delegate-router: multi-model task delegation for Claude Code and Codex CLI

One agent thinks, another types. delegate-router is an agent skill (SKILL.md, the Agent
Skills format) for Claude Code and OpenAI Codex CLI that works as an LLM router for
coding work. It hands bulky, already-decided work (boilerplate, scaffolding, test stubs,
mechanical edits across many files, reading media and long documents) to another model you
already pay for: Codex (GPT models), Gemini through the Antigravity agy CLI, or a
Claude Opus subagent. The host keeps the design and the verification, and saves its own
quota and context.

• Routing by task, not by habit. ROUTING.md maps each kind of
  task to an ordered list of models, chosen from independent benchmarks and cost per task.
  It skips any CLI you don't have and never routes back to the host's own model, except an
  Opus subagent under an Opus host, which keeps the main context clean.
• A judge before every handoff. Four gates (bulk, decided, verifiable, contained) and a
  never-delegate list (debugging, architecture, security code, anything under ~40 lines).
• The host always verifies. A delegate's "done" isn't evidence: the host runs the check and
  reads the diff.
• An update path. model_watch.py notices new model releases in your CLIs and the
  vendors' release notes, and can update the CLI and prove the new model answers.

Interactive diagram: docs/workflow.html (open it locally).

Requirements

• macOS or Linux, git, python3 3.9+.
• Your main agent is Claude Code (claude) or
  Codex CLI (codex), plus at least one more of Codex,
  Gemini through the Antigravity CLI (agy), or Claude. Other
  main agents and other providers (DeepSeek, local models) aren't supported yet. Each CLI
  needs its own subscription or API access; delegate-router adds no accounts and no API keys.
• Optional: uv and a TypeSafe key, only for Jev.

Install

git clone https://github.com/pvjbuilds/delegate-router ~/.delegate-router

~/.delegate-router/install.sh

Clone to a fixed path: the install links to it, so moving the folder later breaks the links.

The installer finds your CLIs, links the skill, and asks a few questions. Enter takes the
default; --yes takes every default. Re-running it keeps your earlier answers.

If claude is installed, the installer also links the opus-implementer agent into
~/.claude/agents: both hosts use it to reach Claude Opus.

It never overwrites a file or folder that isn't its own, and never edits your Claude Code or
Codex settings. It prints two optional snippets instead:

• Codex as host. Codex's sandbox blocks network by default, so calling the claude CLI
  from inside Codex asks for approval. Don't approve it: an approved command runs outside the
  sandbox. Instead add to ~/.codex/config.toml, which keeps the sandbox on:

  [sandbox_workspace_write]
  network_access = true
  

• Claude Code session-start notice. A SessionStart hook that runs model_watch.py status,
  so new-model notices show when a session opens.

Restart Claude Code or Codex after installing.

Use

Ask the host to delegate, or type /delegate:

Delegate this: add typed stubs for the 14 endpoints in api/routes.py, one per function,
each raising NotImplementedError.

The host judges the task, picks the model, writes the spec, runs the delegate, reads the diff
and reports back: what went where, what came back, what it fixed. Name a model to override the
map ("send this to Gemini").

Routing, briefly

The reasons, effort levels, caveats and the council protocol for when two models disagree are
in ROUTING.md. Benchmarks are a prior; a check on the real task
beats them.

Staying current

git -C ~/.delegate-router pull

That brings a newer routing map. Between releases, model_watch.py tells you what's new:

python3 ~/.delegate-router/skills/delegate/scripts/model_watch.py status

• status makes no network calls itself. It lists your delegate CLIs, any new release it
  found, and any model your CLIs list that ROUTING.md doesn't cover yet. With background checks
  on, it also starts run --if-due in the background, which goes online at most every 12 h.
• run does the checking: model catalogs (codex debug models, agy models, Anthropic's
  models page) and five vendor release-note pages. It's capped at 10 names, and every step,
  CLI update and probe included, stops at 15 minutes.
• ack marks everything current as reviewed.

State lives in ~/.local/state/delegate-router/, settings in
~/.config/delegate-router/config.env (see config.example.env).

Jev (optional)

Jev (TypeSafe) is a very cheap classifier, about $0.0001 per call.
Turned on, it does two jobs:

1. Judges release notes. Is a new name a real subscriber release, or API-only or a preview?
2. Scouts files before a spec. It ranks which files a task needs, so the host reads 5
   files instead of 40.

Data gate: Jev sends text to TypeSafe. Use it only on public or low-sensitivity code.
Never on client work, secrets, credentials or personal data. The scout skips secret-looking
files by name and content, but that's a pattern match, not a guarantee.

Key: TYPESAFE_API_KEY in your environment, or on macOS in the login keychain:

security add-generic-password -a "$USER" -s TYPESAFE_API_KEY -w

The key is never read from a file in this repo, never written to config and never printed.

Safety

• Codex delegates run in Codex's sandbox: workspace-write to implement, read-only to
  review. Never full access, never a --dangerously-bypass-* flag. They skip your
  config.toml and execpolicy rules and turn apps off; plugins and admin-managed config can
  still apply.
• Gemini runs through agy in plan mode with --sandbox. It is still an agent: the
  handoff folder isn't a read boundary and your integrations stay on, so send it only packets
  you wrote.
• Claude Opus as an implementer runs with the Claude Code host's own permissions (turn on
  Claude Code's /sandbox for OS-level confinement), or inside Codex's sandbox when Codex is
  the host. As a reviewer it runs --restricted with only Read, Grep and Glob.
• Every Codex packet says "do not remove or weaken any existing test".
• Security, auth and money code is never typed by a delegate. A strong model may review it,
  read-only.
• New-model probes send one fixed prompt from an empty temp folder, with your CLI config,
  rules and integrations turned off where the CLI allows it, and pass only on an exact ok.
• Model names read from web pages are validated before they're printed or passed to a CLI.

Uninstall

~/.delegate-router/uninstall.sh

This removes only its own links. Add --purge to also delete the config and state.

How routes are tested (maintainer only)

You don't run this: these runs are the maintainer's and cost you nothing. When a new model ships, the maintainer runs it through the bench against
the model that holds the route now, and the result reaches you in your next git pull.
Benchmarks pick a new model's first route; the bench decides whether it keeps it.

Ten routine tasks, each with a hidden check that grades a temp copy of the model's work:

python3 bench/bench.py codex gpt-6.1-sol high

python3 bench/bench.py --selfcheck

The first runs one model (codex, claude or agy, then the model ID and effort) and
appends pass, seconds and tokens per task to bench/results.tsv. --task 6 runs one task.
--selfcheck runs no model: for each task with a reference answer in ref/ (tasks 6–10), it
proves the check fails on the untouched files and passes on that answer.

Latest run (2026-10-27): GPT-6.1 Sol high and GPT-6.1 Terra medium both passed 10/10; Terra
took 307 s to Sol's 467 s. The bench doesn't yet find where trackers put Sol ahead on coding.

Tests

python3 tests/test_model_watch.py && bash tests/test_install.sh && python3 bench/bench.py --selfcheck

uv run -q tests/test_jev.py && uv run -q tests/test_scout.py

All offline: stub CLIs, local pages, a temp HOME. No network, no spend.

FAQ

Why a clone and an install script, not a plugin? Both Claude Code and Codex support
plugins, and the folder is already laid out like one. A plugin can't ask setup questions, and
it would install a separate copy per host. A plugin package may follow; the skill itself won't
change.

Does it work with only one agent installed? It installs, but there's nothing to route to.
Delegation needs a second model.

Will it spend my money? Only through subscriptions you already have, when you delegate.
Background checks read public pages and your CLIs' model lists. Probes and Jev run only if you
turn them on.

Keywords

task delegation · delegate coding tasks · AI code delegation · LLM router · LLM routing · model
router · model routing · multi-model routing · multi-model · multi-LLM · cross-model code review ·
Claude Code skill · Claude Code skills · Codex CLI skill · Codex skills · Agent Skills · SKILL.md ·
Claude Code subagent · subagent · coding agent · AI coding agent · agentic coding · multi-agent ·
save tokens · token optimization · reduce AI coding cost · cost per task · subscription quota ·
model release watcher · new model alerts · OpenAI Codex · Google Gemini · Anthropic Claude ·
Antigravity CLI · agy · Claude Opus 5.5 · GPT-6.1 Sol · GPT-6.1 Terra · GPT-6 Luna · GPT-6 Astra ·
Gemini 3.8 Flash · TypeSafe Jev · model benchmark · LLM eval

License

MIT. See LICENSE.