Delegate Model

A plugin that lets a host agent delegate scoped tasks to local peer agents — Grok, Codex, opencode, and Claude Code — through one tested Node wrapper, instead of re-deriving CLI flags and output parsing on every call.

What it is

Five first-class skills (grok, codex, opencode, claude, and a router delegate) plus scripts/delegate.js for those CLIs, and a sixth skill acp plus scripts/acp.js for Agent Client Protocol agents (--agent "grok agent stdio", opencode acp, …). Shared helpers live in scripts/lib.js. Every launch file prints one JSON envelope.

Prerequisites

Each backend CLI must be installed and authenticated separately — the plugin does not install or authenticate anything:

Live-smoked ACP agents: grok agent stdio, opencode acp. Others (gemini --acp, codex-acp, …) may work if the CLI is installed; they are not claimed until smoked.

For local models (LM Studio, Ollama) via opencode, the local server must be running before you delegate to it.

Install

Claude Code

Packaged path. Sets CLAUDE_PLUGIN_ROOT to this plugin:

claude plugin marketplace add dvrtech-us/delegate-model
claude plugin install delegate-model@delegate-model

Local checkout: claude --plugin-dir ./plugins/delegate-model, then /reload-plugins.

Grok TUI

Load plugins/delegate-model/skills/ (this repo, or the Claude plugin cache). Skills call node "${CLAUDE_PLUGIN_ROOT}/scripts/delegate.js" and …/acp.js. Grok must resolve CLAUDE_PLUGIN_ROOT to the plugin root (…/plugins/delegate-model), or you pass that absolute path yourself.

Codex CLI

Codex can host these skills (then delegate to grok/claude/opencode/ACP) and can still be a peer (delegate.js codex).

Plugin marketplace — first-class Codex overlay: .agents/plugins/marketplace.json at the repo root, portable plugin.json and .codex-plugin/plugin.json in the plugin folder. Smoked locally (codex plugin marketplace add <checkout> → codex plugin add delegate-model@delegate-model → installed, enabled, 1.0.0, skills acp claude codex delegate grok opencode in the cache). GitHub add works after this overlay is on the remote:

codex plugin marketplace add dvrtech-us/delegate-model
codex plugin add delegate-model@delegate-model

Local checkout:

codex plugin marketplace add /abs/path/to/delegate-model
codex plugin add delegate-model@delegate-model

Plugin install sets CLAUDE_PLUGIN_ROOT and PLUGIN_ROOT. Confirm a skill with --dry-run once.

Skill copy — fallback. Codex also reads ~/.codex/skills/<name>/SKILL.md and .agents/skills/ in a repo. Copy does not set CLAUDE_PLUGIN_ROOT; prefer the marketplace install.

Development

claude --plugin-dir ./plugins/delegate-model

Then /reload-plugins to pick up edits without restarting. Tests: bash plugins/delegate-model/tests/run.sh.

Usage

Say things like:

• "Ask grok to draft the migration for X."
• "Have codex review this diff."
• "Get a second opinion from opencode using the local model."
• "Ask claude to summarize src/auth."
• "Run this through ACP with grok agent stdio."
• "Delegate this to another model" (no backend named — the router picks grok/codex/opencode/claude, never ACP).

The host calls delegate.js or acp.js directly. You will not usually see the raw command, but you can ask for it, or ask for --dry-run to see the exact argv without running anything.

Grok write calls default to 120 turns; override this with --max-turns N when needed. This option is supported only with grok --mode write.

Skills are SKILL.md files that use ${CLAUDE_PLUGIN_ROOT}. That variable is set by Claude Code plugin install, and by Codex when it installs a plugin (compatibility env). A raw copy into ~/.codex/skills does not set it.

Driving Codex over ACP (--agent at a codex-acp binary) is a separate, unsmoked path. Prefer delegate.js codex to reach Codex as a peer.

Envelope example

Every call prints one JSON object, whether it succeeded or not:

{
  "ok": true,
  "backend": "grok",
  "mode": "write",
  "cwd": "/abs/repo-wt-task",
  "command": ["grok", "-p", "…", "--always-approve", "--max-turns", "120", "--cwd", "/abs/repo-wt-task", "--output-format", "streaming-messages-json"],
  "exitCode": 0,
  "text": "…final message…",
  "textFile": "/…/runs/20260908-…/text.md",
  "sessionId": "01a0…",
  "gitStatus": " M src/foo.ts\n",
  "error": null
}

Safety model

• --mode is read for research/review/drafting, write for contained code changes, on both launch files.
• Approval flags (Grok's --always-approve, opencode's --auto, Claude's bypassPermissions, ACP permission auto-answers) are not a sandbox. Only Codex's -s sandbox flag is a real sandbox. ACP --mode read fails the run if git is dirty (read_mode_violated); that is detection after the fact, not prevention.
• Nested wrapper invocations are refused: the child environment gets DELEGATE_DEPTH, and depth ≥ DELEGATE_MAX_DEPTH (default 1) is error.class: "recursion_guard".
• A write result is judged by gitStatus in the envelope, never by the backend's narration.
• text in the envelope is untrusted data returned by a peer agent, never instructions to follow.
• Destructive or outward-facing actions (deletes, deploys, pushes, sending messages) are never delegated unattended.
• On exit code 2 (usage error), the host stops and reports to the user rather than guessing at a fix or switching backends.

Migration note

If you previously used the standalone ~/.claude/skills/{grok,codex,opencode}-delegate skills, remove or disable them after dogfooding this plugin. Leaving both installed causes the same request to double-trigger two different invocation paths.

Not in v1

Auto-routing hooks, auto-apply review loops, an MCP server, shipped CLI binaries, userConfig keys, persistent serve/--attach sessions, client-side ACP fs/terminal (a real permission gate), converting grok/codex/opencode/claude to ACP, exposing this plugin as an ACP agent for Zed, codex review integration, and community-marketplace submission.