JinPLu/agent-roundtable

Agent Roundtable is a multi‑agent, cross‑vendor collaboration substrate for Cursor that coordinates Codex/Claude/Cursor subagents to plan → execute → review with full per‑turn auditing to disk. It’s aimed at engineers and teams who want reproducible, auditable multi‑agent workflows (planning, code execution, PR/code review, and automated goal‑driven loops). Key features: - CLI/thread commands: /roundtable-setup, /roundtable-plan, /roundtable-review, /roundtable-execute, /roundtable-goal. - Plan workflow: import plan → generate artifacts/PLAN.md and GOAL.md → blind cross‑vendor reviewers → verdict.json → edit → execute. - Executor enforces plan order, scope_check.py gating, and AskQuestion on stalls/violations. - Cross‑vendor blind review + aggregator to reduce model sycophancy; routing by role to cheaper/stronger models for cost control. - Full audit trail (.roundtable/threads/<slug>/history/) including prompts, stdout, verdicts and budget ledger. - Scripts: backend.sh, import_plan.sh, route.sh, codex_turn.sh/claude_turn.sh. Credential isolation and permission modes (read‑only/sandbox).

AI-related keywords found in this repo's description, topics, or README summary — grouped by category. Each badge links to the corresponding ranking detail page.