sethdford/claude-code-teams

claude-code-teams is a drop-in upgrade for Anthropic Claude Code (~/.claude/) that adds production-oriented orchestration, evaluation, telemetry and prompt-evolution primitives. Intended for engineers, researchers and teams running Claude Code agent workflows and automated evals. Key features: - 16 custom skills and 20 specialized agents (verifier, critic, product-owner, scrum-master, tech-lead, auditor, domain specialists). - Full SCRUM-runner with adversarial audit and sprint auditor for scope/drift/DoD detection. - Eval harness with programmatic and LLM-judge scoring, regression detection, rewards.jsonl and per-agent value logs. - RL-style tooling: best-of-N with USC and confidence weighting, Reflexion-based prompt evolution, A/B gating. - Execution-grounded best-of-N and OS sandboxing for safe code runs; UI verification by screenshot+vision-LLM. - Auto-firing hooks, transcript mining to propose new skills, A-MEM/Zettelkasten memory discipline, cache and cost telemetry. - Uses only native Claude Code primitives; not RLHF/fine-tuning, not a framework, not vector DB backed. Includes install scripts, tests, architecture docs and MIT license.

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.