artale
/
fable-agent
Template
1
1
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
fable-agent/SKILLS/fusion/SKILL.md

4.1 KiB
Raw Blame History

name description
fusion Answer a hard question by fanning it out to a PANEL of models running in parallel — each answering independently, none seeing the others' work — then having the judge model evaluate every response into a structured analysis (consensus, contradictions, partial coverage, unique insights, blind spots) and write a final answer grounded in it. Panel options: opus4.8-4.8 — two independent Opus 4.8 runs (no external CLI needed) opus4.8-gpt5.5 — Opus 4.8 + GPT-5.5 via codex CLI opus4.8-gpt5.5-gemini — Opus 4.8 + GPT-5.5 + Gemini 3.1 Pro sonnet-haiku-opus — Sonnet 4.6 drafts, Haiku checks, Opus fuses openrouter-fusion — single call to OpenRouter Fusion API The judge (always Opus 4.8 unless overridden) writes the final answer. The pipeline cannot be reversed: panelists cannot call back out to spawn the judge. Best for: high-stakes research, architectural decisions, debugging, and any question where being confidently wrong is expensive.

Fusion Skill

Mechanism

Fusion turns one prompt into a panel. The question goes to several models at the same time, each answering independently — with full tool access and no knowledge of the others. Then the judge reads every answer, extracts the structure of the panel's reasoning (what they agree on, where they conflict, what only one saw, what they all missed), and writes a final answer grounded in that analysis.

The core principle is independence, then synthesis. Every panelist gets the task verbatim. No assigned "lenses" or personas — the diversity comes naturally from independent execution.

Procedure

Step 1: Fan Out Panel

Dispatch the task to N panelists in parallel:

for each panelist in parallel:
  spawn sub-agent with clean context
  panelist receives: task (verbatim)
  panelist executes: research, tool calls, analysis
  panelist returns: full answer

Step 2: Judge Analysis

The judge (Opus 4.8) evaluates every panelist answer and produces a structured analysis:

Consensus:       what all/most panelists agree on
Contradictions:  where panelists disagree (with resolution or flag)
Partial:         aspects only some panelists addressed
Unique Insights: things only one panelist found
Blind Spots:     what no panelist addressed

Step 3: Grounded Final Answer

The judge writes the final answer, grounded in the analysis:

Synthesis of consensus into a coherent position
Resolution of contradictions
Incorporation of unique insights
Acknowledgment of blind spots
Actionable conclusion

Panel Configurations

Slug Panelists Cost Savings vs Fable 5
opus4.8-4.8 2× Opus 4.8 (~$0.09) ~40% cheaper
opus4.8-gpt5.5 Opus 4.8 + GPT-5.5 (~$0.12) ~20% cheaper
sonnet-haiku-opus Sonnet + Haiku + Opus (~$0.06) ~60% cheaper
openrouter-fusion OpenRouter Fusion compound (~$0.08) ~50% cheaper

CLI Usage

# Default panel (opus4.8-4.8)
node dist/index.js fusion run "Complex research question"

# Specific panel
node dist/index.js fusion run "Design review" --panel sonnet-haiku-opus

# OpenRouter Fusion API
node dist/index.js fusion run "Architecture analysis" --openrouter

# List available panels
node dist/index.js fusion panels

Integration with fable-agent

This skill integrates with:

  • FusionExecutor (src/examples/executors/fusion-executor.ts) — PhaseExecutor-compatible
  • OpenRouterFusionExecutor (src/examples/executors/openrouter-fusion-executor.ts) — API-based fusion
  • DynamicWorkflows.fusionPanel() — workflow-level fusion dispatch
  • CompoundStack — fusion as a stakable layer
  • Fable5PromptEngine — tone/citation/refusal rules from leaked Fable 5 prompt

References