claude-cordyceps

claude-cordyceps

Cross-model consultation as a Claude Code skill. When you ask Claude for a council, a panel, a peer code review, a brainstorm with multiple LLMs, a critique of a plan, a cross-check of a confident claim, or any task where input from a non-Claude model would help — this plugin teaches Claude to drive @ellyseum/cordyceps to spawn Codex, Gemini, Ollama, and additional Claude variants in parallel and consume their output. The point is heterogeneity: a panel from different model families catches blind spots any single family shares.

The skill is a capability, not a single workflow. It teaches Claude how cordy works and lets the model pick the right shape for whatever the user actually asks for — code review, plan critique, design red-teaming, brainstorm, decision exploration, fact-check, meta-review, etc. New patterns fall out of the same primitives without the skill needing to enumerate them.

Why cross-model input matters

Claude is right most of the time and confidently wrong some of the time. The same is true of every other frontier model — but their failure modes are uncorrelated. A bug Claude misses is often a bug Codex catches. A factual claim Gemini gets wrong is often one a local Ollama model nails. On any task that rewards multiple perspectives, running a panel from different families is a measurable quality lift.

This plugin makes that lift one phrase or one slash command away.

Prerequisites

1. Cordyceps: install the harness globally.

   npm i -g @ellyseum/cordyceps
   

   Tested with ^0.5.2. Older 0.5.x works (the skill falls back gracefully) but loses the --no-chair host-as-chair optimization.

2. At least one peer driver besides Claude, for the diversity benefit. Each is optional; install whichever you want as council members.

   Driver	Install	Auth
   claude-code	already on PATH if Claude Code is installed	inherits Claude Code session auth
   codex	npm i -g @openai/codex	codex login (one time)
   gemini	Google's gemini CLI	export GEMINI_API_KEY (or Vertex / GCA)
   ollama	Ollama; run ollama serve; ollama pull qwen2.5-coder:7b	none — fully local

   Cordyceps probes which of these are reachable on every session start; the plugin builds whatever council the user asks for from the available drivers.

Install (the plugin)

/plugin marketplace add ellyseum/claude-cordyceps
/plugin install claude-cordyceps

Reload plugins (or restart your session). On the next session start, the plugin's preflight runs cordy doctor + cordy --ephemeral drivers --json and caches the result. The skill reads that cache when activated.

For local development:

git clone https://github.com/ellyseum/claude-cordyceps
cd claude-cordyceps
claude --plugin-dir .

Use

The plugin works two ways. Both are first-class.

1. Natural-language activation

Just ask Claude. The skill activates on phrases like:

• "consult peers" / "consult with peers"
• "spawn council" / "convene council" / "ask the panel"
• "code review with peers" / "review this file with peers"
• "brainstorm with peers" / "brainstorm with multiple LLMs"
• "critique my plan with peers"
• "ask Codex / Gemini / Ollama about this"
• "what would think about this"

The skill picks the shape (single peer query, silo'd panel + chair, adversarial pair, brainstorm panel, tiered escalation, persistent sidecar, meta-review) that fits the user's actual ask.

Phrases like "second opinion", "peer review", "plan review", and "red team this" are softer triggers — the skill activates on those only when there's a clear cross-model signal alongside (a target artifact and explicit peers/panel framing). Otherwise it asks one clarifying question first.

2. Slash command

Free-form passthrough — describe what you want; the skill handles the rest:

/cordy review src/core/bus.ts
/cordy brainstorm names for an open-source agent harness
/cordy critique my migration plan: [paste plan]
/cordy cross-check "Postgres is faster than SQLite for embedded use"
/cordy what would Codex and Gemini think of this design?
/cordy diff --staged
/cordy do a peer review on the diff vs main

The slash command is especially useful for testing or when the natural-language triggers feel like overreach.

How it works