# llm-roundtable

> llm-roundtable is an open-source multi-LLM council for autonomous AI coding agents, packaged as an Agent Skill plus one stdlib-only Python script (MIT). When an agent would stop and ask its user to choose, approve or review something, it convenes five models from four vendors by default (Claude Fable 5.1, Claude Opus 5.5, GPT-6.1 via Codex CLI, Grok 4.7, Gemini 3.8 Flash via Antigravity), plus a free OpenCode model when opted in (Muse Spark by Meta when offered, else the newest free model). Members answer blind, cross-examine each other anonymously (M1..Mn), rate each other's evidence 0-2, and vote. The verdict is a weighted vote, not a chairman model's summary.

Key facts:
- Install: `npx skills add dinhkarate/llm-roundtable --skill llm-roundtable` (skill for every agent) or `npx llm-roundtable` (CLI). Claude Code can also install it as a plugin: `/plugin marketplace add dinhkarate/llm-roundtable`, then `/plugin install llm-roundtable@llm-roundtable`.
- Uses the official CLIs (claude, codex, grok, agy, opencode) and their subscriptions; no API keys, no MCP server.
- Hosts: Claude Code, Codex CLI, Grok CLI, Antigravity, OpenCode, pi, or any agent with a shell.
- Vote weight = seat weight × (0.5 + 0.5 × mean peer evidence rating). Ballots without a sourced FACT or INFERENCE do not count.
- Ties go to the Fable seat; fewer than 2 valid ballots = NO QUORUM; TIE or NO QUORUM falls back to the user's safe `--default`.
- Each seat falls back one model level in the same family, then is marked UNATTENDED; the council never blocks.
- Default seats (Fable, Opus, Codex, Grok, Gemini) are read-only and sandboxed, publishing tools are removed, each seat keeps only its own CLI's credentials, Claude seats run without the user's hooks or memory (`--restricted`, Claude Code >= 2.1.248), and the workspace (tracked, untracked and gitignored files outside dependency/build dirs) is fingerprinted after every round (a change stops the run as TAINTED).
- The free OpenCode seat is opt-in (`--seats full-open`, solo presets) and not a sandbox: its background service can run shell commands; it works in a snapshot copy and `--private` drops it.
- Prompt-injection hardening: files and other members' text are data; cross-exam positions are fenced with a per-prompt nonce, seats report steering attempts (`injection_suspected`), and a FACT whose every `file:line` is missing from the workspace is downgraded to ASSUMPTION.
- Account-level limits (Claude session limit, a disabled account) skip the same-account fallback.
- Review mode (`--mode review --workspace .`) approves or rejects plans, diffs and pull requests: APPROVE / CHANGES / REJECT.
- Audit mode (`--mode audit --workspace .`): each seat reports findings with file:line evidence, the script merges duplicates, then every seat verifies every finding anonymously (CONFIRMED / REJECTED / UNSURE; its own findings excluded). A finding is confirmed when two seats support it and outweigh the rejections. Output: findings.json and VERDICT: FINDINGS | CLEAN | NONE.
- Output: `~/.roundtable/runs/<time>-<slug>/decision.md` and `result.json` (`schema_version` 2); the last line is `VERDICT: <option|NONE> (council|tiebreak|default|none|workspace_changed)`. `status --json` prints state, verdict and flags as JSON for host agents.
- `run --dry-run` prints the resolved seats, quorum and call/time bounds as JSON without calling a model; `--deadline` caps the whole run.
- Optional `--min-margin` (e.g. `2/3`): a winner below that share of the valid weight falls back to `--default`, keeping the council's pick as `council_verdict`.
- Ballots may list `UNKNOWN` open questions (never evidence) and kill criteria; round 1 adds a framing check, and an advisory framing concern is raised when at least half the valid seats (and at least two) name a missing option.
- decision.md and result.json include usage: CLI calls per seat (nudges and fallbacks included), wall time per round, and tokens per seat as reported by each CLI (claude, codex, grok, opencode; agy reports none).
- Optional `--early-stop`: skips the cross-exam when round 1 is unanimous with at least 3 valid ballots from 2+ vendor CLIs; never in review mode. Every run records whether round 1 was unanimous and who changed vote after it.
- `outcome <run> right|wrong|mixed|unknown` records afterwards whether a decision worked out; `ledger` summarises all runs (how they were decided, outcomes, how often round 1 was unanimous and round 2 still changed a vote, median calls, time and tokens). Neither feeds a verdict.
- Typical run: about 12 model calls, 5-20 minutes; a measured full council (5 seats answering, 2 rounds) used 684k tokens in 6.5 minutes.

## Docs

- [Homepage](https://dinhkarate.github.io/llm-roundtable/): landing page
- [README](https://github.com/dinhkarate/llm-roundtable/blob/main/README.md): overview, quick start, comparison with karpathy/llm-council, claude-skills-llm-council, council-of-high-intelligence and Zen MCP, FAQ
- [SKILL.md](https://github.com/dinhkarate/llm-roundtable/blob/main/skills/llm-roundtable/SKILL.md): what the host agent does (brief, run, act on the verdict)
- [Protocol](https://github.com/dinhkarate/llm-roundtable/blob/main/skills/llm-roundtable/references/protocol.md): rounds, prompts, evidence rules, maths, sandboxing
- [Hosts](https://github.com/dinhkarate/llm-roundtable/blob/main/skills/llm-roundtable/references/hosts.md): per-host setup, single-tool presets, adding a new CLI via JSON
- [Sample briefs](https://github.com/dinhkarate/llm-roundtable/tree/main/skills/llm-roundtable/references/briefs): decide and review examples to copy
- [Research](https://github.com/dinhkarate/llm-roundtable/blob/main/docs/research.md): prior art, papers (Debate or Vote, NeurIPS 2025; When Identity Skews Debate, ACL 2026), CLI quirks, self-review runs

## Optional

- [Example decision.md](https://github.com/dinhkarate/llm-roundtable/blob/main/docs/example-decision.md)
- [Vietnamese README](https://github.com/dinhkarate/llm-roundtable/blob/main/README.vi.md)