hermes-council

An LLM Council plugin for Hermes Agent — karpathy/llm-council semantics as a first-class Hermes plugin.

Four fixed advisor seats answer your question in parallel, anonymously review and rank each other, and a Chairman model synthesizes the final answer. Members borrow the provider credentials you already configured in Hermes — no API keys, no extra setup.

Your question
   │
   ├─ Stage 1 · First opinions (parallel, each seat steered by its duty)
   │    First-Principles Analyst · Skeptic · Pragmatist · Researcher
   │
   ├─ Stage 2 · Peer review (anonymized as "Response A/B/C/D")
   │    every seat ranks the others — or a dedicated Judge ranks them —
   │    strict "FINAL RANKING:" parsing → aggregate avg rank + votes
   │
   └─ Stage 3 · Chairman synthesis
        ranked digest + all answers as private guidance

The council

Seat	Duty (injected into its Stage-1 prompt)
First-Principles Analyst	Reason from fundamentals; ignore convention and popularity.
Skeptic	Assume the obvious answer is wrong; hunt for failure modes, hidden costs and base rates.
Pragmatist	What actually works in practice — constraints, effort, reversibility, second-order effects.
Researcher	Ground claims in verifiable evidence; use web search and code inspection where available.

Seats are fixed — you only pick which of your models sits in each seat (panel or hermes council configure). Stage 2 is fully anonymous: reviewers never see seat names, roles, or model ids (karpathy's anti-favoritism rule, enforced by tests). Unassigned seats are skipped.

Optional slots:

• Judge — when set, one model ranks the anonymized answers in a single call instead of full peer review.
• Chairman — synthesizes the final answer (required).

Install

hermes plugins install girayk/hermes-council
hermes plugins enable hermes-council

or clone into ~/.hermes/plugins/:

git clone https://github.com/girayk/hermes-council ~/.hermes/plugins/hermes-council
hermes plugins enable hermes-council

Requires a working Hermes install with at least two configured provider models (any providers — Qwen, OpenRouter, Anthropic, OpenAI, local…). Restart the desktop app / gateway after enabling.

Configure

Desktop: the plugin ships a /council panel (sidebar → Council): each seat shows its name + duty, you pick provider/model from your configured catalog; Judge and Chairman below; autosaves to council: in config.yaml.

CLI:

hermes council configure    # walk the 4 seats + judge + chairman
hermes council list         # show the roster

config.yaml (what the panel writes):

council:
  default_preset: default
  presets:
    default:
      members:                     # the 4 fixed seats, positionally
        - {provider: qwen, model: qwen3.8-max, enabled: true}       # First-Principles Analyst
        - {provider: qwen, model: deepseek-v4-pro, enabled: true}   # Skeptic
        - {provider: qwen, model: glm-5.3, enabled: true}           # Pragmatist
        - {provider: "", model: "", enabled: false}                 # Researcher (unassigned)
      judge: {provider: "", model: "", enabled: true}   # empty = peer review
      chairman: {provider: qwen, model: qwen3.8-max}
      lite: false                                        # true = skip Stage 2
      enabled: true

Use it

Three surfaces, one engine:

1. Tool — the model can call council_deliberate on hard questions (the bundled skill teaches it when that's worth the cost).
2. Slash command — /council <question> in any chat (CLI, TUI, gateway).
3. Model menu — the plugin registers a Council group in every model picker (its "model" is the council preset): /model default --provider council, or the desktop model dropdown. Every turn of that session runs through the council — the chairman acts as the model (tools + streaming work normally), with the ranked member digest as its private guidance.

Recent deliberations land in ~/.hermes/council/runs/ and show in the panel.

Cost

Per deliberation: N+1 calls with lite: true (N seat answers + chairman), up to 2N+1 with peer review. As a session model, the fan-out runs once per user turn; tool-loop iterations reuse the cached guidance and only re-call the chairman. Every stage has hard wall-clock deadlines — a wedged model is dropped with a failure note, never hangs your turn.

How it works (plugin surfaces only — no core patches)

Surface	Where
Model menu group	plugin ProviderProfile via register_provider — the documented model-provider plugin seam; virtual council://local provider, presets as models
Turn engine	CouncilClient OpenAI-shaped facade via create_client(); stages call agent.auxiliary_client.call_llm(task=…, provider=…, model=…) with YOUR credentials
Tool / slash / CLI	register(ctx): council_deliberate, /council, hermes council
Desktop panel	desktop/plugin.js (plain ESM, plugin SDK) + dashboard/plugin_api.py routes under /api/plugins/hermes-council/
Skill	bundled skills/llm-council/SKILL.md (when to convene a council, anti-sycophancy rules)
Config	council: block in config.yaml (council_config.py normalize/validate/save)

Stage-2 prompts contain only "Response A/B/C" labels — seat names, roles and model ids never reach reviewers (anonymity-guard tests). The chairman's guidance de-anonymizes: Response A (First-Principles Analyst (qwen:qwen3.8-max)).

Development

hermes plugins validate ~/.hermes/plugins/hermes-council   # manifest, security scan, desktop surface
cd tests && HERMES_HOME=/tmp/council-test pytest -q        # 35 tests (engine, config, anonymity guards)

Credits

• Council protocol: karpathy/llm-council
• Seat roster & duties: the llm-council skill protocol (independence before interaction, stake-weighted criticism, anti-sycophancy rules)
• Mechanics follow Hermes's own Mixture-of-Agents design (virtual provider + facade client + call_llm fan-out), implemented entirely through public plugin seams.