Claude Code Pro ⚡

April 25, 2026 · View on GitHub

Token-efficient Claude Code workflow for AI agents.

The Problem

Most Claude Code supervision skills poll tmux every 30-60 seconds, reading hundreds of lines each time. A 20-minute task burns 10,000-20,000 tokens just on supervision.

The Solution

Start task (with callback) → Wait (zero tokens) → 📩 Notification → Read result once

The task itself signals completion via openclaw system event. Your agent sleeps until notified. 80-97% token savings.

Install

# Via ClawHub (for OpenClaw agents)
clawhub install claude-code-pro

# Or clone
git clone https://github.com/voidborne-d/claude-code-pro

Usage

Start a task

bash scripts/start.sh --label auth-refactor --workdir ~/project \
  --task "Refactor auth to use JWT.

When completely finished, run: openclaw system event --text \"Done: JWT refactor\" --mode now"

Monitor (only when needed)

bash scripts/monitor.sh --session auth-refactor --lines 50
bash scripts/monitor.sh --session auth-refactor --json

Send follow-up

bash scripts/send.sh --session auth-refactor --text "Also add tests"
bash scripts/send.sh --session auth-refactor --compact

Manage sessions

bash scripts/list.sh              # human-readable
bash scripts/list.sh --json       # structured
bash scripts/stop.sh --session auth-refactor
bash scripts/stop.sh --all

Attach (human access)

tmux -L cc attach -t cc-auth-refactor

Smart Dispatch

Don't start Claude Code for everything:

Task	Action
< 3 files, simple fix	Just edit directly
Multi-file refactor	✅ Start CC
New feature (5+ files)	✅ Start CC
Needs exploration	✅ Start CC

Token Comparison

Approach	20-min task	Supervision tokens
Poll every 30s	40 reads	~20,000
Poll every 60s	20 reads	~10,000
claude-code-pro	1 read	~500

Modes

Mode	Flag	Behavior
auto (default)	`--mode auto`	`--dangerously-skip-permissions` — full autonomy
plan	`--mode plan`	Asks permission before file changes

⚠️ auto mode skips all permission prompts. Use only in trusted environments with version-controlled code.

Design

Isolated socket (tmux -L cc) — doesn't touch your tmux sessions
cc- prefix — all sessions are namespaced
Safe paste — uses tmux load-buffer + paste-buffer for multi-line prompts
No jq dependency — JSON output via pure bash/python3 fallback
Zero external deps — just tmux + claude CLI

After CCP orchestrates a long task, the next question is whether the output is mergeable. sober-coding is the post-generation pair: 27 checks across 7 dimensions (placeholder TODOs, fabricated APIs, AI-noise comments, scope bloat, FP-prone patterns) targeted at the failure modes that show up because the code was AI-drafted — orthogonal to what ESLint/SonarQube/ruff already cover.

Typical loop:

claude-code-pro --task "..."  →  completion callback fires
                              →  sober-coding lints the diff
                              →  you read it once, with the noise already filtered

Same workshop: also see humanize-chinese (Chinese AIGC detector + rewriter) and lambda-lang (agent-to-agent compression).

License

MIT

Part of Voidborne · Available on ClawHub