Model Routing

Model Routing

Overview

Claude Code orchestrates subagents and workflows, and each spawned agent can run on a different model. Left unguided, everything runs on the session's default model — usually the most expensive one. This skill installs a model-routing section into a project's CLAUDE.md: a rankings table (cost / intelligence / taste per model) plus application rules and mechanics, so the orchestrating Claude routes each piece of work to the cheapest model that clears the quality bar — including non-Claude models reached through the Codex CLI.

The pattern (popularized by Theo's multi-model workflow: x.com/theo/status/2072481845363822914, x.com/theo/status/2072482460122964067): score each available model on three axes, state routing rules as defaults-not-limits with standing permission to escalate, and document the exact mechanics (codex exec, Agent/Workflow model parameter) so routing actually happens instead of staying aspirational.

When to Use

• The user wants Claude Code to delegate bulk/mechanical work to a cheaper model (e.g. GPT via Codex) and reserve the expensive model for hard problems
• The user asks to add model-selection guidance to a project's CLAUDE.md
• The user is burning rate limits on token-hungry work (computer use, codebase analysis) that a cheaper model could absorb
• Setting up a new project that will use multi-model workflows or agent teams

Prerequisites

• Required: none — a Claude-models-only routing section still works via the Agent/Workflow model parameter.
• For GPT routing: the Codex CLI (codex --version), authenticated (codex login or OPENAI_API_KEY). Check ~/.codex/config.toml for the default model and model_reasoning_effort.

Workflow

1. Discover the environment

codex --version                     # Codex CLI installed?
grep -E '^(model|model_reasoning_effort)' ~/.codex/config.toml

Also note which Claude models the harness exposes for the Agent/Workflow model parameter (check the Agent tool's model enum in the current session — e.g. sonnet, opus, haiku, fable).

2. Interview the user

The rankings are personal — cost especially, since it reflects subscription headroom and rate limits rather than list price. Ask (briefly, with sensible defaults offered):

• Which subscriptions/plans do they have, and which is closest to its limits?
• Any model they never want used? Any they consider effectively free?
• What kinds of work do they want delegated away from the expensive model?

3. Instantiate the template

Copy the section from references/claude-md-template.md and customize the table rows, scores, rules, and mechanics per the interview and discovery. Only reference models and commands that actually exist in the user's setup.

4. Insert into CLAUDE.md

• Target the project-local CLAUDE.md at the repo root by default; use ~/.claude/CLAUDE.md only if the user wants it globally.
• If a ## Picking the right models (or equivalent) section already exists, replace it in place — don't append a duplicate.
• If CLAUDE.md doesn't exist, create it with just this section.
• Show the user the final section before or right after writing it.

5. Verify the mechanics

If GPT routing is enabled, smoke-test the delegation path end-to-end:

timeout 120 codex exec -s read-only "Reply with exactly: OK"

(Add --skip-git-repo-check if the target directory isn't a git repository — codex refuses to run outside one by default.)

If this prompts for auth or fails, fix that now — a routing section that references a broken command silently degrades to "use the default model for everything".

Codex CLI mechanics

For the exact flags (sandbox modes, output capture, model override, session resume) see references/codex-cli.md. The short version:

The timeout wrapper is mandatory, not decoration: codex exec can hang indefinitely at zero CPU on a dead network wait, and a hung exec is indistinguishable from a quietly-working one. On expiry, retry once or do the chunk without codex.

Codex prompts must be self-contained: codex shares none of Claude's conversation context, so include absolute paths, the full task spec, and acceptance criteria.

Troubleshooting

• Routing never happens: the section is guidance, not enforcement — it must live in a CLAUDE.md that's actually loaded (project root, or global). Verify with /context or by asking Claude what routing rules it sees.
• Codex "fails" on implementation tasks: codex exec never prompts for approval — commands either run inside the sandbox or fail. If it can't write files, the sandbox was read-only (the default); rerun with -s workspace-write.
• A delegated task goes silent: check CPU time vs process age (ps aux | grep "codex exec") — near-zero CPU on an hours-old process is a hung network wait, not slow work. Kill the whole process chain and re-run under timeout; see references/codex-cli.md.
• Wrapper agents summarize instead of relaying: the wrapper prompt must say "return the codex output verbatim as your final message".