"The eyes of common perception do not see far. Too often we make the most important decisions based only on superficial information."
— Cogitor Kwyna
🇹🇷 Türkçe
When a critical architecture choice, security review or product decision is on the line, consulting a single AI is risky: a model cannot see its own hallucinations, biases and blind spots, and it can defend flawed assumptions with complete confidence.
The Cogitors brings three model families — Codex (OpenAI), Claude (Anthropic) and Antigravity (Google) — to one deliberation table. Every model receives the whole task. Over four rounds they analyze independently, read each other's views and answer objections with evidence. In round five the Elder synthesizes one reasoned decision.
Core principle: Consensus is not proof of correctness. Reasoned minority views and uncertainties are never suppressed; they are preserved in the final decision.
/<model> <task> calls Codex, Antigravity or Claude onto a task, and it gets the job done in the background.
/cogitors <task> is different: it starts the five-round deliberation protocol with all three agents.
/voltron <task> forms Voltron: the Elder splits the job into parts, every other model and account builds one part as a lion at the same time, and the Elder joins them and runs the checks. /lions <task> sends every lion to the same task instead, each from its own view, and the Elder merges what they find; it suits a code review or a second opinion. The Elder picks a model for each lion and starts at once; their output is for the Elder, which gives you one summary. Work that changes files runs in a separate git worktree per lion. Extra accounts (for example a second Claude or Codex login) are listed under accounts in ~/.cogitors/config.json; see installation.
- /antigravity Find the contradicting clauses in this technical spec.
- /claude Analyze the performance bottlenecks in this PR diff.
- /codex List the weaknesses in the authentication flow without changing code.
- /cogitors Review this architecture proposal together. Do not change any code, and state unresolved objections explicitly.
- /cogitors --git-diff HEAD~1 Deliberate on this PR's changes for security and performance.
- /voltron Build the checkout flow: API, UI and tests.
- /lions Review this repository for bugs.
Note
These are not new terminal commands but standard Agent Skills invocations. The agent that starts the session becomes the Elder, runs it and takes part in the deliberation itself; no extra fourth model is launched.
┌───────────────────────┐
│ /cogitors <task> │
└───────────┬───────────┘
Round 1: Initial Analysis ▼
────────────────────────────────────────────────────────
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Codex │ │ Antigravity │ │ Claude │
└─────────────┘ └─────────────┘ └─────────────┘
───────────────────────────┬────────────────────────────
▼
┌───────────────────────┐
│ Drafts Unsealed │
└───────────┬───────────┘
Round 2: Reconsideration ▼
────────────────────────────────────────────────────────
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Codex │ │ Antigravity │ │ Claude │
└─────────────┘ └─────────────┘ └─────────────┘
───────────────────────────┬────────────────────────────
Round 3: Debate ▼
────────────────────────────────────────────────────────
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Codex │─x─│ Antigravity │─x─│ Claude │
└─────────────┘ └─────────────┘ └─────────────┘
───────────────────────────┬────────────────────────────
Round 4: Final Position ▼
────────────────────────────────────────────────────────
┌─────────────┐ ┌─────────────┐ ┌─────────────┐
│ Codex │ │ Antigravity │ │ Claude │
└─────────────┘ └─────────────┘ └─────────────┘
───────────────────────────┐
Round 5: Synthesis ▼
────────────────────┬──────────────┐
│ Elder │
└──────────────┘
- Initial Analysis: Each agent examines the task alone, without seeing the others.
- Reconsideration: Agents read each other's drafts and may update their views.
- Debate: Each agent defends its view with evidence and objects where it disagrees.
- Final Position: Objections are answered and each agent sets its final stance.
- Synthesis: The Elder gathers agreements, minority views and uncertainties into one decision.
Every session is saved in the task's language under your project's docs/cogitors/ directory. Working files go to the session's temp/ folder, which git ignores and which is deleted once the report is finished:
<project-root>
docs/cogitors/<time-stamp>-<slug>/
├── temp/ # working files; removed when the session finishes
├── antigravity-<slug>.md
├── brief-<slug>.md
├── claude-<slug>.md
├── codex-<slug>.md
├── decision-<slug>.md
├── decision-<slug>.json # only with --export-json
├── decisions-<slug>.html
├── metrics-<slug>.json # durations, calls, reported tokens, revisit date; no model text
└── followup-<slug>.md # answers to follow-up questions, after the first one
- Evidence check: Quotes the Elder cites in the synthesis are checked against the shared files and marked as matched or not. A match proves the quote exists, not that the claim is right.
- Metrics:
npx cogitors metricssummarizes time, calls and per-agent tokens across finished sessions; tokens show asunknownwhen a provider reported none. - Follow-up questions: Every part of the HTML report has an Ask button. Ask one question or collect several, paste them into the chat, and the Elder answers from the session record without new model calls. Answers appear under the part you asked about; the decision itself never changes.
- Revisit: The Elder can set a revisit date and what would show the decision was right;
npx cogitors revisitlists the sessions whose date has come. Each reasoned dissent can also name the evidence that would settle it.
Run npx skills add erayendes/cogitors and install all six skills. Details here.
- Cost: A session makes 8 external model calls (plus any timeout retries you approve), and you approve the scope, including every shared file's path, size and hash, before any call starts.
--max-calls Nsets a hard ceiling, retries included. - Source code: Deliberation never edits code; if a source file changes, the session stops.
- Participation: If an agent fails, you are asked; a session never continues with fewer than 2 participants.
- Time: 180 seconds per call and 1800 seconds per session (
--timeout,--total-timeout).
Cogitors are ancient minds in the Dune universe that left their bodies behind and, preserved in a protective fluid, devote thousands of years to thought. They reach a conclusion only after thinking and debating at length, never in haste. This project takes its name from them: several models think through the same question without haste, questioning themselves and each other.
Report a bug · Request a feature · Support · Security · Contributing | MIT © Eray Endes