The Council

Parallel minds. Singular clarity.

Security advisory (2026-05-25): Versions 1.2.0 and 1.2.1 do NOT enforce read-only on the Gemini advisor as their docs claim — agy's --dangerously-skip-permissions auto-approves tool calls rather than withholding them, and a 2026-05-24 incident saw agy ghost-write 8 files into a project directory. Upgrade to 1.3.0 or later for real OS-level sandbox enforcement (sandbox-exec on macOS) and a pre/post-invocation diff safety net. See SKILL.md → Permissions and Safety for the full incident note.

A Claude Code skill that convenes OpenAI Codex and Google Gemini as an advisory board. Both run in parallel via their CLIs — Codex with read-only codebase access, Gemini reviewing the fully-inlined request — and return independent analyses that Claude synthesizes into a unified recommendation.

What It Does

When you say "ask the council" or request a second opinion, Claude:

1. Syncs your project context (CLAUDE.md) to Codex via AGENTS.md
2. Composes a prompt using the appropriate template (code review, architecture, debugging, or general)
3. Invokes Codex and Gemini in parallel, both in read-only sandboxes
4. Synthesizes the responses — highlighting consensus, divergence, and Claude's own recommendation

Codex runs with xhigh reasoning effort by default — the invoke script sets COUNCIL_CODEX_EFFORT=xhigh explicitly rather than trusting ~/.codex/config.toml (which often ships a weaker medium). Codex has read-only filesystem access to the working directory; Gemini reviews the inlined prompt content in a single shot.

Learning Loop

After each council session, Claude reflects on what the advisors revealed — gaps, blind spots, or better approaches — and generalizes those learnings into CLAUDE.md and AGENTS.md. This means future sessions in the same repository start with lessons from previous council deliberations already loaded.

Installation

Via Claude Code Plugin Marketplace

/plugin marketplace add DantesPeak85/the-council
/plugin install the-council@the-council-marketplace

Manual Installation

Clone this repo and register it as a local plugin in your Claude Code configuration.

You need at least one advisor CLI installed and authenticated.

OpenAI Codex CLI

npm i -g @openai/codex
codex auth

Google Gemini — two backends

The Council picks the Gemini backend automatically (COUNCIL_GEMINI_BACKEND=auto):

• Antigravity CLI (agy) — the default backend. Install Google's Antigravity client; it authenticates via ~/.gemini/oauth_creds.json on first run. Runs pty-wrapped, sandboxed, single-shot.
• gemini CLI (optional, preferred when available) — the single-shot JSON backend. Selected automatically only when the gemini binary is on PATH and a paid Gemini API key — GEMINI_API_KEY or GOOGLE_API_KEY (AI Studio) — is exported. gemini-cli's free oauth-personal auth stopped serving 2026-06-18, so this backend requires the paid key.

System dependencies (macOS)

• python3 — parses the gemini-cli JSON response envelope (no jq dependency).
• sandbox-exec and script — ship with macOS. The Council uses sandbox-exec to wrap agy/gemini in an OS-level deny-write sandbox and script to allocate the pty that agy --print needs to emit output. On non-macOS there is no sandbox-exec, so agy refuses to run unless you pass --allow-unsandboxed-gemini (or use --codex-only).

Graceful Degradation

The Council runs a preflight check on first invocation. If only one advisor is available, it operates in single-advisor mode and notes this in the synthesis. If neither is available, it shows installation instructions.

Usage

Just tell Claude:

• "Ask the council to review this PR"
• "Get a second opinion on this architecture"
• "Council review" on any code changes
• "Debug this with the council"

The council is also invoked proactively during high-stakes architectural decisions or security-sensitive code reviews.

Recommended Configuration

Codex (~/.codex/config.toml) — set your preferred model:

model = "gpt-5.6-sol"

Reasoning effort is set by the invoke script (COUNCIL_CODEX_EFFORT=xhigh by default), so you do not need model_reasoning_effort in config. Set COUNCIL_CODEX_EFFORT=config if you'd rather the script defer to config.toml.

Gemini — no config file needed. To pin a model for either backend, export COUNCIL_GEMINI_MODEL (run agy models for agy ids; an AI Studio model id for gemini-cli). Left unset, each backend uses its own default.

Environment Variables

Variable	Default	Description
CODEX_MODEL	unset → ~/.codex/config.toml	Override the Codex model
COUNCIL_CODEX_EFFORT	xhigh	Codex reasoning effort; set config to defer to config.toml
COUNCIL_TIMEOUT	600	Max seconds per advisor (raise to 900 for very large xhigh reviews)
AGY_PRINT_TIMEOUT	8m	agy --print-timeout override; must stay below COUNCIL_TIMEOUT
COUNCIL_GEMINI_BACKEND	auto	Gemini backend: auto, gemini, or agy
COUNCIL_GEMINI_MODEL	unset (backend default)	Pin the Gemini model for either backend
COUNCIL_SNAPSHOT_EXCLUDES	(none)	Comma-separated pathspecs excluded from the safety-net snapshot

Safety

Both advisors run under OS-enforced read-only sandboxes on macOS — neither can modify your project:

• Codex: --sandbox read-only — Codex's built-in OS-level deny-write.
• Gemini (agy): wrapped in sandbox-exec using the deny-write profile at skills/the-council/scripts/council_sandbox.sb, plus a pty (script) so agy --print actually emits output. Writes are allowed only to ~/.gemini/, the per-invocation .council-tmp/ dir, system temp, and agy's cache subdirs — everything else is blocked at the OS layer.
• Gemini (gemini CLI): single-shot headless mode with no tool access, wrapped in the same sandbox-exec deny-write profile.

As defense in depth on all platforms, the script snapshots the working tree before and after the run and fails closed (exit 2) on any unauthorized change. See SKILL.md → Permissions and Safety for the full model.

Author

Created by Tom Wolinitz with Claude Code

More projects at github.com/DantesPeak85

License

MIT