Ccc

/ccc — CC Commander Hub

Click-first entry point to the whole CC Commander surface (<N> ccc-* skills — never hardcode, compute via ls -d ${CLAUDE_PLUGIN_ROOT}/skills/*/ | wc -l — 22 specialist agents, 2 credential-free bundled MCP servers + 16 opt-in via $ccc-connect). The user types /ccc and gets a native visual picker — no typing, no menus.

Workflow-first. CC Commander delegates substantive work to the Workflow tool / fleet; the lead session stays a thin control plane (conclusions in, file-dumps out). See ../../rules/workflow-first.md.

Response shape (EVERY time)

Output exactly four sections in order: 0) onboarding gate, 1) ASCII hero banner, 2) context strip, 3) picker (AskUserQuestion), 4) dispatch on selection. Sections 1-3 are rendered up front; section 4 fires after the user clicks.

0. Onboarding gate (MANDATORY first check)

Run this bash check at the very top of the skill execution:

STATE=~/.claude/commander/state.json
if [ ! -f "$STATE" ] || ! grep -q '"onboardingCompleted": *true' "$STATE" 2>/dev/null; then
  echo "ONBOARDING_NEEDED"
fi

If output contains ONBOARDING_NEEDED, respond with ONLY this (no picker, no context strip):

👋 Welcome to CC Commander — first time? Let me walk you through the basics.

Dispatching $ccc-start…

Then dispatch the ccc-start skill immediately. Do NOT show the main /ccc picker on first run.

If state.json shows onboardingCompleted: true, skip this step and proceed to step 1.

1. ASCII hero (read banner.txt + interpolate version)

Read ${CLAUDE_PLUGIN_ROOT}/lib/banner.txt via Read tool. Replace {{VERSION}} with the version string from ${CLAUDE_PLUGIN_ROOT}/.codex-plugin/plugin.json. Render in a fenced code block (```text ... ```), NO trailing prose on the same block.

This replaces the markdown one-liner header.

2. Context strip (one paragraph, markdown)

Detect project context with four quick reads (cheap, silent on failure):

• git rev-parse --abbrev-ref HEAD → current branch
• Presence of CLAUDE.md in CWD → project recognition
• Presence of tasks/todo.md → active work
• ls -t ~/.claude/sessions/*.tmp 2>/dev/null | head -1 → recent session

Render a one-line summary:

🧭 Context: <branch> · project: <name from CLAUDE.md first line or CWD basename> · open tasks · last session:

If no context detected (not in a project): "🧭 No project detected — pick an intent below, I'll set one up."

3. The picker — AskUserQuestion with 4 intents

Never render a numbered list, never tell the user to "type a number". Always call AskUserQuestion with these 4 options. The preview field gets rich context per option.

question: "What's the move?"
header: "CC Commander"
multiSelect: false
options:
  - label: "🔨 Build something"
    description: "Scaffold a project or feature — web, API, CLI, mobile."
    preview: "(per-context note — see rules below)"
  - label: "🔍 Review my work"
    description: "Audit current branch, security, perf, or run full x-ray."
    preview: "(per-context note)"
  - label: "🚀 Ship it"
    description: "Pre-flight checks + deploy. Tests, tag, push, release."
    preview: "(per-context note)"
  - label: "⋯ More tools"
    description: "Plan, design, learn, x-ray, Linear, fleet, connect apps."
    preview: "<N> more skills — click to expand (N = total plugin skill count minus the 3 shown above; compute via `ls -d ${CLAUDE_PLUGIN_ROOT}/skills/*/ | wc -l`, never hardcode)"

Recommendation logic (prepend ⭐ to the label of ONE option based on context):

• Branch has uncommitted changes + recent diff → ⭐ "Review my work"
• Tests green + branch ahead of main → ⭐ "Ship it"
• Empty repo OR no CLAUDE.md → ⭐ "Build something"
• Session recent and file changes in tasks/ → context-specific preview text

4. Handle the selection

Based on user pick, invoke the matching skill. Do NOT prompt again; dispatch immediately:

• Build → invoke ccc-build skill
• Review → invoke ccc-review skill
• Ship → invoke ccc-ship skill
• More → invoke ccc-more skill (which presents the second-page picker)

Each target skill handles its own sub-picker. Never unroll the full skill catalog in one /ccc call — that's anti-pattern (burns context, overwhelming UX).

Anti-patterns — DO NOT do these

• ❌ Render a numbered list "1. Build, 2. Review, ..." and tell the user to type a number
• ❌ Output raw HTML in fenced code blocks expecting it to render as an interactive artifact — Cowork Desktop shows HTML as literal code
• ❌ Render more than 4 options in a single AskUserQuestion — not supported (max 4)
• ❌ Load references/main-menu.json and dump 18 options — nested flow is required
• ❌ Reference legacy /commander:ccc namespace — the plugin now ships /ccc as a plain skill
• ❌ Hardcode skill/agent counts — compute them live via ls -d ${CLAUDE_PLUGIN_ROOT}/skills/*/ | wc -l (same rule as ccc-start/SKILL.md)

Dispatching after selection

When the user picks an option, use the SendMessage pattern by invoking the target skill inline. Example for "Build":

Loading the build workflow — ccc-build routes you to web / API / CLI / mobile / from-spec with one more click.

Then proceed as if the user had invoked $ccc-build directly. Read ${CLAUDE_PLUGIN_ROOT}/menus/ccc-build.json for its options and call AskUserQuestion with them.

When user passes an argument

If /ccc build → skip the root picker, invoke ccc-build directly. If /ccc review / ship / design / more → same. If /ccc <anything else> → show the root picker with the argument echoed in the context strip: "🧭 You said: <arg> — picking an intent below to route you."

Plugin detection (for context strip)

Scan quickly (parallel Bash calls with timeouts):

• ls ~/.claude/skills/ | wc -l — skill count
• claude mcp list 2>/dev/null | grep -c '^ ' — connected MCPs

If either >0, include in context strip: "🧰 X skills · Y MCPs connected".

Brand rules

• Always read VERSION from plugin.json — never hardcode.
• Emoji-forward, concise (PM Consultant voice): decision up front, reasoning terse. Use the semantic emoji palette from rules/common/response-style.md (🎯 focus · 💡 idea · 🟢 approve · 🟡 caution · 🔴 block · ⏭️ next) — emoji are anchors, not decoration; recommendations close with 🟢 my call + one-line rationale.
• Never mention the CLI (ccc npm binary) in /ccc flow — this is the Desktop-plugin audience.
• Recommended option gets ⭐ and a one-line "Recommended because …" in its preview field.

Tips for the agent executing this skill

1. The whole flow is ≤4 Claude turns: banner+picker → user clicks → dispatch → target skill runs. Don't overthink.
2. If Cowork Desktop caches the marketplace and the version strip shows an old version, that's a client cache issue — not our problem. Note it in the context strip: "⚠️ Plugin v detected; latest on GitHub is — click Update in marketplace".
3. For ls / git / file-check operations, run in parallel in a single Bash call (cmd1 && cmd2 && cmd3) to save turns.

Bottom line: three elements — header, context, picker — in that order. User clicks. We route. No typing, no text menus, ever.

⚙️ Fable contract: plan before build · verifier ≠ worker · prove before alarm · loops need gates · leave durable state — rules/fable-method.md

(On Codex, present these options as a numbered list and ask the user to reply with a number — AskUserQuestion is Claude-only.)