Delegate

🔀 /delegate — Multi-Model Orchestration & Delegation

Turn Claude Code into an orchestrator: it hands the right work to the right model — cheap models for bulk, premium for taste, local for private, other providers for breadth — then reports results back. Two ideas underpin it:

1. Right model for the right job — route by cost / intelligence / taste, and escalate rather than cheap out when output misses the bar.
2. Delegate via CLIs — Claude Code shells out to model CLIs and reports back; an opt-in subagent wraps each lane.

This skill is self-customizing. It does NOT hardcode one person's model roster. It discovers what you actually have and builds your config around it.

🛑 FIRST RUN — SELF-CUSTOMIZATION (do this before anything else)

If ~/.claude/CLAUDE.md has no "Model orchestration" section and ~/.claude/agents/ has no delegate wrappers, this skill is unconfigured for this machine. Execute the four steps in order. Do not skip discovery and paste a generic roster — the whole point is to fit the user's actual setup.

Step 1 — DISCOVER what is on THIS machine

Follow references/discovery.md. Detect, without assuming:

• External model CLIs on PATH: codex, gemini, agy, ollama, llm, aider, and any others.
• Local models: ollama list (if present).
• Which Claude models this Claude Code can use (current aliases + any versioned models exposed in the picker/config).
• Provider API keys present — env var NAMES ONLY. Never read, print, echo, or store a key's value.

Report a short table of what's available. That table is the raw material for the routing config.

Step 2 — ASK the user to calibrate (don't invent their economics)

Model rankings are personal — they depend on the user's plans and limits, not list price. Ask:

1. Which subscriptions/plans do you have? (e.g. a Gemini subscription, a ChatGPT/Codex plan, OpenRouter credits, local GPU.) This decides what is "effectively free" vs metered for them.
2. Priority when models conflict — cost, or output quality?
3. Any models to avoid (unproven/untrusted for them), or a hard "never"?
4. Scope: global config (~/.claude/, every project) or one project (.claude/)? Global is convenient; project-scoped keeps unrelated/brownfield repos unaffected.

Fill the routing table's cost/priority from these answers. Do NOT copy example numbers from the template.

Step 3 — BUILD the customized setup

• For each CLI found in Step 1, copy the matching template from templates/agents/ into the chosen agents dir. Skip lanes for tools the user does not have.
• Write the routing framework from templates/CLAUDE.md into the chosen CLAUDE.md, filled with THIS user's models + priorities.
• Pin wrapper subagents to the user's cheapest proven Claude model (ask which — the wrappers only craft prompts and summarize, so they should be cheap).
• (Windows) if codex or another tool isn't on PATH, see the shim note in references/provider-setup.md.

Step 4 — VERIFY

Smoke-test one round-trip per configured lane (see references/cli-invocations.md → "Smoke test"). Each lane should return a correct, model-identified answer before you declare the setup done. Report the results.

Core principles (the routing ethos)

• Defaults, not limits. Judge the OUTPUT, not the price tag. If a cheaper model underdelivers, rerun on a stronger one without asking — escalating costs less than shipping mediocre work.
• When axes conflict for anything that ships: intelligence > taste > cost.
• Bulk / mechanical / clear-spec (implementation to spec, data wrangling, migrations, investigation) → cheapest capable model (often a GPT/Codex or a local model).
• User-facing (UI, copy, API design) or top-quality output → highest-taste model.
• Reviews → a couple of strong models, optionally one from a different provider for an independent lens.
• Research / large-context / web → a big-context model (e.g. Gemini).
• Never route to a model the user flagged as untrusted or "never".

Delegation mechanics (hard-won — full detail in references/cli-invocations.md)

• Close stdin or many CLIs hang. Append </dev/null (bash / macOS / Linux / Git-Bash) or <nul (Windows cmd). This is the #1 cause of "it just hangs".
• Prefer file output over stdout when a CLI drops output on a non-TTY pipe (some agentic CLIs do): have it write to a file and read the file.
• Keep delegation OPT-IN. Subagents should fire only on explicit request and never auto-delegate — this protects governed/brownfield repos from silently routing work to an external model.
• Dynamic spawns take aliases only. Claude Code's on-the-fly subagent spawn accepts current model aliases; to run a subagent on a versioned model, PIN it in an agent file.
• Sandbox awareness. Claude Code's tool shell may be sandboxed/isolated from the host: tools installed by the agent may not reach the user's real machine, and interactive logins done in a terminal may not be visible to the agent's shell. Have the user install host tools and authenticate themselves; verify.
• Report back, don't dump. A wrapper returns a tight synthesis + the model used, not the raw transcript.

Invoking (after setup)

• Natural language: "delegate this to codex", "have gemini research X", "use a local model for this".
• Explicit: @codex-delegate, @gemini-research, @model-delegate.

Extending — add a new provider lane

Copy templates/agents/model-delegate.md, swap in the new CLI + flags, keep it opt-in and reliable-stdout, then add a row to your CLAUDE.md routing table. The llm CLI (one tool, many providers) is the lowest-effort way to add breadth.

Credits

Inspired by @theo (t3.gg)'s posts on model-tiering for agent orchestration — keeping a CLAUDE.md section that prioritizes different models for different work, and teaching Claude Code to use Codex (and other CLIs) as delegation fallbacks for token-hungry tasks (implementation, computer-use, codebase analysis) while the primary model orchestrates.

This skill generalizes that idea into a self-discovering setup (it detects each user's own models/CLIs instead of hardcoding a roster) and hardens the CLI delegation with the non-TTY / stdin-hang / versioned-model-pinning / sandbox- isolation / opt-in lessons learned making it reliable in practice.