claude-delegate

name	claude-delegate
description	Delegate a coding task to Claude Code as a background sub-agent.

You are the Codex skill wrapper for the cc dispatcher.

Resolve <plugin-root> as the parent directory of the skills/ directory that contains this file (so <plugin-root>/scripts/cc.mjs is the dispatcher). Confirm <plugin-root>/scripts/cc.mjs exists before running.

Run:

node "<plugin-root>/scripts/cc.mjs" delegate -- "<task prompt>"

Forwarded flags go before --; everything after -- is the prompt verbatim. Example:

node "<plugin-root>/scripts/cc.mjs" delegate --model claude-opus-4-8 --name audit -- "<task prompt>"

Return the dispatcher's stdout verbatim. If the command exits non-zero, show stderr/stdout to the user and explain that the dispatcher failed. Do not reimplement the command logic yourself.

Behavior rules:

• Treat the user's remaining text after the skill invocation as the task prompt. If empty, ask the user for the task prompt before running.
• Forward only these flags when the user explicitly requests them: --name, --model, --effort, --permission-mode, --add-dir, --bypass-permissions, --dangerously-skip-permissions, --allow-dangerously-skip-permissions, --mcp-config, --agent, --agents, --allowedTools, --allowed-tools, --disallowedTools, --disallowed-tools, --tools, --settings, --setting-sources, --strict-mcp-config, --append-system-prompt, --system-prompt, --plugin-dir, --plugin-url, --bare, --safe-mode, --ide, --chrome, --no-chrome, --disable-slash-commands, --exclude-dynamic-system-prompt-sections, --verbose, --allow-edit, --json.
• --name is a human-readable label, not a session key. Every job gets a unique session name regardless: the dispatcher appends a random suffix, so the real session name is <your-name>-<random> (since v0.3.4). This means reusing the same --name does NOT resume a job — it starts a fresh, isolated session. To continue an existing job, use $claude-followup <jobId> (the jobId from delegate's output), never a reused --name. Omitting --name is fine; the auto-generated name is already unique and collision-safe even under heavy concurrency.
• The --effort flag accepts low, medium, high, xhigh, or max (Claude CLI valid set). The ultracode value is TUI-only and is silently ignored when passed via --effort. To trigger Claude Code's auto-orchestration workflow planning, use $claude-workflow instead — it injects the ultracode: keyword that activates the same behavior.
• The user may also pass --yes to skip the first-run privacy acknowledgement. Do not inject --yes for ordinary jobs. Exception: if the user explicitly asks for trusted unattended Claude work and you forward --bypass-permissions, --dangerously-skip-permissions, or --permission-mode bypassPermissions, include --yes in the same command so the job stays inside the cc status/result surface instead of falling back to native claude. If the dispatcher reports that an acknowledgement is required in any other case, surface that message to the user instead of retrying with --yes.
• For unattended local QA lanes that intentionally inspect the repo with shell commands, the user may explicitly request --bypass-permissions, --permission-mode bypassPermissions, or the Claude Code alias --dangerously-skip-permissions. Treat an explicit user preference for trusted unattended Claude work in the current task/session/project as sufficient permission to use --bypass-permissions for future fresh local shell/tool automation jobs. Do not infer that preference from ordinary delegation requests. The dispatcher translates these aliases to Claude Code's literal --dangerously-skip-permissions flag. If a bypass-launched job still needs interactive input immediately, the dispatcher exits non-zero and marks the job failed instead of returning a blocked worker. If a delegated job later shows a permission wait, report the blocked state and the claude attach <shortId> command from $claude-status --job <jobId> --json --compact; do not try to approve the permission prompt yourself.
• For real Chrome browser work, forward --chrome only when the user explicitly requests real Chrome access. There is no --real flag. Chrome access is Claude Code's real-browser/extension path, not Codex's in-app browser. If Claude reports a browser-selection, passkey, or permission/input prompt, use $claude-status --job <jobId> --json --compact and surface its waiting object plus claude attach <shortId> action. Do not choose among multiple Chrome browsers or approve user-gesture prompts from the background wrapper.

$claude-delegate -- "/<saved-or-bundled-workflow-name> <args>" also invokes saved workflows and bundled slash commands (e.g. /deep-research <question>). See $claude-workflow for details on saved-workflow invocation.

Next steps

After delegating, the user typically wants one of:

• $claude-status — check live progress of all open jobs
• $claude-wait <jobId> --json --compact --timeout 5m — wait for this job to produce a final result, blocked state, or timeout
• $claude-result — read the final answer when the job completes
• $claude-result <jobId> --partial — read recorded partial output from a running or permission-blocked job
• $claude-followup — send an additional instruction to the same job
• $claude-stop — terminate the job early

Fan-out tip

For large parallelizable tasks, phrase the prompt to request explicit parallel decomposition (e.g. "use parallel subagents, one per directory"). See the ## Subagent fan-out patterns (Codex → Claude Code) section of the plugin README for a decision matrix and example prompts. If the task has multiple dependent phases, prefer $claude-workflow instead.