opencode-fusion

name

opencode-fusion

description

Multi-model analysis and synthesis using OpenCode CLI — replicates OpenRouter Fusion style workflows locally. Runs multiple LLMs in parallel, judges their outputs, and produces a single synthesized response. Trigger whenever the user wants to run a task against multiple models, compare model outputs, get a consensus answer, or says things like "fusion run", "multi-model", "run all models", "synthesize model outputs", or "parallel agents". Also trigger when the user wants to reduce hallucinations by cross-checking multiple LLMs, or when high-stakes tasks need more than one model's perspective.

OpenCode Fusion Skill

Implements an OpenRouter Fusion-style workflow entirely through OpenCode CLI: parallel model execution → judge analysis → final synthesis.

Full workflow runs automatically after loading this skill. No external orchestration needed.

---

Architecture Overview

TASK INPUT
    │
    ▼
┌─────────────────────────────────────┐
│  Phase 1: Parallel Model Execution  │
│  ┌────────┐ ┌──────┐ ┌──────┐      │
│  │ Claude │ │ GPT  │ │ GLM  │      │  ← add/remove freely
│  └────────┘ └──────┘ └──────┘      │
│       All run simultaneously        │
└────────────────┬────────────────────┘
                 │ outputs → /tmp/fusion-*/
                 ▼
┌─────────────────────────────────────┐
│  Phase 2: Judge Layer               │
│  Analyzes consensus, contradictions,│
│  gaps, risks, unique insights       │
└────────────────┬────────────────────┘
                 │ judge report → /tmp/
                 ▼
┌─────────────────────────────────────┐
│  Phase 3: Synthesizer               │
│  Merges best ideas, resolves        │
│  conflicts, fills gaps              │
└────────────────┬────────────────────┘
                 │
                 ▼
         FINAL OUTPUT (only this shown to user)

---

Invocation

The skill runner should call run_fusion.sh with the task as argument:

bash run_fusion.sh "<TASK>"

Everything else is automated. The script handles parallel execution, waiting, judging, and synthesis.

---

Phase 1: Model Panel Configuration

The panel is defined at the top of run_fusion.sh as a bash array — easy to add/remove models.

Default models: opencode/claude-sonnet-4-6, opencode/gpt-5.5-pro, opencode/glm-5.1

Each model runs via:

opencode run "<TASK>" \
  -m "<MODEL_ID>" \
  --attach http://localhost:37993 \
  --dangerously-skip-permissions \
  > /tmp/fusion-<MODEL_SLUG>/output.txt 2>/tmp/fusion-<MODEL_SLUG>/err.txt &

All jobs launched with &, then wait collects them. Exit codes checked per-model — failures are noted but don't abort synthesis.

---

Phase 2: Judge Prompt

Read references/judge-prompt.md for the full judge system prompt.

The judge receives all model outputs concatenated with clear === MODEL: <name> === delimiters and produces a structured report with:

• CONSENSUS — points all/most models agree on
• CONTRADICTIONS — where models conflict, with analysis of which is likely correct
• GAPS — important aspects no model covered
• RISKS — identified failure modes or edge cases
• UNIQUE_INSIGHTS — valuable points from individual models
• RECOMMENDED_APPROACH — synthesis direction

---

Phase 3: Synthesizer Prompt

Read references/synthesizer-prompt.md for the full synthesizer system prompt.

The synthesizer receives:

1. The original task
2. The judge's structured report
3. All raw model outputs

It produces the single final response shown to the user.

---

Skill Runner Instructions

When this skill is invoked with a task:

1. Check that opencode CLI is available (which opencode)
2. Check that the OpenCode server is running (curl -s http://localhost:37993/health or similar)
3. Execute bash run_fusion.sh "<TASK>"
4. Stream the final synthesized output to the user
5. Optionally report which models succeeded/failed

If opencode is not available or the server is not running, output a clear error with setup instructions (see references/setup.md).

---

Adding / Removing Models

Edit the MODELS array in run_fusion.sh:

MODELS=(
  "opencode/claude-sonnet-4-6:claude-sonnet-4-6"
  "opencode/gpt-5.5-pro:gpt-5.5-pro"
  "opencode/glm-5.1:glm-5.1"
  # "opencode/claude-opus-4-8:claude-opus-4-8"
  # "opencode/gpt-5.5:gpt-5.5"
  # "opencode/glm-5:glm-5"
)

Format: "<model-id>:<slug>" — the slug is used for temp file naming.

---

Reference Files

• references/judge-prompt.md — Full judge system prompt
• references/synthesizer-prompt.md — Full synthesizer system prompt
• references/setup.md — OpenCode setup and server prerequisites
• run_fusion.sh — Main orchestration script