docs/ now uses the default GitHub Pages minima theme with custom styling. To run it locally:

cd docs
bundle install
bundle exec jekyll serve --source . --livereload

Then open http://127.0.0.1:4000. Changes to Markdown or assets refresh automatically.