Delegate long-running coding work to an inner CC session running in tmux, while the outer agent stays focused on supervision, monitoring, and end-to-end verification. Use when the user asks to run a long task, offload implementation, keep working while an inner coding agent executes, or use cc-use.
Detected risks:
cc-use lets an outer coding agent delegate long-running implementation work to an inner CC (coding command-line agent) session running in tmux, while the outer session stays focused on supervision, monitoring, and end-to-end verification. Here "CC" can mean Claude Code, Codex CLI, or another compatible coding CLI depending on the host and local configuration. The skill solves the problem of a single conversation losing local detail, polluting its context with logs, or drifting from the top-level objective on long tasks: the inner session handles bounded, focused execution while the outer session preserves a compact supervisory context and performs final judgment. The intent is not just parallelism but supervision, keeping the plan, product intent, risks, compatibility concerns, and acceptance criteria in view.
The skill drives an inner session through the scripts/cc-use helper, which exposes commands including delegate (start or reuse the inner session, send one short request exactly as written, and wait for one observation), monitor (observe an existing session and wait for one observation), and project-status (check derived project/session status without sending input, with an optional --json flag). Key options include --project DIR (sets the tmux working directory, derived session name, and .cc-use/state/ location), --agent codex|claude (match the outer agent family; do not cross-delegate), --session NAME, --profile NAME (Codex only), --initial-quiet-seconds N (default 30), --poll-interval N (default 2), and --replace. Monitoring is based on screen stability rather than parsing agent-specific UI; when the screen stays quiet long enough, a snapshot is saved and an inspect observation is emitted, or session_unavailable if the tmux session is gone. For Codex, the inner agent is launched with --dangerously-bypass-approvals-and-sandbox.
cc-use is aimed at engineers running long tasks, task decomposition and implementation from a plan or design document, adversarial and end-to-end testing, and interactive product testing of TUI products, skills, MCP servers, or plugins. The recommended workflow is to break the request into small, concrete steps, send each unchanged, wait for an inspect observation, read the snapshot semantically, then decide to wait, steer, ask the user, or verify. Requirements and constraints include tmux, matching the inner agent family to the outer one, and passing --project "$PWD" when the outer session may change directories.
The scripts/cc-use helper provides delegate (start or reuse an inner session, send one short request, and wait for an observation), monitor (observe an existing session and wait for one observation), and project-status (check the derived project/session status without sending input, with an optional --json flag).
Use --project DIR to set the tmux working directory, derived session name, and .cc-use/state/ location (it defaults to the current directory), and --agent codex or --agent claude to match the outer agent family. Passing --project "$PWD" is recommended when the outer session may change directories.
It monitors by screen stability rather than parsing agent-specific UI. When the screen stays unchanged for the quiet window (--initial-quiet-seconds, default 30, polled every --poll-interval seconds, default 2), it saves a snapshot and emits an inspect observation; if the tmux session is gone it emits session_unavailable.
cc-use launches the inner Codex with --dangerously-bypass-approvals-and-sandbox, the Codex equivalent of Claude Code's --dangerously-skip-permissions, matching cc-use's intent of running unattended automated work. The older --sandbox and --approval flags are retained for backward compatibility but no longer affect the inner session.
No. The skill directs you to keep task decomposition in the outer session and break the request into small, concrete steps, sending each inner request exactly as written without wrapper text or added role/policy instructions, then inspecting and steering between steps.
Quick Setup:
.claude/skills/Repository
zc277584121/cc-use