ccmeta

Give your agent the keyboard.

Claude Code meta commands (/clear, /compact, /model, /loop, …) normally need a human finger on the prompt bar. ccmeta hands that keyboard to Claude itself — a tiny tmux-powered Bash tool that lets the agent drive its own session.

Claude Code gates these commands behind a human keystroke by default; ccmeta removes that gate. Use at your own risk.

Both ship as matching Agent Skills. The ccmeta skill nudges Claude to reach for session-management commands proactively; the ccmeta-dangerous skill teaches it to reserve high-impact commands for tasks that genuinely need them, and to confirm before anything destructive.

Why?

I wanted to experiment with fully autonomous execution, for this you need primitives like /clear, /model, /effort, etc.

Quick start

Prereqs: tmux, Claude Code running inside a tmux pane.

As an Agent Skill (recommended):

npx skills add Alfredvc/ccmeta

Now you can ask claude to /clear itself, change its model, etc.

How it works

Claude ──Bash──▶ ccmeta.sh / ccmeta-dangerous.sh ──tmux send-keys──▶ Claude's own prompt bar

1. Claude calls ccmeta /clear -- start the next task via its Bash tool.
2. The script finds the tmux pane (via $TMUX_PANE, fallback: PPID tty lookup).
3. It backgrounds a subshell, waits 0.5 s for the current Bash call to return, then types the command and Enter into the pane.
4. Claude sees a fresh prompt and executes it like a human typed it.

Both scripts share ccmeta-core.sh for all logic. The only difference is which commands each exposes.

No API, no daemon, no hooks. Just tmux.

---

Usage

ccmeta <cmd> [args...] [-- prompt]
ccmeta-dangerous <cmd> [args...] [-- prompt]
ccmeta -h | --help

• Tokens before -- → the meta command and its arguments (joined with spaces).
• Tokens after -- → a separate follow-up prompt, sent as a new user turn after the command fires.

Commands that take a prompt-like argument (/compact [instructions], /loop [interval] [prompt], /plan [description], …) want it as args, not after --. Use -- when you want to chain a fresh prompt after a no-arg command like /clear.

Examples

# Low-risk (ccmeta)
ccmeta /clear
ccmeta /clear -- take the next task and complete it
ccmeta /model opus
ccmeta /effort max
ccmeta /fast on
ccmeta /rename nightly refactor run
ccmeta /compact focus on the API layer

# Requires ccmeta-dangerous
ccmeta-dangerous /loop keep fixing failing tests until green
ccmeta-dangerous /loop 5m check if the deploy finished
ccmeta-dangerous /remote-control
ccmeta-dangerous /batch migrate all React class components to hooks

---

Supported commands

Smaller blast radius (ccmeta)

These commands manage context, sessions, and model settings. Narrower scope than ccmeta-dangerous — no unattended pushes, no cloud calls, no auth changes.

Command	Arity	Keywords	Notes
/clear	0	—	start a new conversation with empty context
/compact	*	—	optional focus instructions for compaction
/debug	*	—	optional description of the issue to focus on
/desktop	0	—	continue session in the Desktop app (macOS/Windows)
/effort	0-1	low medium high xhigh max auto	max is session-only; available levels depend on the model
/exit	0	—	quit the CLI
/export	0-1	—	optional filename; otherwise opens a dialog
/fast	0-1	on off	toggle fast mode
/model	0-1	—	model name; no arg opens a picker
/plan	*	—	optional task description to start plan mode with
/recap	0	—	one-line summary of the current session
/reload-plugins	0	—	reload active plugins without restarting
/rename	*	—	session name (tokens joined); no arg auto-generates one
/resume	0-1	—	session id or name; no arg opens picker
/review	0-1	—	optional PR ref; local review
/security-review	0	—	analyze pending changes for security issues

Dangerous (ccmeta-dangerous)

These commands can execute autonomously, make persistent changes, incur cloud costs, or affect external systems. Think before granting ccmeta-dangerous in your Bash permissions.

Command	Arity	Notes
/autofix-pr	*	pushes code to remote unattended; optional instructions
/batch	1+	bulk parallel code changes across the repo; opens multiple PRs
/heapdump	0	writes a JS heap snapshot to ~/Desktop; may contain session data
/logout	0	signs out; session dies
/loop	*	unbounded autonomous execution; [interval] [prompt]
/remote-control	0	opens session to external access from claude.ai
/remote-env	0	configures default remote environment; persists across sessions
/schedule	*	creates persistent unattended scheduled tasks
/simplify	*	multi-agent code rewrites across multiple files
/ultraplan	1+	cloud session; can trigger downstream code execution
/ultrareview	0-1	deep cloud review; sends diff externally; costs money

ccmeta-dangerous also includes every command in the smaller-blast-radius set above.

Arity legend

Symbol	Meaning
0	no args
0-1	optional single arg
1+	one or more args (joined with spaces)
*	zero or more args (joined with spaces)

Each SPEC heredoc is the single source of truth for its script's allowed commands. --help and validation both derive from it — add a row, both update.

---

Security

ccmeta types into the running Claude Code session — any command on the allowlist will execute without you pressing Enter. That gate exists by default for a reason; removing it is a trade-off you're making.

• Two tiers by design. Give Claude ccmeta by default. Only add ccmeta-dangerous to the Bash allowlist when you explicitly want autonomous, cloud, or auth commands available. The split narrows blast radius; it is not a safety claim.
• The allowlist exists so a hallucinated /rm-rf-everything fails loudly instead of running.
• -- separator stops a malicious prompt from injecting extra flags into the command.
• Installing any Agent Skill (this one included) drops code into your agent's skill dir. Read all three scripts before adding. They're ~200 lines combined.
• Keep Claude Code's Bash permissions strict if you run third-party skills.

Do not install skills you have not read. Your agent will. Use at your own risk.

---

Development

ccmeta-core.sh is duplicated into each skill dir so every skill stays self-contained (Agent Skill installs don't guarantee files outside a skill dir ship). The canonical copy lives in shared/; the ones under skills/*/ are generated mirrors.

Always edit shared/ccmeta-core.sh, never the copies — any direct edit to a copy is clobbered on next sync.

After cloning, point git at the checked-in hook once:

git config core.hooksPath hooks

Now every commit runs scripts/sync-shared.sh and stages any regenerated copies automatically. To re-sync without committing, just run the script:

./scripts/sync-shared.sh

To share another file across skills, drop it in shared/ and add one line to SYNC_MAP in scripts/sync-shared.sh:

"my-shared.sh:ccmeta,ccmeta-dangerous"

Format: <filename-in-shared>:<comma-separated-skill-dirs>.