A Claude Code plugin that offloads token-heavy, self-contained tasks to local pre-authed CLI
backends — agy (Gemini) and codex (OpenAI Codex CLI) — picking the backend and model by
task size and type, with credit-exhaustion fallback through the chain to native Claude, and a
glanceable statusline HUD.
The core idea:
Offload commodity work (UI/components, scaffolding, CRUD, scripts, SQL, configs, unit tests,
web research, bulk summarization) to a fast/cheap CLI — keep judgment-heavy and systems-hard
work (reverse-engineering, FFI/unsafe, injection, concurrency, protocol design) on Claude.
Every routing decision is config-driven; tune it without touching code.
agy, codex, and native Claude are equal, configurable tools. /team decomposes a task and
assigns each subtask to its best-fit backend; /reasoning fans one question across a panel of all
three and fuses the answers.
⚡ Quick Start
1 · Install the backends (one-time, pre-auth each)
npm install -g node-pty # the one native dep — gives agy a pseudo-terminal (see note below)
npm install -g @openai/codex # then: codex login
# Windows Powershell
irm https://antigravity.google/cli/install.ps1 | iex # then: agy login
# macOS / Linux
curl -fsSL https://antigravity.google/cli/install.sh | bash # then: agy login
2 · Add the plugin. This repo is the plugin — point Claude Code at it as a local plugin (local
marketplace or --plugin-dir). On enable, Claude Code auto-discovers commands/, agents/, and
hooks/hooks.json. Nothing else to wire up.
3 · (Optional) Turn on the HUD. Add a statusLine to your own ~/.claude/settings.json
(the plugin can't register one for you) — see Statusline HUD.
4 · Use it.
/reasoning 2:gemini,opus,codex What's the best caching strategy for a read-heavy API?
/team 3:gemini,1:codex Build a REST CRUD service with tests
/route-test Write a SQL query to list users by signup date ← dry-run, no call
…or just work normally and let Claude reach for the agy / codex agents on its own.
🎛️ Commands
| Command | What it does |
|---|
/reasoning [panel] <question> | Fusion pipeline. Fan one question across a panel of models in parallel → a judge compares them (consensus / contradictions / unique insights / blind spots) → synthesize one unified answer better than any single model's. |
/team [N:gemini,M:claude,X:codex] <task> | Team pipeline. Decompose → dispatch each subtask to its best-fit backend (dependency-aware waves) → verify each result → bounded fix loop → synthesize. |
/route-test <task> | Dry-run the router: prints {backend, model, tier}, detected types, matched rule. No backend call — a tuning tool. |
Both /team and /reasoning have two engines: an Ultracode deterministic Workflow path
(preferred, when the Workflow tool is available) and a parallel Task-agent fallback. Either way the
work runs across parallel agents — never one inline session.
Agents (Claude spawns these on its own for matching work)
| Agent | Use for | Backend |
|---|
agy | Standard, verifiable coding + Gemini's edges (compact/checkable results) | agy |
codex | Code review, test-writing, verification | codex |
There is intentionally no RE/injection agent — that work stays native by default. An explicit
agent spawn is honored as-is (forces that backend; the router's hard line won't bounce it).
🚦 How routing works
src/bin/route.mjs scores the task (char count + keyword types from config/tags.txt), then matches
routes rules in the roster (first match wins; order encodes priority). src/bin/run.mjs runs the
chosen backend with a fallback chain, writes HUD state, and cleans output.
-blue){kind=icon}
