sigma

  ███████╗██╗ ██████╗ ███╗   ███╗ █████╗
  ██╔════╝██║██╔════╝ ████╗ ████║██╔══██╗
  ███████╗██║██║  ███╗██╔████╔██║███████║
  ╚════██║██║██║   ██║██║╚██╔╝██║██╔══██║
  ███████║██║╚██████╔╝██║ ╚═╝ ██║██║  ██║
  ╚══════╝╚═╝ ╚═════╝ ╚═╝     ╚═╝╚═╝  ╚═╝
  σ · personal AI workflow toolkit
  created by Navid Ghayazi

sigma wraps Claude Code with a disciplined, research-first pipeline built for the way AI/ML work actually happens — from classic ML and deep learning to NLP, RL, data engineering, MLOps, LLM engineering, and AI-agent harness design.

It's plugin-first: every pipeline stage is a native slash command, the domain knowledge and the learning layer are native skills. A thin CLI handles only what Claude Code can't do in a single session — real parallel multi-model research, autonomous hands-off runs, a live kanban board, and setup.

"You shouldn't be prompting coding agents anymore. You should be designing loops that prompt your agents."

sigma is that loop.

✨ Why sigma

• 🔬 Multi-model research — fan out a question to Claude + Gemini + GPT in parallel, aggregate into one cited research.md. Real concurrency, not a sequential loop.
• 📋 Spec-driven, BDD-native — specs carry Gherkin Scenario / Given / When / Then acceptance criteria that flow as a contract through implement → verify → review. No vibe-coding into production.
• 🔥 Adversarial grilling — a skeptical /grill gate pressure-tests the blueprint and the spec before any code exists. /grill-loop auto-drives grill → triage → edit → re-grill (mechanical fixes auto-applied, judgment calls surfaced). A logic flaw caught here costs a sentence, not a rewrite.
• 🔁 Closed learning loop — failures ratchet into skills/ and are recalled by domain on the next run. The loop doesn't just record mistakes; it stops repeating them.
• 🧑‍🔧 Maker ≠ checker, enforced — implementer, verifier, and the optional logic-evaluator are always distinct agents. Separation is a ValueError, not a guideline.
• 🤖 Autonomous when you want it — sigma loop --execute runs maker→checker cycles with the correctness axes ON by default: a logic-evaluator axis (--no-logic opts out), a post-pass simplify cleanup (--no-simplify), an advisor that escalates a fail to a distinct opus-tier agent for a correction plan (--no-advisor), and a live-scenario gate that drives each task's mapped BDD scenario against the running app — a real behavioral FAIL blocks the cycle, an unreachable-app ERROR doesn't (--no-e2e). Also routed by model tier by default (mechanical roles → sonnet, logic → opus; --no-route opts out). --tdd (writes the failing test first) and --team (independent tasks in parallel, each in its own real git worktree, merged back on pass, conflicts surfaced — never auto-resolved) stay opt-in — they change the execution model, not just add a check. --codex-verify/--codex-tdd swap the verifier/test-writer role to the codex CLI instead of claude for a genuine cross-provider maker≠checker check — opt-in, and deliberately excluded from --all (needs a second CLI + its own auth). --all turns on every axis including those two. hermes --auto chains whole stages until a human gate — and routes each stage by tier (planning/grill stages → the strong model, execution stages → the mid tier; --no-route opts out). sigma research's cross-referencing synthesis pass runs on the strong tier by default too. The verify + logic checkers receive each task's mapped BDD scenario as acceptance criteria, not just the task title.
• 🛠️ Bring your own design — already have a design, plan, or big spec? /craft is the in-session back-half conductor: hand it the artifact and it drives spec → grill → tasks → loop to verified code, skipping the research → propose → blueprint front half (hermes --auto covers that from a blank start). Same human gates: grill BLOCK, spec approval, verify fail.
• 🎛️ Lean context — only the domain a task needs is loaded, surfaced in-session by the sigma-domains skill. sigma prune cuts loaded-but-unused MCP servers + plugins that tax every turn.