agent-fusion

Tooling for Agent Fusion — running multiple independent agents (Claude/Opus + Codex/GPT) on one task and fusing their output.

• agent-fusion.md — the harness protocol (Mode A: 2-agent self-integrating; Mode B: 2 symmetric workers + a clean ARBITER conductor). Read this first.
• fusion-lib.sh — shared shell helpers (fusion_slug / fusion_log / fusion_wait); synchronous heartbeat, no background process.
• fusion — launcher: scaffolds a run dir and opens a cockpit workspace with one pane per agent, each started with its role header + prompt.
• skill/agent-fusion/SKILL.md — vendored copy of the agent-fusion skill (the bootstrapping half: compose the brief, pick the mode, launch). Source of truth; installed to ~/.claude/skills/agent-fusion/SKILL.md. Re-copy after editing either side to keep them in sync.

fusion --mode A <slug> task.md        # 2 agents: PRIMARY(claude) + SECONDARY(codex)
fusion --mode B <slug> task.md        # 3 roles:  ARBITER + WORKER_A(claude) + WORKER_B(codex)
fusion <slug> task.md --dry-run       # preview scaffold + prompts; spawn nothing

Runs live under runs/<slug>/ (gitignored — the filesystem is the coordination bus).

---

Part of a trio: santa (search & resume your Claude + Codex history) and cockpit (drive many live sessions at once). See agent-tooling for how they fit together.

License

The Unlicense — released into the public domain. Do whatever you want.