opencode Fusion

A main agent that plans and reviews but never edits. A cheaper sidekick that does the editing.

Frontier-level quality at roughly 35–41% lower cost.

npx skills add mihneaptu/opencode-fusion --skill fusion-setup -g -a opencode -y

The main agent turns your task into a plan, delegates code exploration and editing to the sidekick, reviews the returned diff, requests fixes, and verifies the final code.

Strong models are expensive. Weak models cut corners.

Running a top-tier model for every edit and every token gets costly fast. Running a cheaper model on its own tends to drop quality on real, multi-step work.

opencode Fusion splits the roles. You pay the premium only where judgment matters, and let a cheaper model handle the high-volume editing.

How it works

The main agent plans and reviews but cannot edit files. The sidekick does all the editing. That asymmetry is the whole idea.

  1. You

    Describe the task

    Hand the main agent a goal in plain language.

  2. Main agent

    Plan and spec

    It explores, decides, and writes a precise spec. It directs the work but cannot edit files.

    cannot edit
  3. Sidekick

    Execute the edits

    A cheaper agent applies the changes from that spec across the codebase.

    edits files
  4. Main agent

    Verify and iterate

    It runs tests and lint, reviews the diff, and loops until the work is right.

    reviews diff

Built for cost and quality

Every part of the setup earns its place. Nothing is there for decoration.

  • Cost discipline

    The expensive model spends its tokens on judgment: plans, specs, and reviews. The cheaper model handles the high-volume editing, so you pay the premium only where it changes the outcome.

  • Cross-vendor review

    When main and sidekick use different model families, each tends to catch the other’s blind spots. OpenCode’s multi-provider setup makes this easy — pick main and sidekick from different families.

  • Precise spec hand-off

    Work crosses to the sidekick as an explicit spec: objective, files, interfaces, constraints, and how to verify. Not a vague goal.

  • Parallel sidekicks

    Independent edits fan out to several sidekick runs at once, so unrelated work does not queue behind a single thread.

  • Read-only exploration

    Searching and reading the codebase is delegated to cheap read-only agents, keeping the main agent's context lean.

  • Least-privilege permissions

    Edits and freeform shell are denied for the main agent at the OpenCode permission layer. Safety is harness wiring first, prompts second.

Where the savings come from

Illustrative — not a measured benchmark

Frontier model for everything 100%
opencode Fusion −38% ~62%

Cognition has reported the Fusion sidekick pattern reaching roughly 35–41% lower cost on their own evaluation. The bars above only illustrate that idea.

Get started

One command, one prompt, one restart. No clone required.

Your terminal
  1. # 1 · install the skill — global, no clone npx skills add mihneaptu/opencode-fusion --skill fusion-setup -g -a opencode -y
  2. # 2 · open OpenCode — skills are picked up on demand opencode
  3. # 3 · inside OpenCode — or run /fusion-setup set up fusion

fusion configured — restart, then main plans, sidekick edits

Works on Windows, macOS, and Linux. First test: ask the main agent to fix lint, and watch it delegate to the sidekick.
Offline: copy .opencode/skills/fusion-setup into ~/.config/opencode/skills/. OpenCode install docs