CLAUDE.md · git:20260729.304dcac · 2026-07-29 · sha256 b9b0ee5f418c4b49
CLAUDE.md git:20260729.304dcacB
Immutable. This exact content never changes and is served at /api/v1/blob/b9b0ee5f418c4b49.
# CLAUDE.md — CC Commander
## What This Is
CC Commander — by Kevin Zicherman. Guided AI PM to Master Claude Code Instantly. **Core free forever** — affiliate links + consulting + optional Pro community + GitHub Sponsors. Built on Anthropic's Claude Agent SDK sub-agent architecture — CCC is a PM/orchestration UX layer over that primitive. The Desktop plugin is the primary product — **primary surface: Claude Code Desktop (aka Cowork Desktop)**. **82 plugin skills** (13 /ccc-* specialist workflows + 11 ccc-* domain routers + 7 channel/CI/ECC/setup skills + diagnostic/meta + vendor-sourced + lifecycle + /ccc-deploy + /ccc-rollback + /ccc-onboard + /ccc-fleet-viz + /ccc-tuneup + /ccc-update + /ccc-console; the Cockpit's Browse tab separately displays a 14-domain taxonomy for filtering — a display concept, not a routers count), **22 specialist sub-agent personas** (architect · reviewer · builder · security-auditor · debugger · designer · qa-engineer · devops-engineer · data-analyst · content-strategist · product-manager · performance-engineer · researcher · technical-writer · fleet-worker · typescript-reviewer · python-reviewer · go-reviewer · rust-reviewer · java-reviewer · kotlin-reviewer · csharp-reviewer), **23 lifecycle hooks × 44 handlers** (SessionStart [orchestrator], UserPromptSubmit, PreToolUse, PostToolUse, Stop, Notification, PreCompact, PostCompact, SubagentStop, SubagentStart, PermissionRequest, SessionEnd, Elicitation, ElicitationResult, StopFailure, PostToolUseFailure, PostToolBatch, TaskCreated, TaskCompleted, ConfigChange, UserPromptExpansion, InstructionsLoaded, Setup), **2 credential-free bundled MCP servers** (context7 + sequential-thinking) + 16 opt-in via /ccc-connect. Also a comprehensive Claude Code configuration toolkit + interactive CLI project manager: 467 skills, 11 CCC domains, 83+ commands, 24 hooks, 37 prompt templates, 10 themes, 19 vendor packages, 3 starter templates, real-time agent dashboard, OpenClaw native integration, agent-friendly CLI API, tabbed tmux split mode, status updates, continuous improvement pipeline, modular installer. Built by scanning 200+ articles from the Claude Code community and distilling into one install.
**Current Release:** v7.4.1 — `/ccc-update` and `/ccc-doctor` no longer dead-end on **Desktop-managed (account-synced) installs** — the mode with no marketplace clone and no `installed_plugins.json` entry *by design*. Both now detect three install modes (CLI marketplace / Desktop-managed / dev checkout), read the Desktop copy's real version, fetch latest from GitHub directly, and give mode-correct guidance. Added in v7.4.0: **One Commander.** The four published decks (`/ccc-browse`, `/ccc-mission-control`, `/ccc-usage`, `/ccc-safety`) are now tabs of a single **inline console**, `/ccc-console` (82nd plugin skill): Overview · Usage · Safety · **Memory** · **History** · Launch, rendered through the visualize MCP with a prompt bar that reaches the live session via `sendPrompt(text)`. The security model is the feature — chips are fixed template commands that display exactly what they will send, free text is newline-stripped and capped at 500 chars, and **no telemetry-, memory- or log-derived text ever reaches a `sendPrompt` payload** (anything that can append to a JSONL could otherwise compose a command the user appears to have typed). The four deck skills survive as "publish just this tab" exports on **their existing file paths**, so every bookmarked artifact URL keeps updating in place; `/ccc-console publish` absorbs the Cockpit's URL rather than minting a fifth. **Memory** reads your own claude-mem store read-only, titles only — claude-mem is AGPL and deliberately not bundled, so "not installed" is the normal zero-state, never an error. **History** is 30 days of telemetry Commander already writes (`metrics.jsonl` as the durable backbone, detail counted by filename) — no new collector, and per-message content is neither collected nor claimed. The console **opens itself once at session start** (local-only render; publishing stays consent-gated), off via `/ccc-console off`, `{"autoOpen": false}` in `~/.claude/commander/console.json`, or `CCC_NO_AUTOCONSOLE=1`. Under it all: one renderer and one composed model — the three deck libraries shed a net 1,009 lines of duplicated markup, proven byte-identical against pre-refactor goldens. **Fixed:** a read/write race that let identical telemetry produce different Usage numbers run to run (Mission Control persists `metrics.jsonl`; Usage reads it — the composed model now sequences writer before reader), and hook `timeout` values that were milliseconds in a seconds field, so the anti-hang ceilings had never been in force. `plugin.json` now sets `defaultEnabled: true`. Added in v7.3.2: **Anthropic + Codex CLI only.** Commander no longer routes work to Gemini or any local-inference runtime: the Gemini adapter is gone, executor preference is `claude` → `codex`, and `ccc-research/gemini-fallback` is rewritten as `large-context` (use Opus 5's native 1M window → reduce → chunk-and-synthesize → hand mechanical execution to the Codex CLI). The old skill existed to escape "Claude's 200K limit," a premise Opus 5 retired. The `features/local-llm` page is removed. Added in v7.3.1: **Claude Opus 5 adopted end-to-end**, plus two real bug fixes. The default model pin is now `claude-opus-5` (dateless, no `[1m]` suffix — Opus 5 is 1M context by default *and* max). Effort posture follows Anthropic's Opus 5 guidance: default `high`, escalate to `xhigh` for demanding agentic work, `max` only when justified — replacing the old always-on `xhigh` floor. Delegation gained a size/independence threshold (Opus 5 over-spawns without caps) and the Fable Method's verifier-≠-worker gate is now scoped to substantive work, ranking deterministic checks above a spawned verifier. **Fixed:** `/ccc-doctor` had been failing on correctly-configured repos since v6.0 (it demanded an Opus pin for four agents promoted to Fable 5 in v6.0; a synthetic test fixture drifted with the wrong expectation and hid it — the doctor is now asserted against the real repo), and complexity scores ≥93 emitted an invalid `--effort ultra`, so the *hardest* tasks were the ones failing at dispatch. `skills/advisor` also documented an impossible model pair once the executor moved to Opus 5. Added in v7.3.0: Commander finally notices its own updates — a SessionStart hook (`hooks/update-nudge.js`, 24h-cached, offline-silent) compares the installed version against GitHub `main` and, when newer, spells out the exact fix — marketplace update → plugin update → restart. **`/ccc-update`** (81st plugin skill) is the click-first version of that same walkthrough, and `/ccc-doctor` now diagnoses every stale-install layer (marketplace clone behind `main`, old `cache/<version>` dirs, stale Desktop inline copies, catalog-cache age) and offers — with consent — to enable marketplace auto-update. The Cockpit gained an 8th tab, **Prompts** — a searchable browser across four sources (Claude Code docs library, the CCC library, the repo's prompt templates, a ReadyIQ showcase) with Copy and "Enhance this" actions. Every deck now shows a **data-freshness** stamp ("Data through <timestamp>") and warns when telemetry is stale (>24h). The suggestion engine surfaces update-available and stale-telemetry signals into Mission Control's previously-empty Suggestions panel, keyed per-project so state no longer leaks between repos. **Also fixed:** the Codex mirror was silently dead on every install — it rewrote hooks to a variable Codex never sets — and now carries the correct 10 supported hook events (was 6), strips async handlers Codex would otherwise skip, and documents Codex's hook trust gate. Added in v7.2.0: a shared **Commander decks** switcher across every artifact (so users can see and jump between all of them) plus two new decks — **Usage & Cost** (`/ccc-usage`: burn, savings, cost-by-app) and **Safety** (`/ccc-safety`: blocked/auto-fixed actions + tool-failure hotspots). Added in v7.1.0: an ecosystem **Tools** tab in the Cockpit — open Commander surfaces, install the best MIT-licensed companion apps (Nimbalyst, Usage Monitor, Sniffly…), or generate an artifact dashboard. A full redesign of the Commander Cockpit shipped in v7.0.0 on the real commanderplugin.com coral brand: a filterable skill browser (rows/tiles toggle + real-time filter, replacing the one-skill-domain wall), an 11-strategy multi-select prompt enhancer with a GO button, real per-agent token/cost telemetry (recovered from the subagent transcript, deduped by `message.id`, honest "—" when unavailable), and an honest 14-domain taxonomy (no "core"). **Commander Mission Control**'s Claude+Codex rosters gained a Charts strip (cost/day, agents dispatched/day, tasks completed/week, tool failures/day — zero-dep inline SVG, theme-safe, rendered on both the live dashboard and the snapshot artifact from one shared `charts.js`), a claude-mem **History** panel (opt-in, hidden when claude-mem isn't installed), and a "derived" marker on roster rows synthesized from delegation events (Codex work that has no real start record — dimmed gauges + tooltip, never claims token/cost data it doesn't have) — shipped in v6.8.3. The Commander Cockpit — every tool clickable inside one self-contained artifact: skill/agent browser with job-first search, catalog-grounded idea generator, live prompt enhancer, your own agent analytics (Claude + Codex side by side), an agent-profile manager with edit→export, and a skill-runs logger (42nd handler) so "most-used skill" is real data — shipped in v6.8.2. Claude *and* Codex side-by-side session tracking (`source_app` tagging, Source filter + pills, `sourceApp:name` agent keys) plus the Suggestions panel were added in v6.8.1. Mission Control itself — a visual harness inside Claude Code Desktop (live zero-dep dashboard + never-blocking event feed hook + CSP-safe snapshot artifact) plus a 5-signal proactivity wave — shipped in v6.8.0. Recent additions: /ccc-prompt-fix, claude-code-library, /ccc-spawn, /ccc-broadcast — see CHANGELOG for per-version detail. · npm package `cc-commander@7.4.1` (version tracked in `package.json`)
## Session Defaults
- **Model:** Opus 5 (1M context is default and max, no `[1m]` suffix needed) — enforced via `.claude/settings.json`
- **Mode:** Plan mode by default — SessionStart hook reminds to enter plan mode
- **Effort:** Anthropic's Opus 5 guidance: start at `high` (the default), drop to `low`/`medium` liberally wherever evals hold quality, step up to `xhigh` only for demanding coding/agentic work, `max` only when justified. Set `effortLevel` in `~/.claude/settings.json`. No project-level pin, so global governs — re-sweep any effort setting carried over from Opus 4.8, don't reuse blindly.
- **Thinking:** Summaries visible (`showThinkingSummaries: true`). Note: `thinking:{"type":"disabled"}` errors on Opus 5 at `xhigh`/`max` — only valid at `high` or below.
- **Footer:** Run `node commander/status-line.js` for live session status bar
- **Version:** Single source of truth is `package.json` — `branding.js` reads it at runtime
- **Fable 5 = deep mode:** For architecture, planning, migration, or threat-model work, escalate with `/model claude-fable-5[1m]`. Pay for Fable on the thinking, not the typing.
- **Subagent routing:** Auto-routed by complexity — Haiku → Sonnet → Opus → Fable. Main thread stays Opus until you escalate. Opus 5 delegates more eagerly than 4.8 and over-spawns without explicit caps — delegate to a subagent only for large, genuinely independent, parallelizable work (not things finishable in a handful of tool calls), and keep spawn counts low.
- **Claude Teams (HARD RULE):** Agent Teams is always on — `CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS: "1"` in `.claude/settings.json` env (and in the global `~/.claude/settings.json`). Multi-agent fan-out (`/ccc-fleet`, dynamic workflows) depends on it. Do not disable.
- **Motto:** "Pay for Fable on the thinking, not the typing."
- **Orchestrator/Executor doctrine:** `/ccc-orchestrate` (Fable/Opus plans → GPT-5.6 Sol via `codex` or a Sonnet subagent executes → Fable/Opus verifies), `/ccc-plan-exec` (same idea, Claude-only), `/ccc-handoff` (Matt Pocock-style context-reset — hand off to a fresh chat before quality degrades). Full doctrine: `CLAUDE.md.template` § "Orchestrator / Executor Model". Existing (non-CCC) projects adopt these rules via `/ccc-adopt`.
## Workflow-First (default)
The lead session is a **control plane** — decisions, delegations, and verified conclusions only.
- **Delegate substantive work:** multi-file, multi-step, research, audit, migration, or repo-wide tasks go through the **Workflow tool** — but only when the work is large, genuinely independent, and parallelizable. Agents read/search/build and return ONLY conclusions or structured results — never raw file dumps into the lead context.
- **Go solo (inline)** for a conversational reply, a single trivial edit, reading the one file you're about to edit, or anything finishable in a handful of tool calls — Opus 5 over-delegates without this cap, so don't spawn a subagent just because a task touches more than one file.
- **Keep context slim:** reference code as `path:line`, don't re-read files you just edited, route big tool output to disk.
- **Proactive compaction:** at ~70% context, write/refresh a session handoff doc; at ~85%, compact or hand off rather than risk truncation mid-task.
- Full doctrine: `commander/cowork-plugin/rules/workflow-first.md`.
## Critical Rule
**Research the codebase before editing. Never change code you haven't read.** Read files before modifying them. Understand the surrounding context. If you're about to edit a function, read the whole file first.
## Desktop Plugin (PRIMARY PRODUCT)
CC Commander's primary product as of v6.8.2. Claude Cowork Desktop / Claude Code Desktop install:
**Via GUI (Cowork Desktop / Code Desktop):**
1. Settings → Plugin Marketplace → **Add from GitHub**
2. Enter `KevinZai/commander` → Add
3. Find `commander` in the marketplace → click **Install**
**Via CLI (Claude Code terminal only — `/plugin` slash commands don't exist in Desktop):**
```
/plugin marketplace add KevinZai/commander
/plugin install commander
```
**Plugin name:** `commander` · **Marketplace:** `commander-hub` at KevinZai/commander
**v7.4.1:** 82 plugin skills (13 /ccc-* specialist workflows + 11 ccc-* domain routers + 7 channel/CI/ECC/setup skills + diagnostic/meta + vendor-sourced + lifecycle + /ccc-deploy + /ccc-rollback + /ccc-onboard + /ccc-fleet-viz + /ccc-changelog + /ccc-doctor + /ccc-upgrade + /ccc-tuneup + /ccc-orchestrate + /ccc-handoff + /ccc-adopt + /ccc-relay + /ccc-smb-ops + /ccc-broadcast + /ccc-spawn + /ccc-prompt-fix + /ccc-mission-control + /ccc-usage + /ccc-safety + /ccc-update + /ccc-console) + 22 specialist agents + 23 lifecycle hooks × 44 handlers (incl. PermissionRequest + SessionStart orchestrator + 5 ECC hooks ported) + 2 credential-free bundled MCP servers (context7 + sequential-thinking) + 16 opt-in via /ccc-connect — **core free forever**, sustained by transparent affiliate links in /ccc-connect + Kevin's consulting practice. Hosted MCP (v4.1) will be free with a 100-call/mo anti-abuse cap; hosted-infrastructure Pro tier planned later — all content stays free forever.
**The 12 /ccc-* specialist workflows (all click-first via AskUserQuestion native picker):**
- `/ccc` — main hub (6 intents via nested AUQ)
- `/ccc-start` — first-run onboarding + plan file
- `/ccc-browse` — searchable catalog of all skills + agents
- `/ccc-plan` — spec-first feature planning → plan file
- `/ccc-build` — scaffold projects (web / API / CLI / mobile)
- `/ccc-review` — branch audit (diff / security / perf / x-ray)
- `/ccc-ship` — pre-flight checks + release + deploy
- `/ccc-design` — UI/UX workflow (routes to ccc-design domain)
- `/ccc-learn` — skill discovery across 11 CCC domains
- `/ccc-xray` — project health scorecard with fix chips
- `/ccc-linear` — Linear board integration
- `/ccc-fleet` — multi-agent parallel orchestration
- `/ccc-connect` — opt-in MCP connector (Notion/Zapier/Supabase/Slack/GDrive/Figma)
**v4.1 (next):** hosted MCP (free with 100-call/mo anti-abuse cap; hosted-infrastructure Pro tier planned later — all content stays free forever) + Commander Hub marketplace + Anthropic Connectors directory submission.
Plugin directory: `commander/cowork-plugin/` — CONNECTORS.md documents all 10 connector categories.
### Architecture note (since v7.1.0)
Plugin commands are provided as **SKILLS** (not plugin commands) so they appear as plain `/ccc-*` in the autocomplete (no `commander:` namespace prefix). This is the correct primitive for click-first UX in Cowork Desktop. Menus live in `commander/cowork-plugin/menus/*.json` (root + 6 sub-menus); the template at `commander/cowork-plugin/lib/menu-artifact.html.tpl` and renderer at `commander/cowork-plugin/lib/menu-render.js` are kept for future use if/when Claude Desktop adds interactive-artifact support, but the active UX uses AskUserQuestion-native chip pickers (confirmed working in Cowork Desktop).
## CCC CLI
Interactive CLI that sits ABOVE Claude Code sessions. Manages, dispatches, and tracks AI work.
```bash
ccc # Launch interactive mode
ccc --test # 187 tests across 14 suites
ccc --stats # Quick stats
```
Key components:
- `commander/engine.js` — Main interactive loop with arrow-key menus
- `commander/tui.js` — TUI engine (figlet, gradients, 10 themes, spinners)
- `commander/dispatcher.js` — 14-flag Claude Code dispatch with plan-mode-first
- `commander/knowledge.js` — Knowledge compounding (learns from every session)
- `commander/plugins.js` — Auto-detects gstack, CE, Superpowers, sequences them
- `commander/adventures/*.json` — 14 JSON decision tree flows
- `commander/cowork-plugin/` — Desktop plugin (82 skills, 22 sub-agent personas, 23 hooks × 44 handlers, 2 credential-free bundled MCP servers per `.mcp.json` + 16 opt-in via `/ccc-connect`)
- `commander/update-check.js` — Update checker (4h cache, silent on failure, runs at session start)
- `commander/tests/paths.test.js` — 18 E2E path tests
State: `~/.claude/commander/` (never modifies `.claude/`)
Tests: `node --test commander/tests/paths.test.js`
## Project Structure
```
├── commander/cowork-plugin/ # Desktop plugin (primary product)
│ ├── .claude-plugin/plugin.json # Plugin manifest (schema-compliant)
│ ├── skills/ # 82 plugin skills (23 ccc-* workflows + 7 channel/CI/ECC/setup + remaining lifecycle + 2 vendor-sourced + 25 other + /ccc-tuneup + /ccc-update + /ccc-console)
│ ├── agents/ # 22 specialist agents (architect, reviewer, builder, designer, security-auditor, typescript-reviewer, python-reviewer, etc.)
│ ├── hooks/ # 23 lifecycle hooks (SessionStart [orchestrator], UserPromptSubmit, PreToolUse, PostToolUse, Stop, Notification, PreCompact, PostCompact, SubagentStop, SubagentStart, PermissionRequest, SessionEnd, Elicitation, ElicitationResult, StopFailure, PostToolUseFailure, PostToolBatch, TaskCreated, TaskCompleted, ConfigChange, UserPromptExpansion, InstructionsLoaded, Setup)
│ ├── menus/ # 7 menu JSON trees (root + 6 sub-menus)
│ ├── lib/ # Shared artifact scaffold (menu-render.js + template)
│ ├── .mcp.json # 2 credential-free bundled MCP server configs (context7, sequential-thinking) — 16 more opt-in via /ccc-connect
│ ├── CONNECTORS.md # 10 connector categories
│ └── rules/ # 17 persona voice files + common response-style
├── skills/ # 467 skills organized by category
│ ├── ccc-*/ # 11 CCC domains (router + sub-skills)
│ ├── mode-switcher/ # 10 workflow modes
�� ├── init-decision-tree/ # /ccc-init project wizard
│ ├── confidence-check/ # Pre-execution confidence assessment
│ ├── four-question-validation/ # Post-implementation verification
��� └── */ # Individual skills (SKILL.md each)
├── commands/ # 83 slash commands (.md files)
│ └── cc.md # /cc command center (interactive menu)
├── prompts/ # 36+ prompt templates across 6 categories
├── hooks/ # 24 kit-native hooks (JS) + hooks.json + hooks-standalone.json
├── templates/ # 3 starter templates (nextjs, api, cli)
├── lib/ # Terminal art (bash + JS) + statusline.sh
├── docs/ # GitHub Pages landing site
│ ├── index.html # Single-page marketing site
│ └── assets/ # CSS, JS, images
├── kevin/ # Kevin's personal overlay (not installed by default)
│ ├── install-kevin.sh # Layer Kevin-specific config after public install
│ ├── CLAUDE.md.kevin # Kevin's CLAUDE.md (full MCP, OpenClaw)
│ ├── settings.json.kevin # Kevin's settings (all MCP servers)
│ └── hooks/ # Kevin-specific hooks (Paperclip, OpenClaw)
├── compatibility/ # IDE guides + 4 iTerm2 color profiles (Claude Anthropic, OLED, Matrix)
├── tests/ # Hook test harness
├── install.sh # Interactive installer
├── install-remote.sh # One-line remote installer (curl | bash)
├── uninstall.sh # Clean removal with backup restore
├── BIBLE.md # The Kevin Z Method — 7 chapters + appendices
├── CHEATSHEET.md # Daily reference
├── SKILLS-INDEX.md # Searchable skill directory
├── CHANGELOG.md # Version history
└── CLAUDE.md.template # Staff CLAUDE.md template
```
## Aggregator Ecosystem
19 vendor submodules in `vendor/`. Smart orchestrator scores tools: capability 50% + stars 15% + recency 15% + user pref 20%.
Vendor scanner builds capability index across 8-phase pipeline.
Auto-updates weekly via GitHub Actions.
Key vendors: ECC (120K stars), gstack (58K), Superpowers (29K), oh-my-claudecode (17K),
claude-code-best-practice (26K), repomix (22.8K), Claude HUD (15K), RTK (14.6K),
Compound Engineering (11.5K), claude-skills (8.6K),
claude-code-ultimate-guide (2.7K), acpx (1.8K), caliber (300), claude-reflect (860),
claude-code-prompts (142), MengTo/Skills (web design + prompting, 100★).
**Note:** `claude-mem` was previously vendored but was removed — it ships under AGPL-3.0 which is incompatible with CC Commander's MIT license. Users can install `claude-mem` separately via `npm install claude-mem` or `/ccc-connect`; it is treated as an external opt-in MCP, not a bundled vendor.
## Footer Bar
CCC renders a rich status footer on every menu screen and recommends it for `/ccc` mid-session:
```
━━ CCC7.3.0│🔥Opus1M│🔑gAA│🧠▐██45%░░▌│⏱️▐██░░▌6%│📅▐██░░▌34%│💰$2.34│⬆️640K⬇️694K│⏰8h0m│🎯467│📋CC-150│📂~/project
```
Render this at session start and after major actions. Source: `commander/cockpit.js`.
## Nested Command Menus
All `/ccc` sub-commands open rich menus with recommendations:
| Command | Menu | What it does |
|---------|------|-------------|
| `/ccc` | Main menu | 15 options — build, review, learn, infra, settings |
| `/ccc build` | Build wizard | Website, API, CLI, custom — with spec questions |
| `/ccc infra` | Infrastructure | Fleet, Cost, Synapse, AO, CloudCLI, Paperclip |
| `/ccc linear` | Linear board | View issues, pick to build, create new |
| `/ccc skills` | Skill browser | 467 skills by category, preview, try |
| `/ccc domains` | CCC domains | 11 ccc-* domains |
| `/ccc night` | Night/YOLO mode | Autonomous overnight builds |
| `/ccc settings` | Settings | Name, level, cost, theme, Linear setup |
Every menu recommends the best next action based on context.
## CCC Domains
| Domain | Sub-Skills | Focus |
|--------|-----------|-------|
| `ccc-design` | 41 | UI/UX, animation, responsive, accessibility |
| `ccc-marketing` | 45 | CRO, email, ads, social, content |
| `ccc-saas` | 20 | Auth, billing, multi-tenant, schema |
| `ccc-devops` | 20 | CI/CD, Docker, AWS, monitoring |
| `ccc-seo` | 20 | Technical SEO, content, Core Web Vitals |
| `ccc-testing` | 15 | TDD, E2E, coverage, regression |
| `ccc-security` | 8 | OWASP, pen-test, secrets, hardening |
| `ccc-data` | 8 | SQL, ETL, analytics, visualization |
| `ccc-research` | 8 | Competitive, market analysis |
| `ccc-mobile` | 8 | React Native, Expo, mobile patterns |
| `ccc-makeover` | 3 | /xray audit + /makeover swarm |
## Current Status
**Version:** 7.4.1 (Desktop-first plugin, Claude Agent SDK sub-agent architecture — see `package.json`)
**Stats:** 467 skills total, 82 plugin skills, 22 sub-agent personas, 23 lifecycle hooks (44 handlers), 2 bundled MCP servers + 16 opt-in, 83 commands, 24 JS hooks, 19 vendors, 10 themes
**Live:**
- GitHub repo: github.com/KevinZai/commander
- Marketplace: `commander-hub` — `/plugin marketplace add KevinZai/commander`
- Landing page: KevinZai.github.io/cc-commander
- One-line install: `curl -fsSL https://raw.githubusercontent.com/KevinZai/commander/main/install-remote.sh | bash`
- Core free forever: all skills, all agents, all hooks, all MCP servers — sustained by affiliate links + Kevin's consulting
## Key Commands
```bash
./install.sh # Interactive install (matrix rain, ASCII art, progress)
./install.sh --dry-run # Preview without changes
./install.sh --verify # Validate existing installation
./install.sh --force # Skip confirmation prompts
./uninstall.sh # Remove kit components (preserves CLAUDE.md + settings.json)
ccc --split # Tabbed tmux mode
ccc --dispatch "task" --json # Headless agent dispatch
ccc --list-skills --json # Skill catalog for agents
ccc --status # Health check
```
## Development Notes
- Shell scripts source `lib/terminal-art.sh` for all visual output
- Hooks use `lib/terminal-art.js` for consistent branding
- All animations respect `CC_NO_COLOR=1`, `CC_NO_ANIMATION=1`, CI detection
- OG image generated via `docs/assets/og-image.svg` → PNG conversion
- CCC domains use router pattern: one SKILL.md dispatches to sub-skill directories
- Kevin's personal overlay lives in `kevin/` — not installed by the public installer
- Tests run via `node --test tests/hooks.test.js` (Node.js built-in test runner, 61 tests)
- 24 kit-native hooks form the "Proactive Automation Suite" — context-guard, auto-checkpoint, cost-alert, confidence-gate, pre-compact, self-verify, openclaw-adapter, status-reporter, openclaw-sync, etc. (session-coach retired 2026-07-10 — folded into the plugin suggest engine)
- 10 workflow modes via mode-switcher skill — normal, design, saas, marketing, research, writing, night, yolo, unhinged, caveman
- 36+ prompt templates across 6 categories (coding, planning, design, marketing, devops, meta)
- Agency Orchestrator and OpenClaw patterns for multi-agent integration
- Claude Peers integration — multi-instance collaboration with coordinator/swarm/expert/review/research patterns
- Spawn Manager — `/spawn` command for launching and coordinating multiple Claude Code peers
- Task Commander — multi-agent orchestration with P0-P10 scoping, 6 DAG workflows, circuit breaker, cost ceiling
- Contextual Assist — proactive suggestion system with 5 configurable levels
- Cowork + Dispatch integrations for Claude Desktop autonomous mode and background tasks
- OpenClaw Bridge + Paperclip Bridge for platform integrations
- 5 quickstart guides for beginner/frontend/backend/fullstack/mobile developers
- Vanilla dashboard at `dashboard/` — localhost session inspector (port 4690), reads real `~/.claude/sessions/` data, zero deps, no build step. React app archived to `dashboard/_archive/` (was demo-data only).
- `lib/config-reader.js` — shared config utility for bible-config.json with env var overrides
- VS Code snippets at `compatibility/vscode-snippets.json` (20+ snippets)
- Status line (`lib/statusline.sh`) shows live context gauge, model, cost, tokens, account, rate limit countdown
- `/ccc-init` checks `~/.claude/sessions/` and offers to resume prior sessions before the wizard
- 4 iTerm2 color profiles at `compatibility/` (Claude Anthropic, OLED Black, Matrix); legacy `kz-matrix.itermcolors` preserved
- Theme system with 10 switchable themes
- OpenClaw native integration with auto-detection, skill sync, bidirectional event forwarding
- Status update requests for sending progress reports during long sessions
- Continuous improvement pipeline with daily cron scan and proposal queue
- Modular installer with 5 installation modes (full, essentials, scripts, dashboard, config-only)
- Daemon Mode: commander/daemon.js (tick loop), commander/queue.js (task queue), commander/dream.js (knowledge consolidation)
- Intelligence Layer v1: complexity scoring (dispatcher.js), stack detection (project-importer.js), session learning (knowledge.js), skill filtering (skill-browser.js), smart retry (dispatchWithRetry)
- Tiered skill loading via `skills/_tiers.json` — essential (~30), recommended (~100), domain (11 routers), full (all 500+)
- Update checker at `commander/update-check.js` — 4h cache, silent on failure, fires on session start
- Caveman mode (`caveman` skill) — strips markdown/emojis/prose for ~75% output token savings during iteration
## Ecosystem Context
CCC ships 82 focused `/ccc-*` plugin skills as the curated front door, then routes into 467 CCC ecosystem skills across 11 domains. It is a meta-layer over the broader Claude skill ecosystem — not a replacement for it.
**External ecosystems CCC plays well with:**
- `anthropics/skills` — official Anthropic skills (Frontend Design, Theme Factory, Spec Writer)
- `mattpocock/skills` — 15K stars, TypeScript-first (Grill Me, Ubiquitous Language)
- `skillsmp.com` — 66K+ community marketplace
**Install format alternatives:**
| Method | Command | When to Use |
|--------|---------|------------|
| Marketplace (recommended) | `/plugin install commander` | Full plugin with all 82 skills + MCPs + hooks |
| Full CLI | `curl -fsSL https://raw.githubusercontent.com/KevinZai/commander/main/install-remote.sh \| bash` | CLI mode, non-plugin installs |
| Cherry-pick | `npx skills@latest add KevinZai/commander/skills/<name>` | Individual skills via the skills CLI |
**When Claude agents load this CLAUDE.md, they should know:**
- CCC is a meta-layer over the Claude skill ecosystem — curation + guidance + memory, not raw skills
- 82 `/ccc-*` plugin skills are the curated front door; they route into 467+ deeper skills
- `CONNECTORS.md` defines 13 tool-agnostic connector categories via `~~category` placeholders — skills are tool-agnostic
- Everything core free forever: 82 plugin skills, 22 agents, 2 credential-free bundled MCP servers (context7 + sequential-thinking) + 16 opt-in via /ccc-connect, 23 lifecycle hooks × 44 handlers. No feature gating, no paywalls, no license checks.
- Revenue model: core free forever — 4 levers: (1) transparent affiliate links in /ccc-connect (Supabase, Vercel, Neon, etc.) + (2) Kevin's consulting pipeline (plugin = trust engine → consulting funnel) + (3) optional Pro community ($49/mo Discord) + (4) GitHub Sponsors
## UI Components
Check `~/clawd/shared/refs/shadcn-ecosystem.md` before building custom. Key: tremor (charts), auto-form (Zod→forms), magicui (animations).
- **Tech icons:** `developer-icons` (npm) — standard for all projects. Never custom SVGs for tech logos.