cesarrioaldi/use-your-opencode

A skill that turns a coding agent into an **orchestrator**. It stops writing
production code itself: each task goes to a CLI coding agent through a prompt
file, the result is proven with the orchestrator's own measurements instead of
the executor's report, and state lives on disk so nothing is lost to context
compaction. The prompt file also assigns the record-keeping: the executor writes
its own run report, while everything that claims to be true — the status file and
the append-only verification — is written by the orchestrator alone, so a wrong
claim is contradicted on the page instead of overwritten. It is also a cost
decision: the executor does the typing on a cheap or free model, so the costly
session never pays for the edits.
Works with any agent that can read AGENTS.md — Claude Code, Codex, Cursor,
Gemini CLI, opencode, Copilot, Cline, Windsurf, Zed, and the rest of the
skills ecosystem. Stack-agnostic: no hardcoded datastore
client, port, framework, or package manager.
Why: the typing is the part a cheap model can do
Most of what an agent burns in a coding session is not judgement, it is mechanical
work — reading files back, retyping them, running the edit → run → edit loop. That
is where the tokens go, and it is exactly the part you can hand to a model that
costs nothing.
So the split is also a budget. You — the model in the session the owner actually
talks to — keep the context, the plan, the prompt file and the measurement. The
executor burns the bulk tokens on the edits. Point it at a free or cheap endpoint
and the expensive session stops paying for the typing:
opencode reaches free models (opencode/*-free), anything you already pay for, or
a local router, through opencode auth login (opencode providers). Measured on
the machine this was written on, opencode 1.18.35: opencode models lists ten
opencode/*-free models plus 9router/freemodel, and opencode stats reports 228
sessions over 54 days at **$0.00** total cost — 1.0M average tokens per session.
The workflow's own test runs (a two-file Node task, freemodel behind a local
router) took 29–69 s each, and every run produced a report good enough to verify
against.
Two rules keep the saving honest:
- **Cheap runs fail more often.** Free and shared endpoints drop, rate-limit, and
stall. Keep prompts short, one task per file, rerun the unfinished part as a
smaller prompt — and if failures persist, measure the rate (same prompt, N times,
count the failures) instead of blaming the prompt. Never "stabilize" it by
editing the executor's own provider config; that belongs to the owner.
- **The bar does not move.** A free executor does not make its report evidence.
Exit codes, stored rows, rendered state, git diff --stat — all measured by you.
A cheap executor costs reruns, not correctness, but only for as long as you
actually measure.
Three runs, zero cost, and the costly session never typed a line of product code.
Install
These commands use this repository's path. A fork or a rename substitutes the same
way — swap in your own owner/repo.
Project scope by default: the skill lands in .agents/skills/ for the current
project, alongside a skills-lock.json. In a terminal the CLI may ask which
agents and scope you want before it installs; -y answers everything and is what
agents and scripts must use. Add -g to install at user level instead.
| flag | effect |
|---|---|
| -g | user-level install, so the workflow follows you into every project |
| -a '*' | install into every agent the CLI detects, without asking |